Importance based selection method for day-ahead photovoltaic power forecast using random forests

Abstract

With the great recent moves towards green energy exploitation worldwide, the solar photovoltaic (PV) power has gained much attention. Thanks to PV panels' cost drop and recent improvements in energy conversion systems, the PV installations are getting more and more integrated into power plants. Because of high correlation with weather conditions, accurate short-term PV output forecast is highly recommended. An accurate prediction is needed to assess the effective contribution of solar energy in the grid, and to overcome the problems of intermittence. This paper proposes a day-ahead prediction method of PV output, which estimates the power generated by solar panels with and without prior knowledge of solar irradiance. The proposed model is the random forest using bagging algorithm, characterized by built-in cross validation and immunity to irrelevant inputs. A special attention is paid to the choice of most influential weather conditions on future power. The proposed approach is validated through tests on real data from PV sites in Australia.