| The PM2.5 pollution problem has seriously affected people’s daily travel and life health,causing widespread concern in the scientific community.How to use the atmospheric environment monitoring point data to deeply and effectively excavate the time series characteristics of historical PM2.5 concentration data,and timely and accurate early warning of PM2.5 concentration values in the future period has become a study with strong academic significance and application value.problem.However,the current data processing methods of state-controlled monitoring points have insufficient accuracy for PM2.5 concentration prediction,and the time-linear linear characteristics of PM2.5 concentration values in the time dimension and the bidirectional depth of the nonlinear relationship between PM2.5 concentration values and other indicators.Learning effect is not good.To this end,this paper proposes a PM2.5 concentration prediction method based on a dual hidden layer depth LSTM(long-short-term memory neural network)hybrid model based on CEEMD(improved empirical mode decomposition method)and Pearson correlation test screening.The main work of this article includes:(1)Firstly,the Pearson correlation test method is used to screen the spatial correlation of historical data,and the index data with higher correlation value of PM2.5 concentration value at the next moment is selected as effective data,which effectively improves the input index.Resolution.Next,according to the time series,the CEEMD empirical mode decomposition method is used to decompose the various indicators in the longitudinal direction of the multi-modality,and decompose to obtain multiple decomposition waves in different modes.The processed decomposition wave retains the nonlinearity of the original data,and the timing variation of the data is more stable,which is beneficial to further mining the timing characteristics.After the previous step,the traditional CEEMD mode decomposition method will directly remove the high-frequency wave as clutter.In this paper,the modal decomposition wave is filtered twice according to the Pearson correlation analysis,and the correlation with the original sequence is selected.A strong decomposition wave sequence is used as the final neural network input sequence group.After the above enhancement and screening process,the model data is effectively regularized in both time and space directions,which can effectively enhance the prediction accuracy of the neural network and optimize the network convergence speed.(2)Secondly,in the neural network part,a deep implicit LSTM(DLSTM)neural network with double hidden layer with Re LU as the activation function is established in this paper.The input data will first be divided into two parts: training set and test set.The training set will be iteratively trained in batch input network to adjust the model parameters.After repeated training by the Adam optimizer algorithm selected in this model,when the loss function converges,the model training is completed.By inputting the test set data into the trained network,the timing prediction of the PM2.5 concentration value can be performed.(3)Thirdly,according to the combination of CEEMD-Pearson data enhancement filters and LSTM and deep DLSTM models,three models to be tested were established: simple single hidden layer LSTM hybrid model(P-LSTM)based on Pearson correlation screening,and CEEMD modalities.A single hidden layer LSTM model(PCP-LSTM)after decomposition and Pearson secondary screening,and a double hidden layer deep LSTM neural network hybrid model(PCP-DLSTM)based on Pearson correlation analysis and CEEMD modal decomposition,and carried out PM2.5 concentration prediction comparison experiment.In the experiment,the performance evaluation indexes of the three models were compared,and the sensitivity and adaptability of the three models to the PM2.5 concentration level,fluctuation,different time scales,and model structural parameters were compared.Experimental data shows that the MAE and RMSE of the simple P-LSTM hybrid model are generally high,and the value of NSE is only 0.549076,indicating that the model has a large prediction error and low reliability.In addition,the prediction accuracy of the PCP-LSTM model after CEEMD modal decomposition has been greatly improved,the prediction accuracy has reached 87%,and various errors have been reduced.Furthermore,the PCP-DLSTM model provided better prediction results with an accuracy of more than 90% and a NSE value of 0.889593.In addition,in the test of the prediction error convergence speed,the prediction error of the P-LSTM model still converged after about 7000 training times,while the convergence speed of the prediction error of the PCP-LSTM model was slightly lower than the convergence rate of the loss function value.After about 800 trainings,it stabilized.Compared with the convergence speed of the loss value,the prediction error of the PCP-DLSTM model converges faster,and it has completely converged in about 650 trainings.The above shows that the CEEMD-Pearson enhanced filter and the deep DLSTM model proposed in this paper have played a positive role in the timing prediction of PM2.5 concentration.Based on the above hybrid model,the PCP-DLSTM structure not only has a high-precision PM2.5 prediction effect,can also effectively improve the convergence speed of model training,greatly improving learning efficiency.And after further evaluation,the PCP-DLSTM proposed in this paper has strong sensitivity and stability to the time-series fluctuations of PM2.5 concentration,concentration levels,and different time scales.It has a good multi-time-span time-series prediction ability,which can Meet the requirements of reliability testing in a variety of practical scenarios.The double hidden layer depth LSTM hybrid model PCP-DLSTM based on CEEMD modal decomposition and Pearson correlation test proposed in this paper has good performance in training effect,convergence speed,prediction accuracy and prediction time span,which can be effective.Solve problems in timing prediction. |
