Application of XGBoost algorithm in the optimization of pollutant concentration
With the rapid development of the economy, urban pollution has become a hot issue of
human concern, and people are put forward higher requirements for the accuracy of
pollutant concentration simulations. Based on the the WRF-Chem model simulation results
and Beijing environmental monitoring data, this study constructed the XGBoost algorithm
through the process of data cleaning, feature selection and super parameter optimization,
and the optimized simulation of PM 2.5 and O 3 concentrations in Beijing was then carried …
human concern, and people are put forward higher requirements for the accuracy of
pollutant concentration simulations. Based on the the WRF-Chem model simulation results
and Beijing environmental monitoring data, this study constructed the XGBoost algorithm
through the process of data cleaning, feature selection and super parameter optimization,
and the optimized simulation of PM 2.5 and O 3 concentrations in Beijing was then carried …
Abstract
With the rapid development of the economy, urban pollution has become a hot issue of human concern, and people are put forward higher requirements for the accuracy of pollutant concentration simulations. Based on the the WRF-Chem model simulation results and Beijing environmental monitoring data, this study constructed the XGBoost algorithm through the process of data cleaning, feature selection and super parameter optimization, and the optimized simulation of PM2.5 and O3 concentrations in Beijing was then carried out based on the constructed algorithm. The results showed that the XGBoost algorithm can better capture the spatial and temporal variation patterns of pollutant concentrations, and has a greater improvement on the simulation results of the WRF-Chem model, and that the XGBoost algorithm shows better optimisation results in urban areas compared to suburban areas. In addition, the overall analysis of the features of PM2.5 and O3 concentrations based on the SHAP value theory showed that the time series and periodic features, aerosol ion concentration of Sodium (NAAJ) and Nitrate concentration (NO3AJ) were the important features affecting the prediction of PM2.5 concentration by the XGBoost algorithm, and the most important factor affecting O3 concentration is temperature. PM2.5 and O3 concentrations were divided into three levels, and several samples were selected for single sample analysis. The analysis showed that at low pollutant concentration, most of the features made negative contributions to the concentration prediction, while at high concentration, most of the features made positive contributions. The contribution values of different features varied greatly and were unevenly distributed. The results of prediction were basically composed of a few features with large feature contributions, and the feature contributions of the same feature to different concentration prediction results were also different. Moreover, the XGBoost algorithm was used to optimize the concentration of pollutant at each grid point in Beijing, and a set of pollutant concentration data set with spatial resolution of 6 km and time resolution of 1 h covering the whole Beijing was established, and the optimized spatial distribution of pollutant concentration was closer to the spatial distribution of observed concentration than WRF-Chem simulation. At last, compared with SVR, LR, DTR and RF algorithms, XGBoost algorithm was better than other statistical algorithms in optimising PM2.5 and O3 concentrations. The results of this study provided a new idea for an in-depth analysis of optimization principle of algorithm model for air pollution and a quantitative study of the influencing factors.
Elsevier