Multi-kernel assimilation for prediction intervals in nodal short term load forecasting

Utilization of intelligent systems for information and decision making is of paramount significance toward implementing a smart and sustainable power grid. Nodal load forecasting is an aspect that can greatly benefit from the use of intelligent methods. In this paper, a multi-kernel method is proposed for load forecasting in power systems. In particular, the method adopts a set of kernel-modeled Gaussian process regressors that are subsequently compounded to provide a predictive distribution over the future values of a node's load. The compound predictive distribution is taken by the assimilation of the individual Gaussian processes using a genetic algorithm. In addition, the forecasting horizon varies at each step and is determined by the amount of uncertainty in the forecasted values. The proposed method is applied on a set of historical real-world load demand datasets taken from a node in US metropolitan area. Results exhibit that the assimilated models provide prediction intervals of less variance forecasts than the individual regressors. In addition, the proposed method provided forecast intervals in which a high number of actual forecasts fall within the limits of the interval.