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

Journal article

Artificial neural networks have frequently been proposed for electricity load forecasting because of their capabilities for the nonlinear modelling of large multivariate data sets. Modelling with neural networks is not an easy task though; two of the main challenges are defining the appropriate level of model complexity, and choosing the input variables. This paper evaluates techniques for automatic neural network modelling within a Bayesian framework, as applied to six samples containing daily load and weather data for four different countries. We analyse input selection as carried out by the Bayesian 'automatic relevance determination', and the usefulness of the Bayesian 'evidence' for the selection of the best structure (in terms of number of neurones), as compared to methods based on cross-validation. © 2009 Elsevier Ltd. All rights reserved. Artificial Neural Network; Automatic relevance determination; Bayesian; Bayesian frameworks; Bayesian model selection; Bayesian neural networks; Bayesian techniques; Cross validation; Electricity load; Input selection; Input variables; Load forecasting; Model complexity; Multivariate data sets; Neural network modelling; Non-linear modelling; Selection of the best; Short term load forecasting; Weather data; Electric load forecasting; Neural networks; Bayesian networks; algorithm; article; artificial neural network; automation; Bayesian learning; controlled study; electricity; information processing; mathematical computing; mathematical model; nonlinear system; prediction and forecasting; priority journal; problem solving; weather; Algorithms; Automation; Bayes Theorem; Forecasting; Neural Networks (Computer); Time Factors; Weather 8936080 10.1016/j.neunet.2009.11.016 English

Authors: Hippert, H; Taylor, J

• Publisher: Elsevier
• Journal: Neural Networks
• Volume: 23
• Issue: 3
• Pages: 386 - 395
• Publication date: 2010-01-01
• DOI: 10.1016/j.neunet.2009.11.016
• ISSN: 0893-6080
• Language: English