Attention scores and peak perception in long-term ozone prediction using deep learning

Abstract

To address the limitations of traditional ozone (O₃) forecasting models, this study established a novel Transformer-based model integrating attention scores and the peak perception. Attention scores dynamically quantify nonlinear relationships between O₃ and influencing factors, while peak perception method penalizes peak O₃ errors, ensuring accurate predictions during O₃ exceedance events. Our results demonstrate that the proposed model significantly improves the prediction accuracy for both hourly and maximum daily 8-h average O₃ concentrations, with the R² value increasing from 0.56 to 0.83 and the mean squared error decreasing from 0.47 to 0.42. Wind direction, wind speed, and carbon monoxide emerged as dominant factors during pollution episodes. Additionally, the model exhibits a high potential for predictability over extended lead times, with a mean absolute error of 0.67 at 24 h, stabilizing at 0.75 at 72 h. This approach enhances simulation accuracy and provides policymakers extended response windows for O₃ control strategies.