Load Forecasting Accuracy through Combination of Trimmed Forecasts

Abstract

Neural network (NN) models have been receiving considerable attention and a wide range of publications regarding short-term load forecasting have been reported in the literature. Their popularity is mainly due to their excellent learning and approximation capabilities. However, NN models suffer from the problem of forecasting performance fluctuations in different runs, due to their development and training processes. Averaging of forecasts generated by NNs has been proposed as a solution to this problem. However, this may lead to another problem as odd forecasts may significantly shift the mean resulting in large forecasting inaccuracies. This paper investigates application of a trimming method by removing the α% largest and smallest forecasts and then averaging the rest of the forecasts. A validation set is applied for selecting the best trimming amount for NN load demand forecasts. Performance of the proposed method is examined using a real world data set. Demonstrated results show that although trimmed forecasts are not the best possible ones, they are better than forecasts generated by individual NN models in almost 70% of the cases.