Electricity load forecasting: A weekday-based approach

Abstract

We present a new approach for building weekday-based prediction models for electricity load forecasting. The key idea is to conduct a local feature selection using autocorrelation analysis for each day of the week and build a separate prediction model using linear regression and backpropagation neural networks. We used two years of 5-minute electricity load data for the state of New South Wales in Australia to evaluate performance. Our results showed that the weekday-based local prediction model, when used with linear regression, obtained a small and statistically significant increase in accuracy in comparison with the global (one for all days) prediction model. Both models, local and global, when used with linear regression were accurate and fast to train and are suitable for practical applications. © 2012 Springer-Verlag.