Preface - JHER 1 No 1 2013 full journal.pdffor hydrological model TOPMODEL for the Biliu River Basin in China. The 2nd paper deals with application of Co-active Neuro Fuzzy Inference

Preface

________________________________________

This is the first issue of Journal of Hydrology and Environment Research (JHER). The JHER is a peer reviewed

international journal which publishes high quality research papers in the fields of water and environment. The JHER is

published from Sydney, Australia.

This issue of JHER contains 6 technical papers and one review paper. The first paper focuses on Sobol’s sensitivity analysis

for hydrological model TOPMODEL for the Biliu River Basin in China. The 2nd

paper deals with application of Co-active

Neuro Fuzzy Inference System for regional flood frequency analysis problem in eastern Australia using data from 452

gauged catchments. The 3rd paper examines the effectiveness of rainwater harvesting system in peri-urban regions of

Greater Sydney, Australia, which uses data from 10 different locations in Sydney region. The 4th paper describes the

validation of TRMM rainfall for the Pangani River Basin in Tanzania. The 5th paper examines the variability in rainfall

temporal patterns using data from 86 pluviograph stations in New South Wales, Australia and applying a Monte Carlo

simulation technique based rainfall runoff modeling. The 6th paper applies principal component analysis in water demand

forecasting using data from Blue Mountains water supply scheme in New South Wales. The 7th paper reviews brick

manufacturing practice in Bangladesh with a focus on energy efficiency and air pollution.

We acknowledge the reviewers who have spent their valuable time in reviewing the papers contained in this issue. We

would like to thank Mr Tauqir Ullah and Mr Imran Rahman for assisting in designing the logo and cover pages, and Mr

Imran Hossain and Mr Imran Rahman for developing the website of the journal. We also thank Mrs Waheda Rahman, Ms

Ayesha Rahman and Dr Amir Ahmed for editorial assistance. We welcome papers for 2nd

issue which will be published in

early 2014. Authors are advised to visit website of JHER (www.jher.org) for information on preparation and submission of

manuscripts for possible publication in this journal.

(Dr Ataur Rahman, PhD, Editor, JHER, Senior Lecturer, School of Computing, Engineering and Mathematics, University

of Western Sydney, New South Wales, Australia: [email protected]; [email protected])

http://www.jher.org/

mailto:[email protected]

Editorial Board

Editor: Dr Ataur Rahman, PhD, Fellow (Engineers Australia), Member ASCE, Member IWA, Member AGU, Member AWA

Associate Editors: Associate Professor Dr Chi Zhang, Dalian University of Technology, China

Dr Khaled Haddad, University of Western Sydney, Australia

Professor Dr Muthiah Perumal, Indian Institute of Technology, Roorkee, India

Editorial Coordinator: Dr Amir Ahmad

International Advisory Board: 1. Professor Dr Salaheddine El Adlouni, Université de Moncton, Canada

2. Associate Professor Dr Ramesh Teegavarapu, Florida Atlantic University, USA

3. Dr Tony Ladson, Australia

4. Professor Dr Zulkifli Yusop, University Technology Malaysia, Malaysia

5. Dr Douglas Bertram, University of Strathclyde, UK

6. Professor Abdallah Shanableh, University of Sharjah, United Arab Emirates

7. Professor R.B. Singh, University of Delhi, India

8. Professor Kadri Yurekli, Gaziosmanpasa University, Turkey

9. Associate Professor Dr A/P Arumugam Sathasivan, University of Western Sydney, Australia

10. Associate Professor Dr Mehdi Yasi, Urmia University, Iran

11. Associate Professor Dr Mushtaque Ahmed, Sultan Qaboos University, Oman

12. Professor Dr Fawzi S Awad, King Saud University, Saudi Arabia

13. Professor Dr Md Rezaul Karim, Islamic University of Technology, Bangladesh

14. Dr Namrata Pathak, University of Delhi, India

15. Dr Monzur Imteaz, Swinburne University of Technology, Australia

16. Professor Dr Mahmoud Mohamed Hegazi, Ain Shams University, Egypt

17. Professor Dr Ata-ur-Rehman Tariq, University of Engineering and Technology, Pakistan

18. Dr John OSullivan, National University of Ireland, Dublin, Ireland

19. Dr Amimul Ahsan, University Putra Malaysia, Malaysia

Published by EnviroWater Sydney, New South Wales, Sydney, Australia

© EnviroWater Sydney, 2013 The publisher does not take any responsibility of the contents of a paper in this journal. Before using information published in this journal in

real design and practice, contents/materials must be verified by the user at his/her own responsibility. The publisher does not take any liability or loss incurred by a user who has used information from this journal.

List of Reviewers

Professor Dr. Rezaul Karim, PhD

Dr Ataur Rahman, PhD

Dr. Khaled Haddad, PhD

Dr. Md. Al-Amin, PhD

Evan Hajany, MEng

Orpita Urmi, MEng

Faruk Kader, MEng

Jalal Uddin, MEng

Md Mahmudul Haque, MEng

Peer Review Process: Each of the published papers has been reviewed independently. Papers are revised as per

reviewers’ comments before being finally accepted for publication. A total of 10 papers were submitted, three of them were

rejected and seven were accepted.

We acknowledge the reviewers for their valuable time in reviewing the papers.


© EnviroWater Sydney, 2013

Table of Contents

Technical papers:

Sobol′’s sensitivity analysis for TOPMODEL hydrological model: A case study for 1

the Biliu River Basin, China

W. Qi, C. Zhang, J. Chu and H. Zhou

Co-Active Neuro Fuzzy Inference System for Regional Flood Estimation in Australia 11

K. Aziz, A. Rahman, A. Y. Shamseldin and M. Shoaib

Effectiveness of Rainwater Harvesting System in Peri-Urban Region: A Case Study 21

for the Greater Sydney

E. Hajani and A. Rahman

Validation of TRMM Rainfall for Pangani River Basin in Tanzania 30

R. M. F. Haque, S. Maskey, S. Uhlenbrook and M. Mul

Variability in Rainfall Temporal Patterns: A Case Study for New South Wales, 41

Australia

W. L. Caballero and A. Rahman

Principal Component Regression Analysis in Water Demand Forecasting: 49

An Application to the Blue Mountains, NSW, Australia

M. M. Haque, A. Rahman, D. Hagare and G. Kibria

Review Paper: Brick Manufacturing Practice in Bangladesh: A Review of Energy Efficacy 60

and Air Pollution Scenarios

K. M. Darain, A. B. M. S. Rahman, A. Ahsan, A. B. M. S. Islam and B. Yusuf



Technical Paper

© EnviroWater Sydney, 2013 Journal of Hydrology and Environment Research, Vol 1, No 1 1

Sobol′’s sensitivity analysis for TOPMODEL hydrological model: A case study for the Biliu River Basin, China1

W. Qi, C. Zhang *, J. Chu and H. Zhou

School of Hydraulic Engineering, Dalian University of Technology, Dalian 116024, China

Abstract: For an in-depth understanding of TOPMODEL performance and influences of TOPMODEL parameters on flood simulation, the global sensitivity analysis methodology, Sobol′’s method, is used in this paper to assess TOPMODEL parameter sensitivity in terms of individual parameter, interaction of parameters and combination of parameters using four flood prediction matrices. The results show that all the TOPMODEL parameters have little influences individually on Nash–Sutcliffe Efficiency (NSE), Relative Error of Runoff Amount (RERA) and Relative Error of Peak Flow (REPF). However, the effective lateral saturated transmissivity has a noticeable influence on Peak Flow Time Error (PFTE). The combinations of parameters, form of exponential decline in conductivity and unsaturated zone time delay have notable influences on NSE, RERA and REPF. The combination of parameters, form of exponential decline in conductivity and effective lateral saturated transmissivity has remarkable influences on PFTE. These findings are important and will assist in understanding the performance of TOPMODEL and its calibration for flood prediction. Keywords: Sensitivity analysis, Hydrological modeling, Sobol′’s method, TOPMODEL, Flood simulation

1 Introduction Hydrological models have been popularly implemented in basin hydrological prediction which is needed for many aspects of water resources management. Developed by Beven in 1979, TOPMODEL was one of the first attempts to simulate distributed hydrological responses. As to parameter sensitivity analysis of TOPMODEL, the Mento Carlo based Hornberger–Spear–Yong (HSY) local sensitivity method is often used; however, the interaction of parameters and their combination effects are not easy to detect. Sobol′’s method is a global sensitivity analysis method and is able to identify the influence of each parameter, interaction of parameters and their combination effects on the model outputs (Sobol′, 1993). Recently Sobol′’s method has become increasingly prevalent in hydrological modeling as its ability to integrate consideration of parameters interaction and the relatively straightforward interpretation of the associated results (e.g., Pappenberger et al., 2008; Van Werkhoven et al., 2008; Yang, 2011; Fu et al., 2012). Tang et al. (2007b) comprehensively compared Sobol′’s method with other tools including local analysis, Regional Sensitivity Analysis (RSA), and Analysis of Variance (ANOVA). It found that Sobol′’s method is the most effective approach to characterize single- and multi-parameter interactive sensitivities for lumped watershed models. Furthermore, Tang et al. (2007a) used Sobol′’s method to a distributed hydrological watershed model named as the Hydrology Laboratory Research Distributed Hydrologic Model (HL-RDHM), and the results obtained reveals that the method provides robust sensitivity rankings. Until now, Sobol′’s approach has not been used to analyze TOPMODEL parameter sensitivity in terms of individual parameter, interaction of parameters and their combination. Therefore, in this paper, Sobol′’s method is implemented to reveal sensitivity for TOPMODEL parameters in the Biliu River basin, China, in terms of four different matrices: Nash–Sutcliffe Efficiency (NSE), Relative Error of Runoff Amount (RERA), Relative Error of Peak Flow (REPF) and Peak Flow Time Error (PFTE). This paper is organized as follows. Section 2 describes some details of

1Paper JHER005 submitted 17/09/2013; accepted for publication after peer review and subsequent revision on 19/10/2013 *Corresponding author may be contacted at [email protected]

“Sobol′’s sensitivity analysis for TOPMODEL hydrological model …” Qi, Zhang , Chu and Zhou

Journal of Hydrology and Environment Research 2

TOPMODEL hydrological model and Sobol′’s method. Section 3 outlines the case study details. Section 4 provides the sensitivity analysis results and discussion. Conclusions are given in section 5. 2 Methodology 2.1 Overview of TOPMODEL hydrological model TOPMODEL is a physically based, variable contributing area model of basin hydrology which attempts to combine the advantages of simple lumped parameter model with distributed effects (Beven et al., 1979). Fundamental of TOPMODEL’s parameterization of the dynamic spatial hydrological response and the utilization of a topographic index of hydrological similarity are three basic assumptions: (1) saturated-zone dynamics can be approximated by successive steady-state representations; (2) hydrological gradients of the saturated zone can be approximated by the local topographic surface slope; and (3) the transmissivity profile with a form of exponentially declining along vertical depth of water table or storage is spatially constant. On the basis of above mentioned assumptions, the index of hydrological similarity is represented as the topographic index, , where is the area per unit contour length and is local slope angle. The greater upslope contributing areas and lower gradients areas are more likely to be saturated. More detailed descriptions of TOPMODEL and its mathematical formulation can be found in Beven et al. (1979). As the relatively simple model structure, fewer model parameters included and easily implemented, TOPMODEL has been popularly utilized in many research studies (Bastola et al,. 2008, Blazkova et al., 1997, Bouilloud et al., 2010, Cameron et al., 1999, Gallart et al., 2008, Hossain et al., 2005). Take China as a particular example, TOPMODEL has been used in climate change study (Liu et al., 2012), DEM resolution influences on hydrological simulation (SUN Li-qun et al., 2008, Xiong et al., 2004, Xu et al., 2007), and other aspects (Guo et al., 2000, Wang et al., 2007). However, none of the above applications include a global sensitivity analysis to advance the understanding of influences of each parameter and parameters combination on the model performance in terms of different flood simulation matrices. 2.2 Sobol′’s method Sobol′’s sensitivity is variance-based global quantitative sensitivity analysis method and has been popularly used in many aspects of hydrological modelling (Fu et al., 2012, Hall et al., 2005, Saltelli, 2002, Sobol', 2001, Tang et al., 2007, Zhang et al., 2013). Sobol′’s method (Sobol′, 1993) assumes that a model could be represented in the following functional form:

pXXfXfY ,,1 (1)

where Y is the goodness-of-fit metric of model output, and pXXX ,,1 is the parameter set. In Sobol′’s method, the total

variance of function f , yD , is decomposed into component variances from individual parameter and their interactions:

pji kji

ijkiji

i DDDDyD 12

(2)

where iD is the amount of variance due to the i th parameter iX , and ijD is the amount of variance due to the interaction between

parameter iX and jX . The sensitivity of single parameter or parameters interaction, i.e. Sobol′’s sensitivity indices of different orders, is

then assessed based on their percentage contribution to the total varianceD :

First-order index D

DS ii (3)

Second-order index D

DS ijij (4)



D

DS ~i

Ti 1 (5)

where ~iD is the amount of variance due to all of the parameters except for iX , iS measures the sensitivity from the main effect of iX ,

ijS measures the sensitivity from the interactions between iX and jX , and TiS measures the main effect of iX and its interactions with all the other parameters.

The variances in Eq. 2 can be evaluated using approximate Monte Carlo numerical integrations, particularly when the model is highly

nonlinear and complex. The Monte Carlo approximations for D , iD , ijD , and ~iD are defined as presented in the following prior

studies (Sobol′,1993, 2001; Hall et al. 2005):

n

ssXfn

f1

0

1ˆ (6)

20

1

2 ˆ1ˆ fXfn

Dn

ss

(7)

2

01

~ˆ,

1ˆ fXXfXfn

Dn

s

ais

bsi

asi

(8)

2

01

,~,~ˆ,

1ˆ fXXfXfn

Dn

s

asji

bsji

as

cij

(9)

jicijij DDDD ˆˆˆˆ (10)

2

01

~~ˆ,

1ˆ fXXfXfn

Dn

s

bis

asi

asi

(11)

where n is the sample size, sX is the sampled individual in the scaled unit hypercube, and superscripts a and b represent two

different samples. All of the parameters take their values from sample a are represented by asX . The variables a

isX and bisX

denote that parameter isX uses the sampled values in sample a and b , respectively. The symbols asiX ~ and

bsiX ~ represent

cases when all of the parameters except for isX use the sampled values in sample a and b , respectively. The symbol a

sjiX ,

represents parameters isX and jsX with sampled values in sample a . Finally, b

sjiX ~,~ represents the case when all of the

parameters except for isX and jsX utilize sampled values from sample b .

Sobol′’s sensitivity indices have been shown to be more effective than other approaches in capturing the interactions between a large number of variables for highly nonlinear models (Tang et al., 2007a and b). Building on the recommendations of Tang et al. (2007a), the Latin Hypercube sampling method (McKay et al., 1979) is used for implementing Sobol'’s method. By using Sobol′’s improvement method (Saltelli, 2002), computing the first-order, second-order and total-order sensitivity indices only requires ( 2)n m model evaluations, where n is the number of Latin Hypercube samples and m is the number of parameters being analyzed.



3 Case study 3.1 Biliu River basin Biliu River basin (2567 km2), locating in coast region between the Bohai Sea and the Huanghai Sea of China, covers longitudes from 122.29°E to 122.92°E and latitudes from 39.54°N to 40.35°N. This basin is characterized by a temperate monsoon marine climate and summer (June to September) is the major rainfall period. The major land cover types are forest and farmland, and the average annual temperature is 10.6°C. Due to effects from ocean, the seasonal temperature difference is small. The basin average elevation is 240m. The maximal elevation is 985m in northern part, which is a mountainous region, and the minimum elevation is 4.5m in southern part. There are ten rainfall stations and a discharge gauge, and their spatial distributions are shown in Figure 1. The main study region is the upper sub-basin of Biliu reservoir, which accounts for about 74.1% of the total catchment area. 3.2 Data set The required data sets for TOPMODEL modeling mainly include: precipitation data, observed flow, evaporation, topography index and cumulative area-distance relationship.

Figure 1 Biliu River Basin In this paper, the basin average rainfall is calculated using thiessen method, and eight flood events of Biliu River basin are considered, including one, two, three, eight, thirty and forty years Average Recurrence Interval (ARI). Evaporation is estimated with equation 12.

0

max

1p

E SRE SR

(12)

where SR0 is the root zone storage deficit; SRmax is the maximum capacity of the root zone; and Ep is potential evaporation which can be obtained from meteorological gauge. Topography index and cumulative area-distance relationship are derived with FORTRAN program, and the spatial distribution of topography index is showed in Figure 2. Percentage distribution of topgraphy index and cumulative area-distance relationship are displayed in Figure 3 and Figure 4, respectively. 3.3 Model parameters Six main parameters are included in TOPMODEL, and their ranges and brief description are listed in Table 1. Usually SR0 and SRmax could be estimated by experience or observation. However, all these methods cannot provide precise parameter values. Therefore, SR0 and SRmax are given by ranges. In this paper, the parameter value ranges are given based on prior research by Liu (2008).



. Figure 2 The spatial distribution of topography index

Figure 3 The distribution of topgraphy index

Figure 4 Cumulative area-distance relation

3.4 Goodness-of-fit matrices The sensitivity analysis for TOPMODEL with Sobol′’s approach considers four goodness-of-fit matrices which are considered as major criteria to evaluate the applicability of flood prediction in China. These four matrices include Nash–Sutcliffe Efficiency (NSE), Relative Error of Runoff Amount (RERA), Peak Flow Time Error (PFTE) and Relative Error of Peak Flow (REPF), and they are calculated using equation 13, 14, 15 and 16, respectively.



2

1

2

1

1

n

pi tiin

ti ti

Q QNSE

Q Q

(3)

1

1

1 100%

n

piin

tii

QRERA

Q

(14)

( ) ( )pp tpPFTE T Q T Q (15)

1 100%pp

tp

QREPF

Q

(16)

where piQ and

tiQ are the simulated and measured flows at time i ; n is the total number of flood time; tQ is the mean measured flows

during the flood period; ( )pp

T Q is the simulated peak flood time of simulation; ( )tp

T Q is the observed peak flood time; ppQ is the

simulated peak flood volume and tpQ is the observed peak flood volume. The NSE should be more close to 1.0 if simulation is better. The

lower the RERA, PFTE and REPF values are the better the simulation is performed. RERA, PFTE and REPF values of 0 indicate a perfect fit.

3.5 Sensitivity analysis Statistical sample size is a vital parameter for Sobol′’s approach. Tang et al. (2007b) employed a sample size of 8192 for Sobol′’s analysis considering 18 model parameters. Fu et al. (2012) adopted a set of 2000 LHS samples for 21 parameters. Tang et al. (2007a) used a sample size of 2000 for 403 variables. On the basis of these research studies, this paper takes a LHS sample size of 3000 into our research which leads to 3000*(6+2) = 24000 model evaluations. The sensitivity analysis for TOPMODEL is based on the average NSE, RERA, PFTE and REPF of eight flood events, which stands for a common sensitivity.

4 Results 4.1 Flood simulation The best simulations of the eight flood events within total 24000 model runs are shown in Figure 5 and the corresponding four matrices are displayed in Table 2. Table 2 and Figure 5 show that satisfactory flood simulation results are obtained. The aim of this paper is to analyze parameter sensitivity with Sobol′’s approach, i.e., to obtain information concerning how individual parameter and parameters combination influence model performances, which provides references to model parameter calibration. 4.2 First and total order sensitivity The first and total order sensitivity indices are presented in Figure 6. In each panel, the x-axis represents parameters, and y-axis represents first and total order sensitivity indices. These indices are defined with a threshold of 1%, which is subjective and their ease-of-satisfaction decreases with increasing parameter interactions or numbers of parameters (Tang et al., 2007a.b).



Table 1 Parameter ranges for TOPMODEL

Parameter Description Lower bound

Upper bound

SZM [m] Form of the exponential decline in conductivity 0.005 0.04

LNT0[m2 h-1] Effective lateral saturated transmissivity -25 10

RV[m2 h-1] Hillslope routing velocity 3500 8000

SRmax [m] Maximum root zone storage 0.001 0.01

SR0[m] Initial root zone deficit 0 0.01

TD [m h-1] Unsaturated zone time delay per unit deficit 0.5 5

Figure 6 reveals that each parameter of TOPMODEL has few individual influences on NSE, RERA and REPF, and LNT0 has many effects on PFTE. LNT0 mainly controls the relative proportion of surface flow and subsurface flow when surface soil is saturated. Usually surface flow has faster routing speed than subsurface flow. As the assumption of TOPMODEL, surface flow and subsurface flow have the same routing speed, therefore different component percentage of surface flow and subsurface flow could influence PFTE. 4.3 Second order sensitivity The second order sensitivity indices are displayed in Figure 7. In each panel, the x-axis represents parameter couples, and y-axis represents sensitivity indices. These indices are defined with a threshold of 1%. Figure 7 reveals that the parameter couple, SZM and TD, has massive influences on NSE, RERA and REPF, and parameter couple, SZM and LNT0, has relatively remarkable effects on PFTE. It is mainly because that discharge of TOPMODEL consists of two sources: surface flow and subsurface flow, and SZM, LNT0 and TD have something to do with the saturated groundwater table variation and the delay time of runoff in soil profile. Therefore, SZM, LNT0 and TD have many influences on runoff volume of every time step and the relative proportion of surface and subsurface flow. Furthermore, as mentioned above, TOPMODEL assumes that surface flow and subsurface flow have the same routing velocity (Beven et al. 2001, Beven et al. (1979), Peters et al., 2003), and hence discriminative proportion of surface and subsurface flow could influence runoff hydrograph. Overall, it can be found from Figure 6 and Figure 7 that SRmax and SR0 which primarily control evaporation have few influences. In the Biliu River basin, runoff volume is relatively high and flood has a fairly short duration and hence evaporation has only minor effects.



Table 2 Flood simulation results for the Biliu River basin

Flood NSE RERA(%) PFTE REPF (%)

19840615 0.96 -1.00 1 10.06

19850818 0.92 -9.78 1 7.52

19940815 0.96 -1.43 1 1.39

19950806 0.90 4.65 0 9.73

19960729 0.91 4.83 0 12.11

19980809 0.92 -1.32 1 1.28

20010816 0.85 -0.67 1 -1.69

20110807 0.97 -6.46 0 8.97

Figure 5 Flood simulation for the Biliu River basin



Figure 6 First order and total order sensitivity index of TOPMODEL parameters using the four goodness-of-fit matrices.

Figure 7 Second order sensitivity index of TOPMODEL parameters using the four goodness-of-fit metrics.

5 Conclusions This paper firstly provides a variance-based sensitivity analysis for TOPMODEL for the Biliu River basin, China. The analysis reveals the individual effects, interaction effects and parameters combination effects on the model performances in terms of four matrices of flood prediction. The results of this paper can provide valuable references for parameter calibration of flood prediction using TOPMODEL. The main findings are summarized below:

i. Individual parameter of TOPMODEL has few influences on Nash–Sutcliffe Efficiency, Relative Error of Runoff Amount and Relative Error of Peak Flow;

ii. One parameter, effective lateral saturated transmissivity, has great influences on Peak Flow Time Error; iii. The parameter couple, form of the exponential decline in conductivity and unsaturated zone time delay per unit deficit, has massive

effects on Nash–Sutcliffe Efficiency, Relative Error of Runoff Amount and Relative Error of Peak Flow; and



iv. The parameter couple, form of the exponential decline in conductivity and effective lateral saturated transmissivity has many effects on Peak Flow Time Error.

Acknowledgements The authors are grateful to the reviews whose comments are helpful to improve the quality of this paper. The authors also gratefully acknowledge financial support provided by the China Scholarship Council. This study was supported by the National Basic Research Program of China (Grant No. 2013CB036400), and the National Natural Science Foundation of China (Grant No. 51279021 and 51079014) References Bastola S, Ishidaira H, Takeuchi K (2008). Regionalisation of hydrological model parameters under parameter uncertainty: A case study

involving TOPMODEL and basins across the globe, Journal of Hydrology, 357(3-4), 188-206. Beven K, Freer J (2001). A dynamic TOPMODEL, Hydrological Processes, 15(10), 1993-2011. Beven KJ, Kirkby MJ (1979). A physically based, variable contributing area model of basin hydrology, Hydrological Sciences Bulletin, 24(1), 43-

69. Blazkova S, Beven K (1997). Flood frequency prediction for data limited catchments in the Czech Republic using a stochastic rainfall model

and TOPMODEL, Journal of Hydrology, 195. Bouilloud L, Chancibault K, Vincendon B (2010). Coupling the ISBA Land Surface Model and the TOPMODEL Hydrological Model for

Mediterranean Flash-Flood Forecasting: Description, Calibration, and Validation, Journal of Hydrometeorology, 11(2), 315-333. Cameron DS, Beven KJ, Tawn J (1999). Flood frequency estimation by continuous simulation for a gauged upland catchment (with

uncertainty), Journal of Hydrology, 219. Fu G, Kapelan Z, Reed P (2012). Reducing the Complexity of Multiobjective Water Distribution System Optimization through Global

Sensitivity Analysis, Journal of Water Resources Planning and Management, 138(3), 196-207. Gallart F, Latron J, Llorens P (2008). Upscaling discrete internal observations for obtaining catchment-averaged TOPMODEL parameters in a

small Mediterranean mountain basin, Physics and Chemistry of the Earth, 33(17-18), 1090-1094. GUO F, LIU X, REN L (2000). A topography based hydrological model: TOPMODEL and its widened application, Advances in Water Science,

11(3): 296-301. (in Chinese) Hall JW, Tarantola S, Bates PD (2005). Distributed Sensitivity Analysis of Flood Inundation Model Calibration, Journal of Hydraulic

Engineering, 131. Hossain F, Anagnostou EN (2005). Assessment of a stochastic interpolation based parameter sampling scheme for efficient uncertainty analyses

of hydrologic models, Computers & Geosciences, 31(4), 497-512. Liu Y, Zhang J, Wang G, Liu J, He R, Wang H, Jin J (2012). Evaluation on the influence of natural climate variability in assessing climate

change impacts on water resources: I. Model and methodology in baseline period, Advances in Water Science, 23(2): 147-155. (in Chinese). Peters NE, Freer J, Beven K (2003). Modelling hydrologic responses in a small forested catchment (Panola Mountain, Georgia, USA): A

comparison of the original and a new dynamic TOPMODEL, Hydrological Processes, 17(2), 345-362. Saltelli A (2002). Making best use of model evaluations to compute sensitivity indices, Computer Physics Communications, 145. Sobol' IM (2001). Globalsensitivityindices for nonlinearmathematicalmodels and their MonteCarloestimates, Mathematics and Computers in

Simulation, 55(1-3). Sun L, Hu C, Chen G (2008). Effects of DEM resolution on the TOPMODEL, Advances in Water Science, 19(5): 699-706. (in Chinese). Tang Y, Reed P, Van WK (2007). Advancing the identification and evaluation of distributed rainfall-runoff models using global sensitivity

analysis, Water Resources Research, 43(6). Wang L, Li Z, Bao H (2007). Application of hydrological models based on DEM in the Yihe basin, Journal of Hydraulic Engineering, S1:417-

422. (in Chinese). Xiong L, Guo S, Tian X (2004). DEM-based distribution hydrological model and its application, Advances in Water Science, 15(4): 518-520.

