Ozone Forecasting in Hangzhou via a Frequency-Domain-Augmented Channel Attention Mechanism

Ground-Level ozone is a hazardous air pollutant; prolonged exposure threatens human health and natural ecosystems. Accurately forecasting future ozone concentrations is therefore essential for effective mitigation. Existing models predominantly operate in the time domain and seldom exploit frequency-domain information. We propose a novel ozone-forecasting framework that integrates a frequency-domain channel-enhanced attention mechanism to predict ozone at eight monitoring stations in Hangzhou. The model refines frequency-channel attention by dynamically amplifying the weights of low-frequency channels according to the anticipated periodicity. A frequency-domain graph neural network is further introduced to capture spatial–spectral dependencies among stations. Experimental results demonstrate that the proposed model surpasses baseline approaches across multiple forecast horizons. For 72-hour forecasts, the frequency-domain channel-enhanced model achieves an average RMSE of 12.2 µg /m³, MAE of 9.3 µg/m³, and R² of 0.74. Comparative visualizations in the frequency and time domains reveal that the frequency-domain predictive network captures clearer global patterns, confirming the superior performance of the proposed channel-enhanced architecture.