New approach to daily and peak load predictions using a random vector functional-link network

Abstract

A functional-link neural network based short-term electric load forecasting system is presented in this paper. The electric load forecasting model is assumed to consist of a load time series and a weather dependent component modelled as a functional expansion of the weather variables like temperature or humidity. The parameters of the functional link neural network model are identified using a weight adjustment algorithm based on Widrow-Hoff delta rule. The non linear weight adjustment algorithm adapts the weights every 24-hour or 168-hour producing a MAPE mostly less than 2% for a 24-hour ahead forecast and 2.5% for a 168-hour ahead forecast. The results of forecast for a period over 2 years indicate that the new model produces more accurate and robust forecasts in comparison to simple adaptive, neural network or statistical approaches.