(in Chinese). Xu J, Ren L, Cheng Y, Yuan F (2007). Study on the translation relationship of DEM spatial scale based on TOPMODEL, Journal of Hydraulic

Engineering, S1: 404-408. (in Chinese). Zhang C, Chu J, Fu G (2013). Sobol′’s sensitivity analysis for a distributed hydrological model of Yichun River Basin, China, Journal of

Hydrology, 480, 58-68.

Technical Paper


Co-Active Neuro Fuzzy Inference System for Regional Flood Estimation in Australia1

K. Aziza*, A. Rahmana, A. Y. Shamseldinb and M. Shoaibb aSchool of Computing, Engineering and Mathematics, University of Western Sydney, Australia

bDepartment of Civil and Environmental Engineering, University of Auckland, New Zealand

Abstract

Regional flood frequency analysis (RFFA) involves transfer of flood characteristics from gauged to ungauged catchments. In Australia, RFFA methods generally focus on the application of empirical methods based on linear forms of model such as the Probabilistic Rational Method, the Index Flood Method and the regression-based techniques. There have been successful applications of non-linear models in RFFA in some other countries such as Co-Active Neuro Fuzzy Inference System (CANFIS), Gene-Expression Programming (GEP) and Artificial Neural Network (ANN). The application of these non-linear RFFA methods in Australia is limited. This study focuses on the application of Co-Active Neuro Fuzzy Inference System (CANFIS) based RFFA models to Australian data. Using data from 452 catchments in eastern Australia (a part of Australian Rainfall and Runoff Revision Project 5 Regional flood methods database), it has been found that the CANFIS based RFFA provides quite accurate regional flood quantile estimates. However, the Bayesian generalised least squares based QRT coupled with the region of influence approach outperforms the CANFIS based RFFA models. Keywords: ANN, CANFIS, QRT, RFFA, Floods, ARR, GLS regression 1 Introduction A design flood estimate is needed in the planning and design of hydraulic structures and in many other water resources management tasks such as flood control measures, flood plain mapping and flood insurance studies. The most direct method of flood estimation is the flood frequency analysis, which requires long period of recorded streamflow data at the site of interest. In the case of Australia, many catchments are ungauged or poorly gauged and hence regional flood frequency analysis (RFFA) methods are widely used in Australia. RFFA is the generic name given to describe techniques which utilises streamflow data from gauged catchments in a region to estimate design floods for poorly gauged or ungauged catchments. The use of RFFA enables the “transfer” of flood characteristics information from gauged to ungauged catchments (Bloschl and Sivapalan, 1997; Pallard et al., 2009). The most commonly adopted RFFA methods have been described in Cunnane (1988) and Hosking and Wallis (1997). RFFA essentially involves two important stages: (1) formation of regions; and (2) development of prediction equations. Traditionally these regions have been formed based on geographic, political, administrative or physiographic boundaries (e.g. NERC, 1975; I. E. Aust., 1987); however, they can also be formed in catchment characteristics data space using multivariate statistical techniques (e.g. Acreman and Sinclair, 1986; Nathan and McMahon, 1990; Rao and Srinivas, 2008; Guse et al., 2010). Moreover, regions can also be formed using a region-of-influence approach where a certain number of catchments based on proximity in geographic or catchment attributes space are pooled together based on some objective function to form an optimum region in RFFA (e.g. Burn, 1990; Zrinji and Burn, 1994; Kjeldsen and Jones, 2009; Haddad and Rahman, 2012). To develop the regional flood prediction equations, the commonly used techniques include the rational method, index flood method (IFM) and Quantile Regression Technique (QRT). The rational method has widely been adopted in estimating design floods for small ungauged catchments (e.g. Mulvany, 1851; I. E. Aust., 1987; Jiapeng et al., 2003; Pegram and Parak, 2004; Rahman et al., 2011). IFM has widely been adopted in many countries which rely on the identification of homogeneous regions (Dalrymple, 1960; Hosking and Wallis, 1993; Bates et al., 1998; Rahman et al., 1999; Kjeldsen and Jones, 2010; Ishak et al., 2011). The QRT, proposed by the United States Geological Survey (USGS) has been applied by many researchers using either an Ordinary Least Square (OLS) or Generalised Least Square (GLS) regression technique (e.g. Benson, 1962; Thomas and Benson, 1970; Stedinger and Tasker, 1985; Tasker et al., 1986; Pandey and Nguyen, 1999; Bayazit and Onoz, 2004; Rahman, 2005; Griffis and Stedinger, 2007; Ouarda et al., 2008; Kjeldsen and Jones, 2009; Haddad and Rahman, 2011; Haddad et al., 2011, 2012).

1Paper JHER007 submitted 18/09/2013 accepted for publication after peer review and subsequent revision on 21/10/2013

Corresponding author may be contacted at [email protected]

“Co-Active Neuro Fuzzy Inference System for Regional Flood Estimation …” Aziz, Rahman, Shamseldin and Shoaib


In case of Australia, the diversity of climatic conditions, site characteristics and a vast area have encouraged the hydrologists to emphasize on different aspects of RFFA. Furthermore, increased computing power and the research in statistical methods have enabled the hydrologists to develop new techniques such as non-linear models for the solution of many complex hydrological problems. Neuro fuzzy based techniques e.g., Co-Active Neuro Fuzzy Inference System (CANFIS) presents more flexible model structure that can easily account for non-linearities between the model input and output and their complex interactions. Neuro-fuzzy modelling refers to the way of applying various learning techniques developed in the neural network literature to fuzzy modelling or a fuzzy inference system. This approach adds the advantage of reduced training time not only due to its smaller dimensions but also because the network can be initialised with parameters relating to the problem domain (Maguire et al, 1998). From the early application of fuzzy logic to hydrology, a large amount of research has been pursued and, at present, fuzzy logic has become a practical tool in many hydrologic analyses and water resources decision making. Fuzzy logic can easily incorporate expert knowledge into standard mathematical models in the form of a fuzzy inference system. A judicious integration of fuzzy system and ANN can produce a functional neural fuzzy system capable of learning, high-level thinking, and reasoning (Loukas, 2001). The hybrid neuro fuzzy models are becoming popular as they gets the benefits of neural networks and fuzzy logic systems and removes the individual disadvantages by combining them on the common features. There have been many applications of neuro fuzzy based techniques in hydrology (Talei et al., 2010a, 2010b; Shiri and Kisi, 2010; Nourani et al., 2011; Nayak and Sudheer, 2004) but their application to RFFA problems is rather limited (Turan and Yurdusev, 2009 and Shu et al., 2008). The hydrological and climatic conditions of Australia are different from rest of the world; hence, it is important to develop new models and techniques based on Australian data for RFFA. In this paper, an overview of the CANFIS is presented first, which follows the description of study area and data. The adopted methodology is presented next, which follows the results and conclusion from the study. CANFIS based RFFA models have been developed for eastern parts of Australia. The database developed in the Australian Rainfall and Runoff (ARR) Revision Project 5 Regional flood methods has been used in this study. At the beginning, CANFIS and QRT based RFFA models are compared, which is then followed by a comparison of these models with the ordinary least squares (OLS) based QRT. The QRT method developed in the ARR Project 5 (Haddad and Rahman, 2012) has also been used in this comparison. 2 Overview and applications of CANFIS to RFFA A number of different neuro-fuzzy algorithms are available in the literature: fuzzy inference network (Keller et al., 1992b), fuzzy aggregation networks (Keller et al., 1992a), neural network driven fuzzy reasoning (Takagi and Hayashi, 1991), fuzzy modelling networks (Horikawa et al., 1992), fuzzy associated memory systems (Kosko, 1992) and the most popular neuro fuzzy system ANFIS (Jang, 1993) and CANFIS (Jang, 1997). The CANFIS model integrates the modular neural network with fuzzy inference system in (FIS) in the same topology. The powerful capability of CANFIS stems from pattern-dependent weights between the consequent layer and the fuzzy associate layer. The CANFIS model integrates adaptable fuzzy inputs with a modular neural network to rapidly and accurately approximate complex functions. In recent years much attention has been given to deriving effective data driven neuro-fuzzy models due to its numerous advantages. ANFIS-based neuro-fuzzy modeling was initially developed by Jang (1993), Jang and Sum (1997), Jin et al. (1995), which has been widely applied in engineering applications. Palit and Popovic (1999, 2000 and 2005) developed and applied neuro-fuzzy network for time series forecasting, Deka and Chandramouli (2003) used a fuzzy neural network model for deriving the river stage-discharge relationship, Kisi (2005) estimated suspended sediment by applying neuro-fuzzy and neural network approaches. Shafie et al. (2007) modeled inflow forecasting of the Nile River at Aswan high dam by using a neuro-fuzzy model. Jacquin and Shamseldin (2006) developed two types of fuzzy rainfall runoff models based on Takagi-Sugeno fuzzy inference systems. The results of the developed models are compared with those of Simple Linear Model. They found that the fuzzy inference systems are a suitable alternative to the traditional methods of rainfall and runoff modelling. Kisi (2006c) and Kisi and Ozturk (2007) applied adaptive neuro-fuzzy computing technique for pan evaporation and evapotranspiration modeling, and recently, Saemi and Ahmadi (2008) modeled permeability from well logs by using Genetic Algorithms and CANFIS. Talei et al. (2010a) evaluated the rainfall runoff modelling for a sub-catchment of Kranji basin in Singapore by using a neuro-fuzzy computational technique adaptive neuro-fuzzy inference system ANFIS. The result of the ANFIS was compared with those of physically based storm water management model (SWMM). They concluded that ANFIS model is comparable to storm water management model (SWMM) in terms of goodness of fit results. It was observed that only a few studies existed in the literature related to the use of neuro fuzzy in RFFA e.g. Shu et al. (2008). They applied ANFIS for RFFA at ungauged catchments in Canada and compared the results with nonlinear regression (NLR) and nonlinear regression with regionalisation (NLR-R) approaches and found that the ANFIS approach has a much better generalization capability than the NLR-R approach and comparable to ANN. In 2009, Aytek (2009) adopted CANFIS for evapotranspiration modelling. He found CANFIS based models to be outperforming the conventional model for



InpuFuzzifier

Inference

Defuzzyfier

Outpu

Fuzzy Knowledge base

evapotranspiration (ET0). Tabari and Talaee (2012) compared the utility of CANFIS for pan evaporation (Epan) modelling with multilayer perception (MLP) and found that MLP provides better results than CANFIS. But to the best knowledge of the authors, a CANFIS based RFFA modeling is not available in literature especially in case of Australia. Hence, this paper applies CANFIS in RFFA in Australia and compares the results with traditional RFFA methods. 3 Working structure of ANFIS and CANFIS Fuzzy Logic provides a different way to approach a control or classification problem. This method focuses on what the system should do rather than trying to model how it works. An adaptive network is a feed forward network which makes use of a collection of modifiable parameters for determining the output of the network. Like other neural networks, an adaptive network also consists of a set of nodes connected through directional links and each node is a process unit that performs a static node function on its incoming signal to generate the signal output. Unlike other neural networks, the links in an adaptive network only indicate the flow direction of signals between nodes and no weights are associated with these links. As introduced by Jang (1993), ANFIS is a novel architecture that uses a neural network learning algorithm for constructing a set of fuzzy if-then rules with appropriate membership functions (MFs) from the stipulated input-output pairs.

Figure 1 Fuzzy Inference System (FIS)

This procedure of developing a FIS using the framework of adaptive neural networks is called an adaptive neuro fuzzy inference system (ANFIS). Consider the example of simple FIS with only two inputs x and y and one output z. Suppose that the rule base contains two fuzzy if-then rules of Takagi and Sugeno (1983). Rule 1: If x is A1 and y is B1, then f1 = p1x + q1y + r1, Rule 2: if x is A2 and y is B2, then f2 = p2x + q2y + r2

Where A1, A2 and B1, B2 are the MFs of input x and y respectively; p1, q1, r1 and p2, q2, r2 are the parameters of the output functions. The node function in the same layer of the same function family is described below: Layer 1: Each node in this layer performs fuzzification and generates membership grade of linguistic label of an input variable. Layer 2: Each node in this layer is denoted by determining the MF of the whole input vector by aggregating the fuzzified results of the individual scalar functions of the every input variable. The output of each node in this layer is obtained by multiplying the incoming signals and represents the firing strength of a rule. Layer 3: Each node in this layer is labelled as N and computes the normalized firing strength. Layer 4: The output of each node in the fourth layer is calculated by the sum of the signals of the third and second layer of the network. Layer 5: There is only single node in this layer labelled as ∑ that calculates the overall output of the ANFIS or CANFIS as the summation of all incoming signals. In case of CANFIS, the fuzzy neuron that applies MFs to inputs is the fundamental component of CANFIS. The general Bell and Gaussian functions are the two commonly used MFs (Principe et al., 2000). The bell shaped membership function is used in this study. The normalized axon/neuron in the network is used to expand the output into a range of 0-1. One of the advantages associated with the fuzzy axon is that their MF can be modified through back propagation during network training and results in the expedition of the convergence. The modular neural network that applies functional rules to



the inputs is the second major component of CANFIS. The number of modular networks equals the number of network outputs, and the number of processing elements in each network corresponds to the number of MFs. The CANFIS also has a combiner axon that applies the MFs outputs to the modular network outputs (Roger et al., 1997 and Alecsandru et al., 2004). Finally, the combined outputs are channelled through a final output layer and the error is backpropagated to both the MFs and the modular networks. There are a total of five layers in the CANFIS similar to ANFIS and each layer functions is summarizes as follows. The fuzzification of the input is performed by each node in layer 1. Each node in this layer is the membership grade of a fuzzy set (A1, A2, B1 or B2) and specifies the degree to which the given input belongs to one of the fuzzy sets. The input to the layer 2 is the product of all the output pairs from layer 1. Two components are present in the next third layer in the network. The upper component of this layer applies the membership functions to each of the inputs, while the lower components is a representation of the modular network that computes, for each output, the sum of all the firing strength. The weight normalization of the outputs of the two components of the third layer is performed in the fourth layer of the network and this produces the final output of the network (Ishak and Trifiro, 2007). The CANFIS model integrates adaptable fuzzy inputs with a modular neural network to rapidly and accurately approximate complex functions. The TSK model fuzzy model proposed by Takagi, Sugeno and Kang (Takagi and Sugeno, 1985; Sugeno and Kang, 1988) is used in the present study, since this type of fuzzy model best fits the multi-input, single output system (Aytek, 2009).

4 Study area and data

This paper focuses on the eastern states of Australia consisting of New South Wales (NSW), Victoria (VIC), Queensland (QLD), and Tasmania (TAS). This part of Australia is selected as the spatial and temporal data of gauged catchments in this region are more comprehensive than other parts of Australia. Data was obtained from 452 stations which was prepared as a part of the ARR Revision Project 5 (Haddad and Rahman, 2012). The data preparation involved filling gaps, checking for trends, outliers and rating curve error in streamflow as detailed in Rahman et al (2012). The selected stations are shown in Figure 2, which include 96 stations from NSW, 131 from VIC, 172 from QLD and 53 from TAS. In this study, the ANFIS and ANN-based models were developed using two catchment characteristics: catchment area (area) and design rainfall intensity (Itc_ARI) (where ARI = 2, 5, 10, 20, 50 and 100 years and tc = time of concentration (hour), estimated from tc = 0.76(area)0.38). The basic design rainfall intensities (I) data for the selected catchments were obtained from Australian Rainfall and Runoff (ARR) (I. E. Aust., 1987, 2001). The T-year flood quantile was estimated by fitting the log Pearson Type 3 (LP3) distributions for each of the selected stations using a Bayesian parameter fitting procedure (Kuczera, 1999). If different predictor variables are prioritized in RFFA, it may be noted that catchment area (A) is the most important predictor variable for a catchment followed by design rainfall intensity (I), slope (S), mean annual rainfall (R) and evaporation (E) (Rahman et al., 1999; Haddad et al., 2010). Based on the same data set of the 452 catchments, a study was conducted by Aziz et al. (2010, 2013) using artificial neural network (ANN) and Gene expression programming (GEP) based RFFA techniques. They found that the set of predictor variables giving the best results consisted of A and Itc_ARI. In this study, these two predictor variables were selected for model development. All the eastern states i.e. NSW, VIC, QLD and TAS were considered as one region as this has been found to be producing the best results with the ANN based RFFA models in previous studies by Aziz et al. (2011, 2013). The catchment sizes of the selected 452 stations range from 1.3 km2 to 1900 km2 with the median value of 256 km2. For the stations of NSW, VIC and QLD, the upper limit of catchment size was 1000 km2; however, for Tasmania; there were 4 catchments in the range of 1000 km2 to 1900 km2. Overall, there are 12% catchments in the range of 1 km2 to 50 km2, 11% in the range of 50 km2 to 100 km2, 53% in the range of 100 km2 to 500 km2 and 24% greater than 500 km2. The annual maximum flood record lengths of the selected stations range from 25 to 75 years (mean: 33 years).

5. Method In case of CANFIS, all the 452 catchments were considered to be one region. The 452 catchments were then divided into 80% for training and 20% for independent testing. The training and testing data sets were selected randomly from the region. For each case, the model was built using the 80% model catchments and then used to predict 2, 5, 10, 20, 50 and 100 years ARI flood quantiles for the 20% independent test catchments. Moreover, the results from CANFIS were compared with two different techniques of QRT. In the first case, all the 452 catchments were considered to be one region to be consistent with the CANFIS based model. Randomly selected 80% of the catchments were used to develop the QRT models and the rest of the catchments (20%) were used for validation of the developed QRT model. An ordinary least squares fitting method was adopted to estimate the QRT model coefficients. This is referred to as ‘QRT’ in this paper. In the second case, the QRT method developed as a part of Project 5 was also considered in the comparison where a Bayesian generalised least squares (BGLS) regression was applied with a region-of-influence (ROI) approach (for details see Rahman et al., 2012). The relative error reported in Project 5 Stage II report by Rahman et al. (2012) has been used directly without redeveloping these models in this study. This is referred to as ‘BGLS-QRT-ROI’.



Figure 2 Location of study catchments (Blue colour represents training catchments and red colour represents test catchments) For the CANFIS model development, model catchments are clustered based on model variables (catchment area and rainfall intensity) into several class values in layer 1 to build up fuzzy rules, and each fuzzy rule was constructed through several parameters of membership function in layer 2. A fuzzy inference system structure was generated from the data using subtractive clustering. This was used in order to establish the rule base relationship between the inputs. In order to obtain the best CANFIS models, the mean squared error was used as ‘fitness function’, which was based on the observed and predicted flood quantiles; the training was undertaken to minimise this error. Lavenberg-Marquardt (LM) method was used as the training algorithm to minimize the mean squared error. CANFIS model was trained with a set of input and output data to adjust the weights and to minimize the mean squared error between the desired outputs and the model outputs. The testing data set was selected randomly to produce a reasonable sample of different catchment types and sizes. Two inputs (A, Itc_ARI) were used in one input layer and one output layer with one output (Qpred). In the case of CANFIS, the Bell membership function and the TSK neuro fuzzy model are used, as this type of fuzzy model best fits the multi-input single output system (Aytek, 2009). LM algorithm is used for the training of CANFIS model. The stopping criterion for the training of the CANFIS network is a maximum of 1000 epochs and training is set to terminate when the mean squared error (MSE) drops to 0.01 threshold value. In QRT, flood quantiles (QT) are regressed against catchment characteristics (predictor variables) (X) using the power form equation (Thomas and Benson, 1970; and Stedinger and Tasker, 1985; Haddad and Rahman, 2012):

...21210 XXQT (1)

where regression coefficients s are generally estimated by using an Ordinary Least Square (OLS) or Generalized Least Square (GLS) regression. In this study, in developing the QRT, both the dependent and independent variables were log-transformed to linearise Equation 5. In this study an OLS regression was adopted to develop prediction equations for each of the six flood quantiles using two predictor variables (A, Itc_ARI). The data sets for building and independent testing of the QRT model were the same as with the ANN and GEP models. The MINITAB 14 software was used to develop the QRT models.



Following evaluation statistics were used for model assessment and comparison:

Ratio between predicted and observed flood quantiles:

Ratio (r) = observed

predicted

Q

Q (2)

Relative error (RE):

RE (%) = Abs

100

obs

obspred

Q

QQ (3)

Where Qpred is the flood quantile estimate from the ANFIS, CANFIS and QRT models, Qobs is the at-site flood frequency estimate obtained from LP3 distribution using a Bayesian parameter fitting procedure (Kuczera, 1999).

6 Results

Table 2 summarises the median ratio values for the models based on CANFIS and QRT. In case of CANFIS based model, the best results are obtained for Q5 and Q50 with median ratio values of 0.95 and 0.93, respectively. The overall values range from 0.79 to 2.81. All the results based on CANFIS model are within acceptable range except for Q2 that shows a significant over prediction. The CANFIS based model shows remarkable improvement for ARIs. The QRT shows 19% to 28% overestimation in case of Q50 and Q100. Overall, the CANFIS based RFFA model outperforms the QRT based RFFA models in terms of median ratio values. Table 3 summarises the median relative error values for the CANFIS and QRT. In terms of median relative error, the CANFIS based model provides values from 34% to 59% except for Q2 which is 180%. In terms of relative error, the CANFIS based model either outperforms the QRT or provides competitive results except for Q2. Best value is obtained for 20 years ARI which is 34% for CANFIS and 42% for QRT based RFFA model. Although both models provide competitive results, if the RE value for Q2 is ignored CANFIS based RFFA model can be given a preference over QRT based model for regional flood estimation. Figure 4 shows the plot of observed and predicted flood quantiles for 20 years ARI from the CANFIS based model, which shows a good fit for the validation data sets. Similar results were found for other ARIs. It should be noted that for majority of the cases the model prediction match very well with the observed quantiles, but there are notable differences in few cases, which is not unexpected in RFFA for Australia as found by similar other studies (e.g. Haddad et al., 2011; Haddad and Rahman, 2012 and Haddad et al., 2011). Now, the better of the two models i.e. CANFIS is compared with BGLS-QRT-ROI method in Table 3. These BGLS-QRT-ROI relative error values were obtained from Project 5 Stage II report by Rahman et al. (2012). It is found that overall the BGLS-QRT-ROI provides better results than the CANFIS model; however, the CANFIS model provides results which are comparable to the BGLS-QRT-ROI in few cases. For Q2, CANFIS based model produces higher relative error values for all the states. The results obtained from CANFIS based model are comparable to the QRT-ROI for Victoria. Hence, it can be concluded that the linear QRT model integrated with BGLS and ROI generally outperforms the CANFIS based non-linear RFFA models for eastern Australia.

Table 1 Median Qpred/Qobs ratio values for different models ARI (years) CANFIS QRT

2 2.81 1.155 0.95 1.0610 0.79 1.3520 1.18 1.1350 0.93 1.19100 1.31 1.28



Table 2 Median relative error (%) values for different models ARI (years) CANFIS QRT

2 180 65 5 48 45 10 51 57 20 34 42 50 59 48 100 42 51

Figure 3 Plot of observed (target) and predicted (output) quantiles for Q20 (CANFIS based model) for 90 independent test catchments

Table 3 Median relative error (%) values for the CANFIS model and best QRT model from ARR Project 5 (BGLS-QRT-ROI, Rahman et al., 2012)

ARI (Years) CANFIS BGLS-QRT-ROI (from Rahman et al., 2012)

NSW+ ACT VIC QLD TAS2 188 40 37 39 305 48 36 35 32 2510 51 36 35 31 2420 34 31 33 29 2750 59 32 40 31 28100 42 35 44 31 29

6 Conclusions

The paper examines the application of Co-Active Neuro Fuzzy Inference System (CANFIS) based regional flood frequency analysis (RFFA) method in eastern states of Australia. The data from four states (NSW, VIC, TAS and QLD) were combined to form one region consisting of 452 stations. The



CANFIS presents better model than the OLS-based QRT method where the CANFIS shows the median relative error values in the range of 34% to 59% (except for Q2) and median ratio of predicted and observed flood quantiles in the range of 0.79 to 1.31 (except for Q2). This study also compares the results obtained from the CANFIS with the BGLS-QRT-ROI method (ARR Project 5 model) and it has been found that the BGLS-QRT-ROI model outperforms the CANFIS based RFFA models. This is important to note that the results obtained from BGLS-QRT-ROI model are based on individual states whereas the results obtained by CANFIS are obtained when the data from all states are combined to form one region. Hence, the ANFIS model is likely to perform better when the larger data set is used. This model may also be used in other parts of the world with larger and good quality data set. Acknowledgements The authors acknowledge Engineers Australia, Australian Bureau of Meteorology and various state water agencies in Australia for providing data.

