中文摘要：

近年神经网络模型的发展为降雨—径流这一复杂的水文非线性过程的模拟提供了一种新的解决思路.本文基于淮河流域下游区滨海站2010—2012年的降雨及水位日资料,应用带外部输入的非线性自回归神经网络模型（Nonlinear AutoRegressive models with eXogenous input Neural Network,NARXNN）,构建了以降雨为外部输入的淮河下游区降雨—水位关系模拟模型.设计了不同参数组合的正交模拟实验,采用相关系数,均方误差和平均绝对差评判模型的拟合优度,对模型参数进行优选,实验结果表明节点数对模型的拟合优度影响最大,当激励函数为logsig,节点数为7,延时阶数为4,隐含层数为9时,模型模拟效果最优.根据优化的参数组合,利用NARXNN模型对淮河下游区滨海站和长江下游区黄桥站的水位过程进行了模拟,结果表明该模型具有很强的鲁棒性.

英文摘要：

In recent years,rapid developments of Artificial Neural Networks（ANN）offer a novel approach for handling complex non-linear system like the rainfall-runoff relationship.This study built a rainfall-water level relationship model based on daily observation data of rainfall and water-level at Binghai station of Huai River during 2010-2012 year as well as Nonlinear AutoRegressive models with eXogenous input Neural Network（NARXNN）taking rainfall as the exogenous input.Reduced orthogonal test was designed and performed to save calculation time,since full factor experiments are highly time-consuming.The model was then calibrated with multiple goodness-of-fit criteria including correlation coefficient,mean square error and mean absolute error.The results show that model fitting is the most sensitive to node number;the model accuracy is the best when the activation function is logsig,the node number of hidden layers is 7,the input delays is 4and the hidden layer number is 9.Then the rainfall-water level re-lationship was set up in the lower reaches of Huaihe River Basin and Yangtze River Basin,respectively,and the simulated results fitted the measurements well.It suggests that our model is robust in different hydro-meteorological environments.