相对于传统的确定性水文预报，水文集合预报包含了水文预报各个环节的不确定性信息，因此在理论上更加科学，在生产实践中也提高了对暴雨、洪水、干旱等事件的认知和预报能力，为水资源管理和防洪抗旱提供了更精确的实用信息。同时，水文集合预报的检验必不可少，一方面可以了解预报质量及发现预报问题，另一方面可以根据用户对预报要求的不同，提供相应的不同信息。目前，针对集合预报已提出许多指标，本论文的研究目标是探索如何选择合适的指标评估水文集合预报，并对现有集合预报评价指标的不足进行改进。本论文的主要创新点如下：(1)比较分析水文集合预报检验方法/指标。根据检验指标的不同属性，将其划分为确定性检验指标及概率性检验指标。并根据不同阈值，将实验数据（径流/降水）进行划分，如洪水、暴雨等，再分别使用不同类型指标对实验数据进行检验评估。比较分析检验结果发现，检验指标倾向于评估水文集合预报整体预报性能（例如可靠性、解析度等），但对于极端事件集合预报质量（例如洪峰区间的部分可靠性）的评估则存在不足，无法有效区分整体可靠性与部分可靠性，导致错误的检验结果。(2)提出等级极坐标图，改进极端事件集合预报的可靠性检验方法。针对现有评价指标的不足，对现有指标进行改进。本研究主要对等级直方图（Rank Histogram）进行改进，提出了等级极坐标图（Rank Polar Diagram）。实验研究发现，等级极坐标图可以同时对预报整体及部分的可靠性进行科学检验和有效区分，更加适用于水文集合预报可靠性检验。

With the increasing practical issues, such as flood estimation, drought forecast and water resources management, the demand for the hydrological ensemble prediction is rising. However, what is a good forecast? Is there any appropriate and accurate efficiency criteria can help us evaluate the forecast as “good”, “bad” or “skillful”? Moreover, among the existing various criteria or verification methods, an in-depth study of evaluating and formulating these criteria is very meaningful. The main contributions of this dissertation are presented as follows:(1)Comparison of different efficiency criteria for hydrological ensemble prediction assessment. Because of the emphasis in the suitability of different criteria may vary according to different situations (such as its reliability or sensitivity to the hydrological regime), the experimental design was constructed in two aspects: 1) the experiment using the artificial-designed streamflow data; 2) the real-data experiment contains the continuous simulation (streamflow/precipitation), the high-flow simulation, the torrential rain simulation and so on. And the criteria were evaluated using the MOPEX streamflow simulation data and the precipitation data generated by the Ensemble Preprocessor (EPP). In this study, the categorical/probabilistic criteria through an experiment were estimated. And we find out they have remarkable advantages in evaluating the quality (e.g., accuracy, reliability, resolution, etc.) of the whole hydrological ensemble forecast. However, they are not capable of verifying the partial reliability in certain frequency intervals, such as the flood. And then we test this idea through the real-data experiment and get the same results. At last, further research of the comprehensive assessment is also constructed on the purpose of more conveniences and effectiveness for different users. (2)Improvement of rank histograms for verifying reliability of extreme events ensemble forecasts. Owing to the increasing competition of flood forecasting and regulation, drought risk management, urban flood control and water resource management, and so on, the demands for forecasts of extreme events like peak flows, low flows or rainstorm are continual to rise. And ensemble forecast is playing an important role in hydrological forecast system. But how forecast reliability is supposed to be verified properly? This study proposed an improved verifying method of Rank Polar Diagram (Rpolar diagram) for evaluating reliability of extreme events ensemble forecasts. Previous use of rank histogram may just provide a simple and rough examination of the entire reliability in the ensemble. However, in practical applications, peak flows, low flows and rainstorm are all crucial integral components in different hydro-climate regimes. From the hydrological perspective, guaranteeing the reliability of extreme events ensemble forecasts is much more important and should not be ignored. For instance, it will be a huge risk for the river going higher than the crest prediction in practical flood forecast. The rank polar diagrams are capable to verify not only the overall reliability but also the partial reliability in any certain intervals concluded the extremes. Note that specific intervals can be set autonomously just according to users’ needs. And its effectiveness was validated using two typical sets of simulative ensembles (section 4.2) and actual streamflow/precipitation ensembles (section 4.3). As we expected, both streamflow and precipitation applications results exhibit the suitability of rank polar diagrams for verifying reliability of extreme events in the ensemble. Besides, the comprehensive assessment can also be constructed based on the modified root mean-square error (mRMSE).