Sparse Variational Gaussian Process with Dynamic Kernel for Electricity Demand Forecasting

A maximized accuracy and minimized prediction interval (PI) are desirable features of a forecast model. Especially, energy companies utilize these attributes to efficiently manage power systems, maintain grid stability and maximize their profit gain. In this paper, we focus on the short-term demand forecast, based on the Sparse Variational Gaussian Process (SVGP). We, first, propose a dynamic kernel selection algorithm that simplify the search for a valid covariance matrix. Then, we investigate the dependence of accuracy and PI based on the size of training data and inducing variables used for approximation. The overall model performance is evaluated using prediction interval coverage probability (PICP) and mean prediction interval width (MPIW). In addition to that, we conducted a model evaluation studies on SVGP and long short-term memory (LSTM) model based on these metrics. The results show the proposed algorithm provides a valid kernel to data that exhibit a known and unknown patterns. The simulations reveal SVGP achieves higher prediction accuracy and minimized PI than LSTM. However, the SVGP model suffers from the curse of generalization. The variational inference applied When building SVGP ultimately ignores rare moments that happen periodically. Their forecast is approximated or lumped as similar With the most frequent moments. Finally, utilizing SVGP and aggregated LSTM models, we demonstrated the degree of compromise one could make with respect to the selection of a predictive model, size of training data and inducing variables to achieve the same level of forecast accuracy and minimized PI.