Predicting user comfort level using machine learning for Smart Grid environments

Smart Grid with Time-of-Use (TOU) pricing brings new ways of cutting costs for energy consumers and conserving energy. It is done by utilities suggesting the user ways to use devices to lower their energy bills keeping in mind its own benefits in smoothening the peak demand curve. However, as suggested in previous related research, user's comfort need must be addressed in order to make the system work efficiently. In this work, we validate the hypothesis that user preferences and habits can be learned and user comfort level for new patterns of device usage can be predicted. We investigate how machine learning algorithms specifically supervised machine learning algorithms can be used to achieve this. We also compare the prediction accuracies of three commonly used supervised learning algorithms, as well as the effect that the number of training samples has on the prediction accuracy. Further more, we analyse how sensitive prediction accuracies yielded by each algorithm are to the number of training samples.