Variable Selection in the Kernel Regression Based Short-Term Load Forecasting Model

The short-term load forecasting is an essential problem in energy system planning and operation. The accuracy of the forecasting models depends on the quality of the input information. The input variable selection allows to chose the most informative inputs which ensure the best forecasts. To improve the short-term load forecasting model based on the kernel regression four variable selection wrapper methods are applied. Two of them are deterministic: sequential forward and backward selection and the other two are stochastic: genetic algorithm and tournament searching. The proposed variable selection procedures are local: the separate subset of relevant variables is determined for each test pattern. Simulations indicate the better results for the stochastic methods in relation to the deterministic ones, because of their global search property. The number of input variables was reduced by more than half depending on the feature selection method.