An evaluation of predictor variables for photovoltaic power forecasting

Abstract

Accurate forecasts of the electric power generation by solar Photovoltaic (PV) systems are essential to support their vast increasing integration. This study evaluates the interdependence of 14 predictor variables and their importance to machine learning (ML) models that forecast the day-ahead PV power production. To this purpose, we use two feature selection models to rank the predictor variables and accordingly, examine the performance change of two ML forecast models when a growing number of variables is considered. The study is performed using 3 years of data for Utrecht, the Netherlands. The results show the most important variables for PV power forecasting and identifies how many top variables should be considered to achieve an optimal forecast performance accuracy. Additionally, the best forecast model performance is found when only a few predictor variables are considered, including a created variable that estimates the PV power output based on technical system characteristics and physical relations.