Short-Term Wind Generation Forecasting

7/23/2019 Short-Term Wind Generation Forecasting

1/70

Short-Term Wind Generation Forecasting

Using Artificial Neural Networks

Technical Report


2/70


3/70

EPRI Project ManagerP. Zhang

EPRI 3412 Hillview Avenue, Palo Alto, California 94304 PO Box 10412, Palo Alto, California 94303 USA800.313.3774 650.855.2121 [email protected] www.epri.com

Short-Term Wind GenerationForecasting Using Artificial Neural

Networks

1009219

Final Report, October 2003


4/70

DISCLAIMER OF WARRANTIES AND LIMITATION OF LIABILITIES

THIS DOCUMENT WAS PREPARED BY THE ORGANIZATION(S) NAMED BELOW AS ANACCOUNT OF WORK SPONSORED OR COSPONSORED BY THE ELECTRIC POWER RESEARCHINSTITUTE, INC. (EPRI). NEITHER EPRI, ANY MEMBER OF EPRI, ANY COSPONSOR, THEORGANIZATION(S) BELOW, NOR ANY PERSON ACTING ON BEHALF OF ANY OF THEM:

(A) MAKES ANY WARRANTY OR REPRESENTATION WHATSOEVER, EXPRESS OR IMPLIED, (I)WITH RESPECT TO THE USE OF ANY INFORMATION, APPARATUS, METHOD, PROCESS, ORSIMILAR ITEM DISCLOSED IN THIS DOCUMENT, INCLUDING MERCHANTABILITY AND FITNESSFOR A PARTICULAR PURPOSE, OR (II) THAT SUCH USE DOES NOT INFRINGE ON ORINTERFERE WITH PRIVATELY OWNED RIGHTS, INCLUDING ANY PARTY'S INTELLECTUALPROPERTY, OR (III) THAT THIS DOCUMENT IS SUITABLE TO ANY PARTICULAR USER'SCIRCUMSTANCE; OR

(B) ASSUMES RESPONSIBILITY FOR ANY DAMAGES OR OTHER LIABILITY WHATSOEVER(INCLUDING ANY CONSEQUENTIAL DAMAGES, EVEN IF EPRI OR ANY EPRI REPRESENTATIVEHAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES) RESULTING FROM YOURSELECTION OR USE OF THIS DOCUMENT OR ANY INFORMATION, APPARATUS, METHOD,PROCESS, OR SIMILAR ITEM DISCLOSED IN THIS DOCUMENT.

ORGANIZATION(S) THAT PREPARED THIS DOCUMENT

Energy Insight

ORDERING INFORMATION

Requests for copies of this report should be directed to EPRI Orders and Conferences, 1355 WillowWay, Suite 278, Concord, CA 94520, (800) 313-3774, press 2 or internally x5379, (925) 609-9169,

(925) 609-1310 (fax).

Electric Power Research Institute and EPRI are registered service marks of the Electric PowerResearch Institute, Inc. EPRI. ELECTRIFY THE WORLD is a service mark of the Electric PowerResearch Institute, Inc.

Copyright 2003 Electric Power Research Institute, Inc. All rights reserved.


5/70

iii

CITATIONS

This report was prepared by

Energy Insight2016 Klamath Avenue, #5Santa Clara, CA 95051

Principal InvestigatorW. Zhao

This report describes research sponsored by EPRI.

The report is a corporate document that should be cited in the literature in the following manner:

Short-Term Wind Generation Forecasting Using Artificial Neural Networks, EPRI, Palo Alto,CA: 2003. 1009219.


6/70


7/70

v

REPORT SUMMARY

Wind power is a highly intermittent power output resource that cannot be bid competitively in atraditional market due to scheduling problems associated with the resource. The CaliforniaIndependent System Operator (CAISO) has proposed a unique market arrangement that makessuch bidding possible. The central part of the arrangement is a provision that deviations betweenmetered and scheduled energy for participating intermittent renewable resources will beaveraged across a calendar month, and paid or charged based on the weighted average marketclearing price over that month. On March 27, 2002, the Federal Energy Regulatory Commission

(FERC) approved the proposal and ordered CAISO to report on the status of the program in 16months. This project investigated the application of artificial neural network (ANN) methods toperform short-term wind generation forecasting that will facilitate competitive marketparticipation.

