天气发生器(Weather Generators, WGs)是一种统计模型，可基于观测天气数据计算的输入参数，随机生成任意长度、保留观测数据主要统计特征的模拟天气数据。降雨是全球水循环的重要组成部分，降雨过程特征，包括雨量、历时、峰值雨强及时程雨型等，是流域水文过程、土壤侵蚀过程的关键参数。然而，降雨过程观测数据非常有限，基于WGs的模拟降雨数据成为众多地表过程模型输入参数的重要来源之一。WGs也是常用的统计降尺度方法，可对GCMs(Global Climate Models, GCMs)输出的较粗时空分辨率的未来气候数据进行降尺度。CLIGEN(CLImate GENerator)模型是目前国际上少数几个能模拟降雨过程的WGs之一，输入参数较少，降雨模拟原理相对简单，应用性较强，被广泛应用于水文、侵蚀模型，且在降尺度方面也有广阔的应用前景。观测降雨过程数据的有限性制约了CLIGEN模型降雨过程模拟精度的评估与改进，也制约了CLIGEN模型在中国地区的进一步应用。 本研究基于我国东部季风区(也是我国主要水蚀区)18个气象站的逐分钟降水数据，使用指数法对次降雨事件进行划分，并分析我国次降雨过程的主要特征。在此基础上，分日雨量等级评估CLIGEN模型对我国降雨过程特征的模拟精度，建立基于小时降水准备CLIGEN模型降雨过程输入参数的方法。最后，利用收集到的2405个气象站的逐日气温、逐日降水和逐小时降水数据，建立覆盖中国大陆地区的CLIGEN模型单站输入参数库及1km×1km精度的二维网格输入参数库。本研究主要结论如下： 基于分钟降水数据，东部季风区18个气象站采用指数法确定的最小降雨间歇值(MITexp)的范围在7.6h~16.6h之间。小时数据计算结果与分钟数据相近，18站的平均相对误差小于10%。计算MITexp的降雨资料序列越长，所得的MITexp值越稳定，建议使用10~15年以上的资料计算该指标。大部分次降雨指标特征对MIT取值较为敏感，随着MIT值增加，降雨过程次数和平均雨强减少，次降雨量、降雨历时、I30(最大30min雨强)增加，到达峰值雨强的时间(tp)对MIT的变化不敏感。18个站MITexp均值为10h，用10h与各站的MITexp划分得到的次降雨特征均值间的差异不显著。推荐在我国东部季风区采用指数法确定MITexp；当缺乏实测数据时，可以直接使用10h作为MIT。 未率定的CLIGEN模型对日(次)、月、年尺度的雨量、晴雨转移概率和tp的模拟精度较高，对降雨历时与峰值雨强的模拟精度相对较低，其中降雨历时整体以及中雨、大雨、暴雨的均值被低估；对峰值雨强的模拟效果与所选取的峰值雨强持续时段有关，I1(最大1min雨强)整体均值呈高估趋势，I30整体均值呈低估趋势。CLIGEN对历时和峰值雨强的模拟精度都随日雨量等级的增加而降低。利用中国降雨资料率定CLIGEN内置参数A的变异系数和?分别为0.38和6.13，替换CLIGEN原内置参数后，可改善历时标准差的模拟精度，但均值的模拟精度降低。尽管未率定的CLIGEN模型对降雨历时和峰值雨强的模拟效果不理想，但CLIGEN模拟降雨数据估算的降雨侵蚀力均值、极值以及降雨强度-历时-频率(Intensity-Duration-Frequency, IDF)指标，与观测数据计算的对应指标间存在较好的线性相关关系，通过系统纠偏后，CLIGEN的模拟降雨数据可应用于估算降雨侵蚀力及部分IDF指标。 建立了用小时降雨数据计算CLIGEN模拟降雨过程的输入参数TimePk与MX.5P的方法。结果表明，以小时降雨计算的MX.5Ph相对分钟降雨计算的MX.5Pm整体偏低，需要乘以1.4进行修订。降雨历时、峰值雨强的模拟对输入参数TimePk不敏感，在缺乏分钟降雨降雨时可直接应用小时降雨计算输入参数TimePk。降雨历时和峰值雨强的模拟对输入参数MX.5P更敏感，与基于MX.5Pm的模拟结果相比，18个站基于MX.5Ph模拟得到的历时、峰值雨强的基本统计特征平均相对误差小于7%；模拟得到的多年平均降雨侵蚀力及10年一遇EI30的平均相对误差小于4%。小时降雨可在缺乏观测间隔小于或等于30min的观测降雨资料时用于生成TimePk和MX.5P，其模拟结果可应用于估算降雨侵蚀力。 利用收集到的中国大陆2405个气象站的逐日气温、日降水和小时降水数据，建立2405个站的CLIGEN单站输入参数库，并用克里金插值法对输入参数进行插值。交叉验证法对插值精度的评估结果表明，利用泛克里金法(UK)插值的参数精度高于普通克里金法(OK)。将基于UK插值得到的参数与观测参数同时输入CLIGEN模型，生成同序列的模拟数据，对比结果表明：日最高、最低气温的均值、标准差和偏度系数以及晴雨转移概率模拟精度最高，模型有效系数NSE不低于0.94；日雨量、降雨历时、峰值雨强的均值和标准差精度较好，模型有效系数NSE不低于0.87；日雨量、降雨历时和峰值雨强的偏度系数精度较差，模型有效系数NSE变化于0.26 ~ 0.72之间。插值参数基本保留CLIGEN对气温、降雨的模拟精度，因此将2405站的输入参数插值到中国大陆，形成1km×1km的网格输入参数库，则在中国大陆任意地区，可基于参数库，实现CLIGEN模型对降雨过程的模拟。

Long-term precipitation is one of the most basic and indispensable inputs commonly required by hydrologic and soil erosion models. When observed data are not sufficient in spatial and temporal coverages, simulations of the data may be required. Synthetic sequences of daily precipitation depth generated by stochastic weather generators that preserve the statistical characteristics of measured data are commonly used in such situations. Weather generators also plays an important role in making climate science accessible to impact studies and other applications. CLIGEN (CLImate GENerator) is a stochastic weather generator which can simulate long-term continuous precipitation depth and the corresponding storm patterns on a daily/event scale. CLIGEN was initially incorporated as part of the WEPP (Water Erosion Prediction Project), and now it is used as a general weather generator to simulate precipitation inputs for other models. CLIGEN has been widely adopted to downscale output from GCMs to study climate change impacts. In this study, the exponential method was adopted to determine the minimum inter-event time (MITexp) for identifying statistically independent storms using 1-min resolution data collected from 18 weather stations distributed in the eastern monsoon region of China. Characteristics of the divided storms were analysised. On this basis, the performances of CLIGEN were systematically evaluated in terms of precipitation-related variables for all storms as well as four storm categories grouped by precipitation depth. For wider evaluation and application of CLIGEN in China, 1-min precipitation data were aggregated into hourly data to explore effective methods to estimate input parameters TimePk and MX.5P for CLIGEN storm pattern generation with hourly data. Eventually, the daily temperature, daily precipitation and hourly precipitation data were collected from 2405 weather stations distributed across the mainland China to prepare the corresponding 2405 CLIGEN input files. Then the Oridinary Kriging and Universal Kriging methods were compared to regionalize the temperature and precipitation input parameters to the whole mainland China to establish a two-dimensional inputs dataset with a