ARIMA-BPNN模型在结核病发病率中的预测研究  

Prediction study of ARIMA-BPNN model in tuberculosis incidence rate

作　　者：许弘佳 崇英之 刘媛媛[1] 曹娃 许达 张佳伟 赵俊鹏[1,3] XU Hongjia;CHONG Yingzhi;LIU Yuanyuan;CAO Wa;XU Da;ZHANG Jiawei;ZHAO Junpeng(Tangshan Center for Disease Control and Prevention,Tangshan,Hebei 063000,China;不详)

机构地区：[1]唐山市疾病预防控制中心,河北唐山063000 [2]华北理工大学公共卫生学院 [3]唐山市食品药品综合检验检测中心

出　　处：《医学动物防制》2024年第10期970-973,978,共5页Journal of Medical Pest Control

基　　金：2022年度河北省医学科学研究重点课题计划(20221806)。

摘　　要：目的探讨差分自回归移动平均(autoregressive integrated moving average,ARIMA)-BP神经网络(back-propagation neural network,BPNN)(ARIMA-BPNN)模型预测结核病发病率的准确性,为结核病发病率的短期预测提供依据。方法以河北省结核病发病率数据为例,2005年1月—2016年12月结核病发病率数据作为训练集,2017年1月—2018年12月结核病发病率数据作为验证集。赤池信息量准则(Akaike information criterion,AIC)、贝叶斯信息准则(Bayesian information criterion,BIC)、ljung-box检验以及过度拟合检验确定最优ARIMA模型,继而构建ARIMA-BPNN模型。采用均方根误差(root mean square error,RMSE)、平均绝对误差(mean absolute error,MAE)和平均绝对百分比误差(mean absolute percentage error,MAPE)评价模型预测精度。统计分析软件为R 3.4.0,检验水准α=0.05。结果ARIMA(1,1,2)(0,1,1)_(12)模型中AIC=153.32、BIC=167.70、χ^(2)=3.38,P=0.760,各参数估计结果比较,差异均有统计学意义(t=-2.86、-1.93、-6.21、-1.75,均P<0.05),在训练集中拟合度最优。在此模型的基础上,取RMSE最小值(RMSE=0.06)的训练结果作为参数取值(size=2,decay=0.01),建立ARIMA-BPNN模型。验证集中进行预测精度的比较,ARIMA模型的RMSE、MAE和MAPE分别为0.94%、0.24%和6.52%,ARIMA-BPNN模型的RMSE、MAE和MAPE分别为0.10%、0.02%和0.55%。结论相比ARIMA模型,ARIMA-BPNN模型可提供更高的预测精度,在结核病短期发病率预测中有更好的预测效果。Objective To investigate the accuracy of autoregressive integrated moving average(ARIMA)-back-propagation neural network(BPNN)(ARMI-BPNN)model in predicting the incidence of tuberculosis and to provide a basis for short-term prediction of tuberculosis incidence.Methods Taking the data of tuberculosis incidence rate in Hebei Province as an example,the data of tuberculosis incidence rate from January 2005 to December 2016 were used as the training set,and the data of tuberculosis incidence rate from January 2017 to December 2018 were used as the validation set.Akaike information criterion(AIC),Bayesian information criterion(BIC),ljung-box test,and overfitting test were used to determine the optimal ARIMA model,and then the ARIMA-BPNN model was contructed.The root mean square error(RMSE),mean absolute error(MAE)and mean absolute percentage error(MAPE)were used to evaluate the accuracy of model prediction.The statistical analysis software was R 3.4.0 with a test level ofα=0.05.Results In the ARIMA(1,1,2)(0,1,1)_(12)model,AIC=153.32,BIC=167.70,χ^(2)=3.38,P=0.760.The differences were statistically significant between comparison of the estimation results of each parameter(t=-2.86,-1.93,-6.21,-1.75,all P<0.05),and the fit degree was optimal in the training set.Based on this model,the training result with the minimum value of RMSE(RMSE=0.06)was taken as the parameter values(size=2,decay=0.01)to build the ARIMA-BPNN model.Comparison of prediction accuracy was performed in the validation set,and the RMSE,MAE,and MAPE of the ARIMA model were 0.94%,0.24%,and 6.52%,respectively,while the RMSE,MAE and MAPE of the ARIMA-BPNN model were 0.10%,0.02%,and 0.55%,respectively.Conclusion Compared with the ARIMA model,the ARIMA-BPNN model provids higher prediction accuracy and has better prediction in the prediction of short-term incidence of tuberculosis.

关 键 词：结核病 ARIMA-BPNN ARIMA 时间序列 模型预测 传染病预测 研究 

分 类 号：R725.1[医药卫生—儿科]