为探究 SWAT 模型参数优化过程与方法，降低参数估计不确定性，采用敏感性分析方法遴选关键参数，针对 关键参数采用拉丁超立方抽样构建参数样本集，进而结合各组关键参数组合下的模拟精度指标构建聚类指标集， 采用 SOM 聚类算法进行聚类，并基于模拟精度较高且波动较小类别识别各关键参数取值范围，形成一种 SWAT 模型关键参数优化系统方法。以石头口门水库流域为例，选取 1980—2016 年（1980—1986 年为预热期，1987— 2009 年为率定期，2010—2016 年为验证期）的月径流实测资料，建立流域 SWAT 模型，引入 SOM 聚类算法进行参 数优化，不断缩小模型关键参数合理取值区间，并应用 SUFI-2 算法进行模拟结果对比。结果表明：SWAT 模型适 用于石头口门水库流域，且参数优化前验证期的决定系数 R2为 0.79，纳什效率系数 ENS为 0.74，P-factor 为 0.65，R-factor 为 0.56；参数优化后验证期 R 2为 0.88，ENS为 0.83，P-factor 为 0.70，R-factor 为 0.50，模拟效果较好。故 应用 SOM 算法进行 SWAT 模型参数优化可以降低模型不确定性，提高径流模拟精度，为水文模型参数优化算法 的选择提供思路，对水资源管理政策制定与水库优化调度具有重要意义。

Parameter optimization is a crucial part of hydrological model simulation, which determines the accuracy of simulation and forecast.The Soil and Water Assessment Tool (SWAT) model is a distributed basin hydrological model based on the physical mechanism, and in the study of basin runoff simulation using the SWAT model, the model parameter optimization is mainly determined by the SWAT Calibration Uncertainties Program (SWAT-CUP) automatic rate method, including the SUFI-2 algorithm, Particle Swarm Optimization (PSO) algorithm, Parasol algorithm and General Language Understanding Evaluation (GLUE) algorithm. SWAT models often require thousands of hydrological model runs through the parameter optimization process to find a satisfactory combination of parameters. In recent years, in-depth studies on parameter optimization have concluded that the uniqueness of the optimal parameters is difficult to achieve, the problem of different reference effect is inevitable, so a common practice is to consider the parameter solutions as probability distributions that can be solved by applying mathematical methods.When the SUFI-2 algorithm is used for parameter optimization, it is found that the results often show uncertainties caused by different reference effect, or the runoff extremes are not well simulated.SOM algorithm is a neural network algorithm based on unsupervised learning, which can achieve intelligent clustering of data through self-organized competitive learning without understanding the interrelationship between sample data, and is also for achieving clustering through data mining technology. In practical engineering problems, taking the clustering analysis method can simplify the complex data system, make the data computation processing efficient, and make the internal laws of things clear. The SOM clustering algorithm was applied to the process of SWAT model parameter optimization to reduce the parameter estimation uncertainty. The sensitivity analysis method was used to select key parameters, Latin hypercube sampling was used to construct a parameter sample set for key parameters, and then combining the simulation accuracy indexes under each group of key parameter combinations to construct clustering index sets, the SOM clustering algorithm was used for clustering, and a SWAT model key parameter optimization system method was formed by identifying the range of each key parameter taking values based on the higher simulation accuracy and less fluctuation category. Taking the Shitoukoumen Reservoir basin as an example, the monthly runoff actual measurement data from 1980 to 2016 (1980 to 1986 as the preheating period, 1987 to 2009 as the rating period, and 2010 to 2016 as the validation period) were selected to establish the SWAT model, used the SOM clustering algorithm for parameter optimization, continuously narrow the reasonable value interval of the key parameters, and compare with the simulation results using the SUFI -2 algorithm. The results showed that the SWAT model is suitable for the Shitoukoumen Reservoir basin,R2 is 0.79,ENSis 0.74,P-factor is 0.65 and R- factor is 0.56 in the validation period before parameter optimization. The R2 is 0.88,ENS is 0.83,P-factor is 0.70 and R -factor is 0.50 in the validation period after parameter optimization. The simulation effect was satisfactory.Therefore, the application of SOM algorithm for SWAT model parameter optimization can reduce model uncertainty, improve runoff simulation accuracy, and provide ideas for the selection of hydrological model parameter optimization algorithm, which is important for water resources management policy formulation and reservoir optimization scheduling.