Solar energy forecasting in Gwangju using PM10 data and Physics-Informed Neural Networks

Accurate solar power forecasting is essential for stable grid operation and the efficient integration of renewable energy sources. While traditional data-driven models often neglect atmospheric influences such as particulate pollution, this study investigates the impact of PM10 on solar energy generation in Gwangju, South Korea. We constructed a novel dataset merging solar, weather, and air quality data from 2017 to 2024 and evaluated multiple deep learning models, including a Multi-Layer Perceptron (MLP), Long Short-Term Memory network (LSTM), Physics-Guided Neural Network (PGNN), and a Physics-Informed Neural Network (PINN). PGNN and PINN incorporated the Beer–Lambert law into the loss function to achieve physical consistency between solar irradiance, pollution concentration, and power output. Results indicated that the MLP achieved the lowest mean squared error, outperforming the physics-augmented models, while PINN exhibited the highest error due to over-constraining. This work is among the first to integrate PM10 data and physics priors into a solar energy forecasting pipeline for Korea, offering a reproducible methodology that balances accuracy and interpretability.