Reducing forecast errors of a day-ahead photovoltaic power forecast by ensemble learning

This paper introduces a stacking ensemble as a correction model that reduces forecast errors of a day-ahead photovoltaic (PV) power forecast by incorporating minimal power measurement data. The correction model reduces the forecast error for a 2.24 kWp PV system to 7.67% compared to 8.76% for a PV power forecast without correction for a test period of 231 days. The effect of the correction model becomes especially clear on winter days when the PV system is partially shadowed. The model identifies the recurring error caused by shadowing and reduces it. Only about six months of chronologically collected PV power measurement data for training the correction model is sufficient to optimize the forecast for different weather conditions. Therefore, the model is deemed practicable in real life applications.