Research On Air Quality Prediction Method Based On Deep Learning

In recent years,with the rapid deterioration of air quality and the continuous improvement of people’s pursuit of living standards,the society has paid more and more attention to air quality.The PM2.5 in the atmosphere has always been a hot area of air quality prediction research due to its high degree of harm,large fluctuation and wide distribution.In order to help improve the air quality and improve the prediction accuracy of PM2.5 concentration,this paper proposes several deep learning methods and hybrid learning models to predict PM2.5 concentration.Experiments show that the prediction effect of the model proposed in this paper is better than most of the existing models,and it can better avoid the influence of nonlinear and other factors in the time period with poor stability.The specific research contents are as follows:(1)Considering that existing deep learning methods cannot accurately predict PM2.5 concentration changes in the air,this paper proposes a bidirectional nested long short-term memory network(Bi NLSTM)model to predict PM2.5 concentration changes.The Bi NLSTM model built in this paper is composed of a bidirectional long short-term memory neural network(BiLSTM)and a nested long short-term memory neural network(NLSTM).BiLSTM contains NLSTM units to obtain the forward and backward features of the current time point.Due to the unique nested structure of NLSTM units,Bi NLSTM deepens the learning ability of features and better captures historical data features.Experiments show that compared with other machine learning models and LSTM extended time series models,Bi NLSTM can better adapt to the characteristics of strong instability and complex frequency domain in PM2.etc.are very prominent.(2)Considering the problems caused by the large fluctuation and instability of air quality over time to predict the time domain changes of PM2.5,this paper constructs a hybrid deep learning model VMD-BiLSTM model.First,the original PM2.5 time series data is decomposed into multiple modal components(IMF)according to the size of the frequency domain using Variational Modal Decomposition(VMD),and then the bidirectional long short-term memory network(BiLSTM)is used for prediction of different modal components.The experimental exploration of the ACC curve and the Loss curve proves that the model iteration reaches the optimum,and then the exploration of the model can effectively reduce the prediction error.The comparison experiments in this chapter are compared with machine learning prediction models,time series prediction models combined with empirical modal decomposition EMD,and other LSTM extended models based on VMD,and then use a variety of evaluation indicators and prediction effect charts to show a comprehensive display The superiority of the model in forecasting.Experiments show that the prediction ability of the model combined with signal decomposition is better than the single prediction model,the prediction model based on VMD decomposition is better than the prediction model based on EMD decomposition,and VMD-BiLSTM is the most ideal training effect among several models using VMD decomposition.(3)Considering the influence of data features on the long-term and short-term memory network expansion model and retaining more convolutional features,this part designs a two-layer convolutional neural network(CCNN)model composed of CNNs for feature extraction.The model inputs the extracted features into the LSTM model for prediction,and combines the attention mechanism(Attention)to strengthen the processing ability of the CNN extracted features,which further strengthens the learning ability and feature extraction ability of the CCNN-ALSTM prediction model.The experimental results show that the CCNN-ALSTM model further extracts features by fusing the information extracted by the two convolution channels,so that the prediction effect is stable and can learn more meaningful latent variables,and the prediction effect is better than most machine learning.By changing the weights of different features,Attention makes its output more focused on the extracted feature vectors,so that the model has better prediction performance.