Improving efficiency of artificial neural networks in electricity demand forecasting

With the deregulation of the electricity industries in many countries and the increase in complexity of power systems, electricity demand forecasting has become even more important and challenging. The tandard Backpropagation algorithm used in conventional Multilayer Feed-Forward Neural Networks (MFNN) has been found to be time-consuming and too slow for practical problems like short-term electricity demand forecasting. This paper examined performance of three other numerical optimisation techniques, namely, Scaled Conjugate Gradient, Quas-Newton and Levenberg-Marquardt, in short-term electricity demand forecasting. The effectiveness of Bayesian Regularisation on improving networks generalisability is also examined. Load demand patterns of 12 consecutive months with 30-minute resolution are used for evaluation and analysis. Corresponding data of temperature and other important parameters are also considered. The data is preprocessed to enhance the performance of the algorithms.