Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

2009 International Conference on Computational Intelligence and Natural Computing

Back Propagation Neural Network for Short-term Electricity Load Forecasting with Weather Features
Yong Wang1,2 Dawu Gu1 Jianping Xu2 Jing Li2
2. Dept. of Computer Science and Technology Shanghai University of Electric Power Shanghai, China

1. Dept. of computer Science and Engineering Shanghai Jiao Tong University Shanghai, China E-mail: wy616@126.com
AbstractIn this paper, we expand previous work and present an accurate electricity load forecasting algorithm with back propagation neural networks. It contributes to short-term electricity load forecast methodology with neural network with weather feature such as max centigrade, min centigrade and weather types. The original electricity load is from shanghai district, which is composed of original every 5 minutes load records. Through the data transform the every hour data average 12 original records and weather feature become double value. After training and simulation, the prediction errors fit for the application needs. The algorithm besides other algorithms has been used in electricity load forecasting software. Many results confirm that the proposed method is capable of forecasting load efficiently. Keywords-neural network; forcast; electricity load; short term; weather features

I.

INTRODUCTION

Electricity load forecasting has become an important task for the Demand Side Management (DSM) of power system. The time of use electricity load is one of methods in DSM, which can decrease peak load in summer without building more power plants. Because the importance of electricity load forecasting, many technology is used for predict short-term or long-term prediction, such as linear regression, time-series techniques, data mining approach, fuzzy theory, neural network and so on. The neural network model can be used in many areas [11], which also widely used in electricity load forecasting. There are many application researches in the area around the world. A simple yet accurate model to forecast electricity load with neural networks is used after analyze the problem domain and choose the most adequate set of attributes [2]. An approach to predict the market behaviors uses the historical prices, loads and other required information for short-term electricity price and load forecasting [3]. The wavelet neural networks are used in electricity forecasting. The use of linear and nonlinear regression fed with wavelet-transformed data can aid in capturing useful information on various time scales in neural network training procedure [1]. The wavelet neural networks with genetic algorithms have the ability of optimization of the overall search capabilities [8]. As the power load forecasting sequence has stochastic growth and nonlinear wave characteristics, grey neural network model can effective reflect the growth properties of the sequence and fit the nonlinear relation.
978-0-7695-3645-3/09 $25.00 2009 IEEE DOI 10.1109/CINC.2009.107

Gray neural network utilizes the accumulation generation operation of grey prediction to transform the original load data to better regularity data, making it easier for neural network to model and forecast. Gray neural network is suitable for middle and long term load forecasting, and case study shows that its forecasting accuracy is better than that of neural network and grey prediction method [9]. According to the random-increase and non-linearity relationship, the improved ant colony method is used as the basis of combination weight making of gray neural network, so as to achieve the goal of optimizing the whole forecasting precision and find the combination weight that can exhibit the high consistency and high precision for the series values, finally the whole forecasting accuracy can be improved obviously [4]. Markov chain can easily embody the random characteristic of system by complex factors. The whole forecasting precision of the sequence was optimized by Markov chain error correction, and then the accuracy for power load forecasting was greatly improved [5]. According to the chaotic and non-linear characters of power load data, the model of artificial neural networks neural network based on Lyapunov exponents was established [6]. The neural network has been applied to the area of power load forecast successfully, but it has such disadvantages of local optimization and slow convergence speed. A new kind of neural networks forecast model based on culture particle swarm optimization was proposed for overcoming those disadvantages [10]. The deviation of gray neural network can be optimized step by step with less deviation based on particle swarm optimization [7]. This paper contributes to short-term electricity load forecast methodology with neural network, in which load curves are forecasted by the input information of the past every average hour electricity load from original every five minutes load records, max centigrade, min centigrade and weather feature. Our electricity forecasting software structure with multi predication model is described. Many results confirm that the proposed method is capable of forecasting load efficiently. II. NEURAL NETWORK ARCHITECTURES FOR ELECTRICITY LOAD FORECASTING

Our electricity load forecasting software uses linear regression, time-series techniques, data mining approach, fuzzy theory and neural network algorithm to predict short- term and long-term electricity load. The neural network is main part of the forecasting function. The frameworks of the neural network prediction are as follows:
58

