You are on page 1of 4

2011 Third International Conference on Measuring Technology and Mechatronics Automation

Training Data Selection for Short Term Load Forecasting
Yang Yuhang, Lu Yingliang, Meng Yao, Xia Yingju, Yu Hao
Fujitsu Research & Development Center Co., LTD. Beijing, 100025, China yyh@cn.fujitsu.com

Abstract—Short term load forecasting (STLF), which aims to predict system load over an internal of one day or one week, plays a crucial role in the control and scheduling operations of a power system. Most existing techniques on short term load forecasting try to improve the performance by selecting different prediction models. However, the performance also rely heavily on the quality of training data. This paper focuses on training data selection which is rarely considered in the previous studies. A novel approach is proposed to filter out abnormal data by analyzing historical load curves. Experiments conducted on the real load data show significant improvement over the baseline method using the same data for training. Furthermore, the approach is feasible and thus can be applied in different situations. On one hand, the approach achieves promising results by using only historical load data which is shown in this study. On the other hand, more information can be easily integrated in the approach to select more appropriate training data for further performance improvement. Keywords- Short Term Load Forecasting; Training Data Selection; Load Curve Analysis

I.

INTRODUCTION

Load forecasting, which anticipates the future load by analyzing the historical data, plays a crucial role in the efficient planning, operation and maintenance of a power system. Short term load forecasting (STLF), which aims to predicts system load over an internal of one day or one week, is necessary for the control and scheduling operations of a power system. Further analyses such as load flow are also based on the results of the short term load forecasting. Most existing techniques on short term load forecasting try to improve the performance by selecting different prediction models, such as linear regression[1], exponential smoothing, stochastic process, auto-regressive and moving average (ARMA) models[2], data mining models, and the widely used artificial neural networks (ANN)[3,4]. However, different models suffer from different problems. Linear regression does not perform well due to the lack of selflearning capability. Furthermore, weather changes cannot be easily integrated in linear regression models. In time series methods, load data are regard as cyclical time series changed by season, week, day or hour. Auto-regressive (AR) models, moving average (MA) models and ARMA are typical time series methods which have been widely used for load forecasting. But time series methods, which are vulnerable to dirty data, rely heavily on the quality and size of historical data. In the recent studies, the ANN has been popular used for short term load forecasting with satisfactory results.
978-0-7695-4296-6/11 $26.00 © 2011 IEEE DOI 10.1109/ICMTMA.2011.830 1040

Since the input variables of ANN are difficult to be selected, choosing appropriate training data becomes crucial in ANN based STLF methods. As a result, some previous studies concentrate on selecting historical data for training. Similar days to the forecasted day are extracted in different ways and thus the data of the specific days are used for training. Fuzzy logic and case-based reasoning (CBR) are applied to depict different clusters of days[5]. According to the usage of weather changes, STLF techniques can be divided into three major groups[6]: non-weather sensitive models, weathersensitive models, and hybrid models. Most similar day selection based STLF methods, which consider weather variables (such as temperature and humidity) and day type, are weather-sensitive. Load is closely related to the weather changes. But weather is not the only factor which leads to load changes. Load curves themselves are the most direct way to characterize the abnormal days. However, load curves, which have a greater impact on load forecasting, are rarely considered in recent studies. In this paper, a novel approach based on training data selection is proposed for short term load forecasting. The motivation of the proposed approach is to identify and filter out the abnormal data by analyzing load curves. Thus more appropriate training data are selected for STLF. Experiments are conducted on the real load data to verify the efficiency of the approach. The proposed approach can work by using only historical load curves if the weather information are not available. On the other hand, weather information can be easily integrated in the model. Thus the performance can be further improved by using more information. The rest of the paper is organized as follows. Section 2 describes the proposed algorithms. Section 3 explains the experiments and the performance evaluation. Section 4 is the conclusion. II. METHODOLOGY

As described in Section 1, selecting similar days to the forecasted day is important for short term load forecasting. The main purpose of the proposed approach is to filter out abnormal days by analyzing historical load curves. Thus the accuracy of load forecasting could be improved by using more appropriate training data. The framework of the proposed model for short term load forecasting is shown in Figure 1. It consists of 5 steps listed below.

