Improved Feature Selection Algorithm of Electricity Price Forecasting using SVM

Electricity price is of essential to all market participants due to its substantial applications. Because of its importance, accurate day ahead forecasting price methodology has become the major interest for energy market researchers. However, considering the characteristics of electricity price such as non-linear, non-stationary and volatile, it is arguable that accurate electricity price forecasting indeed difficult and challenging to perform. Nevertheless, there remain an importance in having a more accurate electricity price forecasting methods. Therefore, in this paper, feature selection algorithm consist of threshold based mutual information, inequality correlations of symmetrical uncertainty and pairwise evaluation performance methods is proposed as a process to perform electricity price forecasting. Mutual information method is used to retrieve relevant features then continued with symmetrical uncertainty method to remove the redundant from relevant features to obtain non-redundant features, and finally, pairwise evaluation performance method is performed to obtain final optimal features to be used as an input data for forecasting using Support Vector Machine. The proposed feature selection algorithms are applied to Information gain, Relief, Best-first and Backward search method and provide forecasting errors of 4.11%, 4.11%, 4.23% and 4.32% also improve the forecasting performance by 1.22%, 1.24%, 1.25% and 1.16%, respectively.