A. Architecture of Elctricity Load Forecasting with Neural Network from Out Datset The electricity data is stored in SQLserver2005, which is 45M large of five years everyday records. In order to import the data to Matlab workspace, the ODBC connection is needed. We need firstly create system or user DSN connection by data source tools of control panel. After the data imported into the Matlab workspace, the data needs to be transformed into double type, which fits for the neural network demand. Finlay, train the input data until it meet with the criteria, simulate the electricity load, compared with the forecasting data with the real past electricity data. The accuracy stands for the software efficient. In order to improve the accuracy, we used to using muti-algorithm for prediction. The architecture of electricity load forecasting with neural network is as Figure 1:
Start Electricity Load Data Set ODBC Connection Import Data to Matlab Data Transform Train Neural Network Input Vectors Meets Criteria Change Weights Y

III.

NEURAL NETWORK DATA PROCESSING

Simulate Electricity Load Forecast Electricity Load End

Figure 1. Architecture of electricity load forecasting with neural network from out dataset

A. Import Original Data from ODBC Connection to Matlab Workspace Open Database Connectivity (ODBC) is parts of relation al database of windows open services architecture, which defined a group of application program interface. The Stand Query Language (SQL) can be used to connect or search the different database. In order to import data from SQL server2005 to Matlab workspace, the following codes are widely used in the program. % create connection connectionPowerLoad=database('powerload','',''); ping(connectionA); %get ODBC information cursorPD=exec(connectionPowerLoad,'SELECT clock0,clock1,clock2,clock3,clock4,clock5,clock6,clock7,cl ock8,clock9,clock10,clock11,clock12,clock13,clock14,cloc k15,clock16,clock17,clock18,clock19,clock20,clock21,cloc k22,clock23 FROM EveryDay'); %get ODBC pointer cursorPD=fetch(cursorPD); %get data from pointer P=cursorPD.Data; % assigned to input data close(cursorPD); %close pointer cursorWeather=exec(connectionPowerLoad,'select maxTemprature, minTemperature, weather, from EveryDay ); cursorWeather=fetch(cursorWeather); Pweather=cursorWeather.data; close(cursorWeather); close(connectionPowerLoad); P=cat(2,P,Pweather); The cursor is used to locate the selected records. After open database and search operation, should close the database to avoid wrong operation. In the end connected the hour electricity load database and the weather feature database into a whole input data vector matrix. B. Original Data Ttransform with Weather Features The input data in Matlab workspaces is greater than needed number. The electricity load arranges from 494.9139MW to 751.8833MW. The weather records include minimum, maximum centigrade and weather feature. The neural network needs input data range from 0 to 1. There are lots of methods that can transform the original data to Matlab data. The widely used transform method is as Equation 1:

B. Structure of BP Neural Network for Electricity Load Forecasting The Back Propagation (BP) neural network forecasting electricity is a single hiding layer network. The input vectors include 27 nodes. There 24 every hour electricity load with max temperature, min temperature and weather feature. According to the Kolmogorov theorem, the number of inside layer neural network is square of input vectors and added another one in the end. So the hide layer number is 55 nodes. The output vectors have 24 nodes for every hour electricity load of every day. The structure of neural network forecasting electricity load is as Figure 2.

x=

x xmin xmax xmin

(1)

The original electricity is recorded every 5 minutes, which is averaged for one hour input records. The input electricity load is transformed by the Equation 1 as table 1:
TABLE I.
Date

SHANGHAI DISTRICT ELECTRICITY LOAD TRANSFORM


Electricity Load (MW) Transformed

2008-7-21 1:0 2008-7-21 2:0 Figure 2. Structure of neural network forecasting electricity load 2008-7-21 3:0 2008-7-21 4:0 2008-7-21 5:0 2008-7-21 6:0

560.9933 539.6968 517.8898 505.0693 502.3033 507.2348

0.2572 0.1743 0.0894 0.0395 0.0288 0.0479

59

The minimum and maximum centigrade are transformed by divided by 100 centigrade. The weather features transform rules are as Table 2.
TABLE II.
Min Centigrade

0.14 0.12 0.1 0.08 errors 0.06 0.04 0.02 0 -0.02 -0.04 -0.06 output vectors 1 3 5 7 9 11 13 15 17 19 21 23

