Monthly load forecasting using kernel based gaussian process regression

Abstract

Forecasting of electricity load for a month is crucial for power system planning and safe operation. Monthly demand is subject to various factors such as season and climate effects, thus making accurate load demand forecasts a challenging task. In this paper a data driven machine learning approach is applied on monthly load morphing. More particularly, kernel based Gaussian Process Regression (GPR) is adopted for forecasting the amount of load demand for each month in a year-ahead-horizon. The GPR model is equipped with a valid kernel function (four different kernels are tested) and is trained on a set of historical datasets that contain the recorded monthly demands of the prior four to the tested years. It is extensively benchmarked on a period of five years and its performance is measured using the Mean Average Percentage Error (MAPE). Furthermore, it is compared to the performance taken with simple linear regression model. Results demonstrate the superiority of the GPR forecasting over that of the linear regression, however, a dependence of GPR on selection of kernel is observed.