在风蚀、风能、林火风险等许多地表过程的定量计算和评估中，风是一个关键气象要素。同时，风具有高度时间变率和空间异质性，站点观测风数据的时间分辨率、空间密度和序列长度往往无法满足应用需求。随机天气发生器基于站点观测资料的统计参数，通过随机模拟过程，可以生成任意序列长度、与观测序列具有相似统计特征的天气序列，是一种重要的统计降尺度方法。以往研究中的天气发生器通常无法很好地模拟小时极值风速。本研究提出HWGEN（Hourly Wind stochastic GENerator）单站点小时尺度风随机模拟模型，可以实现逐日风速、风向、逐时风速、日最大10 min风速的模拟，重点是改进极值风速的模拟，从而提升风蚀力和风功率密度的模拟精度。HWGEN目前包括两个版本，HWGEN_D同时模拟风向和风速；HWGEN_ND模拟不考虑风向的风速。基于北方风蚀区388个气象站2000—2020年观测小时风数据，对HWGEN进行校准和验证，并与美国农业部风蚀预报系统WEPS中的随机风模拟模型WINDGEN（WIND GENerator）进行比较；将HWGEN参数进行区域化，实现在没有气象观测地区使用随机模型模拟风速序列。结果表明：

（1）北方风蚀区2000—2020年日风速平均值为2.4 m s^-1，日风速标准差的区域平均值为1.0 m s^-1。第95%分位日风速、≥3 m s^-1日风速频率、日最大10 min风速D_max的区域平均值分别为4.2 m s^-1、24.6 %和5.6 m s^-1。小时风速平均值为2.5 m s^-1，小时风速标准差的区域平均值为1.6 m s^-1。第95%分位小时风速、≥5 m s^-1小时风速频率、日最大小时风速U_max、小时风速达到一日内最大值的时刻H_rmax区域平均值分别为5.6 m s^-1、10.1%、5.2 m s^-1和14:30 h（地方时下午2:30）。春季日和小时尺度风速平均值、标准差和极值均为全年最大；小时风速达到一日内最大值的时刻H_rmax在夏季比其他季节晚。

（2）在日风速模拟方面，HWGEN_D很好地模拟了日风速的平均值和标准差，均方根误差RMSE分别小于0.02 m s^-1和0.04 m s^-1。与WINDGEN相比，HWGEN_D将≥3 m s^-1日风速频率的绝对误差百分比误差MAPE从15.9%降低到10.9%。在小时模拟风速方面，HWGEN_D模拟小时风速平均值和标准差的RMSE分别为0.02 m s^-1和0.09 m s^-1，改进了WINDGEN对标准差的明显低估。与WINDGEN相比，HWGEN_D将≥5 m s^-1风速频率的MAPE从62.4%降低到25.7%；将第95%分位风速的RMSE从1.02 m s^-1降低到0.31 m s^-1。同时，HWGEN_D模拟平均侵蚀风功率密度AWPD和平均风蚀力Wf的RMSE分别为34.06 W m^-2、4.72 kg m^-1s^-1，与WINDGEN相比，将AWPD和Wf的MAPE分别从61.2%降低到26.5%、从72.8%降低到36.5%。HWGEN_ND模拟AWPD和Wf的RMSE分别为44.03 W m^-2和6.61 kg m^-1s^-1，略差于HWGEN_D，但明显优于WINDGEN；且模型的输入参数总数为HWGEN_D的十分之一。

（3）由于HWGEN参数的空间异质性较大，泛克里金（Universal Kriging，UK）、普通克里金（Ordinary Kriging，OK）和样条函数（Spline）对参数的插值效果不是非常理想且精度较接近。留一交叉验证表明与日风速模拟相关的参数，比与小时风速模拟相关的参数插值效果略好。整体来看，研究区西部地区的插值误差比中部和东部更大。使用UK方法插值得到的参数驱动HWGEN模型随机模拟风速序列，并与观测序列统计特征的对比结果显示：在日风速模拟方面，UK插值参数模拟日风速平均值和标准差的精度优于偏度系数，MAPE分别为16%、23.1%和31.9%。UK插值参数模拟≥3 m s^-1和≥5 m s^-1日风速频率的MAPE分别为114.3%和111.1%。在小时风速模拟方面，UK插值参数模拟小时风速平均值和标准差的精度优于偏度系数，MAPE分别为16%、16.1%和30.6%。UK插值参数模拟≥5 m s^-1和≥8 m s^-1小时风速频率的MAPE分别为82.3%和129.7%；AWPD和Wf的RMSE分别为81.52 W m^-2和11.31 kg m^-1s^-1。

Wind is a key climate element in quantitative calculations and evaluations of many earth surface processes, such as wind erosion, wind energy, forest fires, and so on. At the same time, wind is a climate element with high temporal and spatial heterogeneity, making the temporal resolution, spatial coverage and series length of observation wind data often can not meet the application requirements. Stochastic weather generator is based on the statistical parameters of observations. It can generate weather sequences of any length with similar statistical characteristics to the observed sequence through the stochastic simulation process, and is an important statistical downscaling method. Previous wind generators were generally not capable of simulating hourly extreme wind well. In this study, we proposed a single-site stochastic hourly wind simulation model, named as HWGEN (Hourly Wind stochastic GENerator), for daily wind speed, wind direction, hourly wind speed and daily maximum 10 min wind speed simulation, focusing on the improvement of the extreme wind, thereby improving the simulation accuracy of wind erosion force and wind power density. HWGEN includes two single-site versions, HWGEN_D simulating wind direction and wind speed together and HWGEN_ND simulating wind speed without considering wind direction. HWGEN was calibrated and validated based on hourly wind observations during 2000 to 2020, from 388 meteorological stations over northern wind erosion area of China, and was compared with WINDGEN (WIND GENerator) developed by US Department of Agriculture. Regionalizing the HWGEN parameters to achieve that use stochastic wind speed sequences generated by weather generator in areas without meteorological observations. The results indicate that:

