Hybrid Day-ahead Load Forecasting with Atypical Residue based Gaussian Process Regression

The prediction accuracy of electric power consumption plays a crucial role for the efficiency of a smart grid. Hybrid approaches that jointly account for the linear and nonlinear portions of the electric load have shown promising performance because of the mixture of memory effects and random environmental perturbations. Especially for day-ahead short-term prediction, the potentially long time gap between the measurements and prediction point degrades the linear prediction performance, while the nonlinear prediction based on the weather forecast may supplement the degradation. This paper proposes a residue-based hybrid model that uses linear prediction by auto-regressive modeling and nonlinear prediction by Gaussian process regression with atypical residue of the weather forecast, particularly the difference of weather station forecasted and linear predicted local temperatures. Since the typical memory effect of the temperature can be double counted by both models, atypical residue without its linear prediction contribution is employed for the Gaussian process regression step. To verify the performance of the proposed scheme, a GIST campus electric power consumption dataset is evaluated. As expected, the linear prediction residue shows larger correlation to the atypical residue of temperature than the temperature itself. Consequently, hybrid model with the atypical residue temperature based Gaussian process regression shows improved performance in the day ahead load prediction.