ON THE PREDICTION ACCURACY OF THE MID-TERM ELECTRICITY PRICES USING OPTIMIZED SUPPORT VECTOR MACHINE

Abstract

In the modern electricity market, it is crucial to have precise electricity price forecasting. However, few studies have focused on this area. Mid-term electricity price forecasting (MTEPF) has numerous applications, such as scheduling future power plant maintenance, risk management, planning future contracts, purchasing raw materials, and determining market pricing. To forecast electricity prices, some factors are especially significant, such as choosing the most useful price features that influence the market price, and choosing the proper prediction model that is able to predict price behavior using historical data. In forecast modeling, feature selection techniques are an important step in data pre-processing prior to creating the prediction model. Selecting the most relevant input features increases the prediction accuracy and minimizes the data and training time. In this research, various feature selection techniques are compared and analyzed. The techniques are then used as filters prior to electricity price forecasting and their influence on prediction accuracy and mean absolute percentage error (MAPE) of each selected subset is compared. The proposed SVM method and other forecasting methods are evaluated using data from the New England ISO, which is published on their official website. Optimization of SVM parameters and kernels has also been proposed in this thesis to further improve the prediction accuracy obtained by the presented SVM model. The results obtained in this research indicate that with the same input data, the optimized SVM model achieved the highest prediction accuracy. Furthermore, our research findings show that using the SVM Regression model and an optimization of its parameters can improve overall system prediction accuracy compared with other forecasting models investigated in this thesis.