Electricity Demand Forecasting with Gaussian Process Regression
Abstract
For participants in the energy industry, it is vital to have access to reliable forecasts of future energy demands. The predictive routines should necessarily cope with local, transient fluctuations as well as considerable changes in the electricity consumption landscape. In this thesis, a solution for electricity demand forecasting based on Gaussian process regression is presented. To account for the choices taken in the process of designing functional predictive models, a thorough background theory for long-term time-series forecasting with Gaussian processes is established. The models are tested on a real-world data set, and while complying with the limitations that apply for participants in the Norwegian energy market, the results were found superior to a commercially employed model designed for the same task. The models have, in addition, many useful properties such as no user-defined parameters, a quantification of predictive uncertainties, and low time complexity.