A Comparative Study on Feature Selection Based Improvement of Medium-Term Demand Forecast Accuracy

Abstract

Use of the proper demand forecasting method and data set is very important for reliable system operation and planning. In this study, we compare performances of various feature selection methodforecasting algorithm pairs in terms of forecast accuracy for medium-term demand forecasting case. We utilize correlation, recursive feature elimination, random forest, multivariate adaptive regression splines (MARS), stepwise regression and genetic algorithms as feature selection methods. As for forecasting algorithms, we use linear and non-linear methods such as multiple linear regression (MLR), generalized additive model (GAM), MARS, k-nearest neighbors (KNN), classification and regression trees (CART), neural networks (NN) and support vector machines (SVM) methods. In the end, the MARS and stepwise regression as feature selection methods, and MARS model as the forecasting algorithm are found to be showing superior performance. However, with a suitable feature selection technique such as stepwise regression, linear models can yield satisfactory level of performance considering their low computation requirement.