Predicting Photovoltaic Power Production using High-Uncertainty Weather Forecasts

A growing interest in renewable power increases its impact on the energy grid, posing significant challenges to reliability, stability, and planning. Although the use of weather-based prediction methods helps relieve these issues, their real-world accuracy is limited by the errors inherent to the weather forecast data used during the inference. To help resolve this limitation, we introduce the SolarPredictor model. It uses a hybrid convolutional architecture combining residual connections with multi-scale spatiotemporal analysis, predicting solar power from publicly available high-uncertainty weather forecasts. Further, to train the model, we present the SolarDB dataset comprising one year of power production data for 16 solar power plants. Crucially, we include weather forecasts with seven days of hourly history, allowing our model to anticipate errors in the meteorological features. In contrast to previous work, we evaluate the prediction accuracy using widely available low-precision weather forecasts, accurately reflecting the real-world performance. Comparing against 17 other techniques, we show the superior performance of our approach, reaching an average RRMSE of 6.15 for 1-day, 8.54 for 3-day, and 8.89 for 7-day predictions on the SolarDB dataset. Finally, we analyze the effects of weather forecast uncertainty on the prediction accuracy, showing a 23 % performance gap compared to using zero-error weather. Data and additional resources are available at cphoto.fit.vutbr.cz/solar.