A compositional kernel based gaussian process approach to day-ahead residential load forecasting

Abstract

Load forecasting is an expected ability of electric power networks to enable effective capacity planning. This paper proposes a probabilistic approach to short-term load forecasting (STLF) of residential power consumption. The proposed method is based on Bayesian regression modeling. It utilizes an additive Gaussian Process (GP) to estimate climate-sensitive and calendar factors of power demand. The GP model is constructed by using a set of compositional kernels that represent the most significant interactions between input variables. Such collection is built up through a sampling method, capable of selecting the n-upmost order-based interactions. Moreover, a technique is performed to deal with challenges related to multivariate input and large dataset training complexity. The forecasting model is applied to actual power consumption data of a set of houses, located in Quebec, during winter. The results demonstrate that the suggested scheme is highly efficient to model and predict residential electricity use. Furthermore, it is competitive with other forecasting algorithms, as manifested by a comparative analysis.