摘要
目的 应用季节性差分自回归滑动平均(SARIMA)模型以及与广义回归神经网络的组合模型(SARIMAGRNN)预测广东省登革热的月发病数,比较其预测效果,为登革热的预测预警和防控提供科学依据。方法 该研究使用广东省2004年1月至2012年12月登革热的逐月发病资料,分别构建两种模型,并使用2013年1月至12月的数据对模型进行预测验证。结果 登革热疫情呈现明显的周期性和季节性,周期为1年,8~10月份为高发期,在爆发年份发病人数急剧增多。SARIMA(1,1,3)(1,1,0)12模型为SARIMA预测部分的最优模型;神经网络的最优光滑因子为0.04。两种模型对2013年疫情预测的均方根误差(RMSE)为SARIMA(105.76)〉SARIMA-GRNN(92.77),平均绝对百分比误差(MAPE)为SARIMA(2.78)〉SARIMA-GRNN(2.15),平均绝对误差(MAE)为SARIMA(64.75)〉GRNN-ARIMA(58),模型的决定系数(R2)为SARIMA(0.92)〈SARIMA-GRNN(0.95)。结论 两种方法均有较佳的预测效果。在SARIMA模型的基础上,结合GRNN模型可进一步提高预测精度。
Objective To apply Seasonal Autoregressive Integrated Moving Average (SARIMA)model and SARIMAGRNN hybrid model to forecast monthly number of Dengue Fever, and compare the prediction performance of these two models. Methods Based on data of monthly number of Dengue Fever from January 2004 to December 2012 in Guangdong Province, we constructed the SARIMA (p, d, q) ( P, D, Q) s model and SARIMA-GRNN hybrid model, and data from January to December in 2013 were used to assess the predictive validity of models. Results The incidence of Dengue Fever is characterized by an apparent cyclic pattern with a one-year seasonal cycle, with a peak occurring during August to October. The epidemic strength and peak differed by years. In SARIMA section, SARIMA ( 1,1,3 ) ( 1,1,0 ) 12 model is the optimal model. The optimal spread of GRNN model is 0. 04. The root mean square error (RMSE)of these two models was SARIMA (105.76) 〉 SARIMA-GRNN (92. 77 ) ; the mean absolute percent error (MAPE) was SARIMA ( 2.78 ) 〉 SARIMA-GRNN ( 2. 15 ) ; the mean absolute error (MAE) of the two models was SARIMA ( 64. 75 ) 〉 SARIMA-GRNN ( 58 ) ; the determination coefficient ( R2 ) was SARIMA (0. 92) 〈 SARIMA-GRNN ( 0. 95 ). Conclusion Both of the two models had satisfactory prediction capacity. Relatively, the SARIMA-GRNN hybrid model is the optimal model to predict the incidence of Dengue Fever.
出处
《中国卫生统计》
CSCD
北大核心
2016年第5期746-748,751,共4页
Chinese Journal of Health Statistics
基金
广东省科技计划项目(2013B021800041)
国家自然科学基金项目(81573249)
广东省自然科学基金(2016A030313530)