感染性腹泻周发病例数的 PCA-SVM 回归预测研究

摘要

We proposed a regressive prediction method for the weekly cases number of infectious diarrhea using PCA-SVM,which effectively avoids some defects of the BP neural network model like local extremum,multicollinearity.With the weekly cases of infectious diarrhea in Shanghai from the year 2005 to 2008 being the samples,we built the PCA-SVM regressive model.First,we employed PCA to extract meteorological main principal factors from the statistical meteorological factors and removed the multicollinearity from the predictive factors,derived the explanatory variable of the final model.Secondly,we used SVMregression to build the predictive model for weekly cases number of infectious diarrhea in Shanghai.To illustrate the better prediction effect of the model,we compared it with BP neural network model in terms of fitting and prediction results.Numerical results showed that the MAPE and RMSE (0.2694 and 33.113 respectively) predicted by PCA-SVMregression model were all less than those of BP neural network model (0.3745 and 49.909 respectively).Meanwhile, its determination parameter R2 (0.9089)was further approaching 1 than that of BP neural network (0.8590).As a result,it is demonstrated in this paper that the PCA-SVM regressive model has higher prediction accuracy and stronger generalisation capability in predicting weekly cases number of infectious diarrhea,the prediction of the model is reliable on the weekly cases number of the disease,and has better practical value in publicising the diarrhea prediction.%提出一个使用 PCA-SVM进行感染性腹泻周发病例数回归预测方法，有效避免了 BP 神经网络模型存在局部极值、多重共线性的问题。以上海市2005年至2008年感染性腹泻周发病例数为样本，建立 PCA-SVM回归模型。首先用 PCA 从统计气象因子中提取气象主成分因子，去除预报因子多重共线性，得到最终模型的解释变量，其次采用 SVM方法构建上海市感染性腹泻周发病例数预测模型。为了说明该模型有更佳的预测效果，与 BP 神经网络模型比较拟合及预测结果。数据结果显示 PCA-SVM回归模型预测的平均相对误差 MAPE、均方误差平方根 RMSE （数值分别为0．2694，33．113）均小于 BP 神经网络（数值分别为0．3745，49．909），而决定系数 R2（数值为0．9089）较 BP 神经网络（数值为0．8590）更趋近于1。证明 PCA-SVM回归模型在感染性腹泻周发病例数预测中具有较高的预测精度和较强的泛化能力，模型对于感染性腹泻周发病例数的预测可靠，对于向公众发布腹泻预报有更好的实用价值。