BackgroundSufficiently accurate hour-ahead forecasting is important to ensure wind power producers acompetitive position as energy market bidders. CAISO has gained abundant experience withwind generation forecasting requirements and methods. However, there is a need to integrate thisforecasting technology into energy management systems for real-time dispatching. EPRI andCAISO cosponsored this project to investigate the application of ANN methods to short-term

wind generation forecasting.

Objective

To evaluate ANN methods for performing wind generation forecasting in order to develop a

short-term wind power forecast algorithm.

ApproachInvestigators took the basic approach of applying ANN methods to data analysis in order todevelop ANN models suitable for wind generation forecasting purposes. They performed afeasibility study comparing the performance and accuracy of the ANN models, then based theshort-term wind power forecast algorithm design on the most effective method identified.

By way of explanation, neural networks are composed of simple elements operating in parallel.These elements are inspired by biological nervous systems. As in nature, the network function isdetermined largely by the connections between elements. A neural network can be trained toperform a particular function by adjusting the values of the connections (weights) betweenelements.


8/70

vi

Back propagation networks are developed from the simple Delta rule in which extra hiddenlayers (layers additional to the input and output layers, not connected externally) are added. Thenetwork topology is constrained to be feed-forward, meaning loop-free. Generally, connectionsare allowed from the input layer to the first (and possibly only) hidden layer, from the firsthidden layer to the second, and from the last hidden layer to the output layer.

Results

A short-term wind forecast program based on ANN methods has been tested using the dataprovided by EPRI. Investigation of the factors that have significant impact on wind generationforecast indicates that

The correlation coefficients between wind speed and power output vary from 0.4 to 0.95.Apart from January, the coefficients in other months are all greater than 0.6. This indicatesthe existence of a strong correlation between wind speed and power output.

The coefficients between power output and wind speed one-hour previous vary from 0.4 to0.9. Apart from January, the coefficients in other months are all greater than 0.6. This

indicates that there is a positive correlation between those two parameters, with thecorrelation proving stronger in summer than in winter.

Because forecast accuracy is characterized by relative error, the study results indicate that neural-network-based forecasting models exhibit low forecasting error compared with statisticalanalysis results. In all, ANN-based algorithms can accurately predict hourly energy generation ofa wind power plant up to 24 hours in advance, allowing wind plants to participate in marketactivities.

EPRI PerspectiveAn accurate short-term wind generation forecasting tool will improve resource observability and

reduce forecasting uncertainty. This should, in turn, improve grid operations and reliability,result in tighter load following and ramp planning, and decrease reserve margins. It will also givewind energy suppliers, storage service providers, and power marketers access to more accurateinformation on near-term regional generation, thereby improving market efficiency. A number ofother benefits will follow. Wind energy suppliers will be able to reduce forecasting costs andfinancial risks when entering competitive bidding in real-time and next-day power supply andancillary service markets. In addition, a forecasting tool will enhance plant competitiveness anddispatching ability, which should accelerate investment in new capacity and help expand the roleof this emissions-free resource in regional and state portfolios.

KeywordsWind Power

Wind Energy ForecastingArtificial Neural Networks


9/70

vii

CONTENTS

1 INTRODUCTION ....................................................................................................................1-1

2THE WIND ENERGY..............................................................................................................2-1

Theoretically Available Wind Energy.....................................................................................2-1

Actual Available Wind Energy ...............................................................................................2-1

3NEURAL NETWORKS...........................................................................................................3-1

Fundamentals of Artificial Neural Networks ..........................................................................3-1

Backpropagation Networks ...................................................................................................3-2

Network Structure.............................................................................................................3-2

Number of Neurons ..........................................................................................................3-3

Transfer Functions............................................................................................................3-4

Training Functions ............................................................................................................3-5

Training Styles..................................................................................................................3-5

Incremental Training ....................................................................................................3-5

Batch Training..............................................................................................................3-5

4DATA ANALYSIS AND PRE-PROCESSING ........................................................................4-1

Preliminary Data Analysis .....................................................................................................4-1

Data Visualization.............................................................................................................4-1

Correlation Studies ...........................................................................................................4-5

Data Processing....................................................................................................................4-7

5RESULTS - STAGE I..............................................................................................................5-1

