基于广义线性模型和NCEP资料的降水随机发生器

天气发生器可以用来插补历史缺测气象数据或生成未来天气情境, 近年来被普遍应用于对气象变量的降尺度研究, 为陆面的水文、 生态模拟提供外强迫输入。广义线性模型 (GLM) 是近年来用于建立大尺度气象变量与地面气象因子之间的一种有效方法, 基于GLM的天气发生器具有一定的应用前景。本文以NCEP再分析资料中的单格点气温、 500 hPa位势高度、 位温、 相对湿度、 海平面气压等5个变量作为影响降水变化的大尺度因子建立模拟逐日降水量的广义线性模型。模型中对降水概率的描述采用Logistic模型模拟, 而对降水量则分别试用Gamma分布、 指数分布、 正态分布和对数正态分布来模拟, 试图比较和揭示这些基于不同理论分布的模型的能力。模型中待定参数的估计及对研究区逐日降水量的模拟采用了完全相同的实测逐日降水数据和同期NCEP再分析资料。参数的最大似然估计用遗传算法来实现, 对山东省临沂地区10个主要气象观测站降水资料的研究表明, Gamma分布模型的拟合效果最好, 对数正态分布次之, 指数分布再次, 正态分布最差; 参数估计分月获取的拟合效果略好于不分月的。模型逐日降水模拟表明, 对降水发生概率的模拟会低估各月的多年平均值, 基于指数分布的GLM会低估各月总降水量期望 (为月内每日降水量期望之和) 的多年平均值, 而基于对数正态分布的GLM则会在降水量较大时产生高估现象。由对应的天气发生器模型生成的随机模拟降水序列表明, 基于对数正态分布的模型会高估月降水量较大时的多年平均, 而基于指数分布及Gamma分布的模型则模拟效果较好。总体上看, 这种基于NCEP再分析资料和GLM的天气发生器对降水变率具有很强的解释和模拟能力。

A Stochastic Precipitation Generator Based on Generalized Linear Models and NCEP Reanalysis Data

Weather generators were useful for generating incomplete history weather records and weather scenarios in future. Recently they were used to study downscaling climate variables and provide weather forcings for hydrological and ecological simulations. Generalized linear models (GLM) are powerful tools for relating large-scale weather variables with high resolution variables at the earth surface. The application of generators based on GLM is promising. In this paper, by using five variables derived from single grid NCEP reanalysis data, such as air temperature, 500-hPa geopotential height, potential temperature, relative humidity, sea-level pressure, as large-scale independent variables that affect precipitation variations, GLM for downscaling and simulating daily precipitation were constructed. For determination of precipitation occurrence probability, the logistic model was used, and for simulating daily precipitation amounts, the Gamma, exponential, normal, and lognormal distributions were used respectively. The observed daily precipitation series and the selected NCEP reanalysis data were used for parameter-estimation and simulation. The maximum likelihood estimate of the model parameters was performed by genetic algorithms. The results of estimation showed that the fitting effect of the gamma distribution-based models was the best, the lognormal distribution was the second, the exponential distribution was the third, and the fitting effect of the normal distribution-based models was the worst; on the other hand, the fitting effect of the models with their parameters estimated for each month separately was a little better than that without separation of months. Both the observed and simulated daily precipitation-occurrence probabilities were added for each month and the corresponding precipitation amounts were also averaged for each month in order to verify the models ability. Results showed that the precipitation-occurrence probability was underestimated slightly by the logistic model, and precipitation-amounts expectations were underestimated by the exponential and gamma distribution-based GLM but were overestimated by the lognormal distribution-based GLM when monthly total precipitation amounts were large. The stochastic simulation of precipitation showed that the lognormal distribution-based GLM overestimated the yearly averaged monthly total precipitation and the exponential or gamma distribution-based GLM had a good simulation effect. On the whole, these stochastic models based on GLM and the NCEP reanalysis data are powerful tools for explaining and simulating the occurrence probability and amounts of precipitation.