Dynamic Graph Machine Learning for Mult-Site Solar Forecasting

dc.contributor.authorCarrillo, Rafael E
dc.contributor.authorSchubnel, Baptiste
dc.contributor.authorLangou, Renaud
dc.contributor.authorAlet, Pierre-Jean
dc.date.accessioned2024-02-19T11:10:15Z
dc.date.available2024-02-19T11:10:15Z
dc.date.issued2023-09
dc.description.abstractIn this paper, we present DIGERATI, CSEM’s solution for intraday forecasting (six-hour horizon). DIGERATI is based on graph neural networks (GNN). It extends previous algorithms developed by CSEM and enables their use in real operating conditions. These GNN algorithms outperformed the state of the art in forecasting photovoltaic production, while relying only on real, imperfect power measurements. In previous works the same nodes were present throughout training and evaluation. However, in a commercial application, nodes (e.g., photovoltaic systems) would be frequently added or removed as new customers sign up or physical assets change. To meet this constraint, DIGERATI exploits dynamic GNNs, where nodes and edges can be added or removed over time, thus providing a robust and scalable solution for real-life operations. Additionally, DIGERATI uses advanced GNNs to fuse information from heterogeneous data sources (PV power, irradiance, wind speed and temperature) measured at different locations to produce probabilistic forecasts of irradiance (or PV power) in desired locations. The live demonstrator can produce probabilistic forecasts at any location in Switzerland and the Netherlands every 15 minutes with a temporal resolution of 15 minutes.
dc.identifier.citationEU PVSEC 2023
dc.identifier.doi10.4229/EUPVSEC2023/4CO.8.5
dc.identifier.isbn3-936338-88-4
dc.identifier.urihttps://hdl.handle.net/20.500.12839/1350
dc.language.isoen
dc.titleDynamic Graph Machine Learning for Mult-Site Solar Forecasting
dc.typeProceedings Article
dc.type.csemdivisionsBU-V
dc.type.csemresearchareasData & AI
dc.type.csemresearchareasDigital Energy
Files
License bundle
Now showing 1 - 1 of 1
No Thumbnail Available
Name:
license.txt
Size:
2.82 KB
Format:
Item-specific license agreed upon to submission
Description: