中文摘要：
近年来，随着集合预报的广泛应用，大量新颖的集合预报统计学后处理方法层出不穷，因此有必要对其进行系统地回顾。首先，对单变量集合预报统计后处理方法进行分类介绍，包括基于特定数学分布的参数化方法：逻辑回归、非均匀回归、贝叶斯模型平均、集合敷料法，以及灵活的非参数化方法：排序直方图订正、可靠性曲线订正、衰减平均偏差订正、分位数映射、分位数回归、概率匹配、频率匹配、最优百分位、最优评分法、逐成员订正、相似法、邻域法、基于对象的概率预报方法和虚拟降水法。其次，拓展到须考虑变量依赖性结构的多变量集合预报统计后处理方法，包括参数化的连接方法，以及非参数化的集合连接耦合和Schaake洗牌法。再次，介绍多模式集合和机器学习方法。最后， 总结并讨论了常用的集合预报统计后处理方法使用中需要注意的问题。
Abstract：
With the extensive use of ensemble forecasts, various novel ensemble statistical postprocessing methods have emerged in recent years. This review tries to summarize these methods comprehensively, including univariate and multivariate ensemble statistical postprocessing methods. The univariate methods can be classified as parametric methods based on certain mathematical distributions (such as logistic regression, nonhomogeneous regression, Bayesian model averaging, and ensemble dressing) and flexible nonparametric methods. Many techniques belong to the nonparametric type, including rank histogram calibration, reliability curve calibration, decaying average bias correction, quantile mapping, quantile regression, probability matching, frequency matching, optimal percentile, optimal score, member-by-member calibration, analog, neighborhood, object-based probabilistic forecast and pseudo-precipitation methods. As an extension of univariate methods, multivariate ensemble statistical postprocessing methods need to construct multivariate dependence structure, including the parametric copula, the nonparametric ensemble copula coupling, and Schaake shuffle methods. Then, multimodel ensemble and machine learning methods are introduced. In addition, limits and validity of the widely used ensemble postprocessing methods are discussed.
