Market clearing price prediction using a committee machine with adaptive weighting coefficients

The use of a committee machine composed of multiple neural networks can capture more features in data and overcome the inadequacy of a single network. Different neural networks, however, may capture different features in data, and result in different inferences and predictions. Furthermore, network performance may not be constant as input features change. Without information on prediction quality of individual networks, it is difficult to appropriately combine them in a committee machine. To overcome the difficulty, insightful information called the prediction confidence is introduced and an associated method of determining adaptive weighting coefficients is developed in this paper. The key idea is that the confidence of a prediction is measured by the inverse of a prediction variance. Adaptive weighting coefficients are then derived to depend on past network performance and the confidence of current predictions. Therefore, a committee machine can properly determine coefficients according to past performance and prediction confidence of networks. The effectiveness of a new combination method is illustrated by power market clearing price prediction.