duke-energy. similar days to the forecasted day are extracted by using day type and weather variables (if available). 2. the load curves of 2 days (48 hours) are relatively lower than the others which indicates the difference between load curves of week days and weekends. 2. Furthermore. 3 shows the load curves of all Thursdays and all Sundays. is used for training. the load curves of Thursdays can be divided into two major types according to different seasons.dimensional Euclidean space. load curves have weekly and daily periodicity. there is one peak or two in each day (24 hours) which represent the daily periodicity. The dataset is downloaded from Duke Energy Carolinas The (http://www. As shown in Fig. III. the historical load curves with similar trends and values will be useful for 1041 . The load curves of similar days with distances longer than a predefined threshold are removed from the training data. Step 3: The distance from every extracted similar day to the cluster center is calculated. Fig. Step 4: The days having long distances to the cluster center are recognized as abnormal day and filtered out. Fig. y2. If available. Besides. when we predict the load of 7:00 am of a given day. Thus every load curve is not limited in a single day. the cluster center is obtained by calculating the average value of the similar days. As described before. all Mondays in the historical data are regarded as similar days if the forecasted day is Monday. As revealed by the figures. The last 24 hour load data before this hour. Step 2: The cluster center of the similar days is obtained.com/Load/Load. As shown in Fig. Step 5: A prediction model trained by the load curves of the selected days is applied to forecast the load of the given day. For an instance.htm). A set of experiments can be conducted for parameter tuning while abundant features are obtained. The framework of the proposed load forecasting model Step 1: Given a set of historical load data. The first type has two lower peaks and the second type has a higher peak.Historical data day types are mostly similar according to many previous studies. 3 further reveals the difference between week days and weekends.ferc. The last 24 hour load information is used to predict the next hour load. The reason is that the load curves of the days with the same A. such as the lowest load curve in Fig. The similarity between a similar day and the cluster center is measured by Euclidean distance shown as Equation (1). data of every day consists of 24 hour load information. humidity and air pressure are proved to be effective for load forecasting according to previous studies. When the weather information is not available. x2. from January 2008 to December 2008 of a particular data set used. After similar day selection. The weather variables including temperature. For an instance. The load data of the same hours in similar days are utilized to train the ANN model. Using abnormal data for training is likely to mislead the forecasting results. Similar day extraction d ( X .Y ) = Cluster centroid ∑ (x i =1 n i − yi ) 2 (1) Load curves of similar days Center indentification Abnormal data filtering Filtered training data where X={x1. Data preparation The performance of the method for the short term load forecasting is tested by using the 12 months data. …yn} are two nodes in N . The other one containing the December data is used for testing. The load curves of June and October are shown in Fig. In every 7 days (168 hours). similar day selection can be simplified by choosing days with the same day types. the load of June is much higher than the load of October due to the seasonal variation. Artificial neural network is applied as the prediction model in the proposed approach. is used for prediction. the weather variables are combined with the day type to select similar days. which is from 7:00 am of the day before the given day to 6:00 am of the day. which contains 11 months data from January to November. …xn} and Y={y1. Fig. PERFORMANCE EVALUATION Load forecasting Forecasted results Figure 1. The first one. 3(a). Most load curves have the same trends and similar values except a few abnormal ones. The 12 months data is divided into two data sets. 3 indicates that most load curves of the same day type have similar trend. 3(a).

0 0.7 0.5 1. N denotes the number of the hours of a day in this work (N = 24).4 1.7 1.7 0. Baseline methods An artificial neural networks based method.3 1.6 h104 Load(MW) Load(MW) 2 4 6 8 10 12 14 Hours 16 18 20 22 24 (b) The load curves of all Sundays 0 100 200 300 400 Hours 500 600 700 Figure 3.2 1. is taken as baseline method for comparison based on two reasons.6 h104 Load(MW) 2 4 6 8 10 12 14 Hours 16 18 20 22 24 (a) The load curves of all Thursdays 1042 .6 1.9 1.4 1.3 1.2 1.1 1.8 1.3 1. Results and analysis The forecasted results using the proposed approach compared with the actual load are shown in Fig.9 1. which is in accord with the motivation of this study.6 1.9 1.1 1.9 1.4 1.8 1.0 1. PF denote the actual and forecasted values of the load.8 1.8 0. D.5 1. MAPE (%) = Load(MW) 1 N | PAi − PFi | ∑ P i ×100 N i =1 A (2) where PA.9 0.9 0.7 1.0 0.6 1.1 1. The load curves of the same day types (a) The load curves of June 2. The second reason is that the improvement by using a novel training data selection strategy can be clearly revealed by using the same forecast model.0 1.4 1.6 1. respectively.0 0.3 1.7 1.7 0.load forecasting by filtering out the abnormal data.8 0.9 0.1 1. 4.5 1.0 0. The load curves in a month 2. thus using ANN for comparison is fair enough.5 1. 2.7 1.2 1.6 h104 B.6 h104 2.8 1.0 1.8 0. The load curves of only 2 days are shown in the figure due to the space limitation.9 0. The first reason is that ANN is proved to be efficient for load forecasting in literature.0 1. which uses the load curves of the same day types for training without filtering out any abnormal days. 0 100 200 300 400 Hours 500 600 700 (b) The load curves of October Figure 2. Evaluation metric The forecasted results deviation from the actual values are evaluated in term of Mean Absolute Percentage Error (MAPE) shown in Equation (2).8 0.7 0. C.2 1.