WEATHER RECORDS TRANSFORM


Max Centigrade Weather Types Weather Meaning

0.23 0.24 0.22 0.21 0.22 0.23

0.30 0.31 0.29 0.28 0.30 0.32

0 0.2 0.4 0.6 0.8 1

thundershower shower moderate rain light rain cloudy sunny

Any other weather features rules can be used in electricity forecasting. Its very convenient to redefine more rules as Table 2. C. Back Propagation Neural Network Data Training Back propagation neural network data needs training for electricity load prediction. The transport function of internal layer if tangsig and the output layer transport function is logsig. Because the structure of network is complicated, the network needs increase training frequency and learning rate. The training epochs is 100, goal is 0.01 and learning rate is 0.1. After 78 times training, the result is as Figure 3.

Figure 4. Errors of real electricity load and prediction electricity load with neural network

As the Figure 4, the error between forecasting electricity load and real electricity load is tiny. The biggest error is also below 0.12. The results fit for the application needs. B. Forecasting Electricity Load Results Anlysis The real summer electricity load has magnificent load peak and load valley characteristic. The peak electricity load is at clock 11AM, when the temperature is very high and workers still work. The valley electricity load is at clock 7AM, when the temperature is lower and people wake up without using air conditioner. The result is as Figure 5.
real data
800 700

forecasting data

electricity load

600 500 400 300 200 100 0 1 3 5 7 9 11 13 15 17 19 21 23

hours

Figure 3. Results of neural network training

Figure 5. Comparation with real electricity load and prediction electricity load by neural network

IV.

SIMULATION AND RESULTS ANALYSIS

A. Simulation of Electricity Load The trained neural network needs simulation to judge whether the network fits for the electricity load forecasting application. The test summer data is from shanghai industry one of industry districts. We use the average electricity load of past five years the same day and the weather features to predict next day or next week electricity load. The test environment is 45M electricity load records, Matlab 7.0.0.1(R14), SQLserver2005, Intel Core2 Duo CPU E4600 2.4GHZ, and 2G DDR. The simulation function is used to test the forecasting data and real data. The forecasting error is as Figure 4:

As the Figure 5, the forecasting electricity load is near similar to the real electricity load with tiny errors. The bigger error always happened on the turn point of the electricity load in the morning or the night. C. Application in Software of Electricity Load Forecasting The neural network algorithm is used in electricity forecasting software. The electricity load data is from Shanghai some districts. According to the weather records and original data, the software can get better short-time forecasting results. There are other prediction algorithms used in the application software, such as linear regression, time-series techniques, data mining approach, fuzzy theory, and so on. The neural network training time is longer than other algorithms.

60

Although the prediction errors are tinny, special records always happen on holidays, or in suddenly shorttime thundershower weather. V. CONCLUSIONS AND FUTURE WORK In this paper, we expand previous work and present an accurate electricity load forecasting algorithm with neural networks. It contributes to short-term electricity load forecast methodology with neural network, in which load curves are forecasted by the input information of the past every average hour electricity load from original every five minutes load records, including max centigrade, min centigrade and weather feature. In the end, the algorithm with other algorithms is used in our electricity forecasting software. Many results confirm that the proposed method is capable of forecasting load efficiently. Future work needs much research, because the many prediction algorithms work separately. The electricity load prediction software has the prediction menu, each item can use single forecasting algorithm. Next work, we prepared count all prediction errors from different algorithm and append the main item of multi-forecasting for prediction. The electricity load long-term forecasting is not similar to short-term forecasting as neural network input affection factors. The first affection factors are composed of Gross Domestic Product (GDP), heavy industry GDP, light industry GDP, agricultural GDP, the primary industry, secondary industry and tertiary industry. Besides these affection factors, other international or domestic natural incidents should be considered, such as earthquake, tsunami, volcanic eruption, and so on. The most import affection factors are human incidents, such as financial crisis, inflation, deflation, trade fiction, epidemic, political instability, local war, and so on. As the listed reasons, its difficult for us to predict the long-term electricity load. Although its obviously difficult to predict the earthquake and worldwide financial crisis, long-term electricity load forecasting is needed for national development and reform commission of china. According to the given original electricity load data and all kinds of natural or human affection data, we will improve our longterm electricity load forecasting efficiency. ACKNOWLEDGMENT Sponsored by the Hi-Tech Research and Development Program of China under Grant (No. 2006AA01Z405). Shanghai Postdoctoral Scientific Program No. (No.08R214131). Innovation Program of Shanghai Municipal Education Commission (No. 09YZ346).

