Deep Learning Theory And Validation For Time-series Atmospheric Pollution Diffusion Problems

The prediction accuracy of traditional physical models is affected by the construction precision of external environments,observation frequency,and density.In contrast,deep learning models are effective in dealing with complex time series relations,which is due to the feature representation of massive long sequence data being effectively extracted by deep learning models.However,the transport and diffusion of air pollutants is a very complex environmental science phenomenon.Deep learning models with a simple network structure are used to predict the air pollution transmission process by fitting numerical curves,which lack physical mechanism characteristics and have a large error in predicting extreme points that never appeared in historical data sets.To overcome the limitations of deep learning models,a GD-GRU model,based on the combination of physical mechanism process and time-series deep learning model,was proposed in this paper to improve prediction accuracy,which combined the deep learning theory and the laws of physics.The specific work is as follows:1.Aiming at the problem that poor prediction of extreme points by deep learning model,a physical model based on atmospheric diffusion and transport theory was studied.Firstly,in order to understand the laws of diffusion,the diffusion of pollutants in Guilin city was analyzed.Secondly,to determine the input of the model,the correlation coefficient between related stations and target station,and the correlation coefficient between meteorological variables and PM_2.5 concentration were analyzed,respectively.Then,based on the mathematical description of air pollution transport,the Gaussian diffusion model was studied.Finally,a multi-point source pollutant concentration prediction method based on the Gaussian diffusion model was designed to obtain the estimated PM_2.5 values of the target site.2.Aiming at the poor prediction of traditional physical models in practical nonlinear problems,an error correction model based on a deep learning model was studied.Firstly,based on the estimated PM_2.5 values by the multi-point source Gaussian diffusion model,the root mean square error（RMSE）of the estimated values and the actual observed values were extracted as the error sequence.Secondly,the error values were regarded as output of gate recurrent unit（GRU）with the inputs of weather and pollutant parameters.Then,the model was trained,and the hyperparameters,such as sliding window size,were adjusted according to the training results.Finally,the PM_2.5 error predicted values were given by the error correction model,and the final PM_2.5predicted values were obtained by combining the output error values of the model with the estimated values.3.The PM_2.5 prediction results were used to analyze and compare between GD-GRU model,which was based on the Gaussian diffusion model and GRU model,and baseline models.The established model predicted PM_2.5 concentration with a mean absolute error（MAE）of 10.5971,an RMSE of 12.561,and a symmetric mean absolute percentage error（SMAPE）of 0.2024,which was approximately 19.69%,21.02%,and 18.11%,respectively,better than methods like autoregressive integrated moving average model（ARIMA）,support vector regression（SVR）,recurrent neural network（RNN）,long short-term memory model（LSTM）,and GRU.In conclusion,the GD-GRU model based on the combination of physical mechanism process and deep learning model performed well in extreme values prediction and atmospheric environment data prediction,providing a new method for atmospheric environment data prediction.This paper proposed a deep learning error correction model based on physical model simulation for the first time,which can provide a reference for the prediction of time series data based on the combination of deep learning and physical mechanism process,and provide a physical basis for the optimization of deep learning model in the field of air pollution transmission simulation.