Advancement in the application of neural networks for short-term load forecasting

Abstract

An improved neural network approach to produce short-term electric load forecasts is proposed. A strategy suitable for selecting the training cases for the neural network is presented. This strategy has the advantage of circumventing the problem of holidays and drastic changes in weather patterns, which make the most recent observations unlikely candidates for training the network. In addition, an improved neural network algorithm is proposed. This algorithm includes a combination of linear and nonlinear terms which map past load and temperature inputs to the load forecast output. The search strategy and algorithm demonstrate improved accuracy over other methods when tested using two years of utility data. In addition to reporting the summary statistics of average and standard deviation of absolute percentage error, an alternate method using a cumulative distribution plot for presenting load forecasting results is demonstrated.>