基于GRNN的变权重组合预测模型在传染病发病率预测中的应用

  本文关键词：基于GRNN的变权重组合预测模型在传染病发病率预测中的应用，，由笔耕文化传播整理发布。

基于GRNN的变权重组合预测模型在传染病发病率预测中的应用

基于GRNN的变权重组合预测模型在传染病发病率预测中的应用

影响传染病发生发展的因素众多且相互关系复杂,预测传染病发生发展的模型也是种类繁多。目前对传染病的预测策略可分为两类：一类是线性回归预测、时间序列浅析浅析、灰色模型、人工神经网络等单纯预测模型,另一类是通过将两种或多种预测模型以一定策略组合得到的组合预测模型,组合预测模型可以分成定权重组合预测模型和变权重组合预测模型两种。研究目的运用灰色模型(Grey Model,简记为GM)、差分自回归移动平均模型(Autoregressive Integrated Moving Average Model,简记为ARIMA模型)为基础,构建基于广义回归神经网络(Generalized Regression Neural Network,简记为GRNN)的变权重组合预测模型,分别拟合传染病发病率的情况,对基于GRNN的变权重组合预测模型的拟合效果进行评价,提出模型的优越性和不足,为变权重组合预测模型的研究提供依据。资料和策略本研究以浙中某市1998-2008年的肺结核发病率为研究资料,分别在matlab7.11.0软件和SAS9.2软件中构建了灰色模型和ARIMA模型,通过预测2009年的发病率来比较具体模型的精度,并以这两种模型为基础,在matlab软件中构建了基于GRNN预测算法的变权重组合预测模型,以简单平均组合预测模型、加权平均组合预测模型为对照,来评价变权重组合预测模型在预测中的精度。主要结果1.灰色模型利用浙中某市1998-2008年肺结核月发病率数据构建了GM(1,1)模型和残差修正GM(1,1)模型,并对模型进行评估,发现GM(1,1)模型和残差修正GM(1,1)模型的后验差比值C分别是0.7687和0.6187.结果显示残差修正GM(1,1)模型各项指标较小,认为残差修正GM(1,1)模型的拟合效果优于GM(1,1)模型。2.ARIMA模型利用浙中某市1998-2008年肺结核月发病率数据构建了ARIMA(1,0,0)模型和ARIMA(1,0,1)*(1,1,0)12模型,两个模型的残差值白噪声检验显示：在延迟12阶后,ARIMA(1,0,1)*(1,1,0)12模型有统计学意 义,而ARIMA(1,0,0)模型则无统计学意 义,且前者的AIC值为627.6154,SBC值为630.4982；后者的AIC值为587.4054,SBC值为595.7679,认为后者的拟合效果优于前者。故选择ARIMA(1,0,1)*(1,1,0)12模型来建立组合预测模型。3.组合预测模型以灰色模型和ARIMA模型为基础,构建了基于GRNN的组合预测模型,将此模型和灰色模型、ARIMA模型、简单平均组合预测模型和加权平均组合预测模型比较,发现残差修正GM(1,1)模型的MSE=37.451,MAE=5.692, MAPE=53.69％,MER=48.51％；ARIMA(1,0,1).(1,1,0)12模型的MSE=18.509,MAE=3.761,MAPE=35.13％,MER=32.05％；简单平均组合预测模型的MSE=28.984,MAE=4.736,MAPE=45.4％,MER=40.4％；加权平均组合预测模型的MSE=24.649,MAE=4.274,MAPE=41.0％,MER=36.4％；基于GRNN的组合预测模型的MSE=9.961,MAE=2.571,MAPE=25.6％,MER=21.9％；各项评价指标都满足：基于GRNN的组杏预测模型

【Abstract】 The numerous influence Factors of infectious diseases has complex relationships with each others, so there are various kinds of forecasting model for infectious diseases. There are two kinds of forecasting model:one kind of them are called simple prediction model such as linear regression model, time series models, the gray models, artificial neural network models and so on; the others are called combination forecast model which are comprised of more than one kind of simple prediction models with different ways. The combination forecast model can be divided into fixed weight combination forecasting model and the variable weight combination forecasting model.ObjectiveA forecasting model with variable weight combination based on GRNN was built with the time series models and the gray models for for Infectious Diseases. To provide evidence for variable weight combination, these forecasting models was used to predict the incidence rate of the diseases and evaluated to study the advantages and the weakness of them. Data and MethodsThe monthly incidence rate of tuberculosis between 1998 and 2008 were collected from the Center for Disease Control and Prevention in yiwu.With these data, the grey models and the time series models were built with the software called matlab 7.11.0 and SAS 9.2, respectively. Then these models were used to predict he monthly incidence rate of tuberculosis and evaluated which one is better.A forecasting model with variable weight combination based on GRNN for infectious diseases was made up of the grey models and the time series models in the matlab7.11.0. A simple combination forecasting model and a weighted average combined forecasting model were made up of the grey models and the time series models, respectively. The estimation accuracy of the forecasting model based on GRNN was evaluated by comparing with the other two combination forecasting models.Results1. The grey modelsA GM(1,1) model and a residual-modifying GM(1,1) model were built with the monthly incidence rate of tuberculosis between 1998 and 2008 of yiwu. The value of posterior-variance-test of the GM(1,1) model and the residual-modifying GM(1,1) model was 0.7687,0.4140,respectively. The evaluation index of the residual-modifying GM(1,1) model were smaller than the evaluation index of the GM(1,1) model, so the former have better estimation accuracy.2. The ARIMA modelsAn ARIMA(1,0,0) model and an ARIMA(1,O,1)* (1,1,0)12 model were built with the monthly incidence rate of tuberculosis between 1998 and 2008 of yiwu. The white noise test shows:the residual value of the ARIMA(1,0,1)* (1,1,0)12 model was a white noise sequence after 12 lags, but the residual value of the ARIMA(1,0,0) model was not a white noise sequence after 12 lags. The AIC value and the SBC value of former model was 587.4054,595.7679, The AIC value and the SBC value of later model was 587.4054,595.7679, respectively. The evaluation index of the ARIMA(1,0,1)* (1,1,0)12 model were smaller than the ARIMA(1,0,0) model, so the former have better estimation accuracy.3. The combination forecasting modelA forecasting model with variable weight combination based on GRNN for monthly incidence rate of tuberculosis was made up of the grey models and the ARIMA models. A simple combination forecasting model and a weighted average combined forecasting model were made up of the two simple models. The combination forecasting model based on GRNN was compared with the other four models by comparing the MSE value, the MAE value, the MAPE value and the MER value, the results were as follows:the four values of the residual-modifying GM(1,1) model was 37.451,5.692,53.69%,48.51%, respectively; the four values of the ARIMA(1,0,1)* (1,1,0)12 model was 18.509,3.761,35.13%,32.05%, respectively; the four values of the simple combination forecasting model was 28.984,4.736,45.4%,40.4%, the four values of the weighted average combined forecasting model was 24.649,4.274,41.0%, 36.4%, respectively; the four values of the combination forecasting model based on GRNN was 9.961,2.571,25.6%,21.9%, respectively. The evaluation index of the models showed that the values of the five models are the combination forecasting model based on GRNN< the ARIMA(1,0,1)*(1,1,0)12 model< the weighted average combined forecasting model

  本文关键词：基于GRNN的变权重组合预测模型在传染病发病率预测中的应用，由笔耕文化传播整理发布。