非线性集合四维变分同化方法NLS-4DVar之局地化改进
发布时间：2019-03-11 20:24
【摘要】：四维变分同化可利用同化窗口内所有可能的观测信息优化大气、海洋模式的初始场,从而极大地提高大气、海洋模式模拟性能,而作为4DVar标准算法的伴随方法始终无法避免繁琐与复杂的预报模式伴随方程的编程、维护以及更新。为避免伴随模式的使用,集合四维变分方法,4DEnVar方法被逐渐开发,为4DVar的求解提供了一种便捷的途径。4DEnVar一般通过局地化过程消除样本不足所造成的虚假相关,而局地化方案的不同也必然会影响到其最终的同化效果。本文将一种集合样本扩展的局地化方案引入到基于Gaussian-Newton迭代算法的非线性集合四维变分同化方法NLS-4DVar中,从而避免了原算法中为进行局地化过程而额外需要的线性化假设,使得算法收敛更稳定。另外,通过将原Gaussian-Newton迭代序列进行变形、避免了矩阵的直接求逆,极大地提高了同化算法的计算效率。利用非线性动力模型Lorenz-96所开展的观测系统模拟试验表明:采用新的样本扩展型局地化方案的NLS-4DVar算法,其同化精度略优于NLS-4DVar原始算法,由于避免了矩阵的直接求逆,其计算效率反而有所提高,同化所需时间有所降低,对于大气与海洋数据同化领域的应用具有极大的潜力。
[Abstract]:Four-dimensional variational assimilation can optimize the initial field of atmospheric and ocean models by using all possible observations in the assimilation window, thus greatly improving the simulation performance of atmospheric and ocean models. As a standard algorithm of 4DVar, the adjoint method can not avoid the complicated and complicated adjoint equation programming, maintenance and updating. In order to avoid the use of the adjoint model, the set four-dimensional variational method and the 4DEnVar method have been developed gradually, which provides a convenient way to solve 4DVar. 4DEnVar usually eliminates the false correlation caused by the lack of samples through the localization process. And the difference of localization scheme will certainly affect its final assimilation effect. In this paper, a set sample extended localization scheme is introduced into the nonlinear set four dimensional variational assimilation method (NLS-4DVar) based on the Gaussian-Newton iterative algorithm, thus avoiding the linearization assumption that is needed for the localization process in the original algorithm. The convergence of the algorithm is more stable. In addition, by deforming the original Gaussian-Newton iterative sequence, the direct inversion of the matrix is avoided, and the computational efficiency of the assimilation algorithm is greatly improved. The observation system simulation experiments carried out by using the nonlinear dynamic model Lorenz-96 show that the assimilation accuracy of the NLS-4DVar algorithm using the new sample extended localization scheme is slightly better than that of the original NLS-4DVar algorithm, and the direct inversion of the matrix is avoided. On the contrary, the computational efficiency is improved and the time required for assimilation is reduced, which has great potential for the application of atmospheric and ocean data assimilation.
【作者单位】： 山东农业大学;中国科学院大气物理研究所国际气候与环境科学中心;
【基金】：国家高技术研究发展计划项目(2013AA122002) 国家自然科学基金项目(41575100,91437220) 山东省省级水利科研与技术推广项目(SDSLKY201503)资助~~
【分类号】：P714
，
本文编号：2438592
[Abstract]:Four-dimensional variational assimilation can optimize the initial field of atmospheric and ocean models by using all possible observations in the assimilation window, thus greatly improving the simulation performance of atmospheric and ocean models. As a standard algorithm of 4DVar, the adjoint method can not avoid the complicated and complicated adjoint equation programming, maintenance and updating. In order to avoid the use of the adjoint model, the set four-dimensional variational method and the 4DEnVar method have been developed gradually, which provides a convenient way to solve 4DVar. 4DEnVar usually eliminates the false correlation caused by the lack of samples through the localization process. And the difference of localization scheme will certainly affect its final assimilation effect. In this paper, a set sample extended localization scheme is introduced into the nonlinear set four dimensional variational assimilation method (NLS-4DVar) based on the Gaussian-Newton iterative algorithm, thus avoiding the linearization assumption that is needed for the localization process in the original algorithm. The convergence of the algorithm is more stable. In addition, by deforming the original Gaussian-Newton iterative sequence, the direct inversion of the matrix is avoided, and the computational efficiency of the assimilation algorithm is greatly improved. The observation system simulation experiments carried out by using the nonlinear dynamic model Lorenz-96 show that the assimilation accuracy of the NLS-4DVar algorithm using the new sample extended localization scheme is slightly better than that of the original NLS-4DVar algorithm, and the direct inversion of the matrix is avoided. On the contrary, the computational efficiency is improved and the time required for assimilation is reduced, which has great potential for the application of atmospheric and ocean data assimilation.
【作者单位】： 山东农业大学;中国科学院大气物理研究所国际气候与环境科学中心;
【基金】：国家高技术研究发展计划项目(2013AA122002) 国家自然科学基金项目(41575100,91437220) 山东省省级水利科研与技术推广项目(SDSLKY201503)资助~~
【分类号】：P714
，
本文编号：2438592
本文链接：https://www.wllwen.com/kejilunwen/haiyang/2438592.html
