Improving Ozone Forecasting Across Multiple Sites with Temporal Convolutional Graph Neural Networks for Air Quality Management

With rapid economic and industrial growth, ozone has become a major air pollutant, presenting significant risks to human health, ecosystems, and agriculture. Accurate forecasting of ozone concentrations is critical for effective pollution management and air quality improvement. However, most existing methods focus on short-term forecasts at a limited number of locations, while medium- to long-term predictions are still underdeveloped. This paper introduces a novel Temporal Convolutional Graph Neural Network (TCGNN), which integrates Graph Neural Networks (GNN), Temporal Convolutional Networks (TCN), and a temporal attention mechanism to predict ozone levels for up to 72 hours. In addition to ozone data, the model incorporates relevant meteorological and pollutant variables to improve forecast accuracy. Experimental results show that the TCGNN model outperforms baseline models in terms of prediction accuracy at various time steps. For 72-hour forecasts, the TCGNN achieves an average RMSE of 12.30, MAE of 9.74, and R2 of 0.71. The best performance is observed at the HZ03 station, with RMSE and MAE values of 10.96 and 7.46, respectively. Compared to two advanced models, TCGNN consistently yields better prediction results across all stations in Hangzhou Bay. This study aims to assist in proactive government actions and timely emission reduction strategies.