Short-term electrical load forecasting using predictive machine learning models

Availability of cheap power through alternative means such as energy exchanges and bilateral agreements is resulting in short-term load forecasting gaining importance among industries, residential complexes and corporate buildings. Short-term forecasting over an hour or a day requires non-linear predictive models. Machine learning algorithms such as neural networks are inherently non-linear and are suitable for accurate forecasting. This paper compares neural networks, decision trees and Conditional Restricted Boltzmann Machines algorithms for forecasting short-term demand. The algorithms are tested on power consumption data acquired from two test sites with different consumption profiles.