Physics Informed Graph Neural Networks for Multi-Site Solar Forecasting

Accurate forecasting of photovoltaic (PV) power generation is crucial for efficient electricity management and market trading. Traditional data-driven models, while providing state-of-the-art accuracy for short-term forecasts, often suffer from limited generalization when facing data that deviates from their training distribution. Additionally, these models typically produce smooth forecasts that fail to capture the intricate dynamics of cloud movements, crucial for predicting solar irradiance, a primary driver of PV output. To address these challenges, we introduced a physics-informed graph neural network (PING) model that estimates the particle velocities of the historical input data, in an unsupervised fashion, and forecasts the future particle concentration values of advection-diffusion processes. In this paper we propose the combination of PING with our previously developed model, the Graph Convolutional Long-short term memory (GCLSTM) network, for multi-site PV power forecasting tasks. Numerical results showed that PING + GCLSTM outperforms all benchmarks on the entire horizon showing a daytime normalized root-mean-square error, overall sites, between 7% and 13% for 15 minutes and 6 hours ahead prediction, respectively.