两种样本生成方法对DRP-4DVar同化方法性能影响的比较及分析

A comparison of sensitivity of DRP-4DVar performance to perturbation samples obtained between the two different methods

降维投影四维变分同化方法(DRP-4DVar)利用历史预报的集合来统计背景误差协方差,并将分析变量投影到样本空间下求解代价函数,因而集合样本的质量对DRP-4DVar同化方法的性能有着重要影响。文中尝试使用三维变分(3DVar)控制变量的扰动方法来产生集合样本,并与原来的历史预报扰动方法做比较。历史预报扰动样本具有随流型演变的特性,能够较好地描述大气模式的误差发展和模式变量之间的平衡约束关系。但是,由于历史样本之间的时间间隔较短,且样本数量有限,使得样本之间的离散度不够大,因而很难较全面地包含实际大气可能出现的状态。3DVar控制变量的扰动方法是基于3DVar系统中背景误差协方差的结构特征来生成初始扰动样本,因而能够构建比较合理的背景误差相关结构;且将扰动样本向前预报6或12h后能够使统计的背景误差协方差满足较好的协调性与动力和物理的平衡性。在计算代价方面,历史预报结果在业务预报中心能够直接获取,不需要额外的机时来生成样本,所以非常节省时间;而积分3DVar控制变量的扰动方法得到的初始样本比较费时,但是易于实现其并行计算。通过数值试验发现,用历史预报扰动方法生成样本时,模式变量的分析误差稍小于用三维变分控制变量的扰动方法的情况。但是,经过模式向前积分9h至12h后,利用三维变分控制变量的扰动方法生成样本能得到更优的预报结果,且对降水的模拟有明显的改善。因而,当用三维变分控制变量的扰动方法产生样本时,能够进一步改进DRP-4DVar同化方法的性能,尤其是提高数值预报的精度以及对强降水的模拟能力。

The dimension-reduced projection-four dimensional variational data assimilation（DRP-4DVar）approach utilizes the ensemble of hisiorical forecasts to estimate the background error covariance（BEC）and directly obtains the analysis on the ensemble space.As a result,the quality of ensemble members plays an important role in the performance of the DRP-4DVar approach.In this study,the BEC of the Weather Research and Forecast Model（WRF）3-dimensional variational data assimilation（3DVar）system is employed to produce initial perturbation samples for the DRP-4DVar approach.The historical-forecast-based initial perturbation samples are flow-dependent and able to describe the error-growing pattern in the atmospheric model and the balanced relationship between different model variables.However,the ensemble spread is not big enough because of the short time interval between adjacent historical samples and the limited numbers of ensemble.The perturbation method of control variables based on the structure characteristics of 3DVar BEC is able to produce initial perturbation samples that have reasonable background error correlations.Moreover,the estimated background error covariance also has good consistence and dynamic and physical balance between variables if the initial perturbation samples experience a development through 6 or 12 h model forward integration.From the aspect of computational expense,the historical forecast results can be directly obtained without any additional computational cost from operational numerical weather forecast centers,while the integration of samples from the 3DVar-based control variable method is time-consuming but this difficulty can be alleviated through parallel computing.With the analysis of the DRP-4DVar based on the new sample as the initial condition（IC）,the 9 to 12 h forecasts are better, although the IC is a little worse than that from the DRP-4DVar using the historical-forecast-based sample.Moreover,the simulated precipitation is improved significantly when the new sample is used.As a result,the performance of DRP-4DVar approach can be improved further,especially in terms of the accuracy of numerical forecasts and the ability to simulate severe precipitation,when using the perturbarion method of control variables of the 3DVar system to generate the initial perturbation sample.