中文摘要：

目的探讨广义回归神经网络（GRNN）组合预测模型在肾综合征出血热（HFRS）发病率预测上的优势及应用前景。方法利用1990—2001年辽宁省、丹东市、沈阳市和朝阳市HFRS发病率分别建立GM（1，1）灰色预测模型和求和自回归滑动平均（ARIMA）模型，把2个模型的预测值作为GRNN的输入，实测值作为网络的输出，对样本进行训练和预测，并对3个模型的预测效果进行比较。结果针对辽宁省HFRS发病率建立的GM（1，1）模型、ARIMA模型和GRNN组合预测模型的平均误差率（MER）分别为13．5143％、25．0814％和5．5755％；R^2分别为0．8961、0．6997和0．9837。针对丹东市HFRS发病率建立模型的MER分别为19．7329％、20．6275％和14．0789％；R^2分别为0．8112、0．7628和0．8750。针对沈阳市HFRS发病率建立模型的MER分别为15．1421％、18．0584％和14．3592％；R。分别为0．8757、0．7889和0．8585。针对朝阳市HFRS发病率建立模型的MER分别为51．5090％、28．6593％和28．5927％；R^2分别为0．7863、0．8291和0．7753。GRNN组合预测模型对于辽宁省和丹东市的HFRS发病率预测效果好于2个单一模型；针对沈阳市所建立的HFRS发病率预测模型，GRNN组合预测模型和GM（1，1）模型相当，ARIMA模型最差。朝阳市的HFRS发病率预测模型不适合用上述方法建立。结论GRNN组合预测模型充分体现了它在小样本预测中的优势，预测效果优于GM（1，1）模型和ARIMA模型，对解决时间序列类型的HFRS发病率等资料有很好的实用价值。

英文摘要：

Objective To study the superiority and application of generalized regression neural network （GRNN） combination forecast model in the forecast of hemorrhagic fever with renal syndrome（HFRS） incidence. Methods Establish the GM（1,1） model and auto regressive integrated moving average（ARIMA） model based on the data of HFRS of Dandong, Shenyang and Chaoyang Liaoning province, from 1990 to 2001 respectively. The forecasting values of the two models were used as input of GRNN. Train the sample and forecast the value. Compare the forecasting effect of the three models. Results The mean error rate（MER） of GM（1,1） model, ARIMA model and GRNN combination model for Liaoning province were 13.5143%, 25.0814% and 5.5755% respectively. The R2 of the three models were 0.8961, 0.6997 and 0.9837 respectively. The MER of GM （1,1） model, ARIMA model and GRNN combination model for Dandong were 19. 7329 %, 20. 6275 % and 14. 0789 % respectively. The RE values of three models were 0. 8112, 0. 7628 and 0. 8750 respectively. The MER of GM（1,1） model, ARIMA model and GRNN combination model for Shenyang were 15. 1421% , 18. 0584% and 14.3592% respectively. The RE values of three models were 0.8757, 0.7889 and 0.8585 respectively. The MER of GM （1,1） model, ARIMA model and GRNN combination model for Chaoyang were 51.5090%, 28.6593% and 28.5927% respectively. The R2 values of three models were 0.7863, 0.8291 and 0.7753 respectively. The forecasting efficacy of combination model for Liaoning province was better than other two single models. For the forecasting efficacy of Shenyang, the GRNN combination model and the GM（1,1） model were similar, and the ARIMA model was the worst. The incidence of HFRS for Chaoyang is not fit for the establishment of the models we mentioned above. Conclusion GRNN combination model had more advantage in the forecast of small sample and the forecasting efficacy was better than GM（1,1） model and ARIMA model, which had practical value in the treatment of time series data su