Application and Comparison of Evolutionary Techniques for Forecasting the Hellenic Grid Electricity Load

Abstract

Abstract. Electric load forecasting is a process that has to be both fast and reliable. An accurate method of load forecasting plays the most crucial role in achieving the aforementioned properties and also is a valuable tool in overcoming a variety of economic and operational problems connected to electrical energy production and distribution. In this study real data is used and the performance of three different techniques for adaptive electric load forecasting is evaluated. The first method is a combination of the well-established multimodel partitioning filter (MMPF) implementing extended Kalman filters (EKF) with Support Vector Machines (SVM), the second is the adaptive MMPF Kalman filters (KF) model and the third one is an artificial three layer feed-forward neural network (ANN). The results indicate that all three methods are reliable, however the combination of MMPF and SVM provides a more accurate load forecasting and at the same time identifies successfully both normal periodic behavior and any unusual activity of the electric grid.Keywords: Adaptive multimodel partitioning filter (MMPF), Support vector machines (SVM), artificial neural networks (ANN), forecasting, Kalman filters (KF), electricity demand load. IntroductionElectrical energy is considered as one of the most important factors that are closely related to both economic and social development. Accurate and fast load forecasting is related to a number of operational functions, such as system safety, power system real-time control, economic planning, grid's maintenance and also fuel planning. The problem of load forecasting has been studied extensively during recent decades. Some of the proposed techniques make use of time series analysis using ARMA [1][2][3][4][5] or ARIMA models [6][7][8][9][10]. Other algorithms achieve load forecasting by adopting evolutionary techniques such as ANN's [11-12], SVM's [13][14] either alone or combined with other methods for the same purpose [15][16][17].The purpose of this study is not to introduce one more load forecasting criterion, but it focuses on applying three different methods to real electric load data and evaluating their performance.The first method is based on a hybrid model that combines the adaptive MMPF [18][19][20], known for its stability, with SVM. The idea of using this method for electric load forecasting came from the fact that it was applied for wind speed prediction with very good results [21]. The data used now is not subjected to any prior -offline manipulation in order to remove weekly and annual seasonality as was done in previous cases [5]. That's why in this case the MMPF implements a bank of extended Kalman filters (EKF) with ARMA models instead of simple Kalman filters (KF) with ARMA models in order to handle data's non-linearities. MMPF with EKF combined with genetic algorithms (GA) were successfully applied in prediction of epilepsy and in the evolution of stock values using biomedical and financial data respectively [22].The second method, presented analytically in [5], impl...