GRAPES全球混合四维变分同化系统的方法研究

A methodological study of GRAPES global hybrid four-dimensional variational assimilation system

中文关键词:  切线性和伴随模式，背景误差协方差，4DVar，集合预报误差，H-4DEnVar

英文关键词:Tangent linear and adjoint models, Background error covariance, 4DVar, Ensemble forecast error, H-4DEnVar

中文摘要:

      [目的]随流型演变的变量间背景误差协方差特征非常重要，而变分系统中传统气候背景误差很难描述这些信息。虽然4DVar能通过切线性和伴随模式隐式演变初始背景误差协方差，但其存在开发维护复杂、计算成本昂贵等问题，而且在高精度可扩展全球大气模式中尤为突出。[资料和方法]为规避切线性和伴随模式，将四维集合预报误差引入GRAPES全球资料同化系统，发展了H-4DEnVar同化方案，开展批量循环同化及其预报试验和台风预报试验，并与4DVar方案对比。[结果]批量预报试验表明，四维集合预报误差的引入改善了分析场，显著提高了同化系统的全球预报能力；台风预报试验表明，H-4DEnVar中随流型演变的背景误差是台风路径预报误差减小的主要原因；与4DVar对比发现，考虑集合预报误差IO成本情况下，H-4DEnVar以4DVar 26%计算成本表现出基本相当的预报能力。[结论]H-4DEnVar同化方案在规避切线性和伴随模式的同时表现出了良好的同化预报效果，为在不使用切线性和伴随模式情况下实现四维同化提供了参考。

英文摘要:

      The flow-dependent background error covariance between variables is important, yet traditional climatic background errors in variational systems are difficult to characterize these information. Although 4DVar can implicitly evolve the initial background error covariance by tangent linear and adjoint models, it suffers from complex development and maintenance, expensive computational costs, and is particularly prominent in high-precision scalable global atmospheric models. To avoid the tangent linear and adjoint models, the four-dimensional ensemble forecast error is introduced into the GRAPES global data assimilation system, and the H-4DEnVar assimilation scheme is developed. The batch cycling forecast experiments and typhoon forecast experiments are conducted and compared with the 4DVar scheme. Batch forecast experiments indicate that the introduction of four-dimensional ensemble forecast error improves the analysis and significantly improves the global forecast performance; typhoon forecast experiments show that the flow-dependent background error in H-4DEnVar is the main reason for the reduction of the typhoon track forecast error; the comparison with 4DVar reveals that H-4DEnVar exhibits essentially comparable forecast capability at 26% of the computational cost of 4DVar when the IO cost of the ensemble forecast error is considered. The H-4DEnVar assimilation scheme shows good analysis and forecast effect while avoiding the tangent linear and adjoint models, which provides a reference for achieving four-dimensional assimilation scheme without the tangent linear and adjoint models.