An evaluation of Bayesian techniques for controlling model complexity and selecting inputs in a neural network for...

Area of Science:

• Artificial Intelligence
• Electrical Engineering
• Statistical Modeling

Background:

• Artificial neural networks (ANNs) are suitable for nonlinear modeling of large datasets in electricity load forecasting.
• Challenges in ANN modeling include defining model complexity and selecting input variables.

Purpose of the Study:

• To evaluate Bayesian techniques for automatic neural network modeling in electricity load forecasting.
• To compare Bayesian input selection and structure optimization with traditional cross-validation methods.

Main Methods:

• Applied Bayesian framework for automatic neural network modeling.
• Utilized Bayesian 'automatic relevance determination' for input selection.
• Employed Bayesian 'evidence' for selecting optimal network structure (number of neurons).
• Tested on six datasets of daily load and weather data from four countries.

Main Results:

• Bayesian methods demonstrated effectiveness in automatic input selection for load forecasting models.
• Bayesian evidence proved useful for selecting the optimal number of neurons, comparable to cross-validation.
• The approach was validated across diverse international datasets.

Conclusions:

• Bayesian frameworks offer robust solutions for automated complexity and input variable selection in neural network-based electricity load forecasting.
• These methods provide a viable alternative to traditional cross-validation for optimizing neural network models in this domain.