Improving the prediction of radial basis function networks for power systems

Radial basis function (RBF) networks approximate an input-output relationship by building localized radial basis functions (clusters), and have been used in various forecasting problems. To better learn local data characteristics, the general form of Gaussian-like clusters is used to have covariance matrices differentially treating input factors in basis functions' exponents. Such step results in a substantially large number of tunable parameters. A network could easily over-fit the data and comprise its prediction quality. A new procedure to overcome the above dilemma is presented in this paper. The key idea is to reduce the number of tunable parameters in each cluster via eliminating insignificant input factors whose standard deviations are too large or too small. Through this procedure, a new network can select significant input factors for clusters, has parsimonious clusters, is less likely to over-fit the data, and leads to improved predictions. The effectiveness of procedure is illustrated by a simple example and by market cleaning price prediction.