Evaluation of the prediction performance of air quality numerical forecast models in Shenzhen

Abstract

As an economically intensive city within the greater bay area (GBA), Shenzhen has suffered from severe ozone pollution in recent years. Numerical modeling is an important tool in daily air quality prediction. However, a combination of factors, such as outdated emission inventories and the uncertainty of weather and numeric models, pose challenges for accurately forecasting pollution periods and levels. Therefore, systematic evaluation of the numerical model performance is crucial for better manual revised prediction of the daily MAD8h and air quality index (AQI). In this paper, a multimodel air quality forecast system, which was built in our public-owned monitoring center station, was evaluated to understand the possibilities of improvements. Based on the 2021 observed data of the national air monitoring station, we utilized a series of metrics to access the predicted air quality index (AQI) and pollution concentration levels from models (including WRF-CMAQ, NAQPMS, and WRF-CHEM) and compared them to the manual prediction. The results indicated that WRF-CMAQ performs outstandingly in comprehensive analysis, with a yearly 24 h (hour) forecast accuracy (A) exceeding 91%. The NAQPMS 24 h forecast presented the lowest accuracy (A = 71%) with a high overprediction rate. The forecast accuracy of WRF-CHEM and NAQPMS in the winter period was 100%. However, frequently changing weather patterns interfered with spring forecasting. Forecasters can accurately predict mild pollution cases (A = 81.8% for 24 h, A = 72.7% for 48 h), whereas numerical models (NAQPMS and WRF-CHEM) showed better sensitivity to moderate pollution cases. For fine particulate evaluation, CMAQ's low biases occurred in winter, and the manually revised air quality forecast (MRAQR) normally gave higher prediction values except in pollution cases. Model underestimation could be explained by missing the long-range dust transport contribution from the northern part of China. The ozone 24 h forecast part revealed that WRF-CMAQ exhibited the highest correlation coefficient (R = 0.70, 19.74% < RMSE < 42.7%), whereas NAQPMS showed relatively poor results with the highest ME (43 μg/m3) and MFE (41.92%). Close examination of synoptic information revealed that ozone pollution cases were subjected to the western Pacific subtropical high (WPSH) system, tropical cycle, and regional strong pollution transport. The NAQPMS ozone overprediction could be explained by the overestimation of troposphere transmission from the surroundings during nonconductive weather patterns. The study investigates a theoretical method for model optimization, which provides solid support for precise emission reduction and targeted pollution control.