Research On Regional Correlation Prediction Method Of Air Pollutants Based On Integrated Neural Network

Urban air pollution is one of the focuses of current people's lives.As the causes of air pollutants become more complex,the kinds of air pollutants are increasing.It is a key problem that many research fields are currently studying for how to use the massive monitored data comprehensively to analyze and predict urban air pollution concentration.The existing methods for predicting the concentration of air pollutants can be divided into traditional prediction methods based on non-deep learning and prediction methods based on deep learning.Traditional prediction methods are mainly for small sample data sets.Some methods still use linear regression methods for prediction,failing to take the trend of nonlinear changes in air pollutant concentrations into account;The deep learning prediction method can solve the bottleneck problem in the traditional method and can effectively integrate big data.And its unique network structure enables linkage analysis of pollutant data in time and space dimensions.However,current research work on predicting air pollutants based on deep learning is still relatively rare.And there are fewer studies that target different neural network integration for the analysis of spatial and temporal association problems.Meanwhile,there is too much effective solutions for the feature dimension of the data set.And most of the existed studies consider the effect of the neighbor cities on the target city is the same,but the effect on the spatial dimension is different because of the distances.In view of the shortcomings of previous research work,this paper focuses on the linkage analysis and prediction of the spatial and temporal dual dimensions of air pollutant concentrations.A Dev-LSTM prediction model based on the integration of the deconvolution network and the LSTM is proposed.And in order to solve the sparsity of the data set,an SDev-LSTM prediction model based on the integration of anti-stack auto-encoders and Dev-LSTM is also proposed.Meanwhile,the Gaussian function is used to incorporate the inconsistency of the neighbor cities into the prediction system.The main work of this paper includes:1)In the Dev-LSTM model,firstly the definition of the problem and the framework of the model are shown;Secondly,the specific structure of the deconvolution network and the LSTM of the integrated model is given.And the manner in which the deconvolution network extracts the spatial dimension feature association and the LSTM extraction time dimension association are described in detail;Finally,the loss function of the model training is defined,and the training steps and fine-tuning process of the model are explained.2)In the SDev-LSTM model,firstly the definition of the problem and the framework of the model are shown;Secondly,the Gaussian function is used to calculate the effect of neighbor cities and incorporate into the prediction system.The specific structure of the anti-stack auto-encoders is given,and the principle of data dimension expansion of the anti-stack auto-encoders and its pre-training mode are explained in detail.Finally,the method of integrating the SDev-LSTM model with the Dev-LSTM model and the process of training and fine-tuning are described.3)Simulation experiments show that the performance of Dev-LSTM and SDevLSTM prediction models based on neural network integration is better than the classical model and has higher application value in air pollutant prediction.Meanwhile,it is analyzed that the Dev-LSTM and SDev-LSTM prediction models can be applied to different situations,thus further distinguishing the performance of the two models.