Research In Spatiotemporal Prediction Of PM_2.5 Concentrations Over The Pearl River Delta Based On BiLSTM

In the process of studying the formation and diffusion of PM_2.5,the spatiotemporal continuous high-precision PM_2.5 concentrations distribution can be important basic data.Aiming at the problem of discontinuity in time and space of traditional PM_2.5 concentrations inversion methods,this paper carried out inversion and prediction research on the PM_2.5 concentrations in the Pearl River Delta region:A bidirectional long-short-term memory network（Bi LSTM）model based on satellite AOD,meteorology,geographical features and socio-economic factors was constructed,which effectively improved the problem of time discontinuity.Based on the inversion model,this paper develops a scrolling prediction model to achieve accurate prediction of future short-term data,making up for the problem that traditional methods only correct past data.In this paper,kriging method was used for interpolation to realize the spatial continuity of PM_2.5 concentrations in the Pearl River Delta region.The main research work is as follows:（1）Dataset construction and verification of PM_2.5 concentrations and influencing factorsIn view of the incompleteness of PM_2.5 influencing factors in the existing research,this paper collects data of 16 influencing factors based on remote sensing images,meteorological monitoring stations and statistical yearbooks.In this paper,1km-level AOD data is used and verified with ground-based AERONET data,which improves the spatial resolution of the dataset.After correlation analysis,various factors such as AOD showed a positive correlation with PM_2.5concentration.（2）PM_2.5 inversion model considering multiple influencing factors and time characteristicsIn view of the defect that the traditional model cannot predict the time series,this paper uses the PM_2.5data set to construct the Bi LSTM model to invert the PM_2.5 concentrations.After taking into account the time series characteristics of PM_2.5 concentration before and after,the overall determination coefficient R² of the Bi LSTM model increased by 0.03 compared with the ordinary LSTM to 0.70,and the RMSE decreased by 1.22μg/m³ to 10.954μg/m³.The interpolation results obtained on the basis of the inversion data fit well with the real data,with an overall R² of 0.73,which can provide accurate spatiotemporal distribution characteristics.（2）Bi LSTM Scrolling model for future short-term predictionAiming at the problem that the PM_2.5 concentrations are difficult to predict in the short term in the future,this paper establishes a scrolling predicting model based on the trained Bi LSTM model.The model adjusts the time limit parameters of the model through performance analysis to determine the 1-month time.The overall R² of the final model is 0.64 and the RMSE is 11.276μg/m³.The verification results of the interpolation in January 2019 show that the spatial-temporal distribution map of PM_2.5 in the Pearl River Delta obtained through the combination of the Bi LSTM prediction results and the Kriging interpolation method meets the accuracy requirements,and can provide guidance for the prevention and control of air pollutants and residents’travel.This paper constructs a spatiotemporal continuous PM_2.5 inversion and prediction model.In addition to correcting the data of the past time,it can also predict the concentrations of PM_2.5 in the short-term future,which can improve data support for the study of the temporal and spatial changes of PM_2.5 to help the government prevent and control air pollutants and provide guidance for residents to travel.