近年来，非点源污染已逐渐超越点源污染，成为导致多数河流、湖泊水体水质恶化的主要污染源。非点源污染来源和排放具有间歇性和随机性，如何从流域角度开展非点源污染治理已成为国内外关注的热点和难题。非点源污染优先控制区指流域内对水环境质量有着决定性影响的敏感区域，而最佳管理措施则是目前通用的非点源污染控制方法。如今，国内非点源污染控制的实践工作才刚刚起步，优先控制区识别技术和措施优选技术显得相对薄弱，缺少理论和方法的支撑。对流域尺度而言，污染源的时空分布特征是非点源污染控制的基础。本研究通过在三峡库区大宁河流域开展案例研究，识别并量化了非点源污染模拟过程中的不确定性来源，提出了基于区间距离的模型率定方法，为非点源污染精细化模拟提供了必要参考。优先控制区识别则是流域非点源污染治理的关键。本研究首先提出了基于马尔科夫链的负荷贡献量计算方法，进而建立了基于水环境达标保证率的多级优先控制区的分级技术，以期为流域非点源污染控制提供更有利的视角。措施优化是非点源污染控制的最终落脚点。本研究耦合了流域模型，成本计算模型，专家系统和非支配占优遗传算法，初步构建了最佳管理措施决策支持系统(BMP-DSS)和基于偏好解的措施优化技术体系，最终使流域非点源污染控制得以落到实处。本研究主要结论包括：(1)非点源模拟不确定性分析表明，协同克里格方法和高程信息的考虑有助于生成更精准的降雨数据；高程、土地利用等空间数据的分辨率存在“阈值效应”，对大宁河而言，其最佳分辨率分别为1：5万和1：25万；土壤特征参数是影响径流模拟的关键，而河道模块则对泥沙等污染物的模拟影响较大。在此基础上，本文提出了基于区间距离的模型率定方法(Interval Deviation Approach，IDA)。案例研究表明，相对于传统方法，IDA方法能更精准的识别模拟不确定性的来源、大小，为非点源污染精细化模拟提供必要参考。(2)研究区非点源污染时空分布模拟结果表明，2000年至2008年研究区的总氮输出负荷量有所增加，而总磷负荷量则呈现下降趋势；年内的氮磷污染负荷呈现出典型的双高峰特征。从空间分布来看，氮磷负荷在枯水年的空间变异性最大；在年内的典型枯水期，氮磷的空间分布主要受降雨等自然因素影响，而在丰水期，氮磷的空间异质性则受人为干扰影响较大。(3)流域尺度多级优先控制区划分方法有助于污染源污染负荷削减与水环境达标保证率联系起来。案例研究表明，各空间单元对水环境的负荷贡献量差异显著，其中贡献量大小主要取决于污染源输出量和各空间单元距离评估点的距离。分析表明，降雨量对高等级和低等级优先控制区的空间分布影响不大，但对中间等级的优先控制区的影响较大。同时，不同污染物的多级优先控制区划分结果差异较大，就总磷而言，丰水年的降雨冲刷作用要大于稀释作用，这表明针对某一种污染物制定的流域管理方案对其他污染物则未必有效。(4)基于上下游关系，多评估点的优先控制区分级结果表明，上游评估点的健康状态对下游流域的分级结果影响显著。就研究区而言，当上游评估点水质处于临界健康状态，下游评估点的水质达标率也相应地提高到100%。与之对应，下游亚流域均被划分为五级优先控制区，不需要进一步治理。而当上游评估点水质达标率为70%或80%时，下游评估点水质达标率最大为90%。基于Pareto占优关系，考虑多污染物的优先控制区分级结果表明，本方法对多种污染物输出量均较大的亚流域具有较好的识别能力。就大宁河研究区而言，总氮是划分一级和二级优先控制区需要重点考虑的目标污染物，而总磷则是三级、四级、五级优先控制区划分的基础。不同情景的结果表明本研究构建的多级优先控制区分级技术对于流域内的不同水文年、多评估点、多污染物都是适用的。(5)本研究耦合了流域模型、成本计算模型、专家系统和非支配占优遗传算法，开发了BMP-DSS技术。案例研究结果表明，对研究区而言，不同的污染控制方案可使总磷和总氮污染负荷分别削减49.74%~88.83%和43.74%~93.38%。为实现水环境90%的达标率，应率先考虑管理措施，其中包括养分管理方案、保护性耕作措施和退耕还林政策，而当要求水质达标率进一步提高到100%时，则应在管理措施基础上辅以植被过滤带和滞留池等工程措施。(6)在引入决策者偏好信息后，BMP-DSS的收敛速度加快，最终生成的偏好解大部分位于参考点(决策者期望水平)的投影点附近，从而为决策者提供更多满意的非点源污染控制方案。

In the last years, nonpoint source (NPS) pollution has become the key factor deteriorating the water quality in most rivers and lakes. The control of NPS pollution has been identified as the prior importance in watershed management and received a growing level of concern worldwide. The prior management areas (PMAs) are defined as a criterion for identifying the critical sources of pollutants and also a basis for an optimized watershed management. A watershed manager needs to gain insights into the PMAs, and to optimize the best management practices (BMPs) in their drainage basins. However, there is no science-based framework and formal techniques available. Therefore, we focused on the field of PMAs identification and BMPs selection.In this study, an integrated framework was proposed for identifying the PMAs and optimized BMPs in a typical watershed in the Three Gorges Reservoir Area, China. To illuminate the spatial distributions of NPS pollution, the sources of uncertainty were quantified and an interval deviation approach (IDA) was estabilished in the framework of model calibration and validation. A Markov chain method was introduced to quantify the contributions of each sub-watershed, and the grey probability approach used for multiple classifications of priority management areas. A watershed Decision Support System (DSS), incorporating a watershed model, a cost estimator, an expert system and a non-dominant genetic algorithm, was built up. To incorporate the preference of decision maker, the selection process of genetic algorithm (GA) was revised and a preference-based DSS was established.The