Improved Neural Network Prediction Performances of Electricity Demand : Modifying Inputs Through Clustering

摘要

Accurate prediction of electricity demand can bring extensive benefits to any country as theforecast values help the relevant authorities to take decisions regarding electricity generation,transmission and distribution much appropriately. The literature reveals that, when comparedto conventional time series techniques, the improved artificial intelligent approaches providebetter prediction accuracies. However, the accuracy of predictions using intelligent approacheslike neural networks are strongly influenced by the correct selection of inputs and the number ofneuro-forecasters used for prediction. This research shows how a cluster analysis performed togroup similar day types, could contribute towards selecting a better set of neuro-forecasters inneural networks. Daily total electricity demands for five years were considered for the analysisand each date was assigned to one of the thirteen day-types, in a Sri Lankan context. As astochastic trend could be seen over the years, prior to performing the k-means clustering, thetrend was removed by taking the first difference of the series. Three different clusters werefound using Silhouette plots, and thus three neuro-forecasters were used for predictions. Thispaper illustrates the proposed modified neural network procedure using electricity demanddata.