References Acreman MC, Sinclair CD (1986). Classification of drainage basins according to their physical characteristics and application for flood

frequency analysis in Scotland, Journal of Hydrology, 84(3), 365-380. Alecsandru C and Ishak S (2004). Hybrid model-based and memory-based traffic prediction system. Transportation Research Record: Journal

of the Transportation Research Board 1879(1), 59-70. Aytek A (2009). Co-Active neuro-fuzzy inference system for evapotranspiration modelling. Soft Computing, 13(7), 691-700. Aziz K, Rahman A, Fang G, Shreshtha S (2013). Application of Artificial Neural Networks in Regional Flood Frequency Analysis: A Case

Study for Australia, Stochastic Environment Research & Risk Assessment. DOI 10.1007/s00477-013-0771-5. Aziz K, Rahman A, Fang G, Haddad K, Shrestha S (2010). Design flood estimation for ungauged catchments: Application of Artificial

Neural Networks for eastern Australia. World Environment and Water Resource Congress, ASCE, Providence, Rhodes Island, USA. Aziz K, Rahman A, Fang G, Shrestha S (2011). Artificial Neural Networks Based Regional Flood Estimation Methods for Eastern Australia:

Identification of Optimum Regions. 33rd Hydrology and Water Resources Symposium, 26 June-1 July 2011, Brisbane, Australia. Bates BC, Rahman A, Mein RG, Weinmann PE (1998). Climatic and physical factors that influence the homogeneity of regional floods in

south-eastern Australia. Water Resources Research, 34(12), 3369-3382. Benson MA (1962). Evolution of methods for evaluating the occurrence of floods. U.S. Geological Surveying Water Supply Paper, 1580-A,

30pp. Blöschl G, Sivapalan M (1997). Process controls on regional flood frequency: Coefficient of variation and basin scale, Water Resources

Research, 33, 2967-2980. Burn DH (1990). Evaluation of regional flood frequency analysis with a region of influence approach, Water Resources Research, 26(10),

2257-2265. Cunnane C (1988). Methods and merits of regional flood frequency analysis, Journal of Hydrology, 100, 269-290. Dalrymple T (1960). Flood frequency analyses. U.S. Geological Survey Water Supply Paper 1543-A, 11-51. Deka P Chandramouli V (2003). A fuzzy neural network model for deriving the river stage–discharge relationship. Hydrol Sci J 48(2):197–

209. Griffis VW, Stedinger JR (2007). The use of GLS regression in regional hydrologic analyses. Journal of Hydrology, 344, 82-95. Guse B, Thieken AH, Castellarin A, Merz B (2010). Deriving probabilistic regional envelope curves with two pooling methods, Journal of

Hydrology, 380(1-2), 14-26. Haddad K, Rahman A, Weinmann PE, Kuczera G, Ball JE (2010). Streamflow data preparation for regional flood frequency analysis: Lessons

from south-east Australia. Australian Journal of Water Resources, 14(1), 17-32. Haddad K, Rahman A, Stedinger JR (2011). Regional Flood Frequency Analysis using Bayesian Generalized Least Squares: A Comparison

between Quantile and Parameter Regression Techniques, Hydrological Processes, 25, 1-14, DOI: 10.1002/hyp.8189. Haddad K and Rahman A (2011). Regional flood estimation in New South Wales Australia using Generalised Least Squares Quantile

Regression. Journal of Hydrologic Engineering, ASCE, 16, 11, 920-925. Haddad K, Rahman A (2012). Regional flood frequency analysis in eastern Australia: Bayesian GLS regression-based methods within fixed

region and ROI framework – Quantile Regression vs. Parameter Regression Technique, Journal of Hydrology, (2012), doi:10.1016/j.jhydrol.2012.02.012.

Horikawa S, Furuhashi T, Uchikawa Y (1992). On fuzzy modelling using fuzzy networks with back-propagation algorithm. IEEE Trans, Neural Network, 3(5), 801-806.

Hosking JRM, Wallis JR (1993). Some statistics useful in regional frequency analysis, Water Resources Research, 29(2), 271-281. Institution of Engineers Australia (I. E. Aust.) (1987, 2001). Australian Rainfall and Runoff: A Guide to Flood Estimation. Editor: D.H.

Pilgrim, Vol.1, I. E. Aust., Canberra.



Ishak E, Haddad K, Zaman M, Rahman A (2011). Scaling property of regional floods in New South Wales Australia, Natural Hazards, 58, 1155-1167. DOI: 10.1007/s11069-011-9719-6.

Jacquin AP, Shamseldin AY, 2006. Development of rainfall– runoff models using Takagi–Sugeno fuzzy inference systems. Journal of Hydrology 329, 154–173.

Jang JSR (1993). ANFIS: adaptive-network-based fuzzy inference system. IEE Transactions on Systems, Man and Cybernetics, 23(3), 665-685. Jang JSR, Sum CT, Mizutani E (1997). Neuro-fuzzy and soft computing, Prentice-Hall, New Jersey. Jin Y, Jiang J, Zhu J (1995). Neural network based fuzzy identification and its applications to modeling and control of complex systems. IEEE

Trans Syst Man Cybern 25(6), 990–997. Jiapeng H, Zhongmin L, Zhongbo Y (2003). A modified rational formula for flood design in small basins, Journal of the American Water

Resources Association, 39(5), 1017-1025. Keller JM, Krishnapuram R, Rhee FCH (1992a). Evidence aggregation networks for fuzzy logic inference. IEEE Trans, Neural Network, 3(5),

761-769. Keller JM, Yager RR, Tahani H (1992b). Neural network implementation of fuzzy logic rules with neural networks. International Journal of

Approximate Reasoning, 6, 221-240. Kjeldsen TR, Jones D (2009). An exploratory analysis of error components in hydrological regression modelling. Water Resources Research,

45, W02407, doi:10.1029/2007WR006283. Kjeldsen TR, Jones DA (2010). Predicting the index flood in ungauged UK catchments: On the link between data-transfer and spatial model

error structure, Journal of Hydrology, 387(1-2), 1-9, doi:10.1016/j.jhydrol.2010.03.024. Kisi O (2005). Suspended sediment estimation using neuro-fuzzy and neural network approaches. Hydrol Sci J 50(4):683–696. Kosko B (1992). Neural Networks and Fuzzy Systems: A Dynamical Systems Approach to Machine Intelligence., Prentice-Hall, Inc.,

Englewood Cliffs, NJ. Kuczera G (1999). Comprehensive at-site flood frequency analysis using Monte Carlo Bayesian inference. Water Resources Research, 35, 5,

1551-1557. Loukas YL (2001). Adaptive neuro-fuzzy inference system: an instant and architecture-free predictor for improved QSAR studies. Journal of

Medicinal Chemistry, 44, 2772–2783. Maguire LP, Roche B, McGinnity TT, McDaid LJ (1998). Predicting a chaotic time series using a fuzzy neural network. Information

Sciences, 112, 125–136. Mulvany TJ (1851). On the use of self-registering rain and flood gauges, Inst. Civ. Eng. (Ireland) Trans, 4(2), 1-8. Nathan RJ, McMahon TA (1990). Identification of homogeneous regions for the purpose of regionalisation, Journal of Hydrology, 121, 217-

238. Nayak PC, Sudheer KP et al. (2004). A neuro-fuzzy computing technique for modelling hydrological time series. Journal of Hydrology,

291(1–2), 52-66. NERC (1975). Flood studies report, 5 Volumes, Natural Environment Research Centre (NERC), London. Nourani V, Kisi O et al. (2011). Two hybrid Artificial Intelligence approaches for modelling rainfall–runoff process. Journal of Hydrology,

402(1–2), 41-59. Ouarda TBMJ, Ba KM, Diaz-Delgado C, Carsteanu C, Chokmani K, Gingras H, Quentin E, Trujillo E, Bobee B (2008). Intercomparison of

regional flood frequency estimation methods at ungauged sites for a Mexican case study, Journal of Hydrology, 348, 40-58. Pallard B, Castellarin A, Montanari A (2009). A look at the links between drainage density and flood statistics, Hydrology and Earth System

Sciences (HESS), 13, 1019-1029. Palit AK, Popovic D (1999). Forecasting chaotic time series using neurofuzzy approach. In: Proc IEEE IJCNN Washington DC, vol 3, pp

1538–1543. Palit AK, Popovic D (2000). Intelligent processing of time series using neuro-fuzzy adaptive genetic approach. In: Proc. of IEEE-ICIT

Conference, Goa, India, vol 1, pp 141–146. Palit AK, Popovic D (2005). Computational intelligence in time series forecasting; theory and engineering applications. Springer, Heidelberg,

363 p. Pandey GR, Nguyen VTV (1999). A comparative study of regression based methods in regional flood frequency analysis. Journal of

Hydrology, 225, 92-101. Pegram GGS, Parak M (2004). A review of the regional maximum flood and rational formula using geomorphological information and

observed floods, Water South Africa, 30(3), 377-392. Principe JC, Euliano NR, Lefebvre, WC (2000). Neural and Adaptive Systems, John Wiley & Sons, Inc.



Rao AR, Srinivas VV (2008). Regionalization of Watersheds: An Approach Based on Cluster Analysis, Springer. Roger JSCTS, Eiji M (1997). Neuro-fuzzy and soft computing, Englewood Cliffs, Prentice Hall. Rahman A, Bates BC, Mein RG, Weinmann PE (1999). Regional flood frequency analysis for ungauged basins in south–eastern Australia.

Australian Journal of Water Resources, 3, 2, 199-207. Rahman A, Haddad K, Zaman M, Kuczera G, Weinmann PE (2011). Design flood estimation in ungauged catchments: A comparison

between the Probabilistic Rational Method and Quantile Regression Technique for NSW. Australian Journal of Water Resources, 14, 2, 127-137.

Shafie EA, Taha RM, Noureldin A (2007). A neuro-fuzzy model for inflow forecasting of the Nile river at Aswan high dam,Water Res. Management. doi:10.1007/s11269-006-9027.

Shiri J and Kisi O (2010). Short-term and long-term streamflow forecasting using a wavelet and neuro-fuzzy conjunction model. Journal of Hydrology, 394(3), 486-493.

Shu C, Ouarda TBMJ (2008). Regional flood frequency analysis at ungauged sites using the adaptive neuro-fuzzy inference system, Journal of Hydrology, 349, 31-43.

Stedinger JR, Tasker GD (1985). Regional hydrologic analysis - 1. Ordinary, weighted and generalized least squares compared, Water Resources Research, 21, 1421-1432.

Tabari H, Talaee PH, Abghari H (2012). Utility of coactive neuro-fuzzy inference system for pan evaporation modeling in comparison with multilayer perceptron. Meteorol Atmos Phys, (2012) 116:147–154.

Takagi H, Hayashi I (1991). Neural Network driven fuzzy reasoning, International Journal of Approximate Reasoning, 5(3), 191-212. Takagi T, Sugeno M (1983). Derivation of fuzzy control rules from human operator’s control actions. Proceedings of the IFAC symposium

on fuzzy information, knowledge representation and decision analysis. Takagi T, Sugeno M (1985). Fuzzy identification of systems and its applications to modeling and control. Systems, Man and Cybernetics,

IEEE Transactions, (1), 116-132. Talei A, Chua LHC et al. (2010a). A novel application of a neuro-fuzzy computational technique in event-based rainfall–runoff modelling.

Expert Systems with Applications 37(12), 7456-7468. Talei A, Chua LHC et al. (2010b). Evaluation of rainfall and discharge inputs used by Adaptive Network-based Fuzzy Inference Systems

(ANFIS) in rainfall–runoff modeling. Journal of Hydrology 391(3): 248-262. Tasker GD, Eychaner JH, Stedinger JR (1986). Application of generalised least squares in regional hydrologic regression analysis. US

Geological Survey Water Supply Paper 2310: 107–115. Thomas DM, Benson MA (1970). Generalization of streamflow characteristics from drainage-basin characteristics, U.S. Geological Survey

Water Supply Paper 1975, US Governmental Printing Office. Turan ME, Yurdusev MA (2009). River flow estimation from upstream flow records by artificial intelligence methods, Journal of Hydrology,

369, 71–77. Zrinji Z, Burn DH (1994). Flood frequency analysis for ungauged sites using a region of influence approach, Journal of Hydrology, 153(1-4),

1-21.

Technical Paper


Effectiveness of Rainwater Harvesting System in Peri-Urban Region: A Case Study for the Greater Sydney 1

E. Hajani a, * and A. Rahmanb

aSchool of Engineering, University of Duhok, Iraq bSchool of Computing Engineering and Mathematics, University of Western Sydney, Australia

Abstract: In recent years, rainwater harvesting has become quite popular in Australia. In main cities, rainwater tanks are used to save mains water, but in peri-urban and rural areas, rainwater tanks are used as a sole water supply means to many households. In this case, the reliability of rainwater tank to meet given water demand is important. This study presents the development of a rainwater tank model that can determine the tank size which can meet most of the water demand in peri-urban areas. The model considered is a continuous simulation type, with a daily time step. The model is applied to 10 different locations in the Greater Sydney area and assesses the feasibility of 8 different rainwater tank sizes, 1 kL to 20 kL at ten study locations. Nine of these ten locations are situated in the peri-urban areas of Greater Sydney, which are Campbelltown, Hornsby, Penrith, Richmond, Castlereagh, Wallacia, West Pennant Hills, Moss Vale and Cataract Dam. The other location (Parramatta) is situated in the approximate centre of the Greater Sydney, which is used for bench marking the water use in peri-urban and metropolitan Sydney regions. The daily rainfall data at each of the ten selected locations was obtained from the Australian Bureau of Metrology. The rainfall data lengths were in the range of 31 years to 150 years, with an average of 73 years. In this study it has been found that a 7.5 kL tank can meet almost 100% of the demand for toilet and laundry use, but in case of the irrigation use a much larger tank is needed. It has been shown that irrigation demand cannot be met 100% even with a tank size of 20 kL. A set of regression equations are developed which can be used to estimate water savings and reliability using the mean annual rainfall data. The method presented in this paper can also be adapted to other Australian states and countries to estimate water savings and reliabilities of rainwater harvesting system using daily rainfall data. Keywords: Rainwater tanks, rainwater harvesting system, reliability, water saving, peri-urban region, water sensitive urban design. 1. Introduction In recent years, rainwater tank has become popular in Australia. This is due to a greater environmental awareness among the general public, a long-lasting drought in the last decade and subsidy provided by various government authorities in Australia to install a rainwater harvesting system (RWHS) (Coombes et al., 2003). Several alternative water sources have been examined in recent years, the most popular one is the RWHS as it is a simple low-cost technique requiring minimum expertise to operate. Most of the local councils in Australia encourage installation of RWHS by increasing rebate for the RWHS (Eroksuz and Rahman, 2010). The major advantages of the RWHS in urban areas include saving mains water and achieving a greater flexibility in water use during water restriction period as rainwater can be used for non-potable purposes such as gardening, washing of cars and hard surfaces during the period of water restrictions imposed by urban authorities during low water level period in reservoirs. However, in rural and peri-urban regions of Australia where mains water is not available, RWHS are used to source water for meeting principal water demand such as drinking, washing and gardening. It should be recognized that, the design of a RWHS is concerned with determining the tank size required to maintain enough water in the tank to ensure water supply for the desired purposes without much reliance on other water sources. In this regard, reliability can be defined as the proportion of days when a RWHS is likely to meet the intended water demand. An unnecessarily large tank is a misuse of resources and space; on the other hand, a RWHS, which is too small, may not serve its intended purpose and there is likely to be a shortage of water from the tank on a regular basis, and users will have to depend on other water sources. Hence, a RWHS should be designed according to the need of the household and a common size is not the best solution in most locations as rainfall varies over space.

1 Paper JHER006 submitted 14/08/2013; accepted for publication after peer review and subsequent revision on 21/10/2013 *Corresponding author may be contacted at [email protected]

“Effectiveness of Rainwater Harvesting System in Peri-Urban Region …” Hajany & Rahman


There have been many studies demonstrating the benefit of having a RWHS in the house and also on design and performance criteria of RWHS. For example, Coombes et al. (2003) examined the performance of a RWHS in an old house in Maryville suburb of Newcastle in Australia. They considered three common water uses i.e. toilet, hot water and outdoor. Ghisi et al. (2009) evaluated the financial viability of installing a RWHS in five towns in Santa Catarina State in southern Brazil where they considered car washing at petrol stations in the City of Brasilia. The results showed that an increase in the tank size enhanced the reliability of the RWHS notably in meeting the desired demand. Domènech and Saurí (2010) examined the efficiency of RWHS for the two main types of buildings prevalent in the Metropolitan area of Barcelona (Spain) by analysing users’ practices and perceptions, drinking water savings and economic costs. They carried out the financial viability of the RWHS in single and multi-family buildings and found that in single-family households an expected payback period would be between 33 to 43 years depending on the tank size and for a multi-family building a payback period was 61 years for a 20 m3 tank. In Australian studies by Van der Sterren et al. (2009, 2012 and 2013) it was found that although RWHS can be very useful to provide required water in peri-urban regions of Sydney, tank water quality may not always meet the drinking water standard in relation to heavy metals and pathogens. They also found that RWHS can enhance the stormwater runoff quality, which in turn would enhance the health of urban waterways. Kus et al. (2011) compared the water quality with the Australian Guidelines for Water Recycling (2009). They used eleven rainwater tanks located in the Sydney metropolitan area and five rainwater tanks located in peri-urban regions of Sydney. They demonstrated that simple yet effective treatments could bring the water quality to desired standards for non-potable uses. Imteaz et al. (2011a) proposed a daily water balance model for optimisation of RWHS within the campus of Swinburne University of Technology in Melbourne, Australia. They demonstrated the usefulness of large rainwater tanks when these are connected with large roof areas. Furthermore, they found that total construction costs of RWHS can be recovered within 15 to 21 years depending on the tank size, climatic conditions and future water price increase rate. Campisano and Modica (2012) presented a dimensionless methodology for the optimal design of domestic RWHS in Sicily, Italy. Seventeen rainfall stations were used in the methodology which was based on the results of daily water balance simulations. The results showed that the economical convenience of large tanks would decrease as rainwater availability decreases. Rahman et al. (2012) examined the water savings potential, reliability of water supply, financial benefits, and the adequacy of the current government rebate for a RWHS in detached house at ten different locations in Greater Sydney, Australia. They suggested that government authorities in Sydney should maintain or possibly increase the rebate for the RWHS to enhance its acceptance. Ghisi and Schondermark (2013) presented an investment feasibility analysis of the RWHS for the residential sector for five towns in Santa Catarina State, southern Brazil. They observed that the ideal tank capacity can be conservative for high rainwater demands and in such cases; an investment feasibility analysis should be carried out in order to obtain a more appropriate tank capacity. It was noted that rainwater use would be economically feasible for most cases; and the higher the rainwater demand, the greater the feasibility. Mehrabadi et al. (2013) assessed the applicability and performance of a RWHS to supply daily non-potable water. They found that in humid climate, it is possible to supply about 75% of non-potable water demand by storing rainwater from larger roof areas. This paper investigates the potential of a RWHS in peri-urban areas of Greater Sydney where semi-rural and rural properties depend on RWHS to meet most of their household water needs including drinking, washing and irrigation. Here, reliability of a RWHS is quite important as there is no mains water to supplement the RWHS. In this case, if a RWHS cannot meet the demand, water is to be bulk supplied into the tank, or an alternative groundwater well/bore needs to be built. The findings of this study would provide useful guidance in selecting the most appropriate RWHS in this region of Sydney. 2. Water balance model In this study, to simulate the performance of a RWHS, a water balance simulation model (WBSM) on a daily time scale was built in Excel. The developed model accounted for various factors for a RWHS such as tank size, daily rainfall, losses, daily water demand, mains top up and tank spillage. The adopted approach was very similar to Su et al. (2009) and Rahman et al. (2012). The rainfall was regarded as inflow and the release and possible spill as outflow. The release was estimated based on the following equations: Rt = Dt if It + St-1 ≥ Dt (1) Rt = It + St-1 if It + St-1< Dt (2) where Dt is the daily demand (m3) on day t, St-1 is the tank storage at the end of the previous day (m3), Rt is release from rainwater tank (m3) and It is inflow (m3). Spill SPt (m

3) is calculated from the following equations:



SPt = It + St-1 – Dt – SMAX if It + St-1 - Dt> SMAX (3) SPt = 0 if It + St-1 - Dt ≤ SMAX (4) where SMAX is the design storage capacity (m3). The tank storage St at the end of day t was calculated using the following equations: St = SMAX if SPt> 0 (5) St = St-1 + It – Rt if SPt ≤ 0 (6) The reliability of RWHS was calculated as the ratio of the number of days when the intended demand was met fully by the available rainwater and the total number of simulated days. 3. Study area and data In this paper, ten different locations in the Greater Sydney are considered as shown in Figure 1. Nine of these ten locations are situated in the peri-urban regions of Greater Sydney, which are Campbelltown, Hornsby, Penrith, Richmond, Castlereagh, Wallacia, West Pennant Hills, Moss Vale, and Cataract Dam. The other location (Parramatta) is situated in the approximate centre of the Greater Sydney, which is used for bench marking the water use in peri-urban and metropolitan Sydney regions. For each of these ten locations, a daily rainfall station was selected and the data was obtained from the Australian Bureau of Metrology (the list of stations is provided in Table 1). The data lengths were in the range of 31 years to 150 years, with an average of 73 years. The average monthly rainfall distribution in the Greater Sydney region is shown in Figure 2, which shows that July to Sep are the lowest rainfall months, which are also the winter season requiring very little irrigation due to minor gardening activity in this region during winter months.

Figure 1 Ten different locations in Greater Sydney selected in this study In this study, three different combinations of water use were considered: (i) toilet and laundry (ii) irrigation and (iii) a combination of toilet, laundry and irrigation (combined use). The feasibility of 8 different rainwater tank sizes, from 1 kL to 20 kL was examined. A hypothesized new development was considered at each of the study locations with a single household having 4 occupants. A total site area of 450 m2 was considered with a roof, lawn and impervious areas of 200 m2, 150 m2 and 100 m2, respectively. The rainwater tank was considered to be located above ground. The water demand data for the toilet, laundry and irrigation use for residential properties were obtained from Sydney Water. The toilet considered in the study was a 6-litre AAA rated dual flush toilet, and a frequency of toilet use of three times per person per day was assumed, which was equivalent to 0.018 kL of water use per person per day. The washing machine considered in this study was rated AAAA. It was assumed that the washing machine had a volume of 50 ltires and it would be used 3 times per week, which was equivalent to 0.0215 kL/day of water use. The irrigation demand per square metre of lawn area was assumed to be 10mm per day as per Sydney Water recommended value in this region.



Figure 2 Average monthly rainfalls in Sydney region

Table 1 Study locations and details of selected daily rainfall data

Location Rainfall station Period of rainfall record Average annual rainfall

(mm) Campbelltown 068007 1900 - 2009 743

Hornsby 066158 1936 - 2009 1,325 Parramatta 066124 1966 - 2009 963

Penrith 063185 1970 - 2009 971 Richmond 067021 1902 - 2003 800 Castlereagh 067002 1939 - 2011 685

Wallacia Post Office 067029 1943 - 2011 847 West Pennant Hills 067098 1943 - 2006 1,028

Moss Vale 068195 1971 - 2009 1,089 Cataract Dam 068016 1935 - 2011 1,080

4. Results The reliability of rainwater tanks of different sizes and water use across all the ten locations are shown in Figures 3 to 5. Figure 3 shows that on average 1 kL rainwater tank can meet the demand for toilet and laundry use for 87% of the days in a year, which increases to 97% for a 5 kL tank size and 100% for a 20 kL tank size. Among the ten locations considered here, Hornsby has the highest reliability (99.26% for a 5 kL tank) and Castlereagh has the lowest one (79% for a 5 kL tank). This reliability is well related with the average annual rainfall values at these two locations (1,325 mm and 685 mm for Hornsby and Castlereagh, respectively). Figure 3 also shows that the differences in reliability across different locations reduces as the tank size increases and a rainwater tank of 7.5 kL size can meet the demand for toilet and laundry use for about 100% at the nine locations; on the other hand, a rainwater tank of 7.5 kL size can meet the demand for toilet and laundry use for about 85% at Castlereagh. This suggests that in peri-urban areas of Sydney, a 7.5 kL rainwater tank should be adopted to meet toilet and laundry demands to ensure reliable supply and to maximize water conservation. Figure 5 presents reliability values for the toilet, laundry and irrigation use (combined use). It is found that the reliability increases with the increasing rainwater tank size as expected. The highest and lowest reliability values for a 5 kL tank are found to be for Hornsby (69%) and Castlereagh (32%), respectively, which is similar to the toilet and laundry use (see Figure 3). The average reliability over the ten locations is very close to the reliability for West Pennant Hills as can be seen in Figure 4. The reliability for combined use does not reach 100% for any of the ten locations even for a 20 kL tank size. Furthermore, there is a notable difference in reliability across different locations in Sydney, which suggests that a common tank size is not appropriate for Greater Sydney as practiced currently.

0

20

40

60

80

100

120

140

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Rainfall (m

m)



Figure 3 Reliability of rainwater tanks at ten selected locations for toilet and laundry use As can be found in Figure 4, the reliability for irrigation use at all the ten locations is smaller than for toilet and laundry use (shown in Figure 3). For a 20 kL tank, the reliability values for Hornsby and Castlereagh are 82% and 41%, respectively. This result shows that reliability value varies significantly at these two locations, which is well related with the average annual rainfalls (mean annual rainfall of 1,325 mm and 685 mm for Hornsby and Castlereagh, respectively). The average reliability for the ten locations is very close to the reliability of West Pennant Hills. These results highlight that even with a 20 kL rainwater tank, the reliability of 100% cannot be achieved for irrigation use at any of the ten study locations. This is due to the fact of high irrigation demand in the Greater Sydney region, which is characterized by periods of drought spells lasting over weeks and even months in almost all the years on record.

Figure 4 Reliability of rainwater tanks at ten selected locations for irrigation use Figure 6 shows the water savings for toilet and laundry use from a 7.5kL rain water tank; after about 10 kL size, the water savings becomes nearly constant for all the ten locations. This is due to the fact that water utilization from the tank depends on the number of users in the house, and for a larger tank there would be much more unutilised water overflowing from the tank. For a 5 kL tank, average annual water savings range from 27 kL (Castlereagh) to 34 kL (Hornsby) with a mean value of 32 kL, which suggests a smaller variation across the ten locations in water savings for toilet and laundry use.

0%

20%

40%

60%

80%

100%

120%

0 5 10 15 20 25

Rel

iab

ilit

y (%

)

Tank size (kL)

Toilet and Laundry ReliabilityPenrith

Campbeltown

Hornaby

Parramatta

Richmond

Castlereagh

Wallacia Post Office

West Pennant Hills

Moss Vale

Cataract Dam

Average

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

0 5 10 15 20 25

Rel

iab

ilit

y (%

)

Tank size (kL)

Irrigation Reliability

Penrith

Campbeltown

Hornaby

Parramatta

Richmond

Castlereagh


West Pennant Hills

Moss Vale

Cataract Dam

Average



Figure 5 Reliability of rainwater tanks at the ten selected locations for combined use

Figure 6 Water savings for toilet and laundry use

Figure 7 shows that water savings for irrigation use increases with increasing rainwater tank size as expected. For a 5 kL tank, the average annual water savings range from 58 kL (Penrith) to 45 kL (Castlereagh) with the mean value for the ten locations is 54 kL. Figure 8 shows that the differences in water savings for combined use for the ten different locations reduce as the size of rain water tanks reduces. The water saving increases with the increasing rain water tank size as expected. For a 5 kL tank, Penrith has the highest average annual water savings (62 kL), and Castlereagh has the lowest annual water savings (50 kL) and Wallacia corresponds very well to the average annual water savings over the ten locations (56 kL). It should be noted here that the reliability reported in this study is based on the long-term daily rainfall data and represents the average reliability at a given location; the reliability will differ from year to year depending on the rainfall pattern of the location, e.g. for a wet year, reliability will be higher than a dry year, as found by Imteaz et al., 2011a, b; 2012; 2013).