main conclusions of this paper are as following:(1) From a practical point of view, a global interpolation method, such as Kriging, as well as elevation data derived from a digital elevation model (DEM), should be included into the watershed models for reliable predictions of rainfall input. A level (threshold) of GIS resolution existed, within which more precise data would not be of any benefit to watershed prediction. To reduce prediction uncertainty, a 1:50,000 DEM and 1:250,000 land use layer should be used. This paper also revealed that the main sources of uncertainty in the flow simulation came from the catchment process while the channel process had a great impact on the sediment simulation. In this study, an IDA was designed and incorporated into likelihood functions with the support of interval theory. Compared with the traditional point estimates of observations and predictions, the IDA incorporated both prediction and measurement uncertainty into the process of model evaluation. In addition, the IDA quantified the possible range of model performances in a real application and differentiated ‘positive’ uncertainty from ‘negative’ uncertainty.(2) For the period from 2000 to 2008, the total nitrogen (TN) while the total phosphorus (TP) decreased. Two high-load points for TN and TP were successfully detected among the year. It was indicated that TN and TP were mainly contributed by the anthropogenic factors in the dry season, while they were primarily exported by the natural factors in the wet season.(3) Compare with the traditional methods, the Markov Chain methods incorporated a synthesized scheme considering two factors: the sources of the NPS pollution and the channel process. An integrated framework for the multiple classifications of the PMAs was indroduced, which considered the relationship between the multi-scale sources and watershed water quality. The case study inferred the amount of rainfall showed little impact on the spatial distributions of the high-class and low-class PMAs among the year. However, the spatial distributions of PMAs areas were different for various pollutants, indicating the control of certain pollutant might be little use for another pollutant.(4) A multiple-evaluation-points method was established based on the relationship between upstream and downstream. The case study showed the water quality in the upstream would have great impact on the re-classification of downstream watershed. A Pareto-based technique was suggested when multiple pollutants existed. The case study showed the TN was the key factor affecting the water quality from 60% goal to 80% goal, while TP was identified as critical pollutant from 80% goal to 100% goal.(5) A DSS was designed and the model simulation, cost estimator, expert system as well as optimization technology were combined. The case study inferred that the optimized schemes resulted in the reduction rates of 49.74%~88.83% for TP, of 43.74%~93.38% for TN while the cost ranged from 0~185.68 million RMB. To reach the 90% goal, the management practices should be considered firstly and while the goal was set up as 100%, filter script and detention ponds should be added.(6) The preference-based GA was introduced and the preference points were integrated into the selection process of GA. The results showed most preference schemes were projected around the reference point, which would be of great benefit to watershed decision maker.