Community Multiscale Air Quality (CMAQ) model can predict air pollutant concentration at every pre-determined location but usually suffers from considerable systematic bias. By fusing the CAMQ simulations and ground observations, first part of this study developed an unsupervised hybrid model, namely K-means-cluster-Observation-examiner-Boundary-denoiser (KOB) model, to segment region and to systematically reduce the bias of CMAQ simulations. The KOB model was then applied to improve prediction of CMAQ predictions of respirable suspended particulates (RSP) concentrations in the Pearl River Delta (PRD) region of southern China in 2019, using observations from 69 air quality monitoring stations in 2019. Compared to the CMAQ simulations, the KOB model significantly improves the 1-hour pred...[ Read more ]

Community Multiscale Air Quality (CMAQ) model can predict air pollutant concentration at every pre-determined location but usually suffers from considerable systematic bias. By fusing the CAMQ simulations and ground observations, first part of this study developed an unsupervised hybrid model, namely K-means-cluster-Observation-examiner-Boundary-denoiser (KOB) model, to segment region and to systematically reduce the bias of CMAQ simulations. The KOB model was then applied to improve prediction of CMAQ predictions of respirable suspended particulates (RSP) concentrations in the Pearl River Delta (PRD) region of southern China in 2019, using observations from 69 air quality monitoring stations in 2019. Compared to the CMAQ simulations, the KOB model significantly improves the 1-hour predictions of RSP at all stations, with index of agreement (IOA) increasing from 0.70 to 0.86 and average root mean square error (RMSE) reducing from 28.4 to 20.0 μg/m³ . Better agreement was also achieved between monthly average RSP from predictions using the KOB model and ground observations (e.g., IOA increase from 0.89 to 0.93, RMSE decrease from 13.8 to 9.5 μg/m³). In addition, the capability of the KOB model predicting polluted hour concentration greatly improved, with RMSE reduced from 76.3 to 50.8 μg/m³. Furthermore, the improvement was identified not only for next 1-hour predictions, but also for the predictions in next 6 hours. Through the data fusion, our hybrid clustering model can achieve robust real-time air quality forecasting.

Second part of this study developed a semi-supervised learning method for region segmentation with a core of constraint k-means guided by bias segmentation results to correct the systematic bias of CMAQ. The Semi-supervised K-means-cluster-Observation-examiner-Boundary-denoiser (Semi-KOB) model is developed with RSP in PRD region of China, using observations from 59 air quality monitoring stations from 2018 to 2019. Compared to the CMAQ simulations, the Semi-KOB model significantly improves the 1-hour predictions of RSP at all stations, with index of agreement (IOA) increasing from 0.71 to 0.86 and average root mean square error (RMSE) reducing from 28.1 to 19.5 μg/m³. Better agreement was also achieved between monthly average RSP from predictions using the KOB model and ground observations (e.g., IOA increase from 0.91 to 0.93, RMSE decrease from 12.8 to 8.4 μg/m³). In addition, the capability of the Semi-KOB model to predict polluted hour concentration greatly improved, with RMSE reduced from 75.3 to 50.0 μg/m³. Furthermore, the improvement was identified not only for next 1-hour predictions, but also for the predictions in next 12 hours. Through the data fusion, our hybrid clustering model can achieve robust real-time air quality forecasting.

Third part of the study, we generalized the Semi-KOB model into other pollutants (PM₁₀, PM_2.5, NO₂ and O₃) in PRD region, as well as RSP in other four domains of CMAQ model (Guangdong Province, South of China and whole China). The Semi-KOB model shows improvement in all the four pollutants and in all the four domains. In 12-hour predictions, the model shows improvement in all 12-hour in terms of averages and outperforming rate by station of all the evaluation metrics. To summarize, in the PRD region of China, the Semi-KOB model performs best with NO₂, then particulate matters and worst with ozone. And the Semi-KOB model performs better in larger spatial coverage with CMAQ data, that is, performing best in China, then the South of China, Guangdong province and the PRD region for the pollutant of RSP.