差分自回归移动平均模型与Elman神经网络及其组合模型对北京市肺结核发病预测效果的比较  

Comparison of ARIMA model and Elman neural network along with ARIMA-Elman combination model in predicting incidence of tuberculosis in Beijing

作　　者：闫银锁 孙闪华[1] 李亚敏[1] 李艳圆 赵鑫[1] 陶荔莹[1] 高志东[1] YAN Yin-suo;SUN Shan-hua;LI Ya-min;LI Yan-yuan;ZHAO Xin;TAO Li-ying;GAO Zhi-dong(Department of Prevention and Control,Beijing Research Institute for Tuberculosis Control,Beijing 100035,China)

机构地区：[1]北京结核病控制研究所防控科,100035

出　　处：《中国防痨杂志》2019年第6期669-675,共7页Chinese Journal of Antituberculosis

摘　　要：目的比较差分自回归移动平均(autoregressive integrated moving average,ARIMA)模型、Elman神经网络及其组合(ARIMA-Elman)模型对北京市肺结核发病趋势的预测效果,探讨最佳预测模型。方法以2010—2017年北京市肺结核月报告发病例数为数据基础,分别采用ARIMA模型、Elman神经网络及ARIMA-Elman组合模型,预测2018年12个月肺结核报告发病例数,以2018年的实际月报告发病例数验证3种模型的预测效果,评价指标使用平均绝对误差和平均绝对百分误差。结果 ARIMA模型、Elman神经网络和ARIMA-Elman组合模型对北京市肺结核月发病例数的预测相对误差多在±10%以内,分别为8个、8个和9个;此外,ARIMA模型预测结果相对误差在±10%~±20%的有3个,超过±20%的有1个;Elman神经网络预测结果相对误差在±10%~±20%的有2个,超过±20%的有2个;ARIMA-Elman组合模型预测结果相对误差在±10%~±20%的有3个。ARIMA模型、Elman神经网络及ARIMA-Elman组合模型的平均绝对误差分别为44.7(536/12)、47.8(574/12)和43.8(526/12),3种模型的平均绝对百分误差分别为8.7%(1.039/12×100%)、8.2%(0.99/12×100%)和7.9%(0.953/12×100%),ARIMA-Elman组合模型的2个预测评价指标均小于单一ARIMA模型和Elman神经网络。结论 ARIMA-Elman组合模型预测精度更高,对北京地区肺结核发病情况有更加理想的预测效果。Objective To compare the predictive effects of Autoregressive Integrated Moving Average (ARIMA) model, Elman neural network model and their combination (ARIMA-Elman) model on the incidence of pulmonary tuberculosis in Beijing, and explore the best predictive model.Methods The data between 2010 to 2017 was used as training data to establish three models (ARIMA, Elman, ARIMA-Elman). Data of 2018 was used as testing data to evaluate the predictive effects of three models. The Mean Absolute Error and the Mean Absolute Percentage Errors were two predictive efficiency indicators.Results The Relative Errors of ARIMA model, Elman neural network and ARIMA-Elman model were mostly within ±10% in predicting the monthly incidence of tuberculosis in Beijing, there were 8, 8 and 9, respectively. Except for the parts within ±10%, the number of Relative Errors within ±(10%-20%) for ARIMA model was 3 and 1 over ±20%;the number of Relative Errors within ±(10%-20%) for Elman neural network was 2 and 2 over ±20%;while the number of ARIMA-Elman combination model was 3 within ±(10%-20%). The mean Absolute Errors of the three models above were 44.7 (536/12), 47.8 (574/12) and 43.8 (526/12), while the Mean Absolute Percentage Errors were 8.7%(1.039/12×100%), 8.2%(0.99/12×100%) and 7.9%(0.953/12×100%) respectively. Both the two indicators of ARIMA-Elman model were smaller than the other two models.Conclusion ARIMA-Elman combination model has higher prediction accuracy and more desirable effect on the prediction of the incidence of tuberculosis in Beijing.

关 键 词：模型 理论 结核 肺 预测 对比研究 

分 类 号：R521[医药卫生—内科学] R181.8[医药卫生—临床医学]