A comparison of neural network backpropagation algorithms for electricity load forecasting

Load forecasting plays a significant role in planning and operation of electrical power networks. Artificial neural networks have been extensively employed for load forecasting over the last 20 years, owing to their powerful non-linear mapping capability. A range of neural network training algorithms have been developed to solve different kinds of problems. Due to different goals of prediction and variation in size of datasets for load forecasting, the choice of algorithm to train the neural network can greatly influence the forecasting result. In this paper we consider different backpropagation training algorithms for medium term load forecasting and analyze each of the characteristics such as parameter setting complexity, training speed, convergence, prediction accuracy and result stability. From our case study, we conclude Bayesian Regulation Backpropagation to be the best overall choice for medium term load prediction. For cases where processing capability is limited, Resilient Backpropagation and Conjugate Gradient Backpropagation may be suitable alternative choices.