Research On On Air Quality Prediction Based On Bidirectional Long Short-Term Memory Network And Attention Mechanism

People’s quality of life has been significantly improved in recent years with the development of science and technology in our country.However,along with the increase of energy mining and usage and the rapid increase of motor vehicles,the air pollution problem is becoming more and more serious.Some areas have already seen photochemical smog phenomenon,many acid rain areas have been formed in the country,and haze and sand storm weather have appeared in many places,which has seriously threatened our living environment and life health.Therefore,many scholars have carried out research on air quality prediction methods,hoping to assist individual residents in making travel decisions and provide theoretical support for government environmental protection departments when formulating environmental protection policies.However,the existing prediction models based on deep learning method have some problems in time series prediction,such as poor processing of data nonlinear relationship and limited feature extraction ability.In order to further improve the prediction ability of the model,the following researches are carried out in this paper:(1)This paper collected air quality data sets of six cities from 2015 to 2022,namely Beijing,Tianjin,Taiyuan,Nanchang,Xi ’an and Shenzhen,and conducted data preprocessing on the data sets.According to the correlation table and scatter matrix graph obtained by correlation analysis,multicollinearity relationships exist among various characteristic variables in the data set.(2)In view of the problems of multicollinearity among the data collected in this paper,poor parallel processing of data,inability to effectively preserve long-term dependence of data,irrelevant feature interference,data noise and other problems in the existing time series prediction model,this paper,based on the deep learning method,Combined with the ability of Multiscale Convolution Neural Network to extract complex multi-scale time feature information,the higher data feature extraction efficiency of Bidirectional Long Short-Term Memory,the ability of Attention Mechanism to improve model accuracy,and the ability of full connection layer to deal with data nonlinear problems,an air quality prediction model based on MConv-Bi LSTM-AM was proposed.Comparative analysis of Convolution Neural Network,Long Short-Term Memory,Bi LSTM,CNN-LSTM,CNN-Bi LSTM,CNn-LSTM-AM and CNN-Bi LSTM-AM models,It is proved that this model has great advantages in the effective extraction of data features.(3)In order to further reduce the amount of training calculation,avoid the local minimum in the gradient descent optimization algorithm,and better capture the long-term dependence between data,Echo State Networks were introduced to propose an air quality prediction model based on MESN-Bi LSTM-AM.By replacing the hidden layer of MConv with a reserve pool,this model adjusts the weights inside the network to realize short-term memory and ensure that the weights of the network reach the global optimal value.On data sets of six cities,a comparative analysis of many deep learning models,including MConv-Bi LSTM-AM,proves the effectiveness of this model.