Cost of photovoltaic power forecast errors based on machine learning, numerical weather prediction and real-time measurements

The purpose of the study is to investigate the importance of real-time data measurement in photovoltaic power forecasting and market bidding. Five scenarios are tested on 12 industrial scale power plants on two years and evaluated by accuracy and financial revenues in the Hungarian electricity market, which is part of the European connected markets.The forecasts are based on three different sources of numerical weather predictions (NWP), which are converted to photovoltaic power with Light Gradient Boosting machine learning model optimized using Optuna hyperparameter tuning framework. In addition to day-ahead forecast, 1.5h ahead intraday forecasts are created in four different ways. First, only the NWP variables are refreshed, then they are supplemented by either the measured power, or CAMS satellite-retrieved irradiance, or both. Day-ahead market bids are placed on day-ahead power forecasts, which can be updated by trading on the intraday market based on improved predictions. Finally, the residual forecast error is covered with balancing energy.The results show that the average cost of forecast error with respect to a perfect day-ahead forecast is 12.52 €/MWh without intraday update. Intraday trading decreases the costs by 2.15 €/MWh, which can be further reduced by 1.47 €/MWh to 8.9 €/MWh with measured power and irradiance feedback data. Ranking the two feedbacks, accuracy and costs show ambiguity, since power provides better accuracy, while irradiance increases the revenues more.