AirQFormer: Improving regional air quality forecast with a hybrid deep learning model

Abstract

Accurate air quality forecasting is crucial in providing reliable early warning information to the public. However, predictions generated by three-dimensional chemical transport models, such as the widely used Community Multiscale Air Quality (CMAQ) model, often exhibit considerable biases compared to observations. Post-processing techniques can substantially enhance the forecasting skill of air quality models. In this paper, a hybrid deep learning model, namely AirQFormer, is proposed as an end-to-end bias correction method to improve the accuracy and reliability of regional CMAQ forecasts over 72 h. The performance of AirQFormer was evaluated based on ozone observations from the Greater Bay Area in Southern China for the year 2023. AirQFormer demonstrated superior accuracy at the temporal scale compared to the CMAQ model and the long short-term memory (LSTM) model over the 72-hour forecasting period. It achieved an average reduction of 35 % (5.2 ppbv) in mean absolute error (MAE) and 33 % (6.5 ppbv) in root mean square error (RMSE) compared to the CMAQ model. Additionally, it showed a 12 % reduction (1.1 ppbv) in MAE and an 11 % reduction (1.4 ppbv) in RMSE compared to the LSTM model. At the spatial scale, AirQFormer outperformed both the CMAQ model and traditional spatial bias correction methods, with MAE values being 31 % (4.5 ppbv) and 5 % (0.5 ppbv) lower than those of the CMAQ model and traditional methods, respectively. Regarding peak value forecasting, AirQFormer exhibited notable improvements compared to the CMAQ model. The false alarm rate of AirQFormer is 12 % lower than that of the CMAQ model, indicating a more accurate identification of episode events. These results demonstrate the effectiveness of our proposed model in improving air quality predictions.