Hybrid probabilistic photovoltaic power forecasting method for bidding on the day-ahead market

Forecasting the photovoltaic (PV) power generation plays a key role in facilitating the grid integration of the technology by reducing the amount and cost of imbalances caused by the fluctuating solar resource. Probabilistic forecasting and the integration of physical knowledge with machine learning are key concepts that are believed to characterize state-of-the-art solar forecasts. This paper presents a probabilistic PV forecasting method that relies on ensemble numerical weather prediction forecasts and combines them with an ensemble of physical model chains to improve the diversity of the ensemble members and account for the uncertainty of the irradiance-to-power conversion. The resulting ensemble PV power forecasts are post-processed using a non-crossing quantile regression neural network to improve their reliability and overall accuracy. The method is applied and verified for a Hungarian PV plant with 1 MWp installed capacity using four years of data, achieving a mean- normalized continuous ranking probability score of 16.7%. The economic value achieved by utilizing the forecasts for bidding on the day-ahead market is also evaluated for different deterministic forecasts elicited from the probabilistic forecasts. More accurate predictions tend to result in higher net income, and the cost of forecast errors compared to perfect foresight is only 6.6% of the revenues with the best forecast method.