模型参数估计是无测站流域径流模拟中需要解决的关键问题。参数区域化是应用最广泛的方法，但模型参数与流域特征间的非线性关系是参数区域化的主要障碍。本文以黄淮海流域内38个小流域为研究对象，进行了径流模拟研究，纳什效率系数（NSE）、决定系数（R²）和百分比偏差（PBIAS）的统计结果表明SWAT模型在各流域径流模拟中具有良好的性能。利用与气候、土壤、植被和地形相关的9个指标来表示与水文过程相关的流域特征。采用6种回归模型分析SWAT模型参数与流域特征之间的定量关系，包括：线性回归方程(LR)、支持向量回归(SVR)、随机森林(RF)、K近邻(kNN)、决策树(DT)和径向基函数(RBF)。首先，将38个流域依次假定为无测站流域。然后，利用其余37个供体流域构建拟合参数的回归模型，估算目标流域的模型参数，进行径流模拟。此外，本文也将基于相似性的区域化方法与基于回归分析的方法进行了对比。结果表明：基于支持向量回归(SVR)的区域化方法估计模型参数时径流模拟精度高。与传统的线性回归方法相比，机器学习算法处理非线性关系的能力突出，因而提高了无测站流域径流模拟的精度。不同区域化方法在湿润地区的表现比较接近，而机器学习算法的优势在干旱区更为明显。当研究区内含有嵌套流域时，由于流域密度高、空间距离短，此时采用基于相似性的区域化方法最好。研究结论可为无测站流域的洪水预报和水资源规划提高参考。

Model parameters estimation is a pivotal issue for runoff modeling in ungauged catchments. The nonlinear relationship between model parameters and catchment descriptors is a major obstacle for parameter regionalization, which is the most widely used approach. Runoff modeling was studied in 38 catchments located in the Yellow-Huai-Hai River Basin (YHHRB). The values of the Nash-Sutcliffe efficiency coefficient (NSE), coefficient of determination (R²), and percent bias (PBIAS) indicated the acceptable performance of the soil and water assessment tool (SWAT) model in the YHHRB. Nine descriptors belonging to the categories of climate, soil, vegetation, and topography were used to express the catchment characteristics related to the hydrological processes. The quantitative relationships between the parameters of the SWAT model and the catchment descriptors were analyzed by six regression-based models, including linear regression (LR) equations, support vector regression (SVR), random forest (RF), k-nearest neighbor (kNN), decision tree (DT), and radial basis function (RBF). Each of the 38 catchments was assumed to be an ungauged catchment in turn. Then, the parameters in each target catchment were estimated by the constructed regression models based on the remaining 37 donor catchments. Furthermore, the similarity-based regionalization scheme was used for comparison with the regression-based approach. The results indicated that the runoff with the highest accuracy was modeled by the SVR-based scheme in ungauged catchments. Compared with the traditional LR-based approach, the accuracy of the runoff modeling in ungauged catchments was improved by the machine learning algorithms because of the outstanding capability to deal with nonlinear relationships. The performances of different approaches were similar in humid regions, while the advantages of the machine learning techniques were more evident in arid regions. When the study area contained nested catchments, the best result was calculated with the similarity-based parameter regionalization scheme because of the high catchment density and short spatial distance. The new findings could improve flood forecasting and water resources planning in regions that lack observed data.