Improving predictive accuracy for photovoltaics
Improving predictive accuracy for photovoltaics lead image
As the cheapest, easiest and fastest technology to install, photovoltaic (PV) generation continues to expand around the world, bolstered by both government incentives and market demand.
A large part of incorporating PV sites into a power grid involves predicting their output in advance, often relying on forecasts of cloud cover along with historical trends. A newer approach incorporates spatial correlations with other nearby PV sites for increased accuracy. However, this combined spatiotemporal modeling is limited, typically relying on either static or dynamic graphs to construct spatial dependencies.
Yu et al. demonstrated the possibility of multiple spatial dependencies existing simultaneously, combining static and dynamic graphs to more accurately model spatial feature interactions across multiple sites.
“Due to the uncertainty in cloud movement directions, some sites may be more adversely affected than others,” said author Yunjun Yu.
The model they developed features a hybrid graph neural network combined with a gating mechanism and an adaptive spatiotemporal fusion module to prevent sites from being affected by unexpected spatial noise.
The experimental results comparing single-site and multi-site models showed improved model accuracy in PV predictions across datasets from two different climate regions. The researchers believe their system has significant advantages in predictive accuracy and generalization across climate types, and will improve energy management and the developing market around it.
“Accurate PV predictions for all sites are highly beneficial for regional energy trading and grid congestion management,” said Yu. “Our work will help bring more accuracy to this important field.”
Source: “Multi-site solar irradiance prediction based on hybrid spatiotemporal graph neural network,” by Yunjun Yu, Zejie Cheng, Biao Xiong, and Qian Li, Journal of Renewable and Sustainable Energy (2024). The article can be accessed at http://doi.org/10.1063/5.0207462 .