resolution of 1 km×1 km. Main conclusions can be drawn from this study: An exponential frequency distribution fits well with that of the observed inter-event times for each tested station. MITexp for the 18 weather stations using 1-min resolution data varied from 7.6h to 16.6h, with an average of 10.7 h, and the corresponding standard deviation was 3.0 h. The exponential method was not sensitive to the data resolution. At least 10 to 15 years of data records were necessary to obtain stable MITexp. Most event storm characteristics were sensitive to MIT values and characterized by large variations as MIT increased from 1h to 24h. Longer MIT values resulted in fewer annual storm numbers, greater individual precipitation amounts and I30, longer durations and effective durations, and lesser mean storm intensities. However, no trend was found for the time to peak (tp) index. Based on results achieved here using both 1-min and hourly data of the 18 stations, MIT of 10h is recommended during application of event storm based studies in the eastern monsoon region of China. As a stochastic weather generator, CLIGEN reproduced the main statistics and probability distributions of daily precipitation depth well for all storms and four storm categories. Storm duration and peak intensity were not well simulated in general, and the relative errors increased as the precipitation depth increased from moderate to intense storm events. Mean storm duration was underestimated for storms overall, as well as for the subset of moderate, heavy and intense storms, but was overestimated for light storms. I1 tended to be underestimated, while I30 tended to be overestimated with CLIGEN. There was no obvious bias found in I5 for the 18 sites. The CLIGEN internal parameters - CV of A and ? calibrated using Chinese rainfall data are 0.38 and 6.13. After replacing the original parameters, the simulation quality of the duration standard deviation has been improved whereas the mean value gets worse. Rainfall erosivity and intensity-duration-frequency (IDF, in longer durations and lower return periods) computed from the CLIGEN-generated values were systematically greater but well correlated with the measured data, indicating CLIGEN output has the potential to predicted rainfall intensity and IDF values by multiplying the corresponding adjusting coefficients. CLIGEN-generated storm pattern was quite insensitive to the TimePk parameter. The MX.5P parameter estimated using the hourly data needs to be multiplied by a factor of 1.40 to correct the bias to match that determined using the original 1-min data. CLIGEN-generated storm patterns using the TimePk and MX.5P parameter values prepared with hourly data were acceptably similar to those using parameter values prepared using 1-min data in terms of storm duration, and selected peak intensities. The mean absolute relative error in the mean, standard deviation and skewness coefficient of storm duration and peak intensities generated from CLIGEN with inputs TimePkm, MX.5Pm and TimePkh, MX.5Ph were all less than 7% over 18 tested stations. Moreover, the R-factors and 10-year storm EI30 calculated with simulated rainfall from TimePkm, MX.5Pm and TimePkh, MX.5Ph were all less than 4% over 18 stations. Two interpolation methods, the Oridinary Kriging and Universial Kriging, were used to regionalize CLIGEN temperature and precipitation inputs to the mainland China. The interpolation accuracy was evaluated using cross validation and results showed that the Universial Kriging with covirables of latitude, longtitude, elevation and annual precipitation improved the interpolation accuracy obviously. Two methods both behaved unsatisfied for SKEW P (skewness of daily preciptiation), but the interpolation accuracry improved when MEAN P (mean daily preciptiation) and S DEV P (standard deviation of daily preciptiation) were used as covariables. The parameter files prepared using observed and interpolated data by Universal Kriging were input to CLIGEN and the generated data were compared. The basic statistics for temperatue and storm occuracne were statistically simulated with NSE not less than 0.94; mean and standard deviation for daily precipitation, duration and peak intensity were well preseverd with NSE not less than 0.87; the skweness for daily precipitation, storm duration and peak intensity simulated worse among all variables, with the NSE raning from 0.26 to 0.72.