In Huan jing ke xue= Huanjing kexue
Based on the pollutant concentration data of Taiyuan City from 2016 to 2020 and the surface meteorological data of the national benchmark meteorological observation station in the same period, the variation characteristics of PM2.5 concentration in Taiyuan City and the effects of meteorological conditions such as humidity, precipitation, wind, and mixing layer thickness on PM2.5 concentration were analyzed. At the same time, the causes of pollutant concentration changes were discussed, and the PM2.5 concentration prediction model based on the LSTM neural network was established. The results showed that the number of days of heavy pollution in Taiyuan City from 2016 to 2020 was the highest in winter, of which the maximum number of days in 2017 was 28 days. The PM2.5 concentration was generally high in autumn and winter and low in spring and summer. The PM2.5 concentration on weekends was higher than that on weekdays. The daily variation in PM2.5 concentration roughly presented a bimodal distribution, which appeared around 09:00 and 23:00 to 01:00 the following day. Except for relative humidity and winter temperature, other air pressure, wind speed, and PM concentration showed negative correlations in the four seasons. The pollution sources affecting the increase in PM2.5 concentration in Taiyuan City were mainly located in the NE-ENE-E direction, and the pollution in the northwest was not relatively apparent. In flood season, when the precipitation reached the level of moderate rain (rainfall ≥ 10 mm), it had an obvious effect on the reduction of PM2.5 concentration. The increase in atmospheric mixing layer height was very beneficial to the diffusion and dilution of PM2.5 in the vertical direction. The strong northwest air flow in winter, low relative humidity, high pressure control on the ground, and high height of the mixing layer belonged to the cluster most conducive to the reduction in PM2.5 concentration. Using the LSTM model for modeling, the R2 of PM2.5 concentration prediction was as high as 0.95, which was significantly better than that of the traditional tree model and linear regression model (R2<0.60). The residual of the prediction results was close to the normal distribution, of which the absolute error of 84.2% prediction results was less than 20 μg·m-3, and the MAE, MAPE, and RMSE of the model were 38.17, 17.19%, and 20.6, respectively.
Li Ming-Ming, Wang Yan, Yan Shi-Ming, Chen Ling, Han Zhao-Yu
2023-Feb-08
PM2.5concentration, Taiyuan, forecast, k-means clustering, long short term memory(LSTM), machine learning, meteorological characteristics
