Study On The Prediction Effect Of LSTM/BILSTM-ARMA Model Based On Signal Decomposition For Influenza In Shanxi Province

Objective:This paper analyzes the time series characteristics of influenza surveillance data in Shanxi Province from the 14th week of 2010 to the 13th week of 2017.Based on the analysis results of time series characteristics,a combined forecasting model based on decomposition is established,which is compared with the single forecasting model without decomposition and the single forecasting model based on decomposition to evaluate the forecasting performance of each model.Finally,the optimal forecasting model is selected to predict the incidence of influenza in Shanxi Province It can provide an effective scientific basis for the high-precision prediction of influenza in Shanxi Province.Methods:We collected the weekly influenza surveillance data in Shanxi Province,analyzed its development trend and seasonal characteristics,and studied the characteristics of the data,such as stationarity,randomness,nonlinearity and long-term memory.Firstly,ARIMA,LSTM and BILSTM models were established based on the time series characteristics of influenza sequences.Secondly,Based on the complexity of influenza sequences,SSA,EMD and WT were used to decompose the influenza sequences in Shanxi Province.By randomly combining SSA,EMD and WT with the above models,a single model based on decomposition was established to verify whether the sequential decomposition method can improve the prediction performance.Finally,ARMA(stationary subsequence)and LSTM/BILSTM(non-stationary subsequence)were selected to construct the combination model according to the different stationarity of subsequence;MAE,MSE and MAPE were used to evaluate the prediction performance of each model,which provided a certain scientific basis for the high-precision prediction of influenza in Shanxi Province.Results:1.The average development rate of influenza cases in Shanxi Province from the14th week of 2010 to the 13th week of 2017 was 103.33%,and the average growth rate was 3,33%;however,the ILI%showed a downward trend year by year,with obvious seasonal characteristics,and the peak was from August 8 to March 16 of the following year.Time series analysis showed that:ADF test was t=-3.6371,P<0.05;KPSS test statistic was x~2=0.8067,P=0.010;Ljung box test statistic was x~2=287.5732,P<0.0001;BDS test rejects the original hypothesis under different embedding dimension and distance judgment parameters,and the Hurst index obtained by R/S method was 0.8545.2.Comparing the prediction errors of ARIMA,LSTM and BILSTM models,it was found that:Compared with LSTM model,the MSE,MAE and RMSE of BILSTM model decreased by 5.8%,4.4%and 3.0%respectively;compared with ARIMA model,the MSE,MAE and RMSE of BILSTM decreased by 80.6%,60.1%and 56.0%respectively.Compared with ARIMA model,the prediction performance of LSTM model is improved by 79.5%,59.2%and 54.7%respectively.3.Comparing the prediction errors of single model based on decomposition and basic model,it was found that:Compared with ARIMA model,The MSE,MAE and RMSE of SSA-ARIMA,EMD-ARIMA and WT-ARIMA were reduced by 42.7%,84.5%,28.6%;28.3%,65.2%,13.7%;24.2%、60.7%、15.5%respectively;Compared with LSTM model The MSE,MAE and RMSE of SSA-LSTM,EMD-LSTM and WT-LSTM were reduced by 34.5%,39.8%,14.1%;19.0%,23.0%,4.5%;19.0%,22.3%,7.4%respectively;Compared with LSTM model,The MSE,Mae and RMSE of SSA-BILSTM,EMD-BILSTM and WT-BILSTM were 26.1%,38.9%,23.9%;12.5%,22.8%,7.0%;13.9%,21.8%,13.0%respectively.4.Comparing the prediction errors of combinatorial and single model based on decomposition,it was found that:Compared with single model based on decomposition model,the improvement rate of the combination model based on decomposition was positive;and in the six combination models,the MSE,MAE and RMSE values of SSA/EMD/WT-BILSTM-ARMA were lower than the corresponding SSA/EMD/WT-BILSTM-ARMA,which are 0.0108,0.0812,0.1038;0.0108,0.0772,0.1037;0.0137,0.0922,0.1169;respectively.Conclusion:1.Influenza in Shanxi Province showed obvious seasonal periodicity,which is non-linear,non-stationary and long-term memory.2.For non-stationary and non-linear sequences,the prediction performance of LSTM and BILSTM model were better than that of ARIMA model,and BILSTM was better than LSTM.3.The prediction performance of single model based on SSA/EMD/WT decomposition was better than that of basic single model.When the same model was used for prediction,EMD was the best,SSA was the second,and WT was the worst.4.The performance of the combined model which makes use of the advantages of different models was better than that of single model based on decomposition.And EMD-BILSTM-ARMA model was the best.