Integrating Satellite Imagery and GNNs for Improving Day-Ahead Solar Irradiance Forecasting
No Thumbnail Available
Author
Schubnel, Baptiste
Simeunovic, Jelena
Tissier, Corentin
Alet, Pierre-Jean
Carrillo, Rafael
DOI
Abstract
The growing supply of renewable solar energy sources poses a challenge for balance group managers (BGM) and distribution system operators (DSO) due to their high variability. As such, accurate short to medium-term forecasting of local solar production is paramount. However, current solutions lack the high spatial and temporal resolution for the forecasting horizon required by BGMs and DSOs. State-of-the-art approaches for solar forecasting combine numerical weather predictions (NWP), satellite images and ground measurements with physical models. The main limitation of such approaches is that they can only achieve precise spatial and temporal resolution at the expense of high computational and storage load. Data-driven solutions that rely solely on data from a network of ground-based sensors have shown state-of-the-art results for intra-day irradiance forecasting while requiring lower computational resources. Yet, extending these solutions to day-ahead forecasts entails providing additional information on cloud dynamics and a broader spatial context. In this paper, we introduce SolarCrossformer, a novel deep learning model for day-ahead irradiance forecasting. SolarCrossformer builds on CSEM’s previous works by combining two sensing modalities: satellite images (visual and infrared channels) and measurements (irradiance, temperature, etc.) from a ground-based network of meteorological stations. Satellite images provide the wider spatial context of the cloud dynamics, while the ground measurements provide information on the local variations. The proposed model uses novel graph neural networks to leverage the inter- and intra-modal correlations of the input data and improve the accuracy of the forecasts. SolarCrossformer uses data of the past 24 hours from the two sensing modalities, without requiring NWP as inputs, to generate probabilistic forecasts of irradiance at any location in Switzerland for the next 24 hours, with a temporal resolution of 15 minutes. SolarCrossformer can incorporate time-series data from a new location to generate forecasts without needing to retrain the entire model, i.e., no need of historical data from the new location. Additionally, it can produce forecasts for a location without any past data for that location by simply using the coordinates of the desired location. The paper will present the model and its benchmark against state-of-the-art methods. Preliminary results, evaluating the model over one year, show a normalized root mean squared error (NRMSE) of 10.89% over the entire forecasting horizon.
Publication Reference
EU PVSEC 2025, Bilbao, Spain
Year
2025