Neural network design for short-term load forecasting

This paper addresses an issue of the optimal design of a neural-network based short-term load forecaster. It describes the process of developing a multilayer, feedforward neural network for load forecasting, and then presents algorithms for performing two important steps of this process, i.e., input variable selection and network structure design. Input variable selection is carried out by forming a set of variables significantly correlated with the forecasted load and then by removing redundant, mutually correlated variables using singular value decomposition techniques. Selection of the optimal number of hidden neurons is based on the observation that oversized networks display near collinearity in the outputs of their hidden neurons. Hence, the presence of redundant hidden neurons can be detected by examining column dependency in the matrix of the hidden neuron outputs computed from the training data. The methodology presented in this paper can be used in the automatic design of an optimal forecaster based on historical data.