Results Presentation .............................................................................................................5-1

Analysis of Sources of Forecast Error ...................................................................................5-5

From Data Collecting Side................................................................................................5-5

From Data Analysis Side ..................................................................................................5-5


10/70

viii

From Model Developing Side ...........................................................................................5-5

Summary...............................................................................................................................5-5

Future Work Suggestions......................................................................................................5-6

6RESULTS - STAGE II.............................................................................................................6-1

Data Analysis ........................................................................................................................6-1

Results Presentation .............................................................................................................6-2

Improve Generalization ....................................................................................................6-2

Find Appropriate Training Size .........................................................................................6-5

Consider Temperature Difference ....................................................................................6-6

Presentation of 10-Month Results ....................................................................................6-6

Analysis of Sources of Forecast Error .................................................................................6-25

From Data Collecting Side..............................................................................................6-25

From Data Analysis Side ................................................................................................6-25

Summary.............................................................................................................................6-25

7 CONCLUSIONS .....................................................................................................................7-1

8 REFERENCES .......................................................................................................................8-1


11/70

ix

LIST OF FIGURES

Figure 4-1 Wind Generation (kW) vs. Modified Wind Speed, July 2002....................................4-3

Figure 4-2 Wind Generation (kW) vs. Modified Wind Speed, January 2002..............................4-3

Figure 4-3 Power Output (W) vs. Temperature (F), July 2002 .................................................4-4

Figure 4-4 Correlation Coefficients between Power Output and Wind Speed ...........................4-5

Figure 4-5 Correlation Coefficients Comparison........................................................................4-6

Figure 5-1 Training Results by using Three Weeks Data .........................................................5-2

Figure 5-2 Comparison of Actual Power to Forecasted Power for July 2001 ............................5-3

Figure 6-1 Correlation Coefficients Comparison by using One-Hour Prior Data........................6-2

Figure 6-2 Comparison of Forecasted Power to Actual Power for July 2002 by usingdifferent Methods .............................................................................................................6-20

Figure 6-3 Comparison of Actual Power to Forecasted Power for December 2001 ................6-20

Figure 6-4 Comparison of Actual Power to Forecasted Power for January 2002....................6-21

Figure 6-5 Comparison of Actual Power to Forecasted Power for February 2002 ..................6-21

Figure 6-6 Comparison of Actual Power to Forecasted Power for March 2002.......................6-22

Figure 6-7 Comparison of Actual Power to Forecasted Power for April 2002..........................6-22

Figure 6-8 Comparison of Actual Power to Forecasted Power for May 2002..........................6-23

Figure 6-9 Comparison of Actual Power to Forecasted Power for June 2002.........................6-23

Figure 6-10 Comparison of Actual Power to Forecasted Power for July 2002 ........................6-24

Figure 6-11 Comparison of Actual Power to Forecasted Power for August 2002....................6-24

Figure 6-12 Comparison of Actual Power to Forecasted Power for September 2002.............6-25


12/70


13/70

xi

LIST OF TABLES

Table 4-1 Sample Data File from Wind Plant.............................................................................4-2

Table 4-2 Wind Generation (kW) vs. Wind Speed and Direction,October 2001~September 2002.........................................................................................4-4

Table 5-1 Forecasting Errors Estimation for July 2002..............................................................5-2

Table 5-2 Forecasting Errors Estimation for January 2002 .......................................................5-3

Table 5-3 Forecasting Errors Estimation for October 2001 .......................................................5-4

Table 5-4 Forecasting Errors Estimation for April 2002 .............................................................5-4

Table 5-5 Highest One-Day Mean Errors in Four different Months............................................5-5

Table 6-1 Error Comparison by using different Training Algorithms for April andMay 2002 ...........................................................................................................................6-3

Table 6-2 Error Comparison by using different Training Algorithms for June andJuly 2002............................................................................................................................6-4

Table 6-3 Sums of the Errors.....................................................................................................6-4

Table 6-4 Error Comparison by using different Training Size for January andAugust 2002 .......................................................................................................................6-5

Table 6-5 Forecasting Errors Estimation for April 2002 by Adding Temperaturedifference ...........................................................................................................................6-7

Table 6-6 Forecasting Errors Estimation for May 2002 by Adding Temperature difference ......6-7

