Short term load forecasting using a hybrid intelligent method

Due to the regulation of electrical power systems, electricity market players need precise information of electrical energy consumption and generation in order to maximize their benefit based on appropriate decisions. In this paper a new hybrid intelligent method is proposed for short term load forecasting. In this method, load and temperature of previous days are used for prediction of the next hour electrical load consumption. Since electrical load signals are non-stationary, Wavelet Transform (WT) as a powerful signal analyzer is applied for the signal decomposing. For elimination of redundant data from input matrices, the Feature Selection (FS) method based on Gram–Schmidt (GS) is used for selection of more valuable features. The elimination of redundant data can speed up learning process and improve the generalization capability of the prediction scheme. Support Vector Machine (SVM) with simple structure and few tuning parameters is applied as a powerful regression tool. Two separate structures are considered for prediction of weekday and weekend electrical load consumption. Besides, in order to increase the forecasting accuracy, indices are determined for each day. The simulation results reveal that the Coiflet wavelet function with 2 decomposition levels lead to the best detection accuracy. Moreover, 30 dominant features of previous 50 days should be used to obtain minimum forecasting error. Comparative results show the priority of the proposed method in aspect of prediction accuracy as compared to some reported algorithms.