【作者】 王玲；

【导师】 朱元生；

【作者基本信息】 河海大学 ， 水文学及水资源， 2002， 博士

【摘要】 水文过程是一个高度复杂、非线性的过程。为实现水资源管理的自动化，必须解决流域径流过程的预报问题，以往普遍受到关注的是分布式物理模型和概念性模型。实际上，分布式模型的实现非常之难，且要求高深的数学工具。由于人类认知水平所限，这种模型常不得不采用简化或假定的形式表述物理规律，同时要求具备一定的经验和知识技能以及大量的观测数据。现有的一些分布式物理模型，对径流形成的理论研究有所贡献，而在实用上，一般同样使用少量的参数来代表流域特性，而且这些参数需要根据观测资料，通过经验途径率定，从而忽略了降雨——径流过程的空间分布，时变特性和随机性，难以真正表达空间分布信息，而沦为参数化准分布式的模型。集总的概念性模型从数据需求、计算时间及模型结构而言比分布式模型有优势，但由于这种模型的率定也需大量时间，而且要求模型使用者必须具备一定的经验和知识技能；且由于无法考虑系统的非均质、非线性，而且这种模型本质上是用率定后的少量参数来代表流域特性，最终得到的模型参数常因率定过程中，为平衡各部分实测值的离差而失去其原有的物理意义，也即无法解释模型率定出来的参数。从而固守传统的水文规律，在解决问题之前先确定一个模式，把思路局限在以往的经验上，在很多时候是行不通的。 由于计算机软、硬件的日益成熟，功能迅猛增强，许多凭借于计算机的高速计算功能而得以发展的数据挖掘技术，如人工神经网络(ANN)、遗传算法、模糊逻辑、决策树等，在水资源管理上的应用受到了广大的科学工作者和工程师们的亲睐。近年来，ANN在水资源管理上已有一些应用。本文将进一步论证ANN在雨洪径流单次模拟和连续模拟上的适用条件，和改进的途径。 单次模拟将选用伦敦北部二个渐进城市化地区的若干场半小时间隔的单次雨洪资料。在考虑和不考虑PIA的情况下进行数值试验，探讨使用人工神经网络模型模拟城市化地区的降雨——径流关系。建立城市化地区的降雨——径流关系模型，并把反映城市化进程的不透水面积比例这一重要参数加入输入模式中，用人工智能手段研究城市化对降雨径流关系的影响。大量的数值实验证明，当神经网络被适当地配置以后，可以再现潜在的降雨——径流关系，用来生成精确且符合实际的预测结果。不透水面积比这一城市化参数对于城市化地区径流过程的预测非常重要，但由于数据本身的原因(使用年平均值)，对径流模拟的精度提高帮助不大。此外本文探讨了如何解决外延问题及资料的移用问题。快速、精确的径流预测可以对城市雨洪实时控制管理提供快速有效的支持。 连续模拟则收集了淮河上游息县段21年的资料，在建立了单一构型的ANN模型后，把径流过程分成4个组，分别建立各组的局域ANN模型，即在对径流过程进行分类的基础上对各个径流子过程用前向神经网络分别建立非线性模拟模型，选出各自最优的网络配型，即确定各自的最优网络结构，而后把这些模型合并成一个混合模型，来进行连续降雨——径流预报，以期提高网络的性能，这种修正后的神经网络将与优选出的最佳配置的单一构型人工神经网络模型的性能加以对比。 通过这两种应用(城市化地区单次雨洪径流模拟，大流域径流连续模拟)，可以看出人工神经网络在径流模拟应用上表现出较高的精度，完全适用于水文预报工作。更多还原

【Abstract】 Hydrological process is a highly complex and non-linear process. Before artificial intelligence extensively applied in hydrological problems, the physical-based distributed and conceptual models are utilised everywhere. The distributed physical model adapts the different patterns of one- or two-dimension Santa Venent Equations, and try to represent precisely every sub-process of hydrological cycle. Due to the limitation of man’s intelligence, this kinds of models have to use simplified way to represent physical rules. And in reality, the implementation of distributed models is very difficult, and complex mathematical tools needed. The users of such models also need some experience and knowledge, amount of data needed to calibrate the models. After the calibration, using less parameters to represent the catchment characteristics, the models omit the spatial distribution, temporal features and stochastically of runoff process, then the models become the semi-lumped models.The lumped conceptual models have advantages compared to distributed models from the data requirement, calculation time and model structure point of view. But the calibration of such models also need amount of time, and the model users should own some experience and knowledge. Such kind of models can not consider heterogeneous and nonlinealities of the system, and essentially, this kind models use few calibrated parameters to represent the characteristics of the catchment. It is difficult to interpret the calibrated parameters for during calibration, in order to balance the observed data, the parameters always lost their physical meaning.Artificial Neural Networks - ANNs is kind of artificial intelligence tool which is widely applied in the fields of time series analysis, pattern identification etc. The possibility of using ANNs models to simulate the rainfall-runoff relationship of urban areas will be explored first. In this application, ANN model will be set up to serve as the rainfall-runoff models of urban areas, then the important parameter Percentage of Imperious Areas - PIA, which represent the process of urbanisation will be add to the input patterns to find the effect to the relationship between rainfall and runoff of urbanisation. Amount of numerical experience illustrated that after proper setting-up of neural network models, the underlying rainfall-runoff relationship will be reproduced to generate precisely and pragmatically forecasting results and then offer quick and effective support to urban storm water real time control.For continues runoff simulation, clustering will be performed upon the runoff sequence and then several local feed-forward neural networks will be formulated for each class. When new data fed into the modified ANN model, a classifier will direct the new data into different non-linear local ANN model. The performance comparison with that of the singular ANN illustrated that the classifier based local ANN rainfall-runoff model had the superior performance.更多还原