[IMC 2024] Graph machine learning for sort-term PV forecasting

Dr. Rafael Carrillo, CSEM
State-of-the-art approaches for photovoltaic (PV) power forecasting combine numerical weather predictions, satellite images and ground measurements with physical or machine learning models. A current limitation of these approaches is that precise high spatial and temporal resolution require a high computational and storage load. To overcome these limitations, CSEM has developed an approach to PV forecasting based on graph neural networks (GNN) that uses past measurements from a ground-based sensing network distributed in the region of interest. GNNs exploit the spatio-temporal relations of the data to improve the forecast accuracy and resolution. The talk will review recent advances on graph machine learning for PV production forecasting and compare them to state-of-the-art methods in PV forecasting for the six hours ahead forecasting horizon. The Digerati solution for short-term forecasting will be presented. Such solution is based on dynamic GNN to address the problem of varying network of sensors due to real-life conditions. Finally, the talk will highlight how the GNN approach can be used for anomaly detection in rooftop PV installations.