Electricity price forecasting with a new feature selection algorithm

In recent years, energy price forecasting has become very important for the participants in a competitive electricity market. However, price signals usually have complex behavior due to their non-linearity, non-stationarity and time variability. Therefore, an essential requirement is an accurate and robust price forecasting method. The hybrid method proposed in this paper is composed of a combination of wavelet transforms and neural networks. Both time-domain and wavelet-domain features are considered in a mixed data model for price forecasting, in which the candidate input variables are refined by a feature selection algorithm. The "Relief" algorithm is used to remove redundancy and irrelevant input variables. The adjustable parameters of the method are fine tuned by a cross-validation technique. The proposed method is examined on the Pennsylvania- New Jersey-Maryland electricity market and compared with some of the most recent price forecasting methods.