Improved neural network approach for weather sensitive short-term load forecasting

Abstract

This paper presents feedforward neural networks with an improved Kalman filter based learning algorithm for forecasting weather sensitive loads. Several parameters like the sigmoid slope, Kalman gain and learning rate are adaptively tuned to provide fast convergence during learning and subsequent predictions. With these modifications, the accuracy of the forecasted load increases considerably in comparison to the earlier neural network based techniques. Extensive studies have been performed for hourly load predictions and dally average load predictions on a typical one - year utility data and forecast results for autumn, winter, spring and summer days are given to validate the effectiveness and efficacy of the above approach.