Research On PM_2.5 Concentration Prediction In The Yangtze River Delta Region Based On Deep Learnin

The large amount of energy consumption and the continuous increase of vehicle ownership lead to the increasingly prominent problems such as air pollution.It is a great threat to people’s health.PM_2.5 also has a direct or indirect impact on climate change and human living environment.Therefore,PM_2.5 concentration prediction is an indispensable work,scientific and accurate prediction of PM_2.5 concentration can guide the national air pollution prevention and control work,social activity organization and urban residents to travel reasonably.At present,relevant researches use various methods to fill data when dealing with missing values.No matter what method is adopted,data pollution or even information leakage will be caused in the future.When selecting auxiliary parameters for prediction,correlation coefficient is adopted,but correlation does not mean causality.Various parameters are added into the deep learning model for PM_2.5 concentration prediction,but the data expression method of PM_2.5concentration diffusion process is ignored.In view of the above problems,on the basis of analyzing the temporal and spatial variation of PM_2.5 concentration in the Yangtze River Delta region and the causal relationship between PM_2.5 concentration data and other pollutant data and meteorological data,Based on hourly air quality monitoring station data from January 1,2019 to December 31,2019 in the Yangtze River Delta region and model reanalysis data during the corresponding period,a deep learning model based prediction method for PM_2.5 concentration considering diffusion process was proposed.A"diffusion collector"was innovatively designed to express the input mode of deep learning parameters for the diffusion process of gaseous pollutants.In addition,several deep learning models are compared to find the better model,and the generalization ability of"diffusion collector"method on different models is evaluated.Then the causality analysis information flow method was used to select the auxiliary prediction parameters,and the sensitivity analysis of single factor prediction was conducted to evaluate the applicability of causality analysis method in the selection of prediction parameters.The results show that:（1）The variation of PM_2.5 concentration presents certain temporal and spatial characteristics.The characteristics of PM_2.5 concentration are not obvious and difficult to capture from the whole time series diagram.The hourly variation of PM_2.5 concentration in the Yangtze River Delta region is in the shape of"sine function",while the daily variation of PM_2.5 concentration is relatively gentle and the monthly variation of PM_2.5 concentration is in the shape of"V".The spatial distribution pattern of PM_2.5 concentration increased gradually from southeast to northwest in the Yangtze River Delta.At the same time,PM_2.5 concentration at air quality monitoring stations has strong spatial autocorrelation characteristics.（2）From the comparison of different prediction models,the deep learning model is proved to be significantly better than the machine learning model,indicating that the deep learning model can effectively capture the time dependence of time series data.This also gives them a unique advantage in time series prediction.In addition,among these deep learning models,LSTM has the highest accuracy and goodness of fit for PM_2.5 concentration prediction,and LSTM has a better effect on predicting PM_2.5 concentration in the future one hour.（3）Using"diffusion collector"method to organize PM_2.5 concentration data of nearby stations to predict PM_2.5 concentration has a better effect,the root mean square error is reduced by 12.99%on average,and the determination coefficient is increased by 0.026 on average,which also reflects that"diffusion collector"has a good generalization ability.It shows that the expression of gas diffusion process plays an important and positive role in model prediction.（4）According to the results of single factor sensitivity analysis,the improvement of prediction accuracy of PM_2.5 concentration by auxiliary parameters is consistent with the size of information flow of PM_2.5 by each factor.Therefore,the causality analysis information flow method is feasible as the auxiliary parameter selection basis for PM_2.5 concentration prediction,and is stronger than the correlation coefficient method.