Table 6-7 Forecasting Errors Estimation for June 2002 by Adding Temperaturedifference ...........................................................................................................................6-8

Table 6-8 Forecasting Errors Estimation for July 2002 by Adding Temperature difference.......6-8

Table 6-9 Power Output Estimation by using different Methods for July 2002 ..........................6-9

Table 6-10 Forecasting Errors Estimation for December 2001................................................6-10

Table 6-11 Forecasting Errors Estimation for January 2002 ...................................................6-11

Table 6-12 Forecasting Errors Estimation for February 2002..................................................6-12

Table 6-13 Forecasting Errors Estimation for March 2002 ......................................................6-13

Table 6-14 Forecasting Errors Estimation for April 2002 .........................................................6-14

Table 6-15 Forecasting Errors Estimation for May 2002 .........................................................6-15

Table 6-16 Forecasting Errors Estimation for June 2002 ........................................................6-16

Table 6-17 Forecasting Errors Estimation for July 2002 ..........................................................6-17

Table 6-18 Forecasting Errors Estimation for August 2002 .....................................................6-18

Table 6-19 Forecasting Errors Estimation for September 2002...............................................6-19


14/70


15/70

1INTRODUCTION

Despite the slowing of the U.S. economy, wind energy is boosting and appears ready for a periodof strong growth through 2003. This is largely due to the extension of the federal ProductionTax Credit as well as growing public awareness of the benefits of renewable energy sources.

Differs from conventional power energy, wind energy, cannot be bid competitively in atraditional market due to scheduling problems associated with the wind resource. CaliforniaIndependent System Operator (CAISO) has proposed a unique market arrangement that makesthis possible. The central part of the arrangement is a provision that deviations between metered

and schedule energy for participating intermittent renewable resources will be averaged acrossa calendar month, and paid or charged based on the weighted average market clearing priceover that month.

A sufficiently accurate hour-ahead forecasting [4, 5] is important to make wind power producerscompetitive as energy market bidders. This project uses sophisticated neural network models topredict the hourly energy generation of a wind power plant up to 24 hours in advance.

As described below, under the original Scope, the Wind Energy Forecasting Project was to bedeveloped in two phases:

Phase 1: Wind generation forecast modeling; Study the factors that have significant impact onthe wind generation production; Preparation of study data.

Phase 2: Algorithm design; Comparison study of various Artificial Neural Network (ANN)techniques for wind generation forecast; Tune the parameters of a few selectiveANN methods based on preliminary study results.

1-1


16/70


17/70

2THE WIND ENERGY

Air is constantly moving under the influence of the continual atmospheric pressure variations.The resulting air current is the wind. There are two parameters used to define the wind, winddirection and wind speed.

The wind blows from high-pressure zones to low-pressure ones theoretically. However atmedium and high latitudes, its direction is modified by the earths rotation. The wind directionis determined by the direction from which it blows.

The wind speed is measured by anemometers. There are three main classes of anemometersavailable: Rotational anemometers, pressure anemometers and others.

Theoretically Available Wind Energy

Theoretically, the available wind energy in the year is proportional to the area situated underthe power-duration curve (See Figure 2-1). The annual energy can be expressed as following.

dtVKPdtEtt

==0

3

0

Eq. 2-1

The unit ofEis kWh/m2and Kis equal to 0.37.

Actual Available Wind Energy

Since limitations are imposed by other factors, the wind machine cannot intercept all the energytheoretically available in the wind. The actual available energy is proportional to the hatchedarea as shown in Figure 2-2, where

Vm is the cut-in speed at which the wind machine begins to supply power.

VN is the rated wind speed at which the wind machine supplies its rated power or nominal

power. Above this value and lower than the cut-out value, the power will be kept constantby regulating system.

VM is the cut-out speed beyond which the wind machine is stopped for reason of safety, and

gives no more power.

2-1


18/70

The Wind Energy

T1 is the annual time during which the value of the speed is above the cut-out speed.

T2 is the annual time during which the speed is greater than the rated speed.