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

0 5 10 15 20 25

Rel

iab

ilit

y (%

)

Tank size (kL)

Combined ReliabilityPenrith

Campbeltown

Hornaby

Parramatta

Richmond

Castlereagh


West Pennant Hills

Moss Vale

Cataract Dam

Average

0

5

10

15

20

25

30

35

40

0 5 10 15 20 25

Wat

er s

avin

g (k

L)

Tank size (kL)

Toilet and Laundry Water SavingsPenrith

Campbeltown

Hornaby

Parramatta

Richmond

Castlereagh


West Pennant Hills

Moss Vale

Cataract Dam

Average



Figure 7 Water savings for irrigation use

Figure 8 Water savings for combined use

Table 2 shows the correlation coefficients of the relation between reliability, water saving and the size of rain water tanks for the ten different locations. Table 2 shows that Parramatta and Castlereagh have the lowest and highest correlation coefficients for toilet and laundry use. Also the combined and irrigation uses of the all ten locations exhibit strong correlation between the reliability, water saving and rainwater tanks size.

Table 2 Correlation coefficients between reliability, water saving and the size of rain water tanks

Locations Toilet and laundry Irrigation Combined use

Reliability Water saving reliability water saving reliability water saving Penrith 0.890 0.892 0.986 0.992 0.987 0.993

Campbelltown 0.930 0.931 0.990 0.995 0.990 0.996Hornsby 0.929 0.931 0.982 0.989 0.991 0.988

Parramatta 0.850 0.853 0.984 0.991 0.993 0.991

Richmond 0.926 0.927 0.987 0.993 0.996 0.993 Castlereagh 0.953 0.956 0.988 0.995 0.990 0.995

Wallacia Post Office 0.937 0.938 0.986 0.993 0.987 0.992

West Pennant Hills 0.910 0.913 0.983 0.990 0.989 0.990

Moss Vale 0.881 0.884 0.983 0.988 0.991 0.988

Cataract Dam 0.893 0.895 0.982 0.990 0.990 0.990

0

20

40

60

80

100

120

0 5 10 15 20 25

Wat

er s

avin

g (k

L)

Tank size (kL)

Irrigation Water SavingsPenrith

Campbeltown

Hornaby

Parramatta

Richmond

Castlereagh

Wallacia PostOfficeWest Pennant Hills

Moss Vale

Cataract Dam

0

20

40

60

80

100

120

0 5 10 15 20 25

Wat

er s

avin

g (k

L)

Tank size (kL)

Combined Water SavingsPenrith

Campbeltown

Hornaby

Parramatta

Richmond

Castlereagh


West Pennant Hills

Moss Vale

Cataract Dam

Average



For 5 kL rain water tank in the ten selected locations in Greater Sydney, the average annual rainfall is strongly correlated with the reliability for the (a) toilet and laundry use than for (b) irrigation and (c) combined use as expressed by the following equations: R1 = -0.394 + 0.195 × ln(AAR), r = 0.677 (7) R2 = -3.129 + 0.535 × ln(AAR), r = 0.891 (8) R3=-2.974 + 0.509 × ln(AAR), r = 0.901 (9) Where R1, R2, R3 are reliabilities (%) for (a) toilet and laundry, (b) irrigation and (c) combined use respectively and AAR is average annual rainfall value (mm) and r is correlation coefficient. Indeed, it is found that the average annual water savings for 5 kL rainwater tank in case of (c) combined use has higher correlations with the average annual rainfall than for (a) toilet and laundry use and (b) irrigation use as shown by the following equations: AWS1 = -13.69 + 6.747 × ln(AAR), r =0.638 (10) AWS2 = 89.09 + 5.91 × ln(AAR), r =0.235 (11) AWS3 = -18.65 + 11.04 × ln(AAR), r =0.614 (12) Where AWS1, AWS2, AWS3 are average annual water savings (kL) for (a) toilet and laundry, (b) irrigation and (c) combined use, respectively. 5. Conclusion

This paper investigates the performance of a rainwater harvesting system (RWHS) in peri-urban regions of Greater Sydney, Australia. Ten locations are selected from the peri-urban regions of Greater Sydney. It has been found that reliability of water supply increases as tank size increases for all the three water use considered here. For toilet and laundry use, a 7.5 kL tank can meet almost 100% of the demand, and hence this is the recommended tank size in peri-urban regions of Sydney. However, irrigation demand cannot be met 100% of the days in a year even with a tank size of 20 kL. It has also been found that reliability of a RWHS is highly dependent on mean annual rainfall of the location. A set of regression equations are developed which can be used to estimate water savings and reliability using the mean annual rainfall data at any arbitrary location in the peri-urban regions of Greater Sydney. The method presented here can also be applied to other Australian states and countries to estimate water savings and reliabilities of RWHS. 6. Acknowledgements Acknowledgment is made to Mr Joseph Keane for providing the program to carry out the analysis and to Australian Bureau of Meteorology for providing the daily rainfall data. References Campisano A, Modica C (2012). Optimal sizing of storage tanks for domestic rainwater harvesting in Sicily. Resources, Conservation &

Recycling, 63, 9-16. Coombes PJ, Kuczera G, Kalma JD (2003). Economic water quantity and quality impacts from the use of a rainwater tank in the inner city.

Australian Journal of Water Resources, 7(2), 25-36. Domenech L, Sauri D (2010). A Comparative appraisal of the use of rainwater harvesting in single and multi-family buildings of the

metropolitan area of Barcalona (Spain): social experience, drinking water savings and economic costs. Journal Cleaner Production, 11, 1-11. Eroksuz E, Rahman, A (2010). Rainwater tanks in multi-unit buildings: A case study for three Australian cities. Resources, Conservation &

Recycling, 54, 1449-1452. Ghisi E, Fonseca DT, Rocha VL (2009). Rainwater harvesting in petrol stations in Brasilia: Potential for potable water savings and investment

feasibility analysis. Resources, Conservation & Recycling, 54, 79-85. Ghisi E, Schondermark PN (2013). Investment Feasibility Analysis of Rainwater Use in Residences, Water Resources Management, DOI

10.1007/s11269-013-0303-6. Imteaz MA, Shanableh A, Rahman A, Ahsan A (2011a). Optimisation of rainwater tank design from large roofs: a case study in Melbourne,

Australia. Resources, Conservation & Recycling, 55, 1022-29. Imteaz MA, Ahsan A, Naser J, Rahman A (2011b). Reliability Analysis of Rainwater Tanks in Melbourne using Daily Water Balance Model.

Resources, Conservation & Recycling, 56, 80-86.



Imteaz MA, Rahman A, Ahsan A (2012). Reliability analysis of rainwater tanks: A comparison between South East and Central Melbourne. Resources, Conservation & Recycling, 66, 1-7.

Imteaz MA, Ahsan A, Shanableh A (2013). Reliability analysis of rainwater tanks using daily water balance model: variations within a large city. Resources, Conservation & Recycling, 77, 37-43.

Kus B, Kandasamy J, Vigneswaran S, Shon HK (2011). Water quality in rainwater tanks in rural and metropolitan areas of New South Wales, Australia. Journal of Water Sustainability, 1(1), 33-43.

Mehrabadi MH, Saghafian B, Fashi FH (2013). Assessment of residential rainwater harvesting efficiency for meeting non-potable water demands in three climate conditions. Resources, Conservation & Recycling, 73, 86-93.

Rahman A, Keane J, Imteaz MA (2012). Rainwater Harvesting in Greater Sydney: Water savings, reliability and economic benefits. Resources, Conservation & Recycling, 61, 16-21.

Su M, Lin C, Chang L, Kang J, Lin M (2009). A probabilistic approach to rainwater harvesting systems design and evaluation. Resources, Conservation & Recycling, 53, 393-99.

Van der Sterren M, Rahman A, Dennis G (2013).Quality and quantity monitoring of five rainwater tanks in Western Sydney, Australia. Journal of Environmental Engineering, 139, 332-340.

Van der Sterren M, Rahman A, Dennis GR (2012). Implications to stormwater management as a result of lot scale rainwater tank systems: a case study in Western Sydney, Australia, Water Science and Technology, 65, 8, 1475-1482.

Van der Sterren M, Rahman A, Shrestha S, Barker G, Ryan G (2009). An overview of on-site retention and detention policies for urban stormwater management in the greater Western Sydney region in Australia, Water International, 34, 362-372.

Technical Paper

©EnviroWater Sydney, 2013 Journal of Hydrology and Environment Research, Vol 1, No 1 30

Validation of TRMM Rainfall for Pangani River Basin in Tanzania1

R. M. F. Haquea,

*, S. Maskeyb, S. Uhlenbrook

b, c and M. Mul

b

aBangladesh Water Development Board, Dhaka, Bangladesh

bUNESCO-IHE Institute for Water Education, P.O. Box 3015, 2601 DA, Delft, The Netherlands

cDelft University of Technology, Department of Water Resources, P.O. Box 5048, 2600 DA, Delft, The Netherlands

Abstract: Tropical Rainfall Measuring Mission (TRMM) rainfall products are one of the increasingly popular remote sensing based rainfall

estimates, which are available at 0.25o x 0.25

o spatial and up to 3 hourly temporal resolutions since January 1998 and covers 50

o S to 50

o N

around the globe. This study assesses the reliability of the TRMM rainfall for Pangani River Basin in Tanzania by comparing it with the gauge

rainfall data. Monthly rainfall data from 1998 to 2005 for 50 gauge stations and daily rainfall data from 1998 to 2003 from 12 stations are used

for the comparison. The comparison of TRMM grid rainfall with point gauge rainfall is achieved by using nearest neighborhood and bilinear

weighted interpolation methods. Application of bilinear weighted interpolation method has brought some improvement in the comparison of

results over the nearest neighborhood method. In both methods, the TRMM rainfall estimates compares better with the gauge rainfall at a

monthly time scale than at daily time scale. The comparison is also carried out at a sub-basin level by dividing the basin into 7 sub-basins. The

sub-basin rainfall is estimated from the gauge rainfall by the interpolation method based on inverse distance and elevation weighting. The

spatial rainfall comparison at sub-basin level has shown a reasonably good result. Comparing the consecutive yearly total rainfall maps, it is

seen that TRMM has reasonably captured both spatial and temporal rainfall pattern but failed to detect higher intensity rainfall close to the

mountainous parts of the basin. Five statistical performance measures namely the coefficient of determination, mean error (bias), root mean

square error, relative root mean square error and Nash-Sutcliffe coefficient are used to quantify the comparison results. In conclusion, TRMM

has performed fairly well on a monthly scale but its hydrological application in mountainous areas and for high intensity rainfall period still

requires validation with the ground based gauge rainfall before taking it as a reliable alternative.

Keywords: Rainfall estimation, TRMM, nearest neighborhood, bilinear weighted interpolation, sub-basin

1 Introduction Rainfall data availability has been highlighted as a major constraint on the effective application of water resource models, and it has been

argued that quality of rainfall inputs to the model is often more important than choice of model itself as discussed by Wilk et al. (2006). The

quality, availability and coverage of rain gauge data are particular obstacles to effective water resource planning in Africa as well as most

developing countries as per Thorne et al. (2001), Grimes and Diop (2003) and Hughes (2006). Although radar estimates of rainfall are

increasingly available in technologically developed countries like the USA and Europe, in large parts of Africa these are almost non-existent.

However, several satellite products provide complete coverage of the African continent at applicable time and space scales, and this data source

seems to be an obvious way forward for regional-to national-scale rainfall estimation.

Recently, rainfall products from Tropical Rainfall Measuring Mission (TRMM) are becoming a popular alternative rainfall data source,

particularly in large scale applications. Satellite observation based rainfall estimates are not a direct measurement of rainfall but are only an

estimate and are subject to uncertainty. It is important that these data are well validated before using for practical applications.

Bowman (2005) compared TRMM rainfall with rainfall data from 26 rain gauges from ocean buoys in tropical pacific and showed that

comparison results can be improved by properly averaging in space and/or time. For individual satellite overpasses averaged over a 1° x 1° box,

his comparison showed Relative Root Mean Square Error (RRMSE) of as high as 200% to 300%. Whereas for 32-day means over 1° x 1° boxes,

the RRMSE reduced to 40% to 70%. Collischonn et al. (2008) compared the aggregated 3 hourly TRMM rainfall estimates to daily values with

1 Paper JHER002 submitted 09/08/2013; accepted for publication after peer review and subsequent revision on 18/10/2013

*Corresponding author can be contacted at [email protected]

mailto:[email protected]

“Validation of TRMM Rainfall for Pangani River Basin …” Haque, Maskey, Uhlenbrook and Mul


the ground-level precipitation gauge data on a daily basis. Both daily TRMM and rain gauge-interpolated rain fields were then used as input to

a large-scale hydrological model for the Tapajo´s river basin, a major tributary of the Amazon; the calculated hydrographs were then compared

to observations at several stream gauges along the River Tapajos and its main tributaries. Results of the rain field comparisons showed that

satellite estimates can be a practical tool for identifying damaged or aberrant rain gauges at basin scale.

Fengge et al. (2008) took the La Plata basin in South America as a good example of a case where the use of satellite-derived precipitation could

be beneficial. He evaluated basin wide precipitation estimates from 9 yr (1998–2006) of TTRMM Multisatellite Precipitation Analysis (TMPA;

3B42 V.6) through comparison with available gauged data and the Variable Infiltration Capacity (VIC) semi-distributed hydrology model

applied to the La Plata basin. He found in general, the TMPA estimates agreed well with the gridded gauge data at monthly time scales, most

likely because of the monthly adjustment to gauges performed in TMPA. The agreement between TMPA and gauge precipitation estimates was

reduced at daily time scales, particularly for high rain rates. The TMPA-driven hydrologic model simulations were able to capture the daily

flooding events and to represent low flows, although peak flows tended to be biased upward. Hazarika et al (2005) also found TRMM

underestimated the rain for peak monsoon period for the wet rain regimes whereas it overestimated the rain for dry rain regimes of Nepal and

suggested the possibility of using satellite data for operational flood forecasting in those river basins for which real time gauge data were not

available. On the other hand, due to the overestimation and underestimation of daily rainfall measured from TRMM V5 3B42 in comparison

with the ground rain gauge data Islam and Uyeda (2007) recommended special attention for the application of V5 3B42 data to short term

hydrological problems such as flash flood. McIntyer et al. (2008) suggested that the relative performance of the satellite-based rainfall

estimation algorithms depend on what aspects of the rainfall regime are being considered. Differences between the products are large and the

use of more than one product for any application is recommended.

Uddin et al. (2008) found good correlation between the computed TRMM and measured data by the application of bilinear weighted

interpolation method. He recommended the usefulness of bilinear weighted interpolation method to derive precipitation information at any

spatial location when direct measurements are not available. TRMM-derived precipitation showed better detection of rain at low altitude

stations as compared with high elevation stations, with good scores for the (Precipitation Radar) PR product for rain rates >0.5 mm/hr as

discussed in Barros et al. (2000).

This study assesses the reliability of the TRMM rainfall for the Pangani River Basin in Tanzania by comparing it with the gauge rainfall data.

The comparison is carried out at two stages. Firstly, the gauge rainfall is compared with the TRMM rainfall estimated at the gauge location. For

this point comparison, the TRMM rainfall of 0.25×0.25 grid is interpolated for the location of the gauge by nearest neighborhood and bilinear

interpolation methods. Secondly, the comparison is carried out at sub-basin level by dividing the basin into 7 sub-basins. The sub-basin rainfall

is estimated from the gauge rainfall by inverse distance and elevation weighting method.

2 Pangani River basin gauge rainfall data The study area selected for this paper is the Pangani river basin, one of the nine river basins in Tanzania as shown in Figure 1. The total

drainage area of the basin is about 43,650 km2 with 95% of the basin area in Tanzania and 5% in Kenya. It is located in the north eastern part

of Tanzania between Latitudes 2o55’S to 5

o40’S and Longitude 36

o20’E to 39

o2’E. The source of the Pangani River originates at Mt Kilimanjaro

(the highest mountain in Africa) and Mt Meru. The Pare and Usambara Mountains, bordering the west side of the catchment, also contributes

substantial amount to the Pangani river flow. The basin is subject to high variability, in terms of altitude and local climate. Rainfall is subject to

high inter- and intra-seasonal variability. There are five main sub-basins: Ruvu, Kikuletwa, Mkomazi, Luengere and Pangani. The Kikuletwa

and Ruvu tributaries form the headwaters of the Pangani River and originate in the high rainfall areas surrounding Mt Meru and Mt

Kilimanjaro, respectively. Elevation in the basin varies from sea level at Pangani (the basin outlet) to 5895 m on top of Mount Kilimanjaro.

There is high spatial variability of rainfall in the basin, mainly characterised by the topography.

The high altitude slopes above the forest line on Mt Meru and Mt Kilimanjaro have an Afro-Alpine climate and receive in excess of 2500 mm

of rainfall per year. The middle slopes of these mountains, and the Pare and Usambara Mountains are characterized by a humid to sub-humid

tropical climate. The lower Mkomazi catchment and the lower Pangani catchment have a sub-humid to semi-arid climate. The central and

western parts of the Basin have a semi-arid to arid climate. Rainfall varies between 300 and 600 mm/year.



Majority of the rain gauges of Pangani river basin are situated on an axis along the mountain ranges of Mount Kilimanjaro, North-South Pare

Mountains and the Usambara Mountains. There is actually no rainfall station in western part of the basin. The data available was either of

short record or contained a lot of gaps.

Figure 1 Location of Pangani river basin in Tanzania

3 TRMM satellite mission and rainfall products The accurate measurement of the spatial and temporal variation of tropical rainfall around the globe remains one of the critical unsolved

problems of meteorology. TRMM, during its mission and broad sampling footprint between 35°N and 35°S, is providing some of the first

detailed and comprehensive dataset on the four dimensional distribution of rainfall and latent heating over the vastly under sampled oceanic

and tropical continental regimes. The TRMM satellite, a joint project between the United States (under the leadership of NASA's Goddard

Space Flight Centre and Japan (under the leadership of the National Space Development Agency), and the first spacecraft designed to monitor

rain over the tropics, was successfully launched on November 27, 1997. It was placed in low earth orbit as the first precipitation radar (PR) in

space, along with a 9-channel Special Sensor Microwave/Imager (SSM/I)-like passive microwave imager (TMI), an Advanced Very High

Resolution Radiometer (AVHRR)-like visible-infrared radiometer (VIRS), a lightning sensor and a cloud sensor as discussed in Zhong et al.

(2002).

3.1 TRMM product review

There are two main rainfall products based called 3B42 and 3B43. The 3B42 estimates are produced in four stages: (i) the microwave estimates

precipitation are calibrated and combined; (ii) infrared precipitation estimates are created using the calibrated microwave precipitation; (iii) the

microwave and (infrared) IR estimates are combined; and (iv) rescaling to monthly data is applied. Each precipitation field is best interpreted as

the precipitation rate effective at the nominal observation time as per JAXA (2006).

The 3B42 is to provide precipitation estimates in the TRMM regions that has the (nearly-zero) bias of the “TRMM Combined Instrument”

precipitation estimate and the dense sampling of geosynchronous IR imagery. The 3B42 is composed of two separate algorithms, which are (i)

to produce monthly IR calibration parameters, and (ii) to calibrate the merged-IR precipitation data to produce the daily adjusted merged-IR

precipitation and (Root Mean Square) RMS precipitation error estimates as per JAXA (2006).

For the 3B43 monthly data, in the estimates it includes the dependent data. It processes the TRMM Microwave Imager (TMI) 1B01, 2A12 and

Combiner (COMB) 3B31 and Merged IR data (3A44). Additionally, Geosynchronous Precipitation Index (GPI) is used to convert Visible and

Infrared Scanner (VIRS) radiance to precipitation rate as discussed in JAXA (2006). Algorithms 3B43 is executed once per calendar month to

produce the single, best estimate precipitation rate and RMS precipitation error estimate field 3B43 by combining the 3 hourly merged HQ/IR



estimates 3B42 with the monthly accumulation Climate Assessment and Monitoring System (CAMS) or Global Precipitation Climatology

Centre (GPCC) rain gauge analysis as per JAXA (2006).

The purpose of algorithm 3B42 and 3B43 is to produce TRMM rainfall retrievals merged high quality (HQ)/ infrared (IR) precipitation and

root mean square (RMS) precipitation error estimate. The data include retrievals from six different algorithms: VIRS, TMI, (Precipitation

Radar) PR and the combination with other satellite (GPI, GPCP and SSM/I).

4 Methodology 4.1 Data used for comparison

4.1.1 Gauge rainfall data

The precipitation data for the Pangani River basin is obtained via the “Pangani Basin Water Office (PBWO) in Moshi”. Due to the long time

series, data from ground based rain gauges is used for this research. The data that has been used could be divided into two groups: (i) 50 rain

gauges with monthly data from 1970 to 2005; and (ii) 32 rain gauges with daily data from 1916 to 2003.

All rain gauges are standard manual rain gauges and part of Tanzania rain gauge network maintained by the Tanzanian Meteorological Agency

in conformance with the World Meteorological Organization (WMO) standard. The rain gauges in the Pangani basin are primarily located

near permanent settlements.

4.1.2 TRMM rainfall data

For this study, eight years (1998 to 2005) observations from TRMM have been used. The data set used for this study is 3-hourly TRMM and

other rainfall estimate (3B42 V6), daily TRMM and other rainfall estimate (3B42 V6 derived) and monthly TRMM and other data sources

rainfall estimate (3B43 V6). 3B43 V6 consists of accumulated rainfall in millimeters, based on multi-satellite precipitation analysis. The data has

an area averaged over 0.250

x 0.250

longitude and latitude grid boxes (approximately 25km x 25km). The Lat-Lon plot type provides a time-

averaged data plot for a specified area. This data could be downloaded as an ASCII output from the principal web page of TOVAS.

The TRMM 3B42 records precipitation every 3 hours and the TRMM imager is saved from 12 UTC Coordinated Universal Time (UTC) of the

previous day to 12 UTC of the analysis day. Normally, the daily data rain gauges record 24-hr precipitation accumulation ending at the time of

observation and this is reported as the rainfall for the day of observation as suggested by Dingman (2002). To create similar dataset for

comparison, the 3-hourly TRMM precipitation data are accumulated to 24-hr depths depending on the rainfall observation time of the rain

gauge located in the basin. The rainfall observation time in Pangani River basin, Tanzania is 9:00 AM local time. The 3-hourly TRMM

precipitation data are accumulated from 9:00 AM on January 1, 1998 to 9:00 AM January 2, 1998 and this 24-hr depth is considered as the

rainfall for January 2, 1998. Those days with missing and suspicious TRMM data for any hour are excluded from the analysis. Daily

precipitation data from 12 daily rain gauges are assembled in a similar format, excluding the dates with missing data. Then the daily rain gauge

data files and TRMM data files are compared to create a new set of data files that included only those dates without any missing data for both

rain gauge and satellite. In this study, data from rain gauges stations over Pangani River basin, Tanzania is used to compare with TRMM

rainfall data using several statistical measures.

4.2 Indicators used for comparison

The comparison results are evaluated using five indicators, namely the coefficient of determination (R2), Nash-Sutcliff coefficient of efficiency

(CoE), root mean square error (RMSE), relative root mean square error (RRMSE) and mean error bias. These indicators are defined below.

R2 is one of the most common and most powerful statistics. A correlation is a single number that describes the degree of relationship between

two variables. It is good measure of linear association or phase error. Visually, how close the points of a scatter plot are to a straight line. It is

possible for a set of estimated values with large errors to still have good coefficient of determination with the observations. It is sensitive to

outliers and goes from 0 to 1 as discussed by Murphy (1995). For this study, the coefficient of determination addresses the question how well

did the TRMM retrievals correspond to the observed rainfall ground values.

Mean error (bias) =

n

i

FiGin 1

1 (1)



RMSE =

n

rFiGi

n 1

2)(1

(2)

RRMSE =__

1

2)(1

G

FiGin

n

i

(3)

CoE = 1.0 -

n

i i

n

i i

GG

FiG

1

2__

1

2

)(

)( (4)

Gi = Rain gauge measurements

Fi = Satellite estimates

__

G = mean of the rain gauge measurement

n = is the number of observations.

4.3 Point to grid comparison

The rain gauges and the satellite instruments make fundamentally different measurements. The gauges provide in situ high time-resolution

measurements at a point. The TRMM instruments make remote, volume-averaged measurements of hydrometeors in the atmosphere, from

which the area averaged surface rain rate over the instantaneous fields-of-view is inferred. By scanning across the orbit track, TRMM can

provide a snapshot of the rain rate over an extended region. Thus, both instruments provide only limited samples of the precipitation falling

within a region. The rain gauges have good time sampling but poor spatial sampling; while the satellite has good spatial sampling but poor

time sampling as observed by Bowman (2005).

It is frequently the case that when TRMM observes rain within a region it is not raining at the gauge. Similarly, it often rains at the gauge

between TRMM overpasses. Using only gauge data from the time of the TRMM overpass neglects a great deal of information collected by the

gauge between satellite overpasses as discussed by Bowman (2005). Only the individual pixel that contains the gauge neglects information

collected by the satellite at nearby locations. Because rainfall is correlated with itself in both space and time, more information can be obtained

by properly averaging the data in space and/or time.

4.4 Nearest neighborhood method

Here the pixel values for the grid within which rain gauges are cited are used to make comparison. Monthly TRMM precipitation data is

downloaded by using the gauge station geographic coordinates (Lat/Lon) for a specific period. When the station’s coordinate (Lat/Lon) is

specified in TOVAS, for a single point, it gives the time series for the nearest grid point and no averaging or interpolating is performed.

4.5 Bilinear weighted interpolation method

As TRMM 3B43 V6 data with a 0.250 grid is a bit coarse for local interpretation, instead of using the grid value for a particular location, or the

averaging of adjacent pixel values on 0.250 grid, the bilinear weighted interpolation is presented here which is more realistic. The data from

TRMM 3B43 dataset is used for bilinear weighted interpolation to resample the data for a specific point of interest as per Brito et al. (2003),

which in the present study is the spatial location of the rain gauge. The logic that has been adopted is to average four adjacent pixels for a

location of interest as dicussed by Kuhel and Sacchi (2003). The location (X, Y) has been kept in a 2 x 2 grid in such a manner that it occupies

a central position in the grid. An illustration is included in Figure 2. In the bilinear interpolation, a simplistic rationale is followed, i.e. for

calculating the pixel value of a particular position (X, Y), four adjacent pixel values are used. The closer the pixel is to the position (X, Y), the

more influence (weight) it will carry. The method is not merely a falling function of distance from the pixel. Rather it considers a weighted

approach based on its spatial locations in a two-dimensional space as discussed by Gribbon and Bailey (2004). The derivation of bilinear

interpolation weights can be expressed as follows as per Arnold et al. (2002):

21 )1( XSXSX xx (5)

12

1

XX

XXSx

(6)

21 )1( YSYSY yy (7)



12

1

YY

YYS y

(8)

where 21,10 XXXS where the pairs iiYX are the (Long, Lat) cordinates of the 4 grid cells as shown in Figure 2.

Figure 2 A graphical representation of weight estimation on 2 x 2 grid

Considering the above, the actual weight at any point (X, Y) in a two-dimensional space can be computed by equation (9) as suggested by

Arnold et al (2002).

