Univariate and multivariate methods for very short-term solar photovoltaic power forecasting

Abstract

We consider the task of forecasting the electricity power generated by a solar PhotoVoltaic (PV) system for forecasting horizons from 5 to 60min ahead, from previous PV power and meteorological data. We present a new method based on advanced machine learning algorithms for variable selection and prediction. The correlation based variable selection identifies a small set of informative variables that are used as inputs for an ensemble of neural networks and support vector regression algorithms to generate the predictions. We develop two types of models: univariate, that use only previous PV power data, and multivariate, that also use previous weather data, and evaluate their performance on Australian PV data for two years. The results show that the univariate models performed similarly to the multivariate models, achieving mean relative error of 4.15–9.34%. Hence, the PV power output for very short-term forecasting horizons of 5–60min can be predicted accurately by using only previous PV power data, without weather information. The most accurate model was univariate ensemble of neural networks, predicting the PV power output separately for each step of the forecasting horizon.