资源环境科技发展态势分析平台

To improve the forecasting effectiveness of numerical air quality models, two deep learning models, DeepPM and APTR, were constructed and trained in this study using PM_2.5 and O₃ monitoring data, and WRF-Chem numerical forecasts in the south-central Beijing-Tianjin-Hebei region. The optimization effects were evaluated using test datasets and various evaluation metrics. The results show that the PM_2.5 and O₃ forecast results optimized by the DeepPM, and APTR models significantly outperform the WRF-Chem numerical model for both proximity forecasts over the next 24 h and short- to medium-term forecasts over the next 144 h. The APTR model achieves the best optimization results in proximity forecasting, whereas the DeepPM model has a better overall performance in optimizing the short- and medium-term forecasts. WRF-Chem is superior to other models in predicting high O₃ concentration. DeepPM and APTR deep learning models are still significantly better than WRF-Chem for forecasting high concentration bands within the proximity forecast time period. For short- to medium-term forecasting, the DeepPM model outperforms WRF-Chem for forecasting high O₃ concentrations. This paper provides a new method and idea for improving the forecasting performance of air quality numerical models.