Electric Load Forecasting in Energy Power Systems Based on Online Gaussian Process Regression Coupled with Multilayer Perceptron Kernel Method

Abstract

Electricity generation must follow the consumption profile to ensure the optimal management and stability of power systems. In order to maximize profits and efficiently utilize their power system infrastructure while reducing the risk of losses, power electric companies need to make their systems work in a configuration that matches supply with customer demand. This requires the development of a predictive robust model that can accurately forecast future electricity consumption. In this study, an approach based on Gaussian process learning with updating of observed values is proposed to perform short-term predictions of electric load. Comparative studies have been performed with standard models based on linear (Lin), square exponential (SE), radial basis function (RBF), rational quadratic (RQ), multilayer perceptron (MLP), periodic (Per) and Materne (Mv) kernels. The observations are injected into the dataset to train the model before each prediction. Each model was tested on the CEB load data. The results show that the proposed model based on the MLP kernel outperforms standard Gaussian process models based on statistical criteria.