Power load forecasting using neural canonical correlates

We (1998, 1999) have previously derived a neural network implementation of the statistical technique of canonical correlation analysis. We have then extended the network so that it may find nonlinear correlations in data sets. In this paper we demonstrate the capabilities of the network (both linear and nonlinear) on an artificial data set and demonstrate that the nonlinear network finds greater correlations than any lineal network. We then use both networks for forecasting the next day's power loading given the previous days' loads and forecasts of the temperature. We show that the nonlinear correlation method performs better than both a standard supervised learning neural network using backpropagation and a recent modification of that algorithm.