Abstract
Long-term forecasting of renewable energy (RE) sources is important for proper energy planning and operations. Machine learning models are now being adopted by more and more researchers for this purpose. In this paper, we propose an improved Bayesian machine learning method for energy forecasting with uncertainty, namely Gaussian process regression with tree-based multiple kernel search (GPR-TMKS). The main objective is to automate the search for the best kernel combination in the well-known Gaussian process model using a tree structure of base kernels such as linear, periodic, squared exponential, and rational quadratic functions. Using both solar and wind energy time series data collected from 2000 to 2022 at a particular site of interest in the Philippines, namely the province of Isabela, we demonstrate the superiority of GPR-TMKS against autoregressive models and the long short-term memory (LSTM) network, in terms of accuracy. In particular, the proposed model achieved root-mean-squared error (RMSE) test values of 0.43 kWh/m2/day and 0.50 m/s for smoothened solar and wind forecasts, respectively. With more accurate predictions and low uncertainties, the intermittency behavior of RE sources can be better understood, leading to more effective integration of RE with existing power grids.