T3 is the annual time during which the speed is greater than the cut-in speed.

t

V3

Figure 2-1Theoretically Available Wind Energy

T 1 T 2 T 3t

Vm3

VN3

VM3

V3

Figure 2-2

Actual Available Wind Energy

2-2


19/70

3NEURAL NETWORKS

Artificial Neural Network (ANN) has seen an explosion of interest over the last few years,and is being successfully applied across an extraordinary range of problem domains, in areasas diverse as finance, medicine, engineering, geology and physics. Indeed, anywhere that thereare problems of prediction, classification or control, neural networks are being introduced.This sweeping success can be attributed to a few key factors:

Power.ANN is very sophisticated modeling techniques capable of modeling extremely complexfunctions. In particular, neural networks are nonlinear.

Ease of use.Neural networks learn by example. The neural network user gathers representativedata, and then invokes training algorithmsto automatically learn the structure of the data.Although the user does need to have some heuristic knowledge of how to select and prepare data,how to select an appropriate neural network, and how to interpret the results, the level of userknowledge needed to successfully apply neural networks is much lower than would be the caseusing some more traditional nonlinear statistical methods.

Neural networks are applicable in virtually every situation in which a relationship betweenthe predictor variables (independents, inputs) and predicted variables (dependents, outputs)exists, even when that relationship is very complex and not easy to articulate in the usual termsof correlations or differences between groups.

Fundamentals of Artificial Neural Networks

The artificial neural network is an important branch of Artificial Intelligence. Motivated in itsdesign by the human nervous system, ANN mimics the human nervous system in its operations.At this extraordinary interface between natural human systems and created electronic ones,ANN is capable of learning by training to generalize from special cases.

Neural networks are parallel-distributed models that have several distinguishing features:

A set of neurons;

An activation state for each unit, which is equivalent to the output of the unit;

Connections between the units. Generally each connection is defined by a weight wjk, which

determines the effect that the signal of unitjhas on unit k;

A propagation rule, which determines the effective input of the unit from its external inputs;

A transfer function, which determines the new level of activation based on the effective inputand the current activation;

3-1


20/70

Neural Networks

An external input (bias, offset) for each unit;

A method for information gathering (learning rule);

An environment within which the system can operate, provide input signals and, if necessary,error signals.

Backpropagation Networks

Backpropagation was created by generalizing the Widrow-Hoff learning rule to mutiple-layernetworks and nonlinear differentiable transfer functions. Input vectors and the correspondingtarget vectors are used to train a network until it can approximate a function. Properly trainedbackpropagation networks tend to give reasonable answers when presented with inputs that theyhave never seen. This feature makes it suitable for forecasting in power systems.

Network Structure

In theory, three-layer backpropagation networks can be used as a general function approximator.It can approximate any function with a definite number of discontinuities, arbitrarily well, givensufficient neurons in the hidden layer.

The work described in this report has been successively using backpropagation networks toget desirable results. Figure 3-1 shows the structure of backpropagation networks employedin this project.

IW1 1

B11

P1

+

A1

N2N1

LW2, 1

B21

+

A2

Input Hidden Layer Output Layer

A1=tansig (IW1,1P1+B1) A2=tansig (LW2,1A1+B2)

Figure 3-1Three-Layer Backpropagation Networks

The main tasks associated in network design are to choose related input parameters,

to decide proper numbers of neuron in the hidden layer and output layer, to choose

appropriate transfer functions and to use training function that provides best training

performance.

3-2


21/70

Neural Networks

Number of Neurons

The numbers of neurons in the input layer and the output layer is much easier to decide.The number of neurons in the input layer is the same as the number of input parameters,which will be discussed in the next section. The number of neurons in the output layer is one

since there is only one output parameter, which is the output energy.

Another important issue in designing a network is how many units to place in each layer.Using too few units can fail to detect the signals fully in a complicated data set, leading tounderfitting. Using too many units will increase the training time. A large number of hiddenunits might cause overfitting, in which case the network has so much information processingcapacity, that the limited amount of information contained in the training set is not enough totrain the network.

The best number of hidden units depends on many factors the numbers of input and outputunits, the size of the training set, the amount of noise in the targets, the complexity of the errorfunction, the network architecture, and the training function.

There are lots of rules of thumb for selecting the number of units in the hidden layers:

- between the input layer size and output layer size],[ nlm

3

)(2 nlm

+= - two third of the input layer size plus the output layer size

- less than twice the input layer sizelm 2

Short-Term Wind Generation Forecasting

Documents