Improving market clearing price prediction by using neural networks

Date of Completion

January 2004


Engineering, Electronics and Electrical|Computer Science




In a deregulating power market, bidding decisions rely on good market clearing price predictions. One of the common forecasting methods is a Gaussian radial basis function (GRBF) network, which approximates input-output relationships by building localized Gaussian functions (clusters). The accuracy of network predictions relies on the degree of correctness of the input-output mapping built by the network. Currently, all input factors are used by each of the clusters. To some clusters, certain input factors may not be important and should be deleted because they mislead local learning and result in the misrepresentation of the underlying relationship in data. Existing pruning methods for neural networks, which examine the significance of connections between neurons, are not applicable to deleting center and standard deviation parameters since these parameters bear little sense of connection to the pruning methods. Based on the finding that the inverses of standard deviations can sever links between neurons, a new training method to identify and eliminate unimportant input factors is developed. ^ For a neural network, the misrepresentation of the underlying relationship in data cannot be completely solved by selecting input factors. In view of the commonly seen situation that there are insufficient data points available to represent all data features, a network in reality may misrepresent part of the input-output relationship that could have been more appropriately represented by different networks. For example, radial basis function networks are effective in exploiting local data characteristics, while multi-layer perceptron networks are good at capturing global data trends. The use of a “committee machine” composed of multiple networks can in principle alleviate the problem of misrepresenting the underlying relationship. Currently, combining the predictions from multiple networks is based on a straight average or the statistics of historical prediction errors though the performance of individual networks is varying and input-dependent. To solve this problem, our idea is to estimate the quality of a prediction, i.e., the prediction variance that is conditioned on the current input and the historical data. Under the Multiple Model framework, a new method that the prediction qualities of networks are utilized to determine better weighting coefficients is developed. ^