Benchmark analysis of day-ahead solar power forecasting techniques using weather predictions

The increasing penetration of distributed renewable energy sources like Photovoltaics (PV) may form a threat to reliable grid operation. PV-systems impede load balancing due to the intermittent and uncontrollable power production. The development of highly accurate forecasting techniques is essential to support a high PV penetration rate in the local electricity grid. This research examines the performance of different machine learning (ML) models that autonomously predict day-ahead power production of individual and aggregated PV-systems. The forecasting models are developed by considering historic power production and regional predictions of weather metrics. The method allows to generate site specific forecasting algorithms that inherently account for site characteristics including size, orientation and shading and is independent of such input. In the research we evaluate the accuracy of the forecasting models for 152 PV-systems individually and for different levels of aggregated systems. With a skill score of respectively 41.4% and 41.3% Gradient Boosting and Random Forests are found to outperform the other models. This is closely followed by the Feed-forward Neural Network and Kernel Support Vector Machine models. Moreover, the results show the value of aggregating PV sites in day-ahead power forecasting as the mean absolute error and the root mean square error of each ML model improve by at least 18% and 20%.