SPN7, University of Sheffield 29/8/13 PREDICTING CSO CHAMBER DEPTH USING ARTIFICIAL NEURAL NETWORKS WITH RAINFALL RADAR DATA Dr. Steve Mounce Mr. Gavin.
Post on 28-Mar-2015
216 Views
Preview:
Transcript
SPN7, University of Sheffield 29/8/13
PREDICTING CSO CHAMBER DEPTH USING ARTIFICIAL NEURAL NETWORKS
WITH RAINFALL RADAR DATA
Dr. Steve Mounce Mr. Gavin SailorDr. Will Shepherd Dr. James Shucksmith and Prof. Adrian Saul
Pennine Water Group, University of Sheffield, UK
Presentation structure
1. Introduction and aims
2. Case study data
3. Methodology
4. Results
5. Conclusions and Further Work
IntroductionCSOs are common assets in the UK’s combined urban drainage systemDesigned to discharge excess water during heavier rainfall events directly to a receiving watercoursePotential for unconsented spill events and pollution at CSOPossible causes include downstream blockageThis work investigates a data driven method for performance assessment to tackle this problem
Background and objectivesIncreasing amounts of hydraulic field data from wastewater networks are being collected via monitors and telemetry systems alongside higher quality weather dataStandard deterministic models require understanding of the hydrological and hydraulic processes to predict performance of the sewer networkPrevious work (Kurth et al. 2008, Guo and Saul 2011) has explored using Artificial Neural Networks with CSO depth and rain gauge data to predict future depthThis work incorporates rainfall radar data for a case study
Case studyCSO is terminal flow control to a treatment works at the bottom of a steep combined urban drainage catchment (~20 km² area)Water level data within the CSO was recorded using an ultrasonic depth monitor (with 100% signifying the spill level) and rainfall intensity data (mm/hr) from 20 rainfall radar pixels, with a resolution of 1 km² (15 min resolution for six month period)
Qi
Qs
Qc
100 %
426
427
428
429
430
399 401400398397396395
1
2 3
4 5 6 7 98
10 11 13 1412
15 16 17 18 2019
CSO
Case studySchematic with rainfall radar squares: river / canal overlay (blue), urban blocks (grey) and tree areas (green).
Example of relationship0
10
20
30
40
50
60
70
80
900
50
100
150
200
18.
8.1
2
18.
8.1
2
18.
8.1
2
18.
8.1
2
19.
8.1
2
19.
8.1
2
19.
8.1
2
19.
8.1
2
19.
8.1
2
19.
8.1
2
19.
8.1
2
19.
8.1
2
20.
8.1
2
20.
8.1
2
20.
8.1
2
20.
8.1
2
20.
8.1
2
20.
8.1
2
20.
8.1
2
20.
8.1
2
21.
8.1
2
Ra
infa
ll in
tens
ity (
mm
/h)
Ch
am
be
r w
ate
r d
ep
th (
% o
f w
eir
he
igh
t)
Date
Water Level (mm)
Rainfall 6
Rainfall 10
Rainfall 18
Time delay due to stormrunoff arriving at CSO chamber
0
10
20
30
40
50
60
70
80
900
50
100
150
200
14:3
0:00
17:3
0:00
20:3
0:00
Rai
nfal
l int
ensi
ty (
mm
/h)
Ch
am
ber
wa
ter
dep
th (
% o
f wei
r he
igh
t)
Date
Water Level (mm)
Rainfall 6
Rainfall 10
Rainfall 18
CorrelationUsed to investigate the lags between different rainfall radar squares and the CSO depth to select model inputsSerial correlation is a measure of the similarity of a variable with a lagged version of itself – used for depth
The correlation values decrease gradually with increasing lag time
CorrelationCross-correlation is a measure of the similarity of two variables (signals) as a function of a time lag between them – used on CSO depth and rainfall data
• Maximum indicates the point in time where the signals are best aligned: either lag -4 or -5
• The larger maximum correlation squares were 1, 3, 6 and 7
• Delay of -5 was observed in the far western grid squares (4, 5 and 10).