22211211 1)1(,11, YXISSYXISSYXISSYXISSYXI yxyxyxyx (9)

where I is the actual pixel value.

In the current study, it was found that some improvement in the estimation of R2 by the use of bilinear weighted interpolation method over the

nearest neighborhood method for 50 monthly rainfall stations data comparison. R2 is increased by 5% to 80% for the 50 monthly rainfall

stations in Pangani River basin by the use of bilinear weighted interpolation method and it ranges from 0.18 to 0.72. Especially those stations

having poor R2 values got better R

2 by the use of bilinear weighted interpolation method. For example, comparison statistics between TRMM

and gauge monthly data for 8 stations of out 50 stations are provided in Table 1.

By the use of bilinear weighted interpolation method, the overestimation (negative mean error) by TRMM monthly data has been significantly

reduced for Mweka College of wildlife and Uru Estate station which are located in Kikuletwa and Ruvu catchment of Pangani River basin. On

the other hand, Osaki forest station and Lyamungo have got underestimation (positive mean error) which are located in Kikuletwa catchment.

By the same method no significant improvement in mean error has been observed for TPC Langasani, Osaki forest, Mwakinyumbi Sisal Estate

and Hale plantations which are located in Kikuletwa and Pangani main stream respectively. It was found some improvement in the R2 for all

the 8 stations mentioned in Table 1.

The standard deviation of bias for the 50 monthly rainfall stations data of Pangani basin is 23.70 mm/month for bilinear weighted interpolation

method whereas it is 31.31 mm/month for nearest neighbourhood method. This is an acceptable improvement over the previous method. It

deserves a special mention here that the bilinear weighted interpolation method has not brought any improvement in daily comparison over

nearest neighborhood method.



Table 1 Comparison statistics between TRMM and gauge monthly data for 8 stations by the use of nearest neighborhood and bilinear weighted

interpolation method

Station Name Altitude Nearest neighborhood method Bilinear weighted interpolation method

Mean error CoE R2

Mean error CoE R2

Shume forest station

ID 9438012

1891 -43.14 -8.99 0.57 -46.47

-9.97 0.57

Mweka College of wildlife ID

9337098

1452 -65.14 -0.22 0.2 -10.2

0.27 0.28

Osaki forest station

ID 9337121

1432 -9.68 0.05 0.1 47.62

0.09 0.18

Uru Estate

ID 9337140

1415 -62.94 -0.37 0.26 -8.87

0.39 0.41

Lyamungo

ID 9337021

1334 -43.48

0.04

0.16 25.28 0.23 0.27

TPC Langasani

ID 9337028

715 -27.26 0.06 0.48 -23.27

0.22 0.52

Mwakinyumbi Sisal Estate

ID 9538019

248

10.03

0.64 0.65 10.38

0.64 0.66

Hale plantations

ID 9538010

202

-17.89

0.46 0.57 -17.37

0.49 0.59

The average gauge monthly rainfall of the station Shume forest is 21.32 mm/month and 64.46 mm/month from TRMM. Figure 3 plots the

variation between the two estimates which is prominent for the whole time period (1998-2005). This station has an altitude of 1891m and

located on the higher part of Usambara mountain. High rainfall occurs around Usambara due to orographic effect. The precipitation in the

upper part could be different from lower part. Besides, Environmental factors such as wind speed or the error in the gauge itself could result a

negative bias of -43.14mm/month. The variation and poor CoE can be attributed to the weaker detection of TRMM at higher altitude station.

The station Osaki forest (ID 9337121) has got the lowest R2 (0.18). The variation between the two estimates is the reason for this poor

correlation as shown in Figure 4.

Figure 3 Variation between monthly gauge and TRMM rainfall for the station Shume forest station

Station 9438012

-100

0

100

200

300

400

Nov-97 May-98 Oct-98 Apr-99 Oct-99 Apr-00 Oct-00 Apr-01 Oct-01 Apr-02 Oct-02 Apr-03 Oct-03 Mar-04 Sep-04 Mar-05 Sep-05

Time

Gauge-T

RM

M [

mm

/month

]



Figure 4 Variation between monthly gauge and TRMM rainfall for the station Osaki forest station

This station is located at the slope of Mount Kilimanjaro where high rainfall occurs due to orographic effect and TRMM has limitation in

detecting orographic enhancement of rainfall as such clouds are warm-topped. Due to the unique location of this gauge more careful gauge

measurement is required for future comparison with TRMM data.

4.6 Sub-basin level comparison

TRMM provides rainfall on 0.250

x 0.250

grids which is bit course to make comparison with point gauge rainfall. Hence the DEM and sub-

basins of Pangani are resampled into 0.050 x 0.05

0 grids. Here the comparison has been done between the 0.05

0 x 0.05

0 interpolated point

gauge rainfall vs. 0.250 x 0.25

0 TRMM monthly estimates.

Algorithm used for deriving spatial rainfall from gauge rainfall

This interpolation method calculates points without data from the sum of weighted observation in the neighborhood. The weights are

proportional to the distance and elevation by equation (11) as suggested by Maskey and Venneker (2006), the influence of a rain gauge reduces

with increasing distance and elevation to it. A disadvantage of the method is the arbitrary choice for the exponent and there is no estimation

for error.

N

i

ijij pwp1

, (10)

N

i

n

ji

n

jiji ffw1

,,, / (11)

jp = station weighted rainfall for the facet j [L/T]

N = No of rainfall stations relevant to the grid cell. [-]

ip = Observed rainfall at station i [L/T]

jiw , = Station weight for the given station i and facet j [-]

The facet parameter f and the value of the exponent n for different facets are given in (Table 2).

Table 2 Parameters of stations weights per facets (in this table d is the distance between the two stations and z is the differences in

elevations between the two stations) as suggested by Maskey and Venneker (2006)

Facet Parameters f Exponent n

Distance d -2

Elevation | z | -1

Station 9337121

0

100

200

300

400

500

600

700

800

900

1000

Nov-97 May-98 Oct-98 Apr-99 Oct-99 Apr-00 Oct-00 Apr-01 Oct-01 Apr-02 Oct-02 Apr-03 Oct-03 Mar-04 Sep-04 Mar-05 Sep-05

Time

Gauge-T

RM

M [

mm

/month

]



The quality of the interpolated data is determined by Jack knife cross validation method. The radius of influence, the distance (Fd) and

elevation (Fz) weighting importance factors (Appendix-A) are determined by this method. Here radius of influence (R), Fd and Fz are adjusted

by R2 between the measured and estimated interpolated monthly rainfall for the 35 stations. Quality of the interpolated data is determined by

R2 between the measured and estimated rainfall of a station. The interpolated monthly rainfall estimates are obtained by interpolation of 35

monthly point rainfall stations for the time period 1998 to 2004 by Hykit (Appendix-A). TRMM 3B43 monthly product is validated against the

interpolated gauge data at sub-basin level. Pangani is divided into seven sub-basins. The comparison statistics (mean error, RMSE, RRMSE,

CoE, R2) are calculated in Table 3 to compare and quantify the accuracy of TRMM retrievals.

Table 3 Comparison statistics between TRMM and interpolated monthly rainfall data for 7 sub-basins

Sub-basin Mean error or

bias

[mm/month]

RMSE

[mm/month]

RRMSE

CoE (Nash-

Sutcliffe)

Co-eff detn.

sub_1 (Kikuletwa) 3.54 37.94 0.61 0.66 0.69

sub_2 (Ruvu) 2.71 35.63 0.56 0.71 0.71

sub_3 (Part of Pangani main stream) 9.44 33.58 0.58 0.68 0.71


sub_5 (Mkomazi)

-0.78 36.94 0.55 0.65 0.68

sub_6 (Luengera) 20.83 57.65 0.59 0.60 0.65


An analysis of the results during the seven years period revealed a negative bias in sub-basin 5, that was the only sub-basin where TRMM

estimated higher rainfall compared to gauge rainfall. It was seen that after sub-basin 6, sub-basin 7 got the lowest C0E (0.60) and coefficient of

determination (0.67) which was easily understandable as there were very few ground rainfall stations in this sub-basin concentrated at the out

let of the basin. The coefficient of determination for 7 sub-basins ranged from 0.65 to 0.77.

5 Conclusion TRMM satellite rainfall data can play an important part in data sparse regions as well as places where point rain gauges are not available due to

lack of access. This can bring a benefit for many countries where point rain gauges are not sufficient to capture the spatial and temporal

variability of rainfall. TRMM has the inability to capture high rainfall periods and the rainfall influenced by the orographic effect in the

mountainous regions. The use of bilinear weighted interpolation has not given a significant change in comparison statistics but it can be

claimed that it is the most realistic way to make comparison between gauge and TRMM rainfall estimates. As the TRMM grid (0.250 x 0.25

0 ) is

bit coarse, in reality it is seen that number of rainfall stations are situated in the same grid and TRMM is giving the same rainfall estimate for

all the stations in the same grid although they are far apart and their individual rainfall estimate is quite different from each other. Use of

bilinear weighted interpolation method gives separate set of estimates for different rainfall stations. Comparison between the interpolated

monthly gauge rainfall and TRMM rainfall data has given better results at sub-basin level. This is particularly interesting in the perspective of

using this data for hydrological applications.

Acknowledgement The authors are highly indebted to the Zeeland Fellowship Program of The Netherlands for providing scholarship to the first author of

the paper for the M.Sc. in Water Science and Engineering at the UNESCO-IHE, Institute for Water Eductaion, Delft, the Netherlands.

They are also thankful to B. M. C Fischer of TU Delft and Dr. Marloes Mul for their support and assistance in data collection. First

author is very much grateful to Dr. Ataur Rahman for his valuable suggestions and comments on the work. We would like to also thank the

anonymous reviewers of this paper for his positive and constructive comments. Authors would like to acknowledge the TRMM Online

Visualization and Analysis System (TOVAS) for TRMM data.

References Arnold DN, Boffi D, Falk RS (2002). Approximation by quadrilateral finite element, Mathematics of Computation, 71, 909–922.

Barros A, Joshi M, Putkonen J, Burbank DW (2000). A study of the 1999 monsoon rainfall in a mountainous region in central Nepal using

TRMM products and rain gauge observations. Geophysical Research Letters, 27, 22, 3683-3686.



Bowman, Kenneth P (2005). Comparison of TRMM precipitation retrievals with rain gauge data from ocean buoys. Journal of Climate, 18,

178–190.

Brito AE, Chan SH, Cabrera SD (2003). SAR image super resolution via 2-D adaptive extrapolation. Multidimensional Systems and Signal

Processing, 14, 83–104.

Collischonn BW, Collischonn W, Tucci CEM (2008). Daily hydrological modeling in the Amazon basin using TRMM rainfall estimates.

Journal of Hydrology, 360, 1-4, 207-216.

Dingman SL (2002). Physical Hydrology. Second edition, Waveland Pr Inc.

Fengge S, Yang H, Lettenmaier DP, (2008). Evaluation of TRMM Multisatellite Precipitation Analysis (TMPA) and its Utility in Hydrologic

Prediction in the La Plata Basin. Journal of Hydrometeorology, 9, 4, 622-640.

Gribbon KT, Bailey DG (2004). A novel approach to real-time bi-linear interpolation. In: Proceedings of the Second IEEE International

Workshop on Electronic Design, Test and Applications, pp. 126.

Grimes DIF, Diop M, (2003). Satellite-based rainfall estimation for river flow forecasting in Africa. Part 1. Rainfall estimates and hydrological

forecasts. Hydrological Sciences Journal, 48, 4, 567-584.

Hazarika MK, Kafle TP, Sharma R, Karki S, Shrestha RM, Samarkoon L (2005). Statistical approach to discharge prediction for flood

forecasting using TRMM data. In: Proceedings of the 5th Annual Mekong Flood Forum, Ho Chi Minh City, Vietnam, May 17–18, 2007.

Hughes DA (2006). Comparison of satellite rainfall data with observations from gauging station networks. Journal of Hydrology, 327, 2-4, 399-

410.

Islam MN and Uyeda H (2007). Use of TRMM in determining the climatic characteristics of rainfall over Bangladesh. Remote Sensing of

Environment, 108, 3, 264-276.

JAXA, Japan Aerospace Exploration Agency (2006). TRMM data users hand book, JAXA, Japan.

Kuhel H, Sacchi MD, 2003. Least squares wave equation migration for AVP/AVA inversion. Geophysics, 68, 262–273.

Murphy AH (1995). The coefficient of correlation and determination as measures of performance in forecast verification. American Meteorological Society, Weather Forecasting, 10, 681–688.

Maskey S, Venneker R (2006). Generation of input fields for large Scale hydrological models. UNESCO-IHE, LSHM Doc 0604, 2006-2005.

Thorne V, Coakley P, Rokke P, Grimes D, Dugdale G (2001). Comparison of TAMSAT and CPC rainfall estimates with rainfall, for southern

Africa. International Journal Remote Sensing 22, 10, 1951-74.

Uddin S, Al-Dousari A, Ramadan A, Al Ghadban A (2008). Site-specific precipitation estimate from TRMM data using bilinear weighted

interpolation technique: An example from Kuwait. Journal of Arid Environments, 72, 7, 1320-1328.

Wilk J, Kniveton D, Anderson L, Layberry R, Todd M, Hughes D, Ringrose S, Vanderpost C (2006). Estimating rainfall and water balance

over the Okavango River Basin for hydrological applications. Journal of Hydrology, 331, 1-2, 18-29.

Zhong L, Chiu H, Rui W, Teng G, Serafno (2002). Online Analysis and Visualization of TRMM and other Precipitation data sets. GSFC

Earth Data and Information Services Centre Distributed Active Archive Centre NASA/ Goddard Space Flight Centre, Greenbelt Maryland,

USA.



Appendix A Special Data Interpolation kit: Hykit

The interface of spatial data interpolation kit Hykit is shown in Figure 5 below.

Figure 5 Interface of Hykit

A station’s influence in the inverse distance interpolation is assumed to decrease as its distance from the target grid cell increases. D is the

distance between the station and the target grid cell and is the distance weighting exponent typically set to 2, which is equivalent to an

inverse-distance-squared weighting function. Elevation weighting allows the model to focus on the vertical range that is specific to a target grid

cell, thereby accommodating climate profiles that may vary in slope across the altitudinal range of the data. Here the elevation exponent is

typically set to 1.0 which is equivalent to a 1-dimentional inverse distance weighting function. The distance (Fd) and elevation (Fz) weighting

importance factors apply a measure of scaling to the vertical dimension by controlling the relative importance of distance and elevation in the

model. The influence of horizontal distance on the inter-station correlation seems to be greater overall than that of vertical distance. Thus, Fd is

typically set to 0.8 and Fz to 0.2 as shown in Figure 5.

Technical Paper


Variability in Rainfall Temporal Patterns: A Case Study for New South Wales, Australia1

W. L. Caballero* and A. Rahman School of Computing, Engineering and Mathematics, University of Western Sydney, Australia

Abstract: In Australian Rainfall and Runoff (ARR) 1987, the recommended rainfall-based Design Event Approach (DEA) of design flood estimation has a number of limitations that are likely to introduce a probability bias in the final design flood estimates. Recently, considerable research has been undertaken on the development and application of the Joint Probability Approach (JPA)/Monte Carlo Simulation Technique (MCST) to design flood estimation to overcome the limitations associated with DEA. The applications of this method with Victoria and Queensland data have shown that MCST can overcome some of the limitations associated with the DEA and MCST can produce more accurate design flood estimates. However, the wider application of the MCST needs regionalisation of various input variables to the runoff routing model which include rainfall characteristics such as duration, intensity and temporal pattern. This paper focuses on the regionalisation of the rainfall temporal patterns in New South Wales (NSW), Australia using data from 86 pluviograph stations. The regionalised temporal patterns are then applied with the MCST to obtain design flood estimates for both gauged and ungauged catchments in NSW. Using the MCST, it has been found that the application of at-site and regional temporal patterns can give up to 10% differences in flood quantile estimates. In this study, rainfall inter-event duration (a new random variable) has been tested along with other random variables (i.e. rainfall duration, rainfall intensity, initial loss, continuing loss and runoff routing model storage delay parameter) with their probability distributions in the MCST, which has been shown to provide more accurate flood quantile estimates than the DEA. The findings of this research will assist to apply the MCST in practice in NSW. The method can be adapted to other parts of Australia and similar other countries. Keywords: Design Event Approach, Australian Rainfall and Runoff, Design rainfall, Complete storm, Joint Probability Approach, Monte Carlo simulation 1. Introduction Design of hydraulic structures and other water resources planning and management tasks often requires design flood estimation which is commonly carried out by at-site flood frequency analysis provided the availability of recorded streamflow data of adequate quantity and acceptable quality. Though regional flood estimation methods (e.g. index flood, rational and regression-based approaches) are usually adopted for ungauged catchments, these are limited to peak flow estimation. When the estimation of complete streamflow hydrograph is needed (e.g., design of volume sensitive hydrologic systems such as reservoir) rainfall-based methods such as unit hydrograph or runoff routing model are adopted in practice. Australian Rainfall and Runoff (ARR 1987), the national guide for flood estimation, currently recommends the Design Event Approach (DEA) as the preferred method in rainfall runoff modelling in Australia (I. E. Aust., 1987). However, this method has limitations as it only accounts for the probabilistic nature of rainfall depth, but ignores the probabilistic nature of other inputs, such as rainfall temporal patterns and initial loss, in the rainfall runoff modelling (Hill and Mein, 1996). To overcome the limitations associated with the DEA, a Joint Probability Approach (JPA) has been proposed in the design flood estimation (e.g. Eagleson, 1972; Beran, 1973; Russell et al, 1979; Diaz-Granados et al, 1984; Sivapalan et al, 1990). However, as reported by Rahman et al. (1998) the majority of these applications were limited to theoretical studies, and mathematical complexity, difficulties in parameter estimation and limited flexibility in application prevented the wider application of these methods to practical situations. Rahman et al. (2002) developed and applied successfully a simplified Monte Carlo Simulation Technique (MCST) for flood estimation based on the principles of joint probability. This method uses a non-linear runoff routing model to simulate streamflow hydrographs from probability-distributed input variables and can make use of most of the commonly adopted flood estimation models and design data in Australia. Carroll and Rahman (2004) tested the MCST to catchments in Victoria and Queensland with URBS model and found that MCST can overcome some of the limitations associated with the DEA, and this can produce more accurate design flood estimates than the DEA. More recently, considerable research has been made on JPA/MCST (e.g. Aronica and Candela, 2007; Muncaster and Bishop, 2009; Kjeldsen et al, 2010; Aronica et al, 2012; Caballero and Rahman, 2013; Charalambous et al, 2013; Mirfenderesk et al, 2013; Svensson et al, 2013). Example includes the MCST method by Aronica and Candela (2007) which reproduced the observed flood frequency curves with reasonable accuracy over a wide range of return periods using a semi-distributed stochastic rainfall runoff model. The method is suitable for ungauged or partially gauged catchments. Though MCST have been applied in many studies in Australia, the method has not been investigated under different hydrologic conditions. Furthermore, the wider application of this method needs

1Paper JHER004 submitted 10/07/2013; accepted for publication after peer review and subsequent revision on 20/10/2013 *Corresponding author may be contacted at [email protected]

“Variability in Rainfall Temporal Patterns …” Caballero and Rahman


regionalisation of various input variables so that the MCST can be applied easily in practice. These various input variables include rainfall duration, intensity, temporal pattern and losses. This paper focuses on the regionalisation of the temporal patterns data for the State of New South Wales (NSW) in Australia so that this can be used with the MCST to obtain design flood estimates for both gauged and ungauged catchments in NSW. This paper also investigates the variability of rainfall temporal patterns in NSW and its impacts on design flood estimates using a MCST. Recently, the temporal patterns of design rainfall have been investigated by many studies (e.g. Rahman et al, 2006; Varga et al, 2009; Ball and Aboura, 2010; Akbari et al, 2011). The design rainfall temporal patterns in the ARR 1987 Volume 2 are commonly adopted in the rainfall runoff modelling using the DEA in Australia. These were derived using the Method of Average Variability (Pilgrim et al., 1969; Pilgrim and Cordery, 1975; Kennedy, 1991). The principal aim of the Method of Average Variability is to derive representative temporal pattern from a set of selected observed temporal patterns (Rahman et al, 2006). For the MCST, the temporal pattern database can be derived using a ‘multiplicative cascade model’ (Hoang, 2001) or dimensionless pooled historical temporal patterns (Rahman et al, 2002). This study has adopted the historic temporal pattern similar to Rahman et al (2002) to develop regional temporal pattern database for the State of NSW. 2. Adopted Methods 2.1 Selection of rainfall events In the DEA, rainfall duration is ‘fixed’ and it does not need stochastic rainfall duration. However, JPA/MCST needs rainfall duration to be a random variable. In previous application of JPA/MCST by Rahman et al. (2002), the three rainfall characteristics (rainfall duration, intensity and temporal patterns) were treated as random variables unlike the DEA. Thus, storm events which can produce rainfall events with rainfall characteristics as random variables need to be defined in the MCST. Hoang et al (1999) defined a rainfall event of random durations as a ‘complete storm’ (CS) in which it is described as the period of significant rain preceded and followed by an arbitrarily defined period of dry hours (e.g. 6 hours, as adopted in this study). While Rahman et al. (1998) defined the storm burst that produces randomly distributed storm burst durations, called ‘storm-core’, which is defined as the most intense part of a complete storm. In this study, the inter-event duration has been considered as a new random variable in the application of MCST following the approach of Kjeldsen et al. (2010). Here, the inter-event duration is defined as the time (in hours) elapsed between two successive complete storm events. Based on an arbitrary threshold value of rainfall intensity, the complete storm events from the selected 86 pluviograph stations are selected in such a way that 2 to 8 rainfall events on average are selected per year from a given pluviograph station. This selection is made by comparing the corresponding complete storm that has the highest rainfall intensity ratio to the 2-year average recurrence interval (ARI) design rainfall value obtained from ARR1987 (I. E. Aust., 1987). These rainfall events are analysed to develop a database of observed temporal patterns, which can then be used in the MCST. Some previous applications of the MCST focused on storm-cores (e.g. Rahman et al, 2002). This study focuses on the analysis of complete storms as it is believed that complete storms are easy to identify and also the rainfall temporal patterns and losses can easily be analysed for a complete storm as compared to a storm-core. A complete storm-based MCST is therefore likely to provide a more comprehensive design flood estimation method, and hence adopted in this study. 2.2 Development of rainfall temporal pattern database for NSW The temporal pattern (TP) of rainfall is a dimensionless representation of rainfall intensity over the sub-durations of the rainfall event. In this study, hourly pluviograph data is used to derive TP. Here, the TP is characterised by a dimensionless mass curve, i.e., a plot of dimensionless cumulative rainfall depth versus dimensionless storm time with 10 equal time increments (see example in Figure 2). The development and application of the TP database in the MCST can be achieved by following ‘multiplicative cascade model’ (Hoang, 2001) or historic TP (Rahman et al, 2002). This study uses the historic TP similar to Rahman et al (2002) to develop TP database using 86 pluviograph stations from NSW. To form the regional TP database, the dimensionless temporal patterns from individual stations are pooled irrespective of the season and total rainfall depth as it was found by Rahman et al (2002) that the TP does not depend strongly on season and total rainfall depth. It was also found that TP depends on storm durations (e.g. Hoang, 2001; Rahman et al, 2002), and hence the pooled TPs in this study are divided into two groups: up to 12 hours durations and more than 12 hours durations. 2.3 Adopted MCST Two catchments (Oxley River and Belar Creek) are selected to test the applicability of the developed TP database in the rainfall runoff modelling using the MCST. Each catchment has a pluviograph station (as shown in Table 2) within or near the catchment centre. For these study catchments, the significant rainfall and runoff events were selected to calibrate the runoff routing model, which is based on the following equation: S = kQm (1) where S is the storage (m3), k is the storage delay parameter (h), Q is the discharge (m3/s) and m is the non-linearity parameter, which is taken as 0.8 in this application. In the adopted runoff routing model, a single storage concentrated at the catchment outlet was considered, which should provide reasonably accurate results given the two selected study catchments, which are 213 km2 and 133 km2 in size. More accurate results would have been



obtained with a semi-distributed runoff routing model like RORB; however, this has not been adopted in this study as RORB currently supports only partial MCST i.e. it does not allow to randomly vary all the model inputs simultaneously. In obtaining the direct runoff from the observed streamflow hydrograph, the baseflow needs to be separated, which was achieved by using Boughton (1988) equation. In this study, the initial loss (IL) and continuing loss (CL) model was used to generate rainfall excess hyetograph as recommended by ARR1987 (I. E. Aust., 1987). The IL is defined as the amount of rainfall that occurs before the start of the surface runoff while CL is calculated as the average rate of loss throughout the remainder of the rainfall event. From the selected rainfall and streamflow events, the rainfall runoff model was calibrated and value of k was obtained for each of the selected events. For the two study catchments, the distributions of rainfall inter-event duration, rainfall duration, rainfall intensity, IL, CL and k were identified as shown in Table 3. In selecting the appropriate distribution for a given input variable, a number of candidate distributions were tested using a number of goodness-of-fit tests including Chi-Squared (C-S) test, Kolmogorov-Smirnov (K-S) test and Anderson-Darling (A-D) test. The selected distributions and their parameter values are presented in Table 1. The selected best-fit distributions are Gamma for rainfall inter-event duration, rainfall duration, IL and k and Exponential for CL. The stochastic values of rainfall inter-event duration, rainfall duration, rainfall intensity, IL, CL and k were generated (considering their correlations) from the identified probability distributions and the developed TP database was sampled randomly (based on the generated duration) in the adopted MCST. Both at-site TP and regional TP were used in the simulation. In deriving the regional TP, the TP set based on nearby pluviograph stations with distances of 30, 50, 100, and up to 200 km from centre of the catchment of interest or up to maximum of 20 nearby pluviograph stations were considered. The resulting number of pluviograph stations for different distances in deriving the TP database for the two selected catchments are presented in Table 2.

