密云水库作为北京市唯一地表饮用水水源地，其库区蓄水量和水质的变化与首都人民的饮水安全问题密切相关，实施南水北调工程可有效缓解北京市水资源短缺问题。研究表明，南水北调工程虽然能有效提高库区蓄水量，却也会给水库水质的变化带来很大不确定性，水体存在着潜在的营养状态改变的风险。 本研究首先对密云水库的水环境现状及变化趋势进行了评估，筛选出了与目标变量（叶绿素a）密切相关的流域环境因子，并选用预测效果最佳的随机森林模型对库区叶绿素a的变化进行了预测，分析了库区叶绿素a的变化与调水时长的关系。分别构建并筛选了适宜描述调水后库区蓄水量与叶绿素a单变量边缘分布和双变量联合分布的模型。通过多情景设置全面探究了调水后库区蓄水量和叶绿素a的双变量联合概率分布的变化规律和水体富营养化趋势，并定量评估了富营养化风险。论文的主要研究结论如下： （1）利用时间序列分解法、Kendall季节性检验法与区域均一性检验法等一系列方法分析了密云水库的水环境变化趋势，结果表明：潮河和白河的径流量有着变化规律极为相似的趋势性曲线、季节性曲线和随机性曲线，而且这两条入库支流的径流量和流域的降雨量均存在着明显的季节性规律，且都在2000年左右发生了显著变异；整体来看，密云水库属于磷控型水库，需严控来水总磷输入并减少总氮输入。内湖监测站的水质指标变化趋势和其余监测点差异较大，水动力条件和水质状况较差。此外，密云水库东西部水质变化趋势已经出现分化，表现为库东、恒河、金沟及水源九厂等东部区域的水质变化趋势趋好，而白河电站、库西、套里等西部区域以及内湖监测站水质变化趋势趋劣，故需要提高库区水位，并加强对西部区域和内湖的局部治理来改善库区水质。 （2）利用机器学习模型，选取对库区叶绿素a影响较大的环境变量来预测其浓度变化。对密云水库而言，随机森林模型（训练精度0.6557、验证精度0.6488）比支持向量机模型（训练精度0.8447、验证精度0.5875）更适于预测南水北调来水后库区叶绿素a的变化。预测结果表明：采取全年均匀调水的调水方式时，短时间（如10年内）调水对库区的叶绿素a影响较小，且各月受影响的程度存在明显差异。然而，随着调水时间的延长，来水所占比例不断增加，水库受到的影响也越发明显。特别地，当调水时长达到10年，库区水体7 ~ 8月份的营养程度将明显提高。可见长远来看，需要密切关注来水对密云水库叶绿素a的影响。本研究中提出的预测调水后库区叶绿素a变化的方法可以被推广到其他区域。 （3）通过单变量分布的拟合优度检验与评价，选用三重高斯混合的分布形式来拟合调水后库区蓄水量的分布，并选用Weibull分布函数来拟合叶绿素a浓度的分布。在对比了Gumbel-Hougaard、 Ali-Mikhail-Haq、Clayton和Frank Copula四种常见的Archimedean Copulas函数的基础上，确定了Frank Copula函数为拟合效果最佳、可描述调水对库区蓄水量和叶绿素a总体影响关系的模型，为分析二者的联合变化规律及发生水体富营养化的潜在风险提供定量化工具。 （4）利用Frank Copula函数分析了多情景下密云水库库区蓄水量和叶绿素a联合分布概率的变化规律，结果表明：当采取4 ~ 11月均匀、共调取2亿m3 的调水方案时，调水后库区水体营养状态基本不存在转变为富营养化（Chla > 0.01）的可能。对照湖库叶绿素a分级标准，可以发现，叶绿素a满足二类水体要求的概率为0.9995。但随着蓄水量的不断增加，水库中叶绿素a浓度增加的可能性也将有所增加，不过其值仍然很小；当统一采取均匀调水、但调水总量不同的方案时，可以发现随着调水总量的增加，各月叶绿素a浓度超出二类水体标准的联合概率将逐渐增大，但其值小于 0.0005。此外，随着调水总量的增大，联合概率受库区蓄水量变化的影响越来越小，全年各月水质营养状态发生变化的概率趋于一致。当统一采取2亿m3的调水总量、但调水时段不同的方案时，可以发现均匀调水情景对应的水体营养状态发生改变的概率（p = 0.00027）小于集中调水情景（p ≈ 0.00045）。总的来说，为了降低调水对整个库区带来影响，应该采取时间长、尽量均匀的调水方案。 本研究通过构建可以预测调水后库区叶绿素a变化的机器学习模型、以及能够反映库区蓄水量和叶绿素a联合概率分布变化的Copula函数，深入地定量分析了南水北调工程来水对密云水库蓄水量和叶绿素a联合概率分布的变化规律，以及给密云水库带来的潜在富营养化风险。本研究可为南水北调来水的规划以及密云水库水环境管理等工作提供科学的决策依据，具有重要的理论意义和实用价值。

Miyun Reservoir is the only surface drinking water source in Beijing. The changes of water storage and water quality in Miyun Reservoir are directly related to the drinking water safety of the people in the capital. The implementation of the South-to-North Water Diversion Project (SNWDP) can effectively alleviate the shortage of water resources in Beijing. The current research showed that although SNWDP can effectively increase the water storage of the reservoir, it will also bring great uncertainty to the water quality. The water body of the reservoir has a high potential risk of nutritional status change. Firstly, the present situation and trend of water environment in Miyun Reservoir were evaluated, and the watershed environmental factors closely related to chlorophyll a (Chla) were screened out. The optimal random forest (RF) model was selected to predict the change of Chla in the reservoir. Therefore, the relationships between the Chla and the duration of water diversion were analyzed. The univariate marginal distribution and bivariate joint distribution which were suitable for describing the distribution of water storage capacity and Chla in the reservoir after the implementation of the SNWDP were constructed and screened. Therefore, the joint probability distributions of water storage and Chla and the eutrophication status of the reservoir were fully explored under multiple water diversion scenarios. Moreover, the eutrophication risk of the reservoir due to the water division from the SNWDP was evaluated quantitatively. The main findings of the paper were as follows: (1) The trend of water environment in Miyun reservoir was analyzed using the methods of time series decomposition, Kendall seasonal and regional uniformity tests. The results showed that: three components of time series decomposition for the runoffs of Chaohe and Baihe Rivers, including a long-term smoothed trend, a seasonal pattern that changes with time, and the remainder, were very similar. Besides, the runoffs of two tributaries and the rainfall of the watershed had obvious seasonal variation, and they both occurred significant variations in 2000. As a whole, the total phosphorus (TP) in Miyun Reservoir exceeded the standard of water quality seriously, indicating that the reservoir was mainly effected by phosphorus. Therefore, it is necessary to strictly control the input of TP and reduce the input of TN in the incoming water for the prevention and control of eutrophication in the reservoir. The variation trend of water quality in Neihu station was different from that of other monitoring points, and its hydrodynamic condition and water quality condition were relatively poor. In addition, the trends of water quality in the eastern and western