A Novel Wind-Aware Dynamic Graph Neural Network for Urban Ground-Level Ozone Concentration Prediction

Ground-level ozone pollution poses significant risks to public health and ecosystems and remains a major environmental challenge worldwide. Accurate forecasting is difficult due to the nonlinear formation mechanisms of ozone and its strong dependence on meteorological conditions. This study proposes a Wind Speed and Direction-Based Dynamic Spatiotemporal Graph Attention Network (WSDST-GAT) for multi-step hourly ground-level ozone prediction. The model integrates a wind-aware dynamic graph to represent anisotropic pollutant transport and a Transformer-based temporal encoder to capture long-range dependencies. Meteorological variables are incorporated to enhance physical interpretability and predictive robustness. A co-kriging module is further employed to reconstruct continuous spatial ozone fields with quantified uncertainty. Using hourly observations from 35 monitoring stations in Beijing, WSDST-GAT achieves a Coefficient of Determination of 0.957, with a Mean Absolute Error of 5.25 μg/m3, and a Root Mean Square Error of 9.58 μg/m3. The prediction intervals demonstrate strong reliability with a Prediction Interval Coverage Probability of 94.01% and a Prediction Interval Normalized Average Width of 0.174. These results indicate that the proposed framework provides an accurate and physically informed solution for ozone forecasting and air quality management.