An evaluation of Bayesian techniques for controlling model complexity in a neural network for short-term load forecasting

Artificial neural networks have frequently been proposed for electricity load forecasting because of their capabilities for the nonlinear modelling of large multivariate data sets. However, there are still no widely accepted strategies for designing the models and for implementing them, which makes the process of modelling by neural networks largely heuristic, dependent on the experience of the user. This paper evaluates techniques for automatic neural network modelling within a Bayesian framework, as applied to a dataset containing daily load and weather data for England and Wales. We analyse input selection as carried out by the Bayesian "automatic relevance determination", and also evaluate the usefulness of the Bayesian "evidence" for the selection of the best structure (in terms of number of neurons), as compared to methods based on cross-validation.