Table 1 Model inputs and probability distributions for MCST application

Model inputs Probability

distributions

Oxley River Belar Creek

Mean Standard deviation

Median Mean Standard deviation

Median

Rainfall inter-event duration (IED) Gamma 92.23 112.21 38.96 41.95 48.70 24.92

Rainfall duration (D) Gamma 37.54 32.93 30.00 21.48 19.51 16.00

Initial loss (IL) Gamma 28.77 27.57 17.04 38.97 20.21 36.30

Continuing loss (CL) Exponential 3.10 2.43 2.31 6.35 4.15 6.07

Storage delay parameter (k) Gamma 19.04 6.09 20.46 21.15 7.57 19.42

Table 2 Number of pluviograph stations to derive TP database

Catchments Distances from catchment centre and number of stations

30 km 50 km 100 km 200 km

Oxley River 5 7 10 13 Belar Creek 1 1 1 13

Model inputs were generated 50,000 times to establish 50,000 possible set of input values to the runoff routing model. With these generated model inputs, 50,000 different streamflow hydrographs were simulated. The peaks of these simulated hydrographs were stored for further analysis to determine a derived flood frequency curve following the approach by Caballero and Rahman (2013). The FORTRAN program developed by Rahman (1999) was modified to implement this MCST analysis. 2.4 Validation of the model The estimated flood quantiles from the MCST (based on the at-site and regional TP data) are compared against the DEA (using the mean and median k values of the observed data set of the respective model input). The resulting estimated flood quantiles from the MCST based on at-site (Qat-site) and regional TP (Qregional) were compared to assess the sensitivity of TP data on the design flood estimates. In this regard, a relative error (RE) is defined as below: RE (%) = [(Qregional – Qat-site)/Qat-site] x 100 (2) 3. Data Preparation In this study, the continuous rainfall data from 86 pluviograph stations in NSW were obtained from Australian Bureau of Meteorology. The rainfall record lengths of the selected stations range from 30 years to 101 years with an average record length of 45 years. As displayed in Figure 1, the selected stations present a good spatial distribution over the eastern part of the State; however, no station is selected from far western NSW due to insufficient



data lengths. Two catchments were also selected (Figure1 and Table 3) to test the applicability of regionalised TP with the MCST. Each catchment has a pluviograph station (within or near the catchment) with a minimum of 30 years of concurrent rainfall and streamflow data as shown in Table 4.

Figure 1 Selected 86 pluviograph stations and 2 catchments in New South Wales, Australia

Table 3 Selected catchments in NSW for rainfall runoff modelling using MCST Station ID Station name Catchment area (km2) Period of record Record length (years)

201001 Oxley River @ Eungella 213 1957 - 2011 55

420003 Belar Creek @ Warkton 133 1976 - 2005 30

Table 4 Nearest pluviograph station for each selected catchment

Catchment station ID

Pluviograph station ID

Pluviograph station name Distance from catchment's centre

(km) Period of record

Record length (years)

201001 58109 Tyalgum (Kerrs Lane) 2.36 1965 - 1996 32 420003 64046 Coonabarabran (Westmount) 7.24 1970 - 2010 41

4. Results A total of 19,718 complete storm events from the 86 pluviograph stations are selected from NSW. On average, 229 complete storm events are selected from a pluviograph station (5 events per year on average). The TP are derived for each of the 86 pluviograph stations (50 samples are shown in Figure 2 to show the expected variability in TP of NSW). As can be seen in Figure 2, the plots of both the at-site and regional TP show a wide variability, which raises the question how to select a representative TP from these to apply with the DEA so that TP is probability neutral. A total of 47 rainfall and runoff events were selected (17 and 30 from Oxley River and Belar Creek catchments, respectively) for estimating IL, CL and storage delay parameter k. Model inputs and the obtained distributions are presented in Table 1. These model inputs and parameters are then applied in estimating flood quantiles for the selected catchments using DEA and MCST. Using the DEA, a number of trial rainfall durations are adopted for each ARI of interest. The corresponding IFD data is obtained from ARR 1987 (I. E. Aust., 1987). The design temporal pattern data from ARR Volume 2 is used. The ARR 1987 IL-CL design values for NSW (IL = 35 mm and CL = 2.5 mm/h) are adopted. The mean and median k values, obtained from the model calibration (Table 3) are adopted. For a given ARI, the duration giving the highest flood discharge is taken as the critical duration and the corresponding peak discharge is taken as the design flood for the ARI. The derived flood frequency curves from MCST were obtained for the two study catchments. Figure 3 shows the results for the Belar Creek catchment. Here, the MCST and DEA provide similar results and both overestimate the observed floods at smaller ARIs.



Figure 2 Samples of at-site and regional temporal patterns for Oxley River

Figure 3 Design flood estimates for Belar Creek catchment



In the MCST, the use of regional TP shows under-estimation in flood quantiles (in relation to the use of at-site TP) for most of the cases of up to 4.5% for the Oxley River catchment as illustrated in Table 5. The maximum over-estimation for this catchment is 3.5%. The regional TP for the Belar Creek catchment was only applied to distances less than 200 km and for 20 stations as no nearby pluviograph station was available for distances less than 100 km. The regional TP for distances under 200 km and for 20 stations show under-estimation of up to 7.1% and over-estimation of up to 10.1% (Table 6). The high degree of variability in the regionalised TP (as shown in Figure 2) show remarkably small differences of up to about 10% in the design flood estimates using MCST for the two catchments.

Table 5 Design flood estimates using regional temporal patterns for Oxley River catchment

ARI (years)

Flood quantiles using at-site TP

(m3/s)

Flood quantiles (m3/s) for pooled TP and percentage difference from at-site TP

Distances (km) for pooled TP 20 stations

30 50 100 200

2 188 191 1.2% 187 -0.5% 183 -2.9% 183 -3.0% 186 -1.3%

5 380 379 -0.4% 375 -1.4% 364 -4.4% 363 -4.5% 366 -3.7%

10 532 533 0.3% 524 -1.5% 510 -4.1% 514 -3.4% 509 -4.2%

20 701 713 1.7% 682 -2.7% 686 -2.2% 687 -2.0% 676 -3.6%

50 938 955 1.8% 941 0.3% 931 -0.8% 925 -1.5% 924 -1.6%

100 1161 1202 3.5% 1169 0.6% 1113 -4.2% 1146 -1.3% 1159 -0.2%

Table 6 Design flood estimates using regional temporal patterns for Belar Creek catchment

ARI (years)

Flood quantiles using at-site TP

(m3/s)

Flood quantiles (m3/s) for pooled TP and percentage difference from at-site TP

Within 200 km distance for pooled TP

20 stations

2 79 76 -3.5% 73 -7.3%

5 160 156 -3.0% 153 -4.9%

10 222 215 -3.1% 214 -3.3%

20 284 280 -1.6% 287 1.0%

50 372 398 6.9% 380 2.1%

100 452 498 10.1% 467 3.2%

5. Conclusion This paper investigates the regionalisation of design rainfall temporal patterns for NSW for application with the MCST. Based on complete storm events, temporal patterns data at 86 pluviograph stations in NSW are derived. Pluviograph stations with distances up to 200 km can be used to obtain regional temporal patterns at any arbitrary location in NSW provided there is a sufficient number (e.g. 20 stations) of nearby pluviograph stations. The at-site and the regionalised temporal patterns data are applied with the MCST to two NSW catchments for estimating design floods. It has been found that the use of at-site and regional temporal patterns can give up to about 10% differences in flood quantile estimates using the MCST. In this study, a new random variable, rainfall inter-event duration has been tested along with other randomly distributed random variables (i.e. rainfall duration, rainfall intensity, initial loss, continuing loss and runoff routing model parameter) in the MCST, which has provided reasonable results. The findings of this study will be useful towards the application of MCST in NSW, Australia. The method can be adapted to other countries. 6. Acknowledgements The authors would like to acknowledge the Australian Bureau of Meteorology for providing the pluviograph data for New South Wales and the New South Wales Office of Water in Australia for stream flow data.



References Akbari A, Othman F, Abu Samah A (2011). Probing on suitability of TRMM data to explain spatio-temporal pattern of severe storms in tropic

region, Hydrology and Earth System Sciences, 8, 9435-9468. Aronica GT and Candela A (2007). Derivation of flood frequency curves in poorly gauged Mediterranean catchments using a simple

stochastic hydrological rainfall-runoff model, Journal of Hydrology, 347 (1-2), 132-142. Aronica G T, Franza F, Bates PD, Neal JC (2012). Probabilistic evaluation of flood hazard in urban areas using Monte Carlo simulation,

Hydrological Processes, 26, 3962–3972. Ball JE, Aboura K (2010). A likelihood approach for modeling spatial and temporal patterns of storms using radar and land data, Ninth

International Conference on Computing Anticipatory Systems. AIP Conference Proceedings, Volume 1303, pp. 345-353. Beran MA (1973). Estimation of design floods and the problem of equating the probability of rainfall and runoff, Proceedings of the Madrid

Symposium: Design of Water Resources Projects with Inadequate Data, Madrid, Spain, June 1973, pp. 459-471. Boughton WC (1988). Partitioning streamflow by computer, The Institution of Engineers, Australia, Civil Engineering Transactions, 285-291. Caballero WL, Rahman A (2013). Development of regionalised Joint Probability Approach to flood estimation: A case study for Eastern New

South Wales, Australia, Hydrological Processes, Published online in Wiley Online Library, 10 pp. Charalambous J, Rahman A, Carroll D (2013). Application of Monte Carlo Simulation Technique to design flood estimation: A case study for

North Johnstone River in Queensland, Australia, Water Resources Management, 13 pp. Carroll DG, Rahman A (2004). Investigation of sub-tropical rainfall characteristics for use in the Joint Probability Approach to design flood

estimation, In Proceedings of Second Asia Pacific Association of Hydrology and Water Resources (APHW) Conference, Singapore, July 5-8, 2004, 8 pp.

Diaz-Granados MA, Valdes JB, Bras RL (1984). A physically based flood frequency distribution, Water Resources Research, 20 (7), 995-1002. Eagleson PS (1972). Dynamics of flood frequency, Water Resources Research, 8 (4), 878-898. Hill PI, Mein RG (1996). Incompatibilities between storm temporal patterns and losses for design flood estimation, 23rd Hydrology and Water

Resources Symposium 1996, Hobart, Australia, May 21-24, 1996, pp. 445-451. Hoang TMT (2001). Joint probability approach to design flood estimation, Unpublished PhD Thesis, Department of Civil Engineering,

Monash University, Queensland, Australia. Hoang TMT, Rahman A, Weinmann PE, Laurenson EM, Nathan RJ (1999). Joint probability descriptions of design rainfalls, International

Hydrology and Water Resources Symposium of the Institution of Engineers Australia and 2nd International Conference on Water Resources and Environmental Research, Brisbane, Australia, July 6-8, 1999, pp. 43-61.

Institution of Engineers Australia (1987). Australian Rainfall and Runoff – A Guide to Flood Estimation, Institution of Engineers, Australia. Kennedy MR, Turner LH, Canterford RP, Pearce HJ (1991). Temporal distributions within rainfall burst. Australian Bureau of Meteorology

Report Series, Report Number 1. Australian Bureau of Meteorology, Melbourne. Kjeldsen TR, Svensson C, Jones DA (2010). A joint probability to flood frequency estimation using Monte Carlo simulation, Proceedings of the

BHS Third International Symposium: Role of Hydrology in Managing Consequences of a Changing Global Environment, Newcastle University, Newcastle upon Tyne, United Kingdom, July 19-23, 2010, 11 pp.

Mirfenderesk H, Carroll D, Chong E, Rahman M, Kabir M, Van Doorn R, Vis S (2013). Comparison between design event and joint probability hydrological modelling, Floodplain Management Association (FMA) National Conference, Tweed Heads, New South Wales, Australia, 28 - 31 May 2013.

Muncaster SH, Bishop WA (2009). Flood estimation in urban catchments with RORB – Monte Carlo Simulation – Use of local historical pluviographic data, 32nd Hydrology and Water Resources Symposium, Newcastle, Australia, 30 November – 3 December 2009.

Pilgrim DH, Cordery I, French R (1969). Temporal patterns of design rainfall for Sydney, Institution of Engineers, Australia, Civil Engineering Transactions, Volume CE11, 9-14.

Pilgrim DH, Cordery I (1975). Rainfall temporal patterns for design floods. Proceedings of American Society of Civil Engineers, Journal of the Hydraulics Division, Volume 100, 81-95.

Rahman A (1999). Determination of derived flood frequency curve by Monte Carlo simulation: Programs, Unpublished Cooperative Research Centre Working Documents, 27 pp.

Rahman A, Islam M, Rahman K, Khan S, Shrestha S (2006). Investigation of design rainfall temporal patterns in the Gold Coast Region, Queensland, Australian Journal of Water Resources, 10, 1, 49-61.

Rahman A, Weinmann PE, Hoang TMT, Laurenson EM (1998). Joint probability approaches to design flood estimation: A review, Technical Report No. 98/8. Cooperative Research Centre for Catchment Hydrology (CRCCH), Australia.



Rahman A, Weinmann PE, Hoang TMT, Laurenson EM (2002). Monte Carlo simulation of flood frequency curves from rainfall, Journal of Hydrology, 256 (3-4), 196-210.

Russell SO, Kenning BFI, Sunnell GJ (1979). Estimating design flows for urban drainage, Journal of the Hydraulics Division, 105 (HY1), 43-52. Sivapalan M, Beven KJ, Wood EF (1990). On hydrological similarity 3, A dimensionless flood frequency model using a generalised

geomorphologic Unit Hydrograph and Partial Area Runoff Generation, Water Resources Research, 26 (1), 43-58. Svensson C, Kjeldsen TR, Jones DA (2013). Flood frequency estimation using a joint probability approach within a Monte Carlo framework,

Hydrological Sciences Journal, 58 (1), 8-27. Varga CI, Ball JE, Babister MK (2009). An alternative approach for developing temporal patterns, 32nd Hydrology and Water Resources

Symposium, Newcastle, Australia, 30 November – 3 December 2009, 146-157.

Technical Paper


Principal Component Regression Analysis in Water Demand Forecasting: An Application to the Blue Mountains, NSW, Australia1

M. M. Haquea, *, A. Rahmana, D. Hagarea and G. Kibriab

aSchool of Computing, Engineering and Mathematics, University of Western Sydney, Australia bSydney Catchment Authority, NSW 2750, Australia

Abstract: Accurate forecast of water demand is very crucial in developing a water resource management strategy to check the balance of future water supply and demand to ensure proper water supplies to the people. In order to forecast water demand, different models have been adopted in the literature. Among these the multiple regression analysis is quite popular for long term water demand forecasting. In spite of their evident success in modelling water demands, it can face difficulties in the case of multicollinearity, which implies highly correlated variables. Since water demand depends on many factors such as population, household size, rainfall, temperature, age of population, education, water price and policy, a multicollinearity problem may arise during the development of a multiple regression model which may lead to the incorrect estimation of future water demand. To avoid multicollinearity problem, principal component regression analysis has been used in several environmental studies which demonstrated its ability to eliminate the multicollinearity problem and to produce better model results. However, application of principal component regression in water demand forecasting is limited. In this study, principal component regression model was developed by combining multiple linear regression and principal component analysis to forecast future water demand in the Blue Mountains Water Supply systems in New South Wales, Australia. In addition, performances of the developed principal component regression model were compared with multiple linear regression model by adopting several model evaluation statistics such as relative error, bias, Nash-Sutcliffe efficiency and accuracy factor. It was found that the developed principal component regression model was able to predict future water demand with a higher degree of accuracy with an average relative error, bias, Nash-Sutcliffe efficiency and accuracy factor values of 3.47%, 2.92%, 0.44 and 1.04, respectively. Moreover, it was found that the principal component regression model performed better than the multiple linear regression model and could be used to eliminate the multicollinearity problem. The method presented in this paper can be adapted to other cities in Australia and the world. Keywords: Principal component regression, multicollinearity, water demand, forecasting, Blue Mountains, principal component analysis

1 Introduction Availability of adequate potable water is becoming an increasing concern around the world due to many factors, including population growth, increased water demand, rapid urbanization, water pollution and changing climate. Therefore, integrated water resources management is very much needed which considers both water supply and demand measures to cope up with the limited water resources. Water demand measures include but are not limited to the installation of water efficient appliances, implementation of water use restrictions, development of awareness programs and use of effective water pricing policy (Adamowski and Karapataki, 2010). In order to implement these demand management programs effectively, an accurate estimate of future water demand such as peak demand, daily demand and long-time demand, is very crucial. Future estimates of peak, daily and weekly demand are considered as short term forecasting, which is mainly required for operation of reservoirs and pumping stations, and maintenance of a water supply system (Jain et al., 2001; House-Peters & Chang, 2011). On the other hand, long term forecasting is usually greater than one year which is required for planning and design of system expansion and future resilience analysis (Bougadis et el., 2005; Nasseri et al., 2011).

1Paper JHR008 submitted 12/08/2013; accepted for publication after peer review and subsequent revision on 21/10/2013 *Corresponding author: E-mail: [email protected]

“Principal Component Regression Analysis in Water Demand Forecasting…” Haque, Rahman, Hagare & Kibria


Different types of models such as time series, regression and artificial neural networks have been adopted in the literature to forecast/model residential water demand. Short term water demand is normally forecasted by a time series model (Zhou et al., 2000; Caiado, 2007) and artificial neural networks models (Ghiassi et al., 2008; Firat et al., 2009; Herrera et al., 2010). In long term forecasting, regression based analysis especially multiple regression analysis (e.g. Lahlou and Coyler, 2000; Babel et al., 2007; Polebitski and Palmer, 2009) is one of the most commonly used techniques in the water demand related literature. Though multiple regression techniques have been quite successful in modelling residential water demand, it can face some severe difficulties in the case of multicollinearity (Rajab et al., 2012). A multicollinearity problem arises when the independent variables experience high correlation. This mulicollinearity problem may produce biased standard error of estimate in the regression analysis and may lead to incorrect identification of the most significant variables in the modelling exercise. Importantly, water demand depends on many factors, such as population number, household size, dwelling type, age of population, household income, water price, temperature, rainfall, evaporation and water conservation and restriction programs (Franczyk and Chang, 2009, Harlan et al., 2009, Arbués et al., 2010, Babel and Shinde, 2011). These variables are often correlated with each other and may introduce a multicollinearity problem in water demand modelling exercise. Principal component regression (PCR) is a type of regression analysis which considers principal components (PC) as independent variables instead of adopting original variables (Pires et al., 2008). The PCs are the linear combination of the original variables which can be obtained by principal component analysis (PCA). The PCA transforms the original set of inter correlated independent variables to a new set of uncorrelated variables (i.e. PCs). The use of these PCs as independent variables is quite useful in the multiple regression models to avoid the mulicollinearity problem and to identify the variables which are the most significant in making the prediction (Abdul-Wahab et al., 2005; Camdev´yren et al., 2005; Sousa et al., 2007 and Rajab et al., 2013). Abdul-Wahab et al. (2005) used the combination of multiple linear regression and PCA technique to model tropospheric ozone and to identify the significant factors that control ozone levels. Camdev´yren et al. (2005) adopted PCs in multiple linear regression analysis in water quality studies. Sousa et al. (2007) developed the PCR model using PCs as inputs to predict ozone concentrations and compared that model with multiple linear regression and a feed forward artificial neural network model. Rajab et al. (2013) combined a multiple regression model with PCA technique to improve the prediction of ozone levels. All of these studies have found that the incorporation of PCs as independent variables in the regression models improved the model prediction as well as reduced the model complexity by eliminating multicollinearity. Although PCR has been adopted in many fields of water engineering, there have been limited applications of PCR in water demand forecasting. Few studies have been found in the literature that have adopted PCA to forecast water demand, such as Koo et al. (2005) and Choi et al. (2010). Koo et al. (2005) used the PCA and cluster analysis to divide the region into two groups and then developed a new multiple regression model for each of the groups to overcome the negative coefficient issue (the regression coefficient for the population of the Jung-gu model had a negative value, which was not sensible) of the population variable with the water demand in the existing single multiple regression model for each district in Seoul, Korea. They found that the newly developed model solved the irrationality of the regression model and were able to produce better water demand estimates. Choi et al. (2010) developed three multiple regression models: a comprehensive multiple regression model with six independent variables, a multiple regression model for each cluster and a multiple regression model adopting two principal components (derived from eight number of original variables) for each cluster to estimate the water demand in 164 regions in Korea. They found that the clustering multiple regression model performed best in modelling the observed water demand values and the multiple regression model based on PCs were better than the comprehensive multiple regression model. Due to the potential of PCR in reducing the complexity of multiple regression models by mitigating the multicollinearity problem, it is worth exploring the application of PCR in the water demand forecasting. In this study, water demand modelling is done by multiple linear regression and PCR techniques to estimate the future water demand in the Blue Mountains Water Supply System in New South Wales, Australia. The objectives of this paper are twofold: (i) to evaluate the effectiveness of PCs as independent variables in multiple linear regression models of water demand forecasting (i.e. checking the applicability of PCR in water demand forecasting); (ii) to evaluate the performance of developed multiple linear regression and PCR models by estimating and comparing several goodness of fit statistics. Here, PCs are obtained by undertaking the PCA of ten water demand variables.

J

Journal of Hydr

2 Study area aThe Blue Mounprovides water Mountains areagathered from Sdetached housesthe rest 6%. In were divided by

Figure 1 Data on water uWCS refers to twater tank, WatIt-Yourself) kitsGovernment im

“Principal Comp

rology and Envir

and data ntains region (Figto a population

a (Sydney CatchSydney Water fs) accounts for athis study, anal

y the number of d

1 Blue Mountain

usage price and he number of dwterFix (installatios (self-installed

mposed three diff

mponent Regressio

ronment Researc

gure 1) of New n of around 48,0hment Authorityfor the period oabout 94% residlysis was done fdwellings to get

ns region in Aust

water conservatwellings that havon of new showflow restrictors

fferent levels of w

ion Analysis in W

ch

South Wales, Au000 from Faulcoy, 2009). Data of 2003-2011. It dential water confor the single dwthe ‘per dwellin

tralia and Casca

tion savings (Wve participated in

werheads, flow res), water efficienwater restriction

DEL_CA

Water Demand F

Australia is selectonbridge to Moon monthly mewas found that

nsumption whilewelling residentiang monthly water

ade and Greaves

WCS) were also gn the water demestrictions and mnt washing ma

n based on the s

ASCADE & DEL

Forecasting…” H

ed as the study unt Victoria, whetered residentiat the single dwee the multiple dwal sector. Collecr consumption’ t

creeks water sup

gathered from Symand managemenminor leak repairachines and toilseverity during

L_GREAVES

Haque, Rahman, H

area. The Blue Mhich are consideal consumption elling residentialwelling sector (i.cted monthly meto be used in the

pply area (Bluem

ydney Water fornt programs in trs undertaken bylets (Sydney Wthe drought per

Hagare & Kibri

Mountains Wateered as Upper aand number of

l sector (i.e. fre.e. apartment bletered water cone modelling.

mountainsaustral

r the same periothe Blue Mountay a licensed plum

Water, 2010). Neriods (2003-2009

ia

5

er Supply Systemand Middle Bluf dwellings were standing/semiocks/units) is fonsumption value

lia, 2013)

od. In this studyains, such as rainmber), DIY (Doew South Wale9) to manage th

51

m ue re i-or es

y, n o-es he



short water supply. Level 1 and Level 3 were the most liberal and the most severe level of restrictions, respectively. Level 1, Level 2 and Level 3 water restriction were imposed to three different occasions during the 2003-2009 in the Sydney region. In this study, three dummy variables were considered for these three levels of water restriction. The value of a dummy variable was considered as 1 (one) when the water restrictions were in place; otherwise its value was considered to be zero in the data matrices. Meteorological data, such as rainfall, number of rain days, temperature, evaporation and solar exposure were collected from Sydney Catchment Authority. 3 Methods In this study, multiple linear regression (MLR) equation and PCA were combined together to perform PCR analysis. This PCR model was adopted to predict the future water demand. Brief description of PCA, MLR and PCR are given in the following sections. 3.1 Principal component analysis Principal component analysis transforms the original data set of n variables which are correlated among themselves to various degrees to a new data set containing n number of uncorrelated principal components (PCs). The PCs are linear functions of the original variables in a way that the sums of the variances are equal for both the original and new variables. The PCs are sequenced from the highest variance to the lowest one. The first PC explains the highest amount of variance in the data. The next highest variance is explained by the second PC and so on for all n PCs. The values of all the PCs can be obtained by the same equation as Equations 1 and 2. These two equations are for PC 1 and PC 2. Although the number of PCs and original variables are equal, normally most of the variance in the data set can be explained by the first few PCs that can be used to represent the original observations (Abdul-Waheb et al., 2005; Olsen et al., 2012). This helps in reducing the dimensionality of the original data set.

n

jjjnn xaxaxaxaPC

111212111 ...1 (1)

n

jjjnn xaxaxaxaPC

122222121 ...2 (2)

Where nxxx ,...,, 21 are the original variables in the data set and jja are the eigenvectors.

The eigenvalues are the variances of the PCs and the coefficients jja are the eigenvectors extracted from the covariance or correlation matrix of the data set. The eigenvalues of the data matrix can be calculated by Equation 3 as shown below:

0 IC (3) Where C is the correlation/covariance matrix, is the eigenvalue and I is the identity matrix. The PC coefficients or the weights of the variables in the PC are then calculated by Equation 4:

0 jjaIC (4) Due to differences in the units of the water demand variables used in this study, a correlation matrix of the variables was used to obtain eigenvalues and eigenvectors. The eigenvectors multiplied by the square root of the eigenvalues produce a nn matrix of coefficients, which are called variable loadings. Importance of each original variable to a particular PC is represented by these loadings. Furthermore, the sum of the products of the variable loadings and the values of original variables produce a new set of data values which are called component scores. These scores can be used in the multiple linear equations as new variables to predict the future water demand.



3.2 Multiple regression analysis Multiple linear regression attempts to model the relationship between two or more independent variables with a dependent variable by fitting a linear equation to the observed data. The general equation of a MLR model can be expressed as below (Montgomery et al. 2001):

nnxaxaxaaY ...22110 (5)

Where Y is the dependent variable, ),...,0( niai are the parameters generally estimated by least squares method and

),...,0( nixi are the independent variables.

3.3 Principal component regression (PCR) In the PCR analysis, MLR and PCA are combined together to establish a relationship between the dependent variable and the selected PCs of the input variables (Pires, et al., 2008). Mainly principal component scores obtained from the PCA are taken as the independent variable in the multiple linear regression equation to perform the PCR analysis. The general equation of PCR model is as follows:

PCnaPCaPCaY n ...21 21 (6) 4 Model evaluation criteria The performances of the developed MLR and PCR models were compared by adopting four statistical performance indices: (i) relative error (RE); (ii) percent bias (PBIAS); (iii) Nash-Sutcliffe efficiency (NSE); and (iv) accuracy factor (Af). RE measures the relative size of error in the modelled values in terms of percentage with respect to observed values. An ideal value of RE is zero, which indicates that the developed model is perfect, which is not possible in practice. However, the lesser the RE value, the better the performance of the model would be. PBIAS measures the percentage of the residuals with respect to observed values which indicate whether the developed model overestimates or underestimates the observed values overall (Gupta et al. 1999). The perfect value of PBIAS is zero. Low values of PBIAS indicate better simulation results by the model where positive and negative values represent underestimation and overestimation bias, respectively in the modelled results. NSE is a dimensionless model performance indicator which measures the relative magnitude of the residual variance with respect to the observed data variance (Nash and Sutcliffe, 1970). The optimal value of NSE is one; however, any value between 0 to 1 is generally considered as an acceptable range of performance (Moriasi et al., 2007). Af demonstrates the spread of the model results with respect to observed values. The best value of Af is one which indicates a perfect agreement between the modeled and observed values. The smaller the value of Af, the more accurate the model results is (Basant et al., 2010). The values of these evaluation statistics are computed by the equations presented in Table 1 from the modeled and observed values of the dependent variable. In this study, these were computed for both the modelling period (2003-2008) and the forecasting period (2009-2011).

