Improving Market Clearing Price Prediction by Using a Committee Machine of Neural Networks

Abstract

Predicting market clearing prices is an important but difficult task, and neural networks have been widely used. A single neural network, however, may misrepresent part of the input-output data mapping that could have been correctly represented by different networks. The use of a "committee machine" composed of multiple networks can in principle alleviate such a difficulty. A major challenge for using a committee machine is to properly combine predictions from multiple networks, since the performance of individual networks is input dependent due to mapping misrepresentation. This paper presents a new method in which weighting coefficients for combining network predictions are the probabilities that individual networks capture the true input-output relationship at that prediction instant. Testing of the New England market cleaning prices demonstrates that the new method performs better than individual networks, and better than committee machines using current ensemble-averaging methods.