Improving electric load forecasts using network committees

Accurate daily peak load forecasts are important for secure and profitable operation of modern power utilities, with deregulation and competition demanding ever-increasing accuracies. Machine learning techniques including neural and abductive networks have been used for this purpose. Network committees have been proposed for improving regression and classification accuracy in many disciplines, but are yet to be widely applied to load forecasting. This paper presents a formal approach to apply the technique using historical load and temperature data spanning multiple years, with individual committee members trained on different years. Correlation among data for successive years is investigated and methods to enhance independence between member models for improving committee performance are described. Both neural and abductive networks implementations are presented and compared. An abductive network three-member committee was developed on data for three successive years and evaluated on the fourth year. Compared to a monolithic model trained on the same full three-year data, the committee reduces the mean absolute percentage error from 2.52% to 2.19%. The corresponding reduction in the mean of the absolute error from 70 MW to 61 MW is statistically significant at the 95% confidence level.