Accurate prediction of pollutant concentration is essential for environmental protection and public health. Among various pollutants, PM2.5 is particularly dangerous due to its severe health impacts, making its precise forecasting a challenging task. In this work, we propose a hybrid Physics-Informed Graph Neural Network–Gaussian Process Regression (PIGNN-GPR) framework for hourly PM2.5 prediction. The physics-informed layer incorporates the reaction-diffusion-advection equation to maintain physical consistency, while the Graph Neural Network captures spatial dependencies using wind speed and direction across locations. To enhance prediction reliability, Gaussian Process Regression is used to refine outputs and provide uncertainty estimates. Additionally, we apply Inverse Distance Weighting to interpolate PM2.5 levels at unmonitored sites. Model interpretability is improved using SHapley Additive exPlanations, identifying the impact of key inputs like latitude, longitude, wind speed, and direction. We validate our model with real-world data from the Delhi region. The PIGNN-GPR framework offers both accurate and interpretable forecasts, aiding better air quality management.
