Regional Predictions of Air Pollution in Guangzhou: Preliminary Results and Multi-Model Cross-Validations

Abstract

A precise air pollution forecast is the basis for targeted pollution control and sustained improvements in air quality. It is desirable and crucial to select the most suitable model for air pollution forecasting (APF). To achieve this goal, this paper provides a comprehensive evaluation of performances of different models in simulating the most common air pollutants (e.g., PM2.5, NO2, SO2, and CO) in Guangzhou (23.13° N, 113.26° E), China. To simulate temporal variations of the above-mentioned air pollutant concentrations in Guangzhou in September and October 2020, we use a numerical forecasting model (i.e., the Weather Research and Forecasting model with Chemistry (WRF-Chem)) and two artificial intelligence models (i.e., the back propagation neural network (BPNN) model and the long short-term memory (LSTM) model). WRF-Chem is also used to simulate the meteorological elements (e.g., the 2 m temperature (T2), 2 m relative humidity (RH), and 10 m wind speed and direction (WS, WD)). In order to investigate the simulation accuracies of classical APF models, we simultaneously compare the simulations of the WRF-Chem, BPNN, and LSTM models to ground truth observations. Comparative assessment results show that WRF-Chem simulated air pollutant (i.e., PM2.5, NO2, SO2, and CO) concentrations have the best correlations with ground measurements (i.e., Pearson correlation coefficient R = 0.88, 0.73, 0.61, and 0.61, respectively). Furthermore, to evaluate model performance in terms of accuracy and stability, the normalized mean bias (NMB, %) and mean fractional bias (MFB, %) are adopted as the standard performance metrics (SPMs) proposed by Boylan et al. The comparison results indicate that when simulating PM2.5, WRF-Chem was more effective than the BPNN but less effective than the LSTM. While simulating concentrations of NO2, SO2, and CO, the WRF-Chem model performed better than the BPNN and LSTM models. With regards to WRF-Chem, the NMBs and MFBs for the PM2.5 simulations are, respectively, 6.49% and 0.02%, –11.96% and –0.031% for NO2, 7.93% and 0.019% for CO, and 5.04% and 0.012% for SO2. Our results suggest that WRF-Chem has superior performance and better accuracy than the NN-based prediction models, making it a promising and useful tool to accurately predict and forecast regional air pollutant concentrations on a city scale.

Keywords

• WRF-Chem
• back propagation neural network
• long short-term memory
• air pollution
• Guangzhou