The average daily wind speed in the northern wind erosion area from 2000 to 2020 was 2.4 m s-1, and the average standard deviation of daily wind speed was 1.0 m s^-1. The average values of the 95^th percentile daily wind speed, ≥3 m s^-1 daily wind speed frequency, and daily maximum 10-min wind speed Dmax are 4.2 m s^-1, 24.6%, and 5.6 m s^-1, respectively. The average hourly wind speed is 2.5 m s^-1, and the average standard deviation of hourly wind speed is 1.6 m s^-1. The average Hrmax values for the 95th percentile hourly wind speed, ≥5 m s^-1 hourly wind speed frequency, daily maximum hourly wind speed Umax, and the time when the hourly wind speed reaches its maximum value are 5.6 m s^-1, 10.1%, 5.2 m s^-1, and 13.3 hours, respectively. The average, standard deviation, and extreme values of daily and hourly wind speeds in spring are the highest throughout the year. The time when the hourly wind speed reaches its maximum is later in summer than in other seasons.

(1) The average daily wind speed in the northern wind erosion area from 2000 to 2020 is 2.4 m s^-1, and the average standard deviation of daily wind speed is 1.0 m s^-1. The average 95^th percentile value, ≥3 m s^-1 wind speed frequency, and daily maximum 10min wind speed D_max are 4.2 m s^-1, 24.6%, and 5.6 m s^-1. The average hourly wind speed is 2.5 m s^-1, and the average standard deviation of hourly wind speed is 1.6 m s^-1. The average 95^th percentile value, ≥5 m s^-1 wind speed frequency, maximum hourly wind speed during a day U_max, and the hour of maximum wind speed are 5.6 m s^-1, 10.1%, 5.2 m s^-1, and 14.5 h. The average, standard deviation and extreme values of daily and hourly wind speeds are the highest in spring. H_rmax is later in summer compared to other seasons.

(2) For daily wind speed simulation, HWGEN_D simulates the mean and standard deviation of daily wind speed very well, with root mean square errors (RMSE) less than 0.02 m s^-1 and 0.04 m s^-1, respectively. Compared with WINDGEN, HWGEN_D decreases the absolute error percentage error MAPE for the frequency of daily wind speed ≥5m s^-1 from 31.8% to 25.3%. For hourly wind speed simulation, HWGEN_D simulates the mean and standard deviation of hourly wind speed with RMSEs are 0.02 m s^-1 and 0.09 m s^-1 respectively, improving the significant underestimation of standard deviation in WINDGEN. HWGEN_D decreases the MAPE for the frequency of hourly wind speed ≥5 m s^-1 from 62.4% to 25.7%, and the RMSE of the 95^th percentile value from 1.02 m s^-1 to 0.31 m s^-1. Meanwhile, HWGEN_D simulates AWPD (average wind power density) and Wf (average wind erosion force) with RMSEs of 34.06 W m^-2 and 4.72 kg m^-1s^-1, respectively. Comparing to WINDGEN, HWGEN_D decreases the MAPEs for AWPD and Wf from 61.2% to 26.5% and 72.8% to 36.5%, respectively. HWGEN_ND simulates AWPD and Wf with RMSEs of 44.03 W m^-2 and 6.61 kg m^-1s^-1, respectively, which is slightly worse than HWGEN_D but much better than WINDGEN, and it with one-tenth of parameters required by HWGEN_D.

(3) Due to the significant spatial heterogeneity of HWGEN parameters, the interpolation performances of universal kriging (UK), ordinary kriging (OK), and spline functions are not very ideal and the accuracies are relatively close. Leav-one-out cross-validation indicates that the interpolation effect of parameters related to daily wind speed simulation is slightly better than that of hourly wind speed simulation. Overall, the interpolation error in the western region is greater than other areas. The observation and UK interpolated parameters are simultaneously inputted into HWGEN to simulate wind speed. For daily wind speed simulation, the accuracy of the average and standard deviation simulated by UK interpolated parameters are better than the skewness coefficient, with MAPEs of 16%, 23.1%, and 31.9%, respectively. The MAPEs of ≥3 m s^-1 and ≥5 m s^-1 wind speed frequencies simulated by UK interpolated parameters are 114.3% and 111.1%, respectively. For hourly wind speed simulation, the accuracy of the average and standard deviation simulated by UK interpolated parameters are better than the skewness coefficient, with MAPEs of 16%, 16.1%, and 30.6%, respectively. The MAPEs of ≥5 m s^-1 and ≥8 m s^-1 wind speed frequencies are 82.3% and 129.7%, respectively. The RMSEs of AWPD and Wf are 81.52 W m^-2、11.31 kg m^-1s^-1, respectively.