REFERENCES
[1] D. Benaouda, and F. Murtagh, Electricity Load Forecast using Neural Network Trained from Wavelet-Transformed Data, Engineering of Intelligent Systems, 2006 IEEE International Conference, IEEE Press, 2006, pp. 1-6, doi: 10.1109/ICEIS.2006. 1703163. [2] D.Ortiz-Arroyo, M.K.Skov, and Q.Huynh, Accurate Electricity Load Forecasting with Artificial Neural Networks, Computational Intelligence for Modelling, Control and Automation, 2005 and International Conference on Intelligent Agents, Web Technologies and Internet Commerce, IEEE Press, Nov. 2005, pp. 94-99, doi: 10.1109/CIMCA.2005.1631248. [3] P.Mandal, T. Senjyu, and T.Funabashi, Neural network models to predict short-term electricity prices and loads, Industrial Technology, 2005. ICIT 2005. IEEE International Conference on, IEEE Press, Dec. 2005, pp. 164-169, doi: 10.1109/ICIT.2005.1600629. [4] Dongxiao Niu, Jialiang Lv, and Jianrong Jia, Gray Neural Network Forecasting Model of Power Load Based on Ant Colony Algorithm Method, Risk Management & Engineering Management, 2008. ICRMEM '08. International Conference, IEEE Press, Nov. 2008, pp. 222-226, doi: 10.1109/ICRMEM.2008.30. [5] Dongxiao Niu,and Jialiang Lv, The Markov Error Correcting Method in Gray Neural Network for Power Load Forecasting, Risk Management & Engineering Management, 2008. ICRMEM '08. International Conference, IEEE Press, Nov. 2008, pp. 202-205, doi: 10.1109/ICRMEM.2008.36. [6] Yongli Wang, Dongxiao Niu, and Jiangyan Liu, Optimization of Artificial Neural Networks Based on Chaotic Time Series in Power Load Forecasting Model, Natural Computation, 2008. ICNC '08. Fourth International Conference, IEEE Press, Oct. 2008, pp. 106110. doi: 10.1109/ICNC.2008.777. [7] Chunjie Li, Tao Chen, Jun Dong, and Wen Chen, A Novel Particle Swarm Grey Neural Network Model for Power Load Risk Forecasting, Wireless Communications, Networking and Mobile Computing, 2008. WiCOM '08. 4th International Conference, IEEE Press, Oct. 2008, pp.1-4, doi: 10.1109/WiCom.2008.2478. [8] Zhengyuan Jia, Li Tian, and Dan Zhao, Application for ShortTerm Power Load Forecasting Using Improved Wavelet Neural Networks Based on GA, Risk Management & Engineering Management, 2008. ICRMEM '08. International Conference on, IEEE Press, Nov. 2008, pp.353-356, doi: 10.1109/ICRMEM.2008.40. [9] Dahai Zhang, Zhiwu Ren, Yanbing Bi, Dazhou Zhou, and Yanqiu Bi, Power load forecasting based on grey neural network, Industrial Electronics, 2008. ISIE 2008. IEEE International Symposium on, IEEE Press, July. 2008, pp.1885-1889, doi: 10.1109/ISIE.2008. 4677085. [10] Dongxiao Niu, Zhihong Gu, and Mian Xing, Research on Neural Networks Based on Culture Particle Swarm Optimization and Its Application in Power Load Forecasting, Natural Computation, 2007. ICNC 2007. Third International Conference, IEEE Press, Aug. 2007, pp.270-274, doi: 10.1109/ICNC.2007.627. [11] Yong Wang, Dawu Gu, Wei Li, Hongjiao Li , and Jing Li, Network Intrusion Detection with Workflow Feature Definition using BP Neural Network, Six international Symposium on Neural Networks ISNN2009 Wuhan China,May. 2009.

61

You might also like