Hybrid prediction model improves photovoltaic power output predictions in variable weather

Solar energy is a premier renewable energy source, but is critiqued for its variable output, which can be affected by weather conditions. Short-term fluctuations in power output from solar photovoltaic systems make it difficult to operate safe and reliable power grids. To better equip power systems to adjust for these power variations, more accurate power generation predictions are needed.

Gaoxuan Chen and Lingwei Zheng developed a method for power generation prediction that applies a hybrid network composed of a graph neural network and a long short-term memory network to define the system in phase space. Realizing from previous research that photovoltaic power output is chaotic, the researchers used chaotic system analysis methods, such as phase space reconstruction, to predict the evolution of the system in phase space. This hybrid approach preserves the nonlinear nature of the system while maintaining model performance.

“The results provide a brand-new relational perspective to conduct predictions on the evolution of complex dynamic systems, representing a deep integration and innovation of interdisciplinary theories,” said Zheng.

When tested on a photovoltaic microgrid system, the model was able to significantly improve predictions on clear, cloudy, and rainy days.

The findings have applications beyond photovoltaic systems, including financial markets, the Internet of Things, meteorological science, biomedicine and other areas with time series predictions. However, as an open frontier, challenges remain, such as dealing with unstable periodic orbits of chaotic systems and injecting physical constraints into the calculations.

“Breaking through each of these challenges could significantly enhance prediction capabilities,” said Zheng.

Source: “Short-term prediction method of PV output sequence based on the phase space reconstruction and GAT-LSTM hybrid model,” by Gaoxuan Chen and Lingwei Zheng, Journal of Renewable and Sustainable Energy (2025). The article can be accessed at https://doi.org/10.1063/5.0281896 .