Artificial Neural NetworkParallel computational models consisting of densely interconnected adaptive processing units which transform a set of inputs into a set of outputs Universal function approximatorsStatic architectures can be used to make a time series prediction
Turns a temporal sequence into a spatial pattern encoded on the input layer of the network using ‘sliding window’No explicit reference to the temporal nature of time.
This work uses a straightforward static ANN: a single layer feed-forward network with single outputCan be trained with ADALINE rule or Moore–Penrose pseudoinverse
Training and testingModelPredictedCSO Chamber Water depth‘n’ time steps forward
Correlation analysis helps to select the lagsRainfall intensity parameter U was always one data step ahead of the chamber water depth parameter YPrediction 1 to 5 time steps ahead (up to 1 hr 15 mins)Six month data set bisected into training and testing sets containing both dry and wet weather periodsVarious ANN models applied
ResultsOne time step ahead prediction for unseen test data
ResultsIncrease in test error as prediction forecast horizon (p) increasesLess than 5% error for predictions 5 time steps ahead (75 minutes) for unseen dataThis improves on previous work which showed less than 5% error for 3 time steps ahead prediction (rain gauges with 5m sampling) but increased above this further into the future.
Radar Architecture Test RMSE
Test %
Grid square
u delay
y delay
p
ANN-1 6 11 8 1 3.97 1.99
ANN-2 6 8 6 1 3.97 1.98
ANN-3 6 15 10 1 3.97 2.00
ANN-4 6 11 8 2 4.54 2.72
ANN-5 6 11 8 3 5.42 3.84
ANN-6 6 11 8 4 6.11 3.97
ANN-7 6 11 8 5 6.58 4.28
ANN-8 5 11 8 1 3.94 1.97
ANN-9 5 11 8 5 6.35 4.32
ANN-10 18 11 8 1 3.98 2.27
ANN-11 18 11 8 5 5.71 4.07
Results
0
5
10
15
20
25
300
50
100
150
200
18
/12
18
/12
18
/12
18
/12
18
/12
18
/12
18
/12
18
/12
19
/12
19
/12
19
/12
19
/12
19
/12
19
/12
19
/12
19
/12
20
/12
20
/12
20
/12
20
/12
20
/12
20
/12
20
/12
20
/12
21
/12
21
/12
21
/12
21
/12
21
/12
21
/12
21
/12
21
/12
22
/12
Ra
infa
ll in
ten
sity
(m
m/h
)
Ch
am
be
r w
ate
r d
ep
th (
% o
f w
eir
he
igh
t)
Date
Water Level
Prediction
Rainfall 6
Weir Height
Prediction output shown four time steps advanced
ANN-1 predicting chamber depth one hour in future – spilling after rainfall
ConclusionsFor the case study, chamber depth was found to be at a correlation maximum with rainfall radar at a lag of 60 to 75 minutesAn ANN model trained with the pseudo-inverse rule to learn the response to rainfall was shown to be capable of providing prediction of CSO depth with less than 5% error for predictions 5 time steps ahead (75 minutes) for unseen data The tool offers the potential benefit of early detection of unexpected or abnormal performance behaviour and the identification of various failure modes in both dry and wet weather conditions thus enabling pollution incidents to be managed more proactively
Future workThe water utility company is exploring a wider roll out of daily download for CSO assets and a six month project to develop an automated online pilot system to incorporate rainfall radar data will shortly commenceOnline data processing could allow the prediction of CSO failures (unconsented spill events) much earlier - potentially in real time Possible deviations between predicted and measured performance signify anomalies which could be highlighted using fuzzy logic, Bayesian inference systems or a BED There is significant potential for application to other sewerage asset types such as Detention Tanks and Sewer Pumping Stations with a view to enabling wider network performance visibility.
Future workCSO Analytics – Phase II System development and trial
CSO telemetry system
Rainfall Radar data
50 CSOs
ANN hydraulic performance prediction model
Daily data import
ANN engine Predicted depth
Classification module
Lower than weir height
Safe
Beyond weir height
Spill
Interfacing from / to existing water company IT infrastructure
Thank you!
Any Questions?
With Thanks To:
top related