A precondition technique with reconstruction of data similarity based classification for short-term load forecasting

This paper proposes an efficient ANN (artificial neural network) based method for short-term load forecasting in power systems. The proposed method classifies input data into clusters and reconstructs them to acquire sufficient learning data so that more exact load forecasting is realized at reconstructed each cluster. The classification of input data makes the construction of the load forecasting model easier due to the data similarity. Also, the reconstruction of clusters is useful for handling a cluster with insufficient data or data dose to the classification boundary. In this paper, the self-organization map (SOM) of Kohonen is used as a classifier since it has a feature that output neurons are assigned to preserve the similar features in the topological map. Based on the output neuron, the clusters are reconstructed to acquire appropriate learning data. As a forecasting model at each cluster, the radial basis function network (RBFN) is employed due to the excellent function of nonlinear approximation. The proposed method is successfully applied to real data of utilities in Japan. A comparison between the proposed and conventional methods is made in terms of average and maximum errors.