“Short-term load forecasting via ARMA model identification including nongaussian process considerations. No. Soliman S. D. Chuangxin Guo. S.Electric Power System Research. Huang and K.42 10. [7] T. [8] S. 4.” IEEE Trans. Transactions of the Institute of Measurement and Control. vol. “Neural network based short term load forecasting..1993.8%.73 10. “Regression based peak load forecasting using a transformation technique. 2003.”IEEE Trans. pp. vol. 4. 4 vividly shows that the actual and forecasted load curves have similar trends and values. no.3 1. Hippert.15 10. vol. 1993. Actual and forecasted load curves Fig. REFERENCES Actual Forecast 2 4 6 8 10 12 Hours 14 16 18 20 22 24 [1] (b) Actual and forecasted load curves of Dec. 6. Shih. “A priority vector based technique for load forecasting. pp.508–514. The performance will be further improved by using more information such as weather variables. 516–521.02 Baseline 4.8 h104 This paper presents a novel approach for short term load forecasting by selecting appropriate training data. 28. 6 Figure 4. MAPE error of different approaches As shown in Table 1. C. 336–342.24 9. “A generalized knowledge-based short term load-forecasting technique. [9] H. no.9 0. 17..1. vol. 1459– 1464.. 1994. Short-term load forecasting based on Kalman filtering algorithm with moving window weather and load model. A Fault Diagnosis Method for Transformer Integrating Rough Set with Fuzzy Rules.1788–1794. 2004.4 1. no.” IEEE Trans. vol. no. 1. EPSR. Actual Forecast CONCLUSIONS 2 4 6 8 10 12 Hours 14 16 18 20 22 24 (a) Actual and forecasted load curves of Dec. Power Syst. 3. 44–55. 18. no.1 1. The approach achieves promising results by using only historical load data which is shown in this study.” IEEE Trans.26 9..07 7.Power Syst. C. It should be pointed out that the proposed approach achieves promising results by using only limited historical load data. 2. 2001. In other words. Vemuri. 1. “A regressionbasedapproach to short-term load forecasting. E.00 Table 1. [10] S.1 1. Hesterberg. “A data mining approach for spatial modeling in small area load forecast.0 0. 1 1.. Experiments are conducted on the real load data to verify the efficiency of the proposed approach.” IEEE Trans. 16.” IEEE Trans. 9. 2. pp. the efficiency of the proposed approach has been initially proved. 2. no. Shrestha.43 10. Lu. 1993.0 0. vol. Souza. 2003. Power Syst. 8. vol.05 7. Pedreira. pp. R. Thus the performance could be further improved by considering more factors such as weather changes.50 12.73 9. Vol.98% higher performance on average compared to the baseline method.” IEEE Trans. 4. [3] C. IV. Rahman and G.” IEEE Trans.. and R. [5] Zhiyong Wang.76 12.5 1. The proposed approach indentifies abnormal days by analyzing the load curves instead of using weather variables in most existing techniques.01 14. pp. no.5. Muto. Haida and S. the proposed approach provides more than 0. The proposed approach outperforms the baseline method using the same prediction model. Quanyuan Jiang and Yijia Cao.. These translate to improvements of A. 1990. Power Syst. Power Syst. Papalexopoulos and T. 1535–1550. MAPE(%) Day Dec 1(Monday) Dec 2(Tuesday) Dec 3(Wednesday) Dec 4(Thursday) Dec 5(Friday) Dec 6(Saturday) Dec 7(Sunday) Overall TDS 3. This indicates that the proposed model is benefit from the training data selection strategy.Wu and C.2 1.96 10. 673–679. [2] S. vol. Abnormal days in the training data would have a serious impact on the forecasting performance. Power Syst. 68: 47-59. “Neural networks for short-term load forecasting: A review and evaluation. Power Syst. no. Hazim.2 Load(MW) 1. 1043 . pp. 8. The performance of the proposed approach based on training data selection (TDS) is also measured in terms of the MAPE error. Power Syst. Rahman and O.86 10. pp. 243-251 (2006) [6] Al-Hamadi H.4 1. [4] H. The proposed approach and the baseline method use the same prediction model and resources. N. The performance is compared with the reference algorithm which is shown in Table 1..9 h104 precision for over 9. More information can be easily integrated in the approach to select appropriate days for training. pp. Lu and S. C.3 Load(MW) 1. The approach is feasible with or without weather variables. J.