Table 1 Equations used to calculate the model performance indices Performance Indices Equation

RE ∑ | |100

PBIAS ∑∑

100

NSE

2

2

)(

)(1

meani

ii

OO

PO

Af

10

∑

O: Observed water demand, P: Model estimated water demand, N: Number of observations.



5. Results Pearson correlation matrices of the water demand variables are presented in Table 2. Statistically significant correlation coefficients (p < 0.05) are highlighted in bold. The linear relationship between two variables and the existence of the collinearity between the independent variables can be identified from these coefficients.

Table 2 Pearson correlation matrix of different variables PDRC RF NRD MMT EVP SE WP WCS RL1 RL2 RL3

PDRC

Pearson Correlation

1 -.150* -.174* .284** .568** .283** -.611** -.728** .020 -.130 -.452**

Sig. (2-tailed) .046 .020 .000 .000 .000 .000 .000 .790 .085 .000

RF

Pearson Correlation

1 .648** .227** .014 .086 .094 .074 -.038 -.012 .063

Sig. (2-tailed)

.000 .002 .858 .254 .214 .327 .613 .872 .408

NRD

Pearson Correlation

1 .239** .039 .167* .256** .189* -.102 -.104 .061

Sig. (2-tailed)

.001 .609 .026 .001 .012 .179 .169 .424

MMT

Pearson Correlation

1 .814** .852** -.014 -.008 .130 .008 .019

Sig. (2-tailed)

.000 .000 .849 .917 .085 .920 .801

EVP

Pearson Correlation

1 .849** -.197** -.223** .207** .015 -.201**

Sig. (2-tailed)

.000 .009 .003 .006 .842 .007

SE

Pearson Correlation

1 .126 .128 .082 .016 .033

Sig. (2-tailed)

.094 .090 .276 .829 .663

WP

Pearson Correlation

1 .906** -.102 -.108 .159*

Sig. (2-tailed)

.000 .176 .154 .034

WCS

Pearson Correlation

1 -.023 .000 .474**

Sig. (2-tailed)

.762 .996 .000

RL1

Pearson Correlation

1 -.059 -.135

Sig. (2-tailed)

.438 .074

RL2

Pearson Correlation

1 -.167*

Sig. (2-tailed)

.026

RL3

Pearson Correlation

1

Sig. (2-tailed)

*Correlation is significant at the 0.05 level (2-tailed). **Correlation is significant at the 0.01 level (2-tailed).



In Table 2, PDRC is per dwelling residential consumption (dependent variable) in kL/dwelling/month; RF is total monthly rainfall in mm; NRD is number of rain days in a month; MMT is monthly mean maximum temperature in 0C; EVP is monthly total evaporation in mm; SE is monthly mean daily global solar exposure (MJ/m2); WP is water price in AUD/kL; WCS is water conservation savings in dwelling numbers; RL1 is dummy variable for Level 1 restriction; RL2 is dummy variable for Level 2 restriction; and RL3 is dummy variable for Level 3 restriction. As can be seen in Table 2, per dwelling residential consumption were negatively correlated with RF, NRD, WP, WCS, RL1, RL2 and RL3. This result was expected as the rainfall and number of rainy days increase; the water requirement for gardening would be less. Hence, total water consumption would be reduced. Furthermore, it is obvious that with the enhancement of water conservation programs and restriction levels, the total water consumption goes down. Water consumption was found to be positively correlated with MMT, EVP and SE (Table 2). As these variables are mostly related with temperature, water consumption will be more if the temperature is relatively high in a day. High correlation coefficients were found between the independent variables, such as MMT and EVP (0.814), MMT and SE (0.852), WP and WCS (0.906), which demonstrate the existence of multicollinearity between the variables. The PCA was done on the ten independent variables to explain per dwelling water consumption level in the Blue Mountains Water Supply systems. Table 3 and 4 summarise the results of the PCA on the ten independent variables with the amount of variance explained by each PC. From Table 3, it can be seen that the first five PCs had eigenvalues higher than 1. Moreover, these first five PCs explained around 94% of the total variation of variables in PCA. These five PCs were selected for PCR. Contribution of a particular variable within a PC is normally judged by its variable loadings value. The higher the loading of a variable, the more contribution is reflected by that variable within a particular PC. The bold marked loads in Table 4 indicate the high existing correlation between the variables and corresponding PC.

Table 3 Variance explained by the PCs Value PC 1 PC 2 PC 3 PC 4 PC 5 PC 6 PC 7 PC 8 PC 9 PC 10

Eigenvalue 2.97 2.68 1.39 1.29 1.04 0.32 0.15 0.08 0.06 0.02% of Var. 29.68 26.81 13.87 12.92 10.44 3.23 1.46 0.77 0.60 0.21Cum. % 29.68 56.49 70.36 83.29 93.73 96.96 98.42 99.19 99.79 100.00

Table 4 Component loadings (correlations between original variables and the firs five PCs)

Variable PC 1 PC 2 PC 3 PC 4 PC 5 RF 0.08 0.52 0.37 -0.61 -0.26

NRD -0.01 0.70 0.25 -0.48 -0.21 MMT 0.82 0.43 0.09 0.11 0.13 EVP 0.92 0.24 -0.11 0.16 0.07 SE 0.78 0.52 -0.03 0.17 0.23 WP -0.48 0.73 -0.44 0.10 0.02

WCS -0.53 0.76 -0.22 0.22 0.10 RL1 0.48 -0.29 -0.22 0.13 -0.77 RL2 0.27 -0.46 -0.25 -0.64 0.47 RL3 -0.15 -0.05 0.90 0.35 0.14

All ten water demand variables were included in the five selected PCs. However, only certain variables showed high loadings within each PC, such as the first PC was heavily loaded on MMT, EVP and SE, and the second PC was heavily loaded with RF, NRD, WP and WCS. Similarly as can be seen in Table 3, PC 3, PC 4 and PC 5 were loaded heavily with Level 3, Level 2 and Level 1 restrictions, respectively. Component score coefficients (eigenvectors) and the values of the original variables were then multiplied to obtain PC score values. These score values were used as independent variables in the stepwise multiple linear regression analysis to determine the most significant PCs for water demand prediction. Data from October 2003 to December 2008 were used to develop the PCR model. Then the model was used to



forecast water demand for the period of January 2009 to September 2011. In PCR analysis, PC 1, PC 2 and PC 5 were found to have significant (p < 0.05) linear relationship with per dwelling water consumption (PDRC) (Table 5).

Table 5 Results of regression analysis Included

Independent Variables

Regression Coefficients

Standard Error

Standardized Beta Coefficients

t Sig. R2

(%)

Constant 12.463 .096 129.568 .000 66.4

PC 1 .463 .050 .720 9.325 .000 PC 2 -.285 .060 -.361 -4.728 .000 PC 5 -.192 .068 -.215 -2.823 .006

As can be seen in Table 5, the three PCs (i.e. PC 1, PC 2 and PC 5) could explain 66% of the variation in water consumption. PC 1 and PC 2 were found to be the most significant independent variables in the regression analysis as the standardized beta coefficients values are the highest and the second highest for these two PCs, respectively. PC 1 had positive impact on water consumption while PC 2 had a negative impact (Table 5) as the sign of the regression coefficients were found to be positive and negative for PC 1 and PC 2, respectively. This implies that if the value of PC 1 increases, water consumption would be expected to increase and water consumption would decrease as the value of PC 2 increases. Therefore, a total increase in significant variables of PC 1, namely MMT, EVP and SE would lead to an increase in the water consumption level. On the other hand, an increase in the significant variables of PC 2 (RF, NRD, WP and WCS) would lead to a decrease in water consumption level, as expected. Dummy variables for Level 3 and Level 2 water restrictions had significant loadings in PC 3 and PC 4, respectively which were excluded from the PCR model as its p values were not statistically significant. However, linear effects of these variables were partially incorporated in the model as RL 3 and RL 2 were also included in PC 1, PC 2 and PC 5. PC 5 is negatively correlated with water consumption. As level of water restriction goes up, the level of water consumption would be expected to decrease. The developed PCR model can be written as:

5192.02285.01463.0463.12 PCPCPCPDRC (7) The comparison of observed and predicted water consumption values by the PCR model is presented in Figure 2. The forecasted monthly water consumption values were found to be close to the observed values. Average relative error values for all of the predicting months were found to be 3.47%, which indicates that the model is capable of forecasting monthly water demand with a high level of accuracy. Three forms of multiple linear regression techniques, linear, semi-log and log-log were adopted to develop the multiple linear regression models. In the linear model, the relationship between the dependent variable, per dwelling water consumption and the independent variables (e.g. rainfall, temperature and water price) were assumed to be linear. In semi-log model, only the dependent variable was in logarithmic form, whereas in the log-log models all the independent variables and dependent variable were entered in logarithmic form in the regression equation. After checking the model performances of these three models, it was found that the semi-log model performed best to model the water demand. Therefore, the semi-log model was taken as the final model to report in this study. The variables were retained in the regression equation for which the regression coefficients were significant at 5% significance level. It was found that only five variables were found to be significant in the equations. Five other variables were found not to be statistically significant, including the number of rain days in a month, monthly total evaporation, monthly mean daily global solar exposure, water conservation savings and a dummy variable for Level 1 restriction due to the high correlation with other variables.



Figure 2 Comparison of predicted and observed water demand The equation of the developed semi-log model is given below:

30331.0

20376.0121.000227.0000079.024.1)log(

RL

RLWPMMTRFPDRC

(8)

where PDRC =per dwelling residential consumption (dependent variable) in kL/dwelling/month; RF = total monthly rainfall in mm, MMT = monthly mean maximum temperature in 0C; WP = water price in AUD/kL; RL2 = Dummy variable for Level 2 restriction; RL3 = Dummy variable for Level 3 restriction. The comparison of the performance of the developed MLR and the PCR model is presented in Table 6 for both the modelling and forecasting period. It was found that the performances of the models were nearly the same during the modelling period. However, the PCR model outperformed the MLR model during the forecasting period. All of the performance statistics were in favour of the PCR model. The PCR model considered the PCs as independent variables which accounted for the contribution of all the original variables without having any multicollinearity problem. On the other hand, in the developed MLR model, half of the original variables had to be discarded due to the multicollinearity problem, which might be the reason for the underperformance with respect to the PCR model.

Table 6 Results of the performance indices for both the MLR and PCR model Performance Indices MLR PCR

Modelling period (2003-08)RE 3.54 3.57

PBIAS 1.07 0.00 NSE 0.65 0.67 Af 1.04 1.03 Forecasting Period (2009-11)

RE 7.91 3.47 PBIAS 8.04 2.92 NSE 0.37 0.44 Af 1.09 1.04

0.00

2.00

4.00

6.00

8.00

10.00

12.00

14.00

16.00

Jan

Feb

Mar

Apr

May Jun

Jul

Aug Sep Oct

Nov

Dec Jan

Feb

Mar

Apr

May Jun

Jul

Aug Sep Oct

Nov

Dec Jan

Feb

Mar

Apr

May Jun

Jul

Aug Sep

2009 2010 2011

Per

Dw

elli

ng

Mon

thly

Wat

er D

eman

d

in k

L

Month, Year

Observed Modelled



6 Conclusions In this study, the principal component regression (PCR) model was developed by combining multiple linear regression (MLR) and principal component analysis to identify the most important variables for water demand modelling and to forecast future water demand. It was found that PC 1 and PC 2 were the most significant independent variables in the PCR model. Therefore, the variables which had significant loadings within these PCs could be considered as important predictor variables for water demand forecasting. Hence, monthly mean maximum temperature, monthly total evaporation, monthly mean daily global solar exposure, monthly total rainfall, number of rainy days in a month, water price and water conservation savings variables were found to be the most important predictor variables as these variables were within PC 1 and PC 2. However, some of these variables were found to be highly correlated with each other. Inclusion of all these variables in the multiple linear regression model might lead to inconsistent estimation of future water demand. Therefore, the developed PCR model was used to forecast the future water demand which showed a high degree of prediction accuracy with an average relative error value of 3.47%. Moreover, the developed PCR model with three PCs as independent variables was able to explain 66% variation in water consumption level. The performances of the developed PCR model were compared to the MLR for both the modelling and forecasting period. Though both models performed similarly during the modelling period, the PCR model outperformed the MLR during the forecasting period. All the performance statistics, relative error, percent bias, Nash-Sutcliffe efficiency and accuracy factor value were found to be in favour of the PCR model. Moreover, it was found that half of the original variables were discarded in the MLR model due to the multicollienarity problem. To avoid these problems, the PCR model could be used to get a better prediction in water demand forecasting. The method presented in this paper can be applied to other water supply systems to develop water demand forecast models. Acknowledgements Water consumption data used in this study was obtained from Sydney Water on 4 May 2012. The best available data at the time of study has been used, which may be updated in the near future. The authors express their sincere thanks to Pei Tillman and Frank Spaninks of Sydney Water for their assistance in collating and providing the data. Further, the authors are grateful to Lucinda Maunsell and Peter Cox of Sydney Water and Mahesh Maheswaran of Sydney Catchment Authority for their cooperation and assistance during data collation and analysis.

References Abdul-Wahab SA, Bakheit, CS, Al-Alawi SM (2005). Principal component and multiple regression analysis in modelling of ground-level ozone and factors affecting its

concentrations, Environmental Modelling & Software, 20(10), 1263-1271. Adamowski J, Karapataki C (2010). Comparison of multivariate regression and artificial neural networks for peak urban water-demand forecasting: Evaluation of different

ANN learning algorithms, Journal of Hydrologic Engineering, 15(10), 729-743. Arbués F, Villanúa I, Barberán R (2010). Household size and residential water demand: an empirical approach, Australian Journal of Agricultural and Resource Economics,

54(1), 61-80. Babel M, Gupta AD, Pradhan P (2007). A multivariate econometric approach for domestic water demand modeling: An application to Kathmandu, Nepal, Water Resources

Management, 21(3), 573-589. Babel MS, Shinde VR (2011). Identifying Prominent Explanatory Variables for Water Demand Prediction Using Artificial Neural Networks: A Case Study of Bangkok,

Water Resources Management, 25(6), 1653-1676. Basant N, Gupta S, Malik A, Singh KP (2010). Linear and nonlinear modeling for simultaneous prediction of dissolved oxygen and biochemical oxygen demand of the

surface water-a case study, Chemometrics and Intelligent Laboratory Systems, 104(2), 172-180. Bluemountainsaustralia.com (nd), Location and maps, viewed 10 February 2013, http://www.bluemts.com.au/info/about/maps/ Bougadis J, Adamowski K, Diduch R (2005). Short‐term municipal water demand forecasting, Hydrological Processes, 19(1), 137-148. Caiado J (2009). Performance of combined double seasonal univariate time series models for forecasting water demand, Journal of Hydrologic Engineering, 15(3), 215-222. Çamdevýren H, Demýr N, Kanik A, Keskýn S (2005). Use of principal component scores in multiple linear regression models for prediction of Chlorophyll-a in reservoirs,

Ecological Modelling, 181(4), 581-589. Choi TH, Koo JY (2010). Water Demand Forecasting by Characteristics of City Using Principal Component and Cluster Analyses, Environmental Engineering Research,

15(3), 135-140. Firat M, Yurdusev MA, Turan ME (2009). Evaluation of artificial neural network techniques for municipal water consumption modeling, Water Resources Management,

23(4), 617-632. Franczyk J, Chang H (2009). Spatial analysis of water use in Oregon, USA, 1985–2005, Water Resources Management, 23(4), 755-774. Ghiassi M, Zimbra DK, Saidane H (2008). Urban water demand forecasting with a dynamic artificial neural network model, Journal of Water Resources Planning and

Management, 134(2), 138-146. Gupta HV, Sorooshian S, Yapo PO (1999). Status of automatic calibration for hydrologic models: Comparison with multilevel expert calibration, Journal of Hydrologic

Engineering, 4(2), 135-143. Harlan SL, Yabiku ST, Larsen L, Brazel AJ (2009). Household water consumption in an arid city: affluence, affordance, and attitudes, Society and Natural Resources, 22(8),

691-709. Herrera M, Torgo L, Izquierdo J, Pérez-García R (2010). Predictive models for forecasting hourly urban water demand, Journal of hydrology, 387(1), 141-150.



House-Peters LA, Chang H (2011). Urban water demand modeling: Review of concepts, methods, and organizing principles, Water Resources Research, 47(5),W054. Jain A, Kumar Varshney A, Chandra Joshi U (2001). Short-term water demand forecast modelling at IIT Kanpur using artificial neural networks, Water Resources

Management, 15(5), 299-321. Koo J, Yu M, Kim S, Shim M, Koizumi A (2005). Estimating regional water demand in Seoul, South Korea, using principal component and cluster analysis, Water Science

& Technology: Water Supply, 5(1), 1-7. Lahlou M, Colyer D (2000). Water conservation in Casablanca, Morocco. JAWRA Journal of the American Water Resources Association, 36(5), 1003-1012. Montogomery, D. C., Peck, E. A., and Vining, G. G. (2001). “Introduction to linear regression analysis.” Third edition, John Wiley & Sons, New York, USA. Moriasi D, Arnold J, Van Liew M, Bingner R, Harmel R, Veith T (2007). Model evaluation guidelines for systematic quantification of accuracy in watershed simulations,

Transactions of the ASABE, 50(3), 885-900. Nash JE, Sutcliffe J (1970). River flow forecasting through conceptual models part I—A discussion of principles, Journal of Hydrology, 10(3), 282-290. Nasseri M, Moeini A, Tabesh M (2011). Forecasting monthly urban water demand using Extended Kalman Filter and Genetic Programming, Expert Systems with

Applications, 38(6), 7387-7395. Olsen RL, Chappell RW, Loftis JC (2012). Water quality sample collection, data treatment and results presentation for principal components analysis-literature review and

Illinois River watershed case study, Water Research, 46(9), 3110-3122. Pires J, Martins F, Sousa S, Alvim-Ferraz M, Pereira M (2008). Selection and validation of parameters in multiple linear and principal component regressions.

Environmental Modelling & Software, 23(1), 50-55. Polebitski AS, Palmer RN (2009). Seasonal residential water demand forecasting for census tracts, Journal of Water Resources Planning and Management, 136(1), 27- Rajab JM, Jafri MZM, San Lim H, Abdullah K (2012). Regression analysis in modeling of air surface temperature and factors affecting its value in Peninsular Malaysia,

Optical Engineering, 51(10). Rajab JM, MatJafri M, Lim H (2013). Combining multiple regression and principal component analysis for accurate predictions for column ozone in Peninsular Malaysia,

Atmospheric Environment, 71, 36-43. Sousa S, Martins F, Alvim-Ferraz M, Pereira M (2007). Multiple linear regression and artificial neural networks based on principal components to predict ozone

concentrations, Environmental Modelling & Software, 22(1), 97-103. Sydney Catchment Authority (2009). Blue Mountains water supply system: Strategic review. Sydney Catchment Authority, Penrith, Australia. Sydney Water (2010). Water conservation and recycling implementation report, 2009-10. Sydney Water, New South Wales, Australia. Zhou SL, McMahon TA, Walton A, Lewis J (2000). Forecasting daily urban water demand: a case study of Melbourne, Journal of Hydrology, 236(3), 153-164.

Review Paper


Brick Manufacturing Practice in Bangladesh: A Review of Energy Efficacy and Air Pollution Scenarios1

K. M. Daraina, A. B. M. S. Rahmanb, A. Ahsanc,*, A. B. M. S. Islamd and B. Yusufc

aArchitecture Discipline, Khulna University, Khulna-9208, Bangladesh

bShelter and WatSan unit, UNHCR- The UN Refugee Agency, Sub-office, Cox’s Bazar, Bangladesh cDepartment of Civil Engineering, and Institute of Advanced Technology, University Putra Malaysia, Malaysia

dDepartment of Civil Engineering, University of Malaya, Kuala Lumpur-50603, Malaysia Abstract: Building structures and major construction works are booming in the urban areas of Bangladesh to keep up the rapid urbanization rate. Brick is an essential construction material for its building industries. To meet the need of construction, brickfields are growing sporadically here and there at the fringe zones and within the urban regions. Eventually, brickfields are producing major environmental pollutants. This paper highlights existing technologies of brick manufacturing as well as their emission which exceeds the tolerable limit and puts a threat to the environment. Treatise is also portrayed regarding techniques which could help minimizing the drawback of existing brick kiln technologies. Vertical Shaft Brick kiln might be an alternative solution for the small investors in Bangladesh which is energy efficient and can minimize the air pollution to achieve a sustainable environment. Keywords: Brick kilns, environment, energy consumption, pollution, green house gas emission. 1. Introduction Brick is a popular construction material for thousands of years. At present, the demand of bricks is soaring, especially in the developing country like Bangladesh, where infrastructure development projects are the top priority. The current population of the country is assessed at around 149 million (BBS, 2011). A large portion of this huge populace are migrating to the urban areas for better living condition. Owing to this rapid urbanization, a sharp rise of 5.6 percent per year has been noticed for the construction industry. This trend eventually directed the brick sector to increase annually at an projected 2 - 3 percent over the next decade for housing construction and commercial sector developments (WB, 2010). Though, the brick manufacturers in Bangladesh is therefore expanding in production, a good number of these producers are not formally recognized as industry and not advancing technologically (MOI, 2010). The kilns and technology remained unchanged for long times back and still consumes energy inefficiently. Biomass, mainly firewood and rice husk, are the main energy sources for the brick firing (Alam, 2009). Brick making is traditionally a cottage industry which produces bricks for local consumption; though its technological development is inadequate. Current technologies for brick production such as clamps, high draught kilns and bull’s trench kilns consume large quantities of fuel such as coal, firewood and other biomass (Begum et al, 2010). In Bangladesh, about 4,500 brick kilns are in operation, producing about 9 billion bricks per year (Gomes and Hossain, 2003). Of the 4,500 kilns, more than 4,000 are of the Bull’s trench kiln (BTK) type. The BTK employs an extremely crude technology to fire bricks. There are also about 400 fixed chimney kilns, 15 zig zag kilns, 25 Hoffman kilns and 5 modern tunnel-type kilns currently in operation in the country. Hoffman kiln uses natural gas and the other three types of kiln use low-grade coal and firewood as fuel (Begum et al, 2011). The devastating effect of the pollution caused by the huge amount of emission from the brick industry has attracted the attention of the regulatory agencies, like the Department of Environment (DoE) in Bangladesh. Bangladesh Government first promulgated the ‘Brickfield Law 1 Paper JHER003 submitted 11/07/2013; accepted for publication after peer review and subsequent revision on 29/08/2013 *Corresponding author may be contacted at [email protected]

“Brick Manufacturing Practice in Bangladesh: A Review …” Darain, Rahman, Ahsan, Islam and Yusuf


(Regulation)’ in 1989 that prohibits the use of biomass in Brickfield. This law was amended in 1992. But the total condition of the Brickfields remained largely unchanged. The study shows a huge number of unregistered brickfield are producing brick without having any license which are mainly responsible for violation of the present regulation (Iqbal, 2004). Till now, the greater part of the brick factories are operating Fixed Chimney Kiln (FCK) technology with coal as the key fuel as it is more economical than the alternatives (Hossain, 2008). FCKs are not energy efficient and consequently pollutants are being emitted by a greater rate. To ease the emissions from brick field industries, the Government is working to propose alterations in FCK technology and newer technologies such as Zig-Zag and Hybrid Hoffman kilns, which are more energy efficient (Hossain, 2008). The problem with the brick industry can be found in all South Asian and several Southeast Asian countries as well as China. The problems related to brick manufacturing using the BTK and other polluting and inefficient technologies have been examined, but often no solution to the problem has been offered. (AIT, 2002a; AIT, 2002b; AIT, 2002c). Well-known energy research institutes such as The Energy and Resources Institute (TERI) and Development Alternatives (DA), both of New Delhi, India, have been trying to popularize VSBKs (Maithel et al, 2003) but with limited success. This goal of the study is to identify the principal technical approaches and some barriers, to assess alternative strategies for pollution control and to explore broad implications for large or small-scale brick kilns in developing countries like Bangladesh. A more complete knowledge of raw materials and their properties, better control of firing, improve kiln designs; exhaust air quality enhancement and more advanced mechanization will result in the development of a progressive, modern industry. 2 Major pollutants emission from brick fields Brick making is an energy intensive process (Croitoru and Sarraf, 2010, Narasimha and Nagesha, 2013). Brickfields are the important contributors of the emission of greenhouse gases in Bangladesh as they burn huge amounts of coal and wood fuel. As a result, rate of deforestation increases in the brick making season of 6 months. If the wood fuel burning is stopped in Brickfields, the Bangladesh Atomic Energy Commission accounts that 2.1 petajoules of energy (Which is equivalent to the energy produced by 2.1×2312×100 tons of refined fossil oil) will be saved. Burning of wood fuel except bamboo is illegal (GOB, 1989) but till now a huge amount of wood fuel is being used. After the Government regulation in Brickfields, the major fuel used now for brick manufacturing is coal. The conventional practice of firing clay bricks in rural country clamps and BTKs (Bull Trench Kiln) consume huge quantities of fuel in terms of coal, firewood and other biomass fuels such as old tires, tainted sawdust, discarded motor oil , plastic and household garbage, dung cakes and agricultural residue. In Bangladesh, the brick industries consume several million tons of coal biomass fuels; kilns are notorious as highly polluting establishments, affecting not only just flora and fauna but also posing threats to human health. Higher energy costs and the inability of the industry to meet the environmental standards have raised serious concerns about the survival and well being of the industry. The main reason for poor emission from brick kilns is the poor quality of coal and uses of biomass mainly firewood. The main pollutants which are emitted from the brickfields are particulate matter (PM), some hazardous gases like CO2, CO, NOX, NO and SO2 . The PM concentration appears to be low but it is expected to have long term massive impact on global environments as well as on human health. The particulate matter consists of dust, smoke, fumes, and fly ash. Ahmed and Hussain (2007) studied the pollutant load within the cluster region of brick kilns in Bangladesh for SO2 and particulate matter. It was found that particulate matter was a major pollutant in that region. 3 Hazards from pollutants Huge amounts of hazardous materials are discharged from the brick manufacturing industry, which would have great impacts on the environment as well as on human health as discussed below. 3.1 Effects of particulates Particulate is one of the major pollutants emitted from brickfields. Two types of particulate, settlable and suspended are produced, both are harmful for human health and the surrounding environment (Croitoru and Sarraf, 2010).