regoins of Miyun Reservoir were different. The water quality of monitoring stations in the east, including Kudong, Henghe, Jingou, and Shuiyuanjiuchang tended to be better than before; while it tended to be worse than before in Neihu station and other monitoring stations in the west, including Baihezhuba, Kuxi, and Taoli. Therefore, it is necessary to take local treatment measures for the western region and Neibu zone. (2) The concentration of Chla could be predicted with related environmental variables using two machine learning models. For Miyun Reservoir, RF model (training precision 0.6557, verification precision 0.6488) was more suitable than the supported vector machine model (training precision 0.8447, verification precision 0.5875) to predict the change of Chla in the reservoir under the influence of the inflow from the SNWDP. The results showed that: when adopting a scheme of uniform water transfering from April to November, we could find that the effect of water diversion on Chla for a short period of time (such as within 10 years) was small, and the degree of impact varied from month to month. However, with the extension of water diversion time, the proportion of incoming water would continue to increase, and the impact on the reservoir would be more obvious. In particular, after 10 years of water transfer, the eutrophication of the reservoir would increase in July and August. In the long run, the effects of SNWDP on Chla in Miyun reservoir still needed close attention. The method proposed in this study could be extended to other areas. (3) Through the test and evaluation of the goodness of fit, the distribution of water storage in the Miyun Reservoir after the implementation of the SNWDP was fitted with the triple normal distribution and that of Chla was fitted with the Weibull distribution function. Based on the comparisons of four common Archimedean Copula functions (i.e. Gumbel-Hougaard, Ali-Mikhail-Haq, Clayton and Frank Copulas), the Frank Copula function was selected as the best suitable model to describe the joint impact of SNWDP on water storage and Chla in Miyun Reservoir. Furthermore, it revealed the joint change of these two variables and the potential risk of eutrophication in the reservoir under various water diversion scenarios. (4) Using Frank Copula function, the variations of joint distribution probability of water storage and Chla in Miyun Reservoir under multiple scenarios were analyzed. The results showed that: when the scheme of uniform water transfering from April to November with a total water quantity of 200 million m3 was adopted, we could find that the likelihood of eutrophication in the reservoir was very small compared with the national classification standard of Chla for the lake. The concentration of Chla could meet the requirement of the second class with the probability of 0.9995. However, with the increase of water storage, the concentration of Chla in the reservoir would also increase, but its probability was very small. When the schemes of uniform water transfering from April to November with different total water quantity were adopted, we could find that with the increase of the total amount of incoming water, the joint probability of Chla exceeding the standard of the second class water would increase gradually, but its value was less than 0.0005. In addition, the influence of water storage on the joint probability would become smaller and the probability of the change of water quality in each month tended to be consistent. When adopting the schemes of same total water quantity (200 million m3) with different water diversion period, we could find that the probability (p = 0.00027) of the change of nutrient state in water body under uniform water diversion scenario was obviously smaller than that (p≈0.00045) of the centralized water diversion scenario. In general, to reduce the impact of water diversion on the entire reservoir area, we should take a long-term water diversion programme for the SNWDP, and the incoming water should be as uniform as possible with small total water quantity. In this study, a machine learning model for predicting the variations of Chla and a Copula function for reflecting the variations of the joint probability distributions of water storage and Chla in Miyun Reservoir after the implementation of the SNWDP were constructed. The potential eutrophication risks brought by the SNWDP under the multiple water diversion scenarios were quantitatively evaluated. This study had an important theoretical and practical value, for it could provide a scientific decision basis for the SNWDP planning and the water environment management of Miyun Reservoir.