Study On Prediction Of Regional Air Pollutant Diffusion Trend Based On Spatiotemporal Feature Learning

Air pollutants have brought a huge impact on people's living environment at present,and the treatment and prediction of pollutants has attracted wide attention from the world.Traditional pollutant prediction methods encounter obstacles in solving the long-term sequence prediction problem.It is difficult to deeply extract the spatiotemporal correlation features between pollutants and weather data,and it is impossible to calculate the influence of location information,pollutants and weather factors on the prediction results at each moment.Therefore,in the process of predicting urban pollutants,the calculation of the influence on regional urban pollution has not been combined in the research of pollutant prediction.At present,with the support of historical pollutants and big weather data,new machine learning technologies have brought new ideas and methods to solve the problems encountered in pollutant prediction.According to the features of pollutant concentration prediction methods used in related research,air pollutant concentration prediction can be basically divided into two main research methods: deterministic and statistical methods.The deterministic approaches can be applied to a limited set of historical data.However,meteorological principles and statistical approaches are needed to simulate the process of real-time emission,diffusion,transformation,and removal of pollutants based on atmospheric physics and chemical reactions.The model structure based on the deterministic method is predefined based on certain theoretical assumptions and prior knowledge,so it is difficult to learn the spatial and temporal dynamics of regional pollutants and to accurately and accurately predict pollutant concentrations.Statistical methods can be subdivided into traditional machine learning methods and new-type deep learning methods in pollutant prediction tasks.The characteristics of traditional machine learning are that it can converge quickly on historical training data with a small amount of data and the prediction accuracy in the prediction task of pollutant concentration is higher than the deterministic method.In recent years,methods based on deep learning have become important technical means for studying environmental pollution.The deep learning method can quickly learn the distribution features and laws among the data in environmental pollution big data,and the model quickly converges through training.According to existing research,deep learning methods have made rapid progress in the prediction of pollutant concentrations.However,in the existing researches,it is rarely possible to correlate the spatiotemporal features of environmental pollution data of regional multi-city sites,and it is difficult to calculate the impact of regional urban pollutant diffusion on the target urban pollutant prediction process.However,the current prediction methods face problems in the task of pollutant prediction:(1)the complex correlation features among meteorological data and air pollution data should be extracted and learned for further prediction and performance improving;(2)the temporal dependency feature among the historical data should be extracted accurately for prediction.That means,the redundant information or feature from passed long-time intervals should be forgotten in prediction,while the useful information or feature should be remembered in some duration for improving prediction;(3)the spatial related features among the adjacent cities in a region should be extracted based on their massive meteorological data and pollution data with temporal series tags.Unfortunately,these problems cause a poor performance in most traditional air pollutant prediction models;and(4)consider the impact of regional urban pollutants on the target research task,and influence in the comprehensive spatiotemporal dimension.Aiming at the shortcomings of the existing research work,this paper focuses on the joint regional multi-site environmental pollution data,and researches on the prediction of regional air pollution concentration and its diffusion trend from two dimensions of spatial and temporal.First,in order to fully study the distribution of historical data and spatiotemporal correlation features and improve the accuracy of pollutant concentration prediction,an RCL-Learning prediction model based on the integration of residual network and convolutional LSTM was proposed.On the one hand,the problem of spatiotemporal correlation of pollutants and meteorological data in multiple cities is solved,and on the other hand,accurate prediction can be made,which can be used as the basic work for regional urban pollution diffusion research.The extracted spatiotemporal correlation features are used as an important support for subsequent research.Secondly,on the basis of maintaining the prediction accuracy of the RCL-Learning model,we conduct a regional pollutant diffusion trend study and propose a cascade-based Attention mechanism model ABL-Learning.On the one hand,the model calculates the contribution of input time-series features to target prediction through time series influence,and on the other hand calculates the influence of pollutants from multiple cities in different regions on target city pollutant prediction.Therefore,based on the RCL-Learning and ABL-Learning models,combining the advantages of the two models,the final research work on the prediction of pollutant concentration and regional urban pollutant diffusion is completed.(1)Taking the data of multiple city sites as input and the residual network as the bottom layer,the spatial features of the input data are deeply extracted,and the output is used as the input of the high-level convolutional LSTM to extract the spatial-temporal correlation features of the data.Finally,the hidden state output at each moment is input to the fully connected layer to produce the final prediction result.(2)Aiming at the problem of the time series pollutant diffusion trend,a self-encode network based on cascade attention is proposed to deeply extract the spatio-temporal correlation features of data and jointly consider the influence calculation problem,which is convenient for the encoder to encode the spatio-temporal cfeatures of data.Attention is implemented in the encode and decode parts,so that the influence of regional urban pollutants and time series pollutant diffusion trends can be found based on the results of attention.(3)The experiments show that the performance of the RCL-Learning model based on neural network integration is better than the classic model,and it has higher application value in the prediction of air pollutants.At the same time,the ABL-Learning model based on the cascade attention mechanism can be applied to the prediction of regional pollutant diffusion trends.