3.2 Effects on human health At high concentration suspended particulate matter poses health hazards to humans, particularly those susceptible to respiratory illness (WHO, 1999). As indicated in Table 1, the nature and extent of the ill effects that may be linked to suspended particulates depend upon the concentration of particulates, the presence of other atmospheric contaminants (notably sulfur oxides) and the length of exposure.

Table 1 Particulate matter effects on health (WHO, 1999) Concentration μg/m3

Accompanied by Exposure Time Effects

750 715 µg/m3 SO2 24-h average Considerable increase in illness 300 630 µg/m3 SO2 24-h average Acute worsening of chronic bronchitis patients200 250 µg/m3 SO2 24-h average Increase absence of industrial workers 100-130 120 µg/m3 SO2 Annual mean Children likely to experience increased incidence

of respiratory diseases 100 Sulfation rate above 30

mg/cm2/mo Annual geometric mean

Increased death rate for those over 50 likely

80-100 Sulfation rate above 30 mg/cm2/mo

2-yr geometric mean Increased death rate for those 50 to 69 yrs

3.3 Effects on plants and animals Dry brick kiln dust appears to cause little damage if deposited on a leaf surface, yet in the presence of moisture; such dust imparts damage and consequential growth inhibition to plant tissues. Dust coating of leaves reduces photosynthesis and the increased plugging of stomata reduces plant growth. Animals who eat plants coated with particulates may suffer from some ill effects (Jan et al, 2013, Fatima, 2011). 3.4 Effects on materials Particulate matter can damage materials by soiling clothing and textiles, corroding metals (especially at relative humilities above 75%),eroding building surfaces, and discoloring and destroying painted surfaces. For example, at particulate matter concentrations ranging from 130 to 180 µg/m3 and in the presence of SO2 and moisture, corrosion of steel and zinc panels can be three to four times greater than in areas close to background particulate levels (60 µg/m3). 3.5 Effects of Sulfur Oxides SO2 emission is one of the concerns of environmental pollution by the brick industry (US EPA, 1986). It may have some significant impact on the environment and human health. SO2 is one of the principal pollutants, which gets emitted from brick industry. Sulfuric acid (H2SO4), Sulfur dioxide (SO2) and sulfate salts tend to irritate the mucous membrane of the respiratory track and foster the development of chronic respiratory diseases, particularly bronchitis, pulmonary emphysema and asthma. Individuals who suffer from chronic respiratory diseases may experience coughing and a difficulty in breathing when the SO2 concentration rises from 0.1 to 0.2 ppm. Also SO2 is the primary component of acid deposition and damages the ecosystem both directly and indirectly. The WHO guideline for SO2 emission is 0.9 ppm at sampling time of 10 minutes. Table 2 shows the effects of SO2 on humans.

Table 2 Effects of SO2 on humans (WHO, 1999) Concentration ppm Exposure Time Effects0-0.6 No detectable response0.15-0.25 1-4d Cardio respiratory response 1.0-2.0 3-10min Cardio respiratory response in healthy subjects 1.0-5.0 Detectable responses, tightness in chest5.0 1 h Choking and increased lung resistance to air flow 10.0 1 h Severe distress, some nose bleedingGreater than 20 Digestive tract affected, also eye irritation 400-500 Dangerous for short period of time



3.6 Effects of oxides of nitrogen (NOx) There are different types of oxides of nitrogen, such as N2O, NO, N2O3, NO2, N2O5 out of these oxides NO2 emissions is too high in the brickfield and it has quite high impact on the environment as well as on human health (IPCC, 1996;US EPA, 1999). The main component of NOx emitted from brickfield to the atmosphere is NO and NO2. NO is a relatively inert gas and only moderately toxic. Although, NO, like CO, can combine with hemoglobin to reduce the oxygen carrying capacity of the blood. NO concentrations are generally less than 1.22 (1ppm) is the ambient air and thus is not considered as a health hazard. This irritates the alveoli of the lungs; the response of the human respiratory system to short term exposure to nitrogen dioxide. Other ill effects include heart tightness, barring of eyes, and headaches. People with asthma and bronchitis are particularly vulnerable to these effects. Recurrent exposure to high concentration of NO is more damaging that constant exposure to low level concentration. The WHO guideline for ambient concentration is 0.2 ppm (400) at a sampling time of 1 hour. 3.7 Effects of CO CO at present ambient levels has little if any effect on property, vegetation or materials. At high concentrations, it can seriously affect human aerobic metabolism, owing to its high affinity for hemoglobin, the component of the blood responsible for transport of oxygen. CO reacts with the hemoglobin (Hb) of blood to form carboxyhemoglobin (COHb) thus reducing the capability of blood to carry oxygen. Hence, more blood need to be pumped to deliver the same amount of oxygen, resulting in strain on the heart. This is fatal when inhaled at very high concentrations, and heart disease patients, pregnant women, infants, senior citizens and those especially susceptible to respiratory problems may exhibit symptoms of poisoning at a relatively lower concentration. The WHO guideline for ambient CO concentration is 87 ppm (100) at a sampling time of 15 minutes. 4 Brick kiln technology Brick manufacturing process in Bangladesh is being carried out in a very primitive way. Still date traditional process is applied in Bangladesh. These age-old methods of manufacturing brick are leading to environmental degradation. The demands for bricks continue to rise, as the population increases and people aspire to having better standards for housing. To supply this growing demand as well as to reduce fuel consumption and air pollution it is necessary to develop new technologies and process of brick manufacturing. In the technological aspects the existing kiln technology should be studied in details. 5 Classification of brick kiln Brick kilns can be classified into four main categories, on the basis of how they are operated:

Intermittent or periodic Kiln: intermittent or periodic kiln that consists of a single firing chamber. The intermittent kilns are loaded with green bricks, which is fired and allowed to cool before unloading, in preparation of next loading and firing. These types of kilns are capable of firing only one loading of brick at a time.

Semi continuous kiln: Semi continuous kiln, where two or more intermittent kilns are interconnected by flues and dampers, to allow the heat from cooling bricks in one kiln to dry and pre-heat the bricks in another. The kilns are alternated being unloaded once the heat from the cooling bricks had been used to dry and pre-heat the bricks in the second kilns, that is then fired up to top temperature.

Continuous kiln: In the continuous kilns, the firing zone moves through the kilns without stopping. Green bricks are loaded in front of firing zone and fired bricks are removed behind it. These kilns run day and night, with the fire never going out except for seasonal or maintenance stoppages.

Tunnel kiln: In Tunnel kilns, the bricks are placed on trolleys and moved through the hottest part of the kiln at a predetermined rate. This is a form of continuous kiln, but with a stationary rather than moving firing zone.

Kilns can be further subdivided into three main classes, based on how they actually work.

Up-draught kilns: Up-draught kilns where the heat travels naturally by convection, from the area of combustion up through the bricks.

Down–draught Kiln: Down–draught Kiln, where the heat combustion is drawn down through the bricks by use of a chimney or forced draught system.



Horizontal/cross draught kilns: Horizontal/cross draught kilns, where the heat of combustion is drawn sideways thought the bricks by the use of a chimney or forced draught system.

6 Kiln technology Different types of kilns in the technological point of view are used in brick manufacturing in Bangladesh. Most of them, which are technically not sound enough to produce brick without polluting the environment. Recently, the researchers have done some technical improvement in the brick kilns. 6.1 Clamps This is a very old aged method of manufacturing bricks. The clamp is the most basic type of kiln since no permanent kiln structure is built (Aziz 1981). The bricks are stacked in alternate layers to reach the desired height, gradually tapering towards the top. The base of the clamp is rectangular. The top surface is covered with earth to prevent the escape of heat. The biomass (fuel wood) is used on clamp as fuel. 6.2 Bull's Trench Kiln (BTK) This type of kiln (Fig. 1), which is generally used in Bangladesh, has some technological drawback. Now a days, a few countries are trying to improve this technology and going to implement in the modern sector.

Figure 1 Schematic diagram of a BTK

The main deficiencies of existing Bull's Trench Kiln are noted below (Aziz, 1981):

Improper kiln construction leads to excessive air leakage from the kiln system thus increasing the losses.

Small size chimney leads to excessive flue gas temperature to give effective draught.

Heat loss from the side and top do not allow attainment of full firing temperature and thus leads to deterioration in quality of fired goods.

Very high loss due to repeated heating of the kiln system because of high thermal mass. 6.3 Hoffman Kiln The Hoffman kiln has fixed roof, which enables bricks to be fired throughout the year. The inside roof of the kiln is arched and has a firebrick lining on the inside surface. The thick walls of the kiln and good insulation minimize heat loss to the surroundings. The bricks are fired from the top by introducing the fuel (natural gas) into the combustion zone through pipe-type burners. The burners are shifted forward from section to section as the fire progresses fired bricks are unloaded at the back while green bricks are stacked in front of the firing zone. The flue gas has been conveyed towards the chimney through a network of channels just below the kiln. Fire is controlled without the aid of any instrument or controllers by merely adjusting the gas flow rate and the opening and closing of dampers located at selected points in the flue gas network. Controlling the fire is the trickiest part of the whole operation. 6.4 Vertical Shaft Brick Kiln (VSBK) An energy efficient and sustainable alternative to traditional brick kilns is Vertical Shaft Brick Kiln (Fig. 2). This technology which is originated in China, is all set to emerge as a viable alternative for small-scale brick manufacturing in developing countries.



Figure 2 Schematic diagram of a VSBK (reference may be needed if it has been taken from other source) Advantages of VSBK Technology are discussed below:

Highly energy efficient methods for firing bricks: VSBK technology results in an energy saving of about 50% compared to clams and 20 to 30% compared to BTKs (TERI, 2009).

Compact kiln: The kiln is very compact. For the same production capacity VSBK needs about 20 to 30% of land required by BTKs (TERI, 2009).

The construction cost: Thought slightly higher compared to that of a BTK, is still reasonably low.

Weatherproof: It can be operated during the monsoon because, unlike other traditional kilns, VSBK has a roof, which provides protection from rain allowing year-round operations. Other brick kilns can operate only 5 to 6 months in a year.

Suitable for internal fuel bricks. The kiln is highly suitable for use where fuel is added internally into the brick. In some countries agricultural residue like rice husks and fuels like coal dust are mixed with clay before molding. This allows the use of inferior quality fuel and in general, results in the reduction of the production cost. VSBK, because of being fast-firing characteristics, are particularly suited for firing bricks containing internal fuel.

Modularity in construction and flexibility in production are important attributes of the technology. Additional shafts (production capacity of 2,000 to 5,000 bricks per day) can be easily added/operated as per the demand for bricks.

7 Comparative Energy Consumption in Different Brick Kiln Specific fuel consumption in brick manufacturing is an important factor for the environmentally friendly brick industry. This depends upon the type of raw clay available for making bricks, kiln used for firing, the fuel used and the operation practice in place. Energy consumption of the different brick kilns is compared in Table 3. In Fig.3, energy consumption is compared amongst different brick kiln technologies which was carried by the Tata Energy Research Institute (TERI). The study depicted that the VSBK kiln consumed the lowest specific energy utilization.



Table 3 Energy consumption in different types of Kiln Process Technologies Specific Energy Consumption (Gcal/'000 nos. Brick) Status

Coal Biomass Firing Bull's Trench Kiln (Movable

Chimney) 1.08 - Existing

Bull's Trench Kiln (Fixed Chimney) 0.86 - ExistingClamps - 1.35 ExistingVertical Shaft Brick Kiln 0.57 - New High Draught Kiln 0.72 - Existing

Source: Development Alternatives, 1995

Figure 3 Comparison of specific energy consumption

8 Emission co-efficient Brich industry consumes not only a lot of coal but also a lot of biomass energy in the form of agricultural residue. However, since the biomass is considered to be carbon neutral, no CO2 emissions are estimated from its combustion. Their individual carbon emissions rates are low while their carbon content is high indicating an inefficient production. Efficiency improvement measures for these units would offer benefits for higher profit margins, better compliance with pollution control norms and in turn could improve the local environment. In Table 4, a comparative data on CO2 reduction potential are demonstrated amongst the three types of brick kiln. 9 Particulate Matter The major pollutant emitted from brick field is the Suspended Particulate Matter (SPM). Increasing the stack height of the kiln, the pollution concentration can be reduced around the area of Brickfields. It is estimated that, while for 60 ft stack height the concentration of pollutants exceed the WHO standard (100 for SO2, 150 for PM and NOx), (WHO, 1999) increasing the stack height to 120 ft, the concentration can be well controlled.

Table 4 CO2 reduction potential of the Hoffman kiln and VSBK (Gomes and Hossain, 2003) BTK Hoffman VSBK Fuel Coal Natural Gas Coal Fuel consumption per brick 0.28 Kg 0.1632 m3 0.17 kg Total fuel 28 t 16320 m3 20 t Total energy consumption 586 GJ 571 GJ 419 GJ CO2 emission 55.4 t 32.0t 39.6 t CO2 emission reduction compared to BTK (%)

- 42 28.5

Basis : 100,000 bricks produced Conversion Factors: Heating value of coal = 20.93 GJ/t ; Heating value of natural gas = 35 MJ/m3 CO2 emissions of coal = 94.6 t CO2/TJ ; CO2 emission of natural gas = 56.1 t CO2/TJ



10 Scope of improvement: perspective of Bangladesh Brick kilns are one of the major sources of emission in Bangladesh. There is wide scope to improve the traditional brick manufacturing process of Bangladesh. The major issues in environmental improvement involve the increasing of the combustion efficiency of existing kilns, and upgrading kilns to newer and more efficient process design, upgrading the design of the chimney so that it emits less hazardous pollutants and the overall process technology upgrading. Process technology upgrades are usually capital intensive. Combustion efficiency improvement can be achieved with relatively lower costs in many kilns. Promoting cleaner production in this industry requires extensive understanding of work with brick making technology changes. Several low cost ways to increase efficiency and to reduce waste and pollution in brick making are described below. Stack fuel around bricks to facilitate preheating: Solid fuel is mixed with the bricks throughout the kiln, either as sawdust mixed into the brick mass or as fuel channels in different levels of the kiln. By doing this, a combustion zone can be generated in the kiln that gradually moves upwards, using the residual heat in the lower, already burnt bricks for preheating of combustion air. The residual heat in the flue gases is used for drying and preheating of the higher levels of crude bricks. Improve brick drying before firing: Extended drying time reduces fuel requirements. Even drying throughout brick stacks reduces defective firing of bricks. Improve airflow control: Stopping air leaks and controlling the kiln opening size allows better control of airflow speed and direction to improve combustion. Switch to propane or natural gas fuel: If available and competitively priced, these fuels have significantly less emission and can increase production quality and speed. New kiln design: Vertical shaft brick kiln allows increased production rate and significantly decreased emissions through improved combustion airflow efficiency. Several kiln designs have also proven to be relatively low-cost and much more efficient than traditional ovens or kilns. Some newly developed technologies may be introduced in Bangladesh. The Bull's Trench Kiln, which is common in Bangladesh, may improve this technology by taking some measures. Most of the technology consists of movable metallic chimney, which can be replaced by fixed chimney. A gravitational settling chamber may be designed and set up at the chimney to reduce suspended particulate in the flue gas passing through the chimney. The following steps may be adopted to take necessary measures to upgrade the existing technology.

The traditional BTK technology could be upgraded by adding limestone scrubbers to reduce sulfur dioxide emissions and filters for entrained particulate removal. (Pembshaw and Smith, 2005).

Mixing and pulverizing the fuel before combustion and adding insulation would also result in higher energy efficiency. Fuels such as natural gas and low sulfur coal could replace the current fuel mix of wood and poor quality coal.

A properly designed chimney of 120 feet height (as per the requirement of the Department of Environment, Bangladesh) along with an integral gravitational setting chamber should be provided.

To supplement the chimney, the flue ducts are to be designed to provide the least amount of resistance to the flue gases.

Alternate building materials such as straw bales, block bricks, cement blocks and reprocessed waste could also be employed to reduce the demand for ceramic bricks.

Properly designed chimney which provides required draught for combustion fuel and transfer from cooling zone to firing zone and then to pre-heating one.

Gravitational settling chamber providing below the chimney which arrests SPM.

Space for valves providing in flue lines, which will fulfill the requirements of factory acts.



11 Conclusion Brick making is an energy intensive process. It should emphasize the need for energy conservation and pollution control in brick industries of Bangladesh. In this paper data for different types of fuel consumption and emitted pollutants from the Brickfields have been compared. Some of the technical options are discussed as an effective means of energy conservation and pollution control. However, existing technologies are not energy efficient. On the other hand, new technologies are not popular in Bangladesh and costly in terms of initial investment. Following notes cannot be made from this study:

If these existing approaches (e.g. Clamp, BTK etc.) are replaced overnight by the modern but expensive and complicated technology like Hoffman’s kiln, fixed chimney kiln or Zigzag kiln, those traditional brick kiln owners would be wiped out and this incident will have a great impact to the national economy of Bangladesh. As a result, consumers will have to buy bricks at a higher rate which will create a negative impact on the rapid growth of the urbanization. Therefore, a transition should be created between the existing and proposed techniques.

The government should come forward to establish a platform for stakeholder interaction (industry representatives, policy makers, researchers and financial institutions) to expedite the adoption of efficient kiln technologies and should offer incentive to the entrepreneurs of the new green approach.

A feasible alternative for the small investors might be Vertical Shaft Brick Kiln (VSBK) method which consumes relatively smaller energy with lower initial investment and can be operated against the natural calamities like rain or flood.

References Ahmed S, Hossain I (2008). Applicability of Air pollution Modeling in a Cluster of Brickfields in Bangladesh, Chemical Engineering Research

Bulletin, 12, 28-34. AIT (2002a). Energy, Environment and Climate Change Issues: A Comparative Study in Asia, a study by the Asian Regional Research

Programme in Energy, Environment and Climate Phase II (ARRPEEC-II), Energy Program, Asian Institute of Technology (AIT), Bangkok, Thailand.

AIT (2002b). Energy, Environment and Climate Change Issues:India, a study by the Asian Regional Research Programme in Energy, Environment and Climate Phase II (ARRPEEC-II), Energy Program, Asian Institute of Technology (AIT), Bangkok,Thailand.

AIT (2002c). Energy, Environment and Climate Change Issues: China, a study by the Asian Regional Research Programme in Energy, Environment and Climate Phase II (ARRPEEC-II), Energy Program, Asian Institute of Technology (AIT), Bangkok,Thailand.

Alam GJ (2009). Environmental pollution of bangladesh–it’s effect and control, Pulp and Paper, 51, 13-7. Aziz MA (1981). A Textbook of Engineering Materials, Madhuban Press, Bangladesh, 24-39. BBS (2011). Bangladesh population and housing census 2011, Bangladesh Bureau of Statistics (BBS), Govt. of Bangladesh. Begum B, Biswas S, Hopke P (2011). Key issues in controlling air pollutants in Dhaka, Bangladesh. Atmospheric Environment. December 22,

2011;45(40):7705-7713. Begum BA, Biswas SK, Markwitz A, Hopke PK (2010). Identification of sources of fine and coarse particulate matter in Dhaka, Bangladesh.

Aerosol Air Qual. Res, 10, 345-353. Croitoru L, Sarraf M (2010). The Cost of Environmental Degradation in the Middle East and North Africa, World Bank. Croitoru L, Sarraf M (2012). Benefits and Costs of the Informal Sector: The Case of Brick Kilns in Bangladesh, Journal of Environmental

Protection, 3(6), 476-484. Development Alternatives (1995). Technology Systems (VSBK). Accessed via http://www.devalt.org/ on 11 Feb, 2005. Fatima I, Bhat GA, Kamili A NG (2011). Impact of Brick Kiln Emissions on the Ambient Air Quality and Vegetation: A Case Study of District

Budgam, M. Phil. Dissertation, University of Kashmir, India. Flamos A, Georgallis PG, Psarras J (2010). Bioenergy options in the industrialized and developing world and opportunities for the clean

development mechanism, International Journal of Green Energy, 7(6), 647-661. Gomes E, Hossain I (2003). Transition from traditional brick manufacturing to more sustainable practices, Energy for Sustainable

Development, 7, 2. GOB (1989). Brick Burning Law (Regulation), Government of Bangladesh (GOB), Bangladesh Gazette. Guttikunda S (2009). Impact analysis of brick kilns on the air quality in Dhaka, Bangladesh, SIM-Air Working Paper Series. Hossain I (2008). Impact of Brickkiln pollution on Dhaka City. Accessed via http://centers.iub.edu. bd/chpd/Pres_Sem_Ijaz Hossain_Aprl

17-08.pdf on 11 Feb, 2008



IPCC (1996). Revised IPCC Guidelines for national Greenhouse Gas Inventories: Reference Manual, Volumes I, II, and III. Intergovernmental Panel on Climate Change.

Iqbal A (2004). Impact of brick kiln on environment: A study on greater Khulna region of Bangladesh, A project thesis paper submitted for the BSc. Degree to Environmental Science Discipline, Khulna University.

Jan FA, Khan S, Ishaq M, Naeem M, Ahmad I, Hussain S (2013). Brick kiln exhaust as a source of polycyclic aromatic hydrocarbons (PAHs) in the surrounding soil and plants: a case study from the city of Peshawar, Pakistan, Arabian Journal of Geosciences, 1-7.

Maithel S, Vasudevan N, Johri R (2003). Status paper on vertical shaft brick kiln (VSBK), The Energy and Resources Institute, New Delhi. Accessed via http://www.teriin.org on 25 Sep, 2003.

Miah MD, Alam SA (2002). Deforestation and Greenhouse gas emission caused by the brickfields: Bangladesh Perspective, Bangladesh Environment 2002, BAPA, Dhaka, 88-96.

MOI (2010). National Industrial Policy, MOI (Ministry of Industries), Bangladesh. Narasimha C, Nagesha N (2013). Energy efficiency in sustainable development of Small and Medium Enterprises: An empirical study, In

Intelligent Systems and Control (ISCO), 2013. 7th International Conference on IEEE, 487-491. Pembshaw N, Smith N (2005). Improvement of the Brick making process in Bangladesh. Accessed via

http://www.chemeng.adelaide.edu.au/research/reports/brick/ on 11 Feb, 2005. TERI (2009). Energy Saving in Brick Industry: Brick-by- Brick approach to Sustainable development, (Tata Energy Research Institue (TERI), e-

Newsletters. Issue 18. Accessed via http://societyarts.tripod.com/04sshe/wp/vsbk_teri.htm on 23 Aug, 2013. US EPA (1986). Brick And Structural Clay Product Manufacturing, Final Report; Compilation of Air Pollution Emission Factor, US EPA

(Environmental Protection Agency). Accessed via www.epa.gov/ttn/chief/ap42/ch11/final/c11s03.pdf on 11 October, 1986. WHO (1999). Global Air Quality Guidelines, Geneva, World Health Organization. WB (2010). Introducing Energy-Efficient Clean Technologies in the Brick Sector of Bangladesh, Report No. 60155-BD, Environment, Climate

Change and Water Resource Unit, World Bank (WB), Washington DC.

Guide to Authors

Guide to authors for preparing manuscript for submission in Journal of Hydrology and Environment Research (JHER):

1. The paper should be prepared using Microsoft Word and emailed to the Editor of JHER ([email protected]). No pdf file is accepted.

2. The margins of the papers are 2.5 cm on top, 2 cm at bottom, 2.2 cm on left and 1.5 cm on right sides.

3. The paper should be written using Times New Roman font, 1.5 lines spaced and justified on both sides.

4. The title of the paper is 18 point Times Roman font size, centred, with the first letters of keywords as capital.

5. The text in the body of the paper is of 11 point Times Roman font size.

6. The major headings are of 12 point font size and bold, the sub-headings are of 11 point font size and bold and the sub sub-headings are 11 point font size

and italics.

7. The headings and sub-headings in the papers are to be numbered.

8. The structure of the paper may vary depending on the type of the paper, but the typical structure of a paper in the JHER consists of Abstract: 250 words,

Keywords: Three to five keywords, Introduction, Methods, Data, Results, Conclusion, Acknowledgements, References and Appendix.

9. The figures and tables should be numbered by Arabic numerals (e.g. 1, 2, 3 …) and titles should be centred. The table texts are of 10 font size.

10. The figures and tables should be placed in the centre of the page.

11. The equations should be numbered by Arabic numerals (e.g. 1, 2, 3). Equations should be left justified, with equation numbers to be written at the far

right end of the margin. The equations should be quoted in the body of the paper following the style: It is found in equation 2 that temperature changes

inversely with wind velocity.

12. References: (i) In the body of the paper: depending on the number of authors or organisations, the following styles are acceptable: Smith (1980), Smith

and Jones (2008), Smith et al. (2010), Sydney Water (2011), ABS (2012). (ii) All the quoted references are to be listed in the reference section of the

paper. Make all journal names in italics. The following styles are acceptable:

Smith KL, Rahman A, Jones PR, Haddad K (1999). Towards a new regional flood methods in Australia, Australian Journal of Water Resources, 4, 23-32.

Dawson JL, Rahman A (2012). A new regional flood estimation technique for north Australia, Technical report, University of Western Sydney, 46 pp.

Yang X, Lu K, Henry R (2011). Water data analysis, 3rd edition, Wiley, 306 pp.

Imrul K, Jones D (2013). Water price in Australia, Accessed via www.waterprice.australia on 22 Jan 2012.

USE OF WORD TEMPLATE:

Authors are encouraged to use the word template in preparing the manuscripts which is available in the download page (see www.jher.org).

http://www.waterprice.australia/

Journal of Hydrology and

Environment Research

Vol 1 No 1

______________________________________________________

Contents

Technical papers:

Sobol′’s sensitivity analysis for TOPMODEL hydrological model: A case study for 1

the Biliu River Basin, China

W. Qi, C. Zhang, J. Chu and H. Zhou

Co-Active Neuro Fuzzy Inference System for Regional Flood Estimation in Australia 11

K. Aziz, A. Rahman, A. Y. Shamseldin and M. Shoaib

Effectiveness of Rainwater Harvesting System in Peri-Urban Region: A Case Study 21

for the Greater Sydney

E. Hajani and A. Rahman

Validation of TRMM Rainfall for Pangani River Basin in Tanzania 30

R. M. F. Haque, S. Maskey, S. Uhlenbrook and M. Mul

Variability in Rainfall Temporal Patterns: A Case Study for New South Wales, 41

Australia

W. L. Caballero and A. Rahman

Principal Component Regression Analysis in Water Demand Forecasting: 49

An Application to the Blue Mountains, NSW, Australia

M. M. Haque, A. Rahman, D. Hagare and G. Kibria

Review Paper: Brick Manufacturing Practice in Bangladesh: A Review of Energy Efficacy 60

and Air Pollution Scenarios

K. M. Darain, A. B. M. S. Rahman, A. Ahsan, A. B. M. S. Islam and B. Yusuf

______________________________________________________

