Feature Selection for Electricity Market Data Through Feature Extraction

This paper investigates how well different combinations of data pre-processing and feature selections perform for filtering and selecting time series data for day-ahead electricity price forecasting with neural networks. We compared these against the selection of a brute force trying all combinations. Publicly available data on the electricity generation capacity of individual electricity producer groups as well as the grid load are utilized as exogenous time series to forecast the electricity price. The pre-processing methods used include signal processing methods and processing according to the Box-Jenkins model. In the feature selection methods, all three typical areas consisting of filter, wrapper and embedded methods - are covered and examined. Additional noise time series are included in the feature selection to evaluate their robustness. The final analysis shows the weaknesses of individual methods and that a general solution is not possible due to the diverse structure of the different approaches.