Abstract: Based on the daily concentration of PM_2.5 data and the surface meteorological data from national meteorological observatory station in Shantou from 2014 to 2017,the variation characteristics of PM_2.5 concentration in Shantou and the influences of meteorological conditions such as wind,mixed-layer thickness and precipitation on the concentration of PM_2.5 were analyzed and the causes of pollutant concentration variation were investigated.On this basis,according to the climatic characteristics of Shantou,the models for predicting PM_2.5 mass concentration in flood and non-flood season were respectively established with BP (Back-Propagation) artificial neural network method.The results show that the daily variation of PM_2.5 concentration in Shantou is unimodal,which is different from most inland cities and is related to the geographic location.More specifically,Shantou is located in the coastal area affected by the land-sea breeze.The daily peak of PM_2.5 concentration appears at around 8 o'clock,which is caused by the lower wind speed and the increase of pollutant emission in the morning.The average concentration of PM_2.5 decreases with the increase of wind speed and is significantly correlated with the thickness of the mixed layer at 8 o'clock (the correlation coefficient is -0.143,p< 0.001).The concentration of PM_2.5 in the non-flood period in Shantou city is higher than that in the flood season,which is related to the subtropical monsoon climate characteristics.In addition,there is no significant difference in the removal effect of PM_2.5 among various magnitudes of precipitation in the flood season (except for rainstorms),while the precipitation during the non-flood period has an obvious effect in decreasing the concentration of PM_2.5.The BP artificial neural network model shows a high hit rate in forecasting the grade of PM_2.5 concentration.More specifically,Rank accuracy of PM_2.5 are 100% and 90.3%,the accuracy coefficients are 87.7% and 89.9% in flood season and non-flood season,respectively.The number and amplitude of positive errors of the forecasted PM_2.5 concentration by the model in different periods are larger than those of the negative errors.Furthermore,the prediction error of the model is larger when heavy rainfall occurs in flood season and cold air invades in non-flood season.