[Revealing Driving Factors of Urban O3 Based on Explainable Machine Learning]
Jia-Qi Dong,Dong-Mei Hu,Yu-Long Yan,Lin Peng,Peng-Hui Zhang,Yue-Yuan Niu,Xiao-Lin Duan
DOI: https://doi.org/10.13227/j.hjkx.202208214
2023-07-08
Abstract:Driven by precursor emissions, meteorological conditions, and other factors, atmospheric ozone (O3) has become the main pollutant affecting urban air quality in summer. The current deductive models driven by physical and chemical mechanisms require a large number of parameters for the analysis of O3 pollution, and the calculation timeliness is poor. The data-driven inductive models are efficient but have problems such as poor explanation. In this study, an explainable model of data-driven Correlation-ML-SHAP was established to reveal the strongly correlated influencing factors of O3 concentration. Additionally, the machine learning ML module coupled with the explainable SHAP module was used to calculate the contributions of driving factors to O3 concentration, so as to realize the quantitative analysis of driving factors. The O3 pollution process in the summer of 2021 in Jincheng City was used as an example to carry out the application research. The results showed that the Correlation-ML-SHAP model could reveal and use strong driving factors to simulate O3 concentration and quantify influence contribution, and the ML module used the XGBoost model to achieve the best simulation accuracy. Air temperature, solar radiation, relative humidity, and precursor emission level were the strong driving factors of O3 pollution in Jincheng City in summer 2021, and the contribution weights were 32.1%, 21.3%, 16.5%, and 15.6%. The contribution weights of air temperature, solar radiation, and precursor emission level increased by 3.4%, 1.2%, and 1.2% on polluted days, respectively, and the contribution weights of precursor emission level rose to third place on polluted days. Each driving factor had a nonlinear interaction effect on O3 concentration. When the air temperature exceeded 24℃, or the relative humidity was lower than 70%, there was a 94.9% and 94.1% probability of positive contribution to O3 pollution, respectively. Under such meteorological conditions, ρ(NO2) exceeded 9 μg·m-3, or ρ(CO) exceeded 0.7 mg·m-3, and there was a 94.9% and 99.3% probability of positive contribution to O3 pollution, respectively. The southeast wind speed was lower than 5.8 m·s-1, or the south wind speed was lower than 5.3 m·s-1, both of which contributed positively to O3 pollution. The model quantitatively analyzed the influence contribution of various driving factors on urban O3 concentration, which could provide a basis for the prevention and control of urban atmospheric O3 pollution in summer.