Research On Urban Air Quality Prediction Based On Temporal And Spatial Optimization Neural Network

Today's social economy is in a state of rapid development,which benefits from the economic effect of a large number of industrial production.However,with the emission of air pollutants in industrial production,air quality has been seriously damaged,which will cause serious harm to people's physical and mental health,so air quality prediction has important practical significance.In this paper,the air quality data and meteorological forecast data of 9 regional urban monitoring stations in 2013-2018 are analyzed systematically.It is found that there is a mutual influence relationship between the changes of pollutant PM2.5 values among multiple stations.The changes of pollutant concentration are affected by the spatial characteristics.In addition,the changes of PM2.5 concentration are affected by other pollutant concentrations and meteorological factors With the passage of time,it shows periodic changes,reflecting that there are time series problems in air quality data and meteorological factors,which have a long-term dependent characteristic relationship.In the previous methods of feature extraction,the convolution neural network single-scale convolution kernel method or artificial selection method is used for feature extraction,which has the problems of large error,low efficiency and low accuracy.In this paper,a PM2.5 prediction model(mscnn-lstm)based on convolution neural network multi-scale convolution kernel(mscnn)and LSTM network is proposed.In order to get more accurate prediction results,the mscnn-lstm prediction model is further optimized.Combined with the global optimization of genetic algorithm(GA),the mscnn-galstm prediction model is proposed.The main research contents are as follows:1)The traditional convolution neural network(CNN)uses the same size convolution kernel for feature extraction,which will make the network reach the bottleneck state to some extent,making the result of feature extraction not so ideal.In this paper,the multi-scale convolution kernel(mscnn)is used to extract features,and the multi-scale features are spliced and fused,and then the deeper and more comprehensive features are extracted,so that the network has stronger generalization ability.In addition,in the process of air quality data feature extraction,the air quality data of multi stations and multi features are simply changed to form a parallel one-dimensional feature map of multi stations and single features.Then,the spatiotemporal features of air quality data are obtained by multi-scale convolution of the changed parallel one-dimensional feature map.Finally,they are spliced and fused to obtain the spatiotemporal feature relationship of multi site and multi feature.2)Combined with the ability of LSTM network to deal with time series problems effectively,mscnn-lstm prediction model is proposed.Specifically,the mscnn-lstm prediction model inputs the spatiotemporal features extracted from mscnn into the LSTM network to mine the long-term dependence of air quality data.Then,the output of the LSTM network is connected to the all connected neural network to get the predicted PM2.5 concentration of the target station.3)Aiming at the problem that it is difficult to train LSTM network parameters in mscnn-lstm prediction model,a new model(mscnn-galstm)is proposed to optimize LSTM network parameters with genetic algorithm(GA).To be exact,mscnn-galstm model can find the optimal parameters through GA,and further improve the prediction accuracy of the model.The experimental results show that the mscnn-lstm and mscnn-galstm prediction models proposed in this paper have better performance in three evaluation indexes compared with the two comparative prediction models,and verify that the mscnn-lstm and mscnn-galstm prediction models have better development potential in PM2.5 concentration prediction.