Studies On Key Technologies And Parallel Computing Of Data Assimilation In Numerical Weather Prediction

Data assimilation is one of the core technologies of numerical weather prediction, and four-dimensional variational data assimilation is accepted as the most promising one, which plays an important role in improving the validity of high resolution numerical weather prediction.Tangent-linear model and adjoint model are the bases and important components to construct a four-dimensional varaitional data assimilation system. In a sophisticated numerical weather prediction global spectral model, the development of tangent-linear and adjoint models is still a puzzling difficulty. To construct the tangent-linear and adjoint models of spectral model, problems in developing the key technologies must be solved through innovative studies. Parallel computation is the important means and foundation for application of numerical weather prediction, so parallel computation of global spectral model and its tangent-linear and adjoint models is the key technology in four-dimensional variational data assimilation.The key technologies and parallel algorithms of data assimilation in numerical weather prediction are studied in this dissertation. A scalable parallel algorithm of data assimilation system based on optimization statistical interpolation scheme is designed and realized. A means of generating tangent-linear and adjoint models is proposed by using automatic differentiation tools and analyzing operator property. The parallel algorithms of numerical weather prediction global spectral model and its tangent-linear and adjoint models are studied. The main achievements are summarized as follows:(1) This dissertation analyzes thoroughly serial algorithm of the data assimilation system based on optimization statistical interpolation scheme, designs a distributed scalable parallel algorithm based on jumping selection strategy, and implements the distributed parallel computation of data assimilation system based on optimization statistical interpolation scheme. The parallel algorithm has good computer-sized scalability and computation scale scalability.(2) This dissertation suggests a distributed parallel algorithm based on two-dimensional data transposition strategy of operational global spectral model, through thoroughly analyzing the algorithm and data dependence of global spectral model, which can obtain a good load balance and has high-efficiency parallel scalability when reduced Gaussian gridding is used in high resolution spectral model. This dissertation also realizes the high-efficiency scalable parallel algorithm of operational global spectral model TL399L31.(3) This dissertation suggests a means of using automatic differentiation tools and analyzing operator property to construct tangent-linear and adjoint models of operational global spectral model TL399L31 through analysis and comparison. A validity test of the tangent-linear and adjoint models is taken and the validity is proved.(4) This dissertation introduces the partition of three-dimensional sphere into region of equal area, a special case of algorithm for partitioning muti-dimensional sphere into equal region, to the domain-decomposition in grid point space. This dissertation also designs a parallel algorithm of pseudo two-dimensional data transposition stragety for spectral model and tangent-linear model based on equal area partition algorithm, which can avoid the requirement of special processor number in two-dimensional partiotion. This algorithm can obtain a load balance and high parallel efficiency no matter how many processors are used, and realize distributed parallel algorithm of tangent-linear model in operational global spectral model TL399L31.(5) This dissertation studies and realizes parallel computation of adjoint model. Since computation and communication is separate in forward model and the same method of data-partition and domain-decomposition is used in forward model, tangent-linear model and adjoint model. Adjont model realization focuses on dealing with key performance subroutines in the communication level and subroutines for distributing data, and the parallel algorithm of adjoint model of TL399L31 is realized. A numerical test proves the success of parallel computation in adjoint model and high relative parallel computation efficiency.(6) This dissertation presents creatively a solution to critical problems arising in the tangent-linear calculation of non-differentiable function, which cannot be handled by the automatic differentiation tools. This method can ensure the calculation precision in the critical problems.