Regional PM2.5 concentration prediction analysis and spatio-temporal mapping incorporating ZWD data

Abstract

In recent years, air pollution events have been frequent and machine learning models have been widely used in research on long-term and short-term PM2.5 concentration prediction due to their excellent predictive capabilities, but most studies have only dealt with a small number of stations or cities, failing to take into account a spatial plane perspective to demonstrate future PM2.5 concentration changes in the region. Therefore, in this paper, 617 air quality monitoring sites in central and eastern China are used as examples. First, the five meteorological elements of wind speed (WIN), temperature (TEM), precipitation (PRE), pressure (PRS), relative humidity (RHU) and zenith wet delay (ZWD) data were interpolated to the air quality monitoring sites (PM2.5, PM10, SO2, NO2, O3, CO) by the inverse distance interpolation (IDW) method. Secondly, the density distribution of each factor was analyzed, and the monthly mean value of each site was interpolation mapped to observe the overall change trend of PM2.5 in the central and eastern regions, and then Spearman correlation coefficient was used to determine the relevant factors for modelling. Finally, four models of back propagation (BP) neural network, random forest (RF), support vector regression (SVR) and long and short-term memory network (LSTM) were constructed to predict the change of PM2.5 concentration at each air quality monitoring station for the next week (7 days), half month (15 days) and month (30 days), selecting the model with the best applicability at each time scale, and mapping the prediction results of stations with R2 greater than 0.5 in the model through spatial interpolation for spatial and temporal mapping. The results show that: (1) except for the differences between the density distribution trends of TEM, PRS and RHU and PM2.5, the remaining elements are consistent with the PM2.5 change trends, and the PM2.5 concentrations in the central-eastern region are high in winter (January, February and December) and low in summer (June, July and August), with a spatial distribution of high in the central-west and low in the east; (2) through the evaluation of the index results, it can be seen that in the three time scales of PM2.5 concentration prediction, RF has better applicability compared with the three models of BP, SVR and LSTM, among which the model accuracy is the best in predicting the future week; (3) By comparing the model interpolation results with the original data interpolation results, it can be seen that the results obtained from the model interpolation basically fit the spatial and temporal trends of future PM2.5 concentrations, and the best interpolation results in predicting the future week, the highest reliability of the spatial and temporal mapping results, and more accurate description of the future PM2.5 concentration changes.