TY - JOUR N2 - The prediction of PM2.5 is important for environmental forecasting and air pollution control. In this study, four machine learning methods, ground-based LiDAR data and meteorological data were used to predict the ground-level PM2.5 concentrations in Beijing. Among the four methods, the random forest (RF) method was the most effective in predicting ground-level PM2.5 concentrations. Compared with BP neural network, support vector machine (SVM), and various linear fitting methods, the accuracy of the RF method was superior by 10%. The method can describe the spatial and temporal variation in PM2.5 concentrations under different meteorological conditions, with low root mean square error (RMSE) and mean square deviation (MD), and the consistency index (IA) reached 99.69%. Under different weather conditions, the hourly variation in PM2.5 concentrations has a good descriptive ability. In this paper, we analyzed the weights of input variables in the RF method, constructed a pollution case to correspond to the relationship between input variables and PM2.5, and analyzed the sources of pollutants via HYSPLIT backward trajectory. This method can study the interaction between PM2.5 and air pollution variables, and provide new ideas for preventing and forecasting air pollution. L1 - http://www.czasopisma.pan.pl/Content/120756/Archives%203_vol47_2021_pp98_107.pdf L2 - http://www.czasopisma.pan.pl/Content/120756 PY - 2021 IS - 3 EP - 107 DO - 10.24425/aep.2021.138468 KW - PM2.5 KW - lidar KW - machine learning KW - air pollution monitoring A1 - Fang, Zhiyuan A1 - Yang, Hao A1 - Li, Cheng A1 - Cheng, Liangliang A1 - Zhao, Ming A1 - Xie, Chenbo PB - Polish Academy of Sciences VL - 47 DA - 19.09.2021 T1 - Prediction of PM2.5 hourly concentrations in Beijing based on machine learning algorithm and ground-based LiDAR SP - 98 UR - http://www.czasopisma.pan.pl/dlibra/publication/edition/120756 T2 - Archives of Environmental Protection ER -