Prediction Of Haze Concentration Based On Recurrent Neural Network

With the development of social economy,haze pollution becomes more and more serious,which has a serious impact on people's lives.PM_2.5,as the main body of haze concentration data,has been widely concerned by academia and industry.At present,a large number of PM_2.5prediction models have been proposed.However,the diversity of sources of PM_2.5brings challenges to accurately predict its concentration.Moreover,the haze pollution data has the characteristics of linearization and differentiation,and ordinary prediction methods can not take these characteristics into account,which also brings challenges to haze prediction.Therefore,from the perspective of haze characteristics,this paper establishes a reliable prediction network by looking for the correlation between each feature and the difference of haze data at different times.In a word,the main contributions of this paper are as follows:1.In view of the high correlation characteristics of haze concentration data,this paper proposes a CD-BGRU network(Convent Dense Bidirectional Gate Recurrent Unit)for haze concentration prediction,which improves the accuracy of haze concentration prediction.Firstly,CD-BGRU uses Pearson correlation coefficient to calculate the correlation coefficient of the original haze concentration data,and selects some characteristics most related to PM_2.5concentration,which reduces the impact of redundant information in the original data on the prediction accuracy.Secondly,the network uses the powerful feature extraction ability of 1D convolution network to further process the filtered features,which not only improves the prediction accuracy of the model,but also shortens the calculation time to a certain extent.Finally,this paper uses the forward and backward prediction ability of bidirectional Gru to design and improve the network.Starting from feature processing,CD-BGRU makes full use of the linear correlation between features,so as to achieve more accurate prediction of PM_2.5.Compared with the prediction accuracy of other networks,experiments show that CD-BGRU has certain prediction accuracy and stability in most scenes.2.In view of the complex characteristics of haze data sequence,this paper proposes TSLSTM network(Trend Supervised Long Short Time Memory)for haze concentration prediction under trend supervision.In TSLSTM,the loss function is optimized sufficiently.In the past,the loss was only limited to the prediction of accuracy,ignoring the impact of trend changes.TSLSTM pays more attention to the change of data trend,and uses the monitored trend to counter monitor the predicted data.Not only that,this paper also uses the idea of ARIMA to differentiate the original data,so as to get more accurate trend changes.Then,TSLSTM improves LSTM unit to make it transparent and easier to monitor data flow.At the same time,TSLSTM changes the input of LSTM,so that LSTM can consciously predict the trend of training data in the process of training,so as to complete more accurate prediction.Experiments show that TSLSTM has good performance in different time intervals and different areas of haze data prediction,which proves the effectiveness and stability of the network.Combined with the above main contributions,this paper will improve the current prediction algorithm from the two directions of network improvement and trend supervision to improve the prediction accuracy,which has a certain practical and theoretical value in the field of air quality prediction.