A Learnable Kernel Machine for Short-Term Load Forecasting

Short-term load forecasting is very important for decision making in power system operation and planning. During the last several years, kernel machines have been widely employed for short-term forecasting. Owing to the inherent limitations the corresponding forecasting accuracy can be impaired. To overcome the limitations, this paper develops a novel kernel machine, hereafter called learnable kernel machine, for short-term load forecasting. The proposed method possesses several appealing properties. First, like all other kernel machines, it handles nonlinearity in a disciplined manner that is also computationally attractive; second, by incorporating both kernel learning and regularization parameter learning, it effectively enhances the overall performance; third, as the optimality criterion, it employs the leave-one-out error, leading to an almost unbiased estimator of the generalization error; forth, using the leave-one-out error as the optimality criterion, it can be also expressed in closed form, making it computationally feasible in practice; fifth, the computational cost to optimize the leave-one-out error can be further reduced by matrix technology. We present experimental results on real-world data sets to demonstrate the effectiveness.