Study And Implementation Of Parallel Variational Data Assimilation

Initial value is one of the key factors to improve numerical weather forecasting effect, so the objective analysis and data assimilation techniques take the important role in numerical weather forecasting research. With the improvement of remote sensing technique, the unconventional observations such as satellite have predominated atmosphere observation. However, there are complex non-linear relations between unconventional observations and analysis variables. It brings about new problems for data objective analysis and assimilation. In order to use unconventional observations in data assimilation system to produce analysis, variational data assimilation techniques that can process these complex non-linear relations effectively need to be researched.Data objective analysis is described asan optimization problem with dynamic model restriction by variational data assimilation algorithm. It uses optimal control theory and minimizes the difference between background provided by model and observation in given time window by adjusting control variables. Variational data assimilation can use observations of different time, different area, different property, including unconventional observations that are difficult used by analysis methods early. Variational data assimilation fundamentally solves the problem that available observations are lack, especially in southern hemisphere, ocean and tableland. Therefore, variational data assimilation is consided as the most potential data assimilation method.In order to build the next generation military numerical medium-range weather forecasting system, this paper researches the key technology of variational data assimilation system, include the essential algorithm, balance transformation and physical transformation, the process and construction of background error covariance matrix, observation operator, system implement technique and parallel computing. The paper researches incremental method, precondition process and optimization algorithm adopted by variational data assimilation system. The increment method is an effective way to reduce computing amount. With the needs of implement variational operation system, the paper researches the general incremental method and simple incremental method. Because condition number of target function is so important to speedup the convergence of minimization algorithm, preconditioner is an important method to reduce condition number and then improve iteration convergence. Introducing preconditioner in the variational data assimilation can not only improve the convergence of minimation algorithm, but also simples the process of the background covariance matrix. The paper implements proconditioner technique using control variables tranformation method; Minimization algorithm is the kernel of the variational data assimilation system, so besides researching truncating Newton method, the paper presents a new method that is similar incremental method, but can ensure convergence, we call it General Truncating Newton Method, and the method has a good application future in the variational data assimilation.Balance transformation and physical transformation is the most important techniques to simplify the analysis variable correlations., this paper discusses the key techniques about the analysis variables choosing, then emphasizes on considering the balance relations of the mass field and wind field and implemental way of balance transformation in analysis system, simultaneously researches the transform method of vorticity/divergence field and wind field. About background covariance matrix process, the paper mainly researches the processing way of background covariance matrix in the variational data assimilation, analyses the importance of background error covariance, gives the construction method of background covariance matrix, researches horizontal transformation and confirmation of horizontal correlative coefficient in the variable frame using iteration filter and spectral express, and the process of vertical transformation using EOF decompose.At present global and region variational data assimilation operation system are implemented separately. Analysis space of global variational data assimilation generally is spectral space. However analysis space of region variational data assimilation is grid-point space. The most important distinguish between global variational data assimilation and region variational data assimilation is processing method of horizontal relation. The paper presents the method that implement global and region variational data assimilation using uniform framework. For global variational data assimilation system design, the paper researches spectral transformation and recursive filter in horizontal space synchronously. That recursive filter is used to the design of global variable system is newly, and there is not an application example in the operation system. The paper researches the way that recursive filter is used to the design of global variational system by separating earth into the two poles area and middle area.The design process of variational data assimilation system is great complicated. Foreign experience indicates that the implement of a variational data assimilation operation system needs tens man-years even under the condition that theory has well resarched. Base on the theory frame of variational data assimilation and relative algorithm, the paper researches three-layer software system framework, key data structure design, registration mechanism and so on in the implement of three-dimensional variational data assimilation system, and gives the whole process of region three-dimensional variational data assimilation operation system, and analyses the experiment result for the process of hunan province typical weather. Based on the reduction of four-dimension variational data assimilation(4D-var) equations with incremental formulation, this paper presents the methodology to implement 4D-var system, including design of computing flow chart, computation of cost-function J and its gradient, implementation of Conjugate Gradient optimization algorithm, computation of innovation, which takes effect to produce the initial state for the global spectral model.This paper also does researches on parallel computation of variational data assimilation system from the following three aspects: Parallel algorithm According to 3D/4D-variational data assimilation algorithm, we design a multi-stage domain-decomposition algorithm to implement parallel computing, and observations are self-adaptive partitioned due to their geographic location. Four-dimension variational data assimilation parallel computation is implement by means of a coupler. Parallel implementation supporting kits Considering the common techniques to parallel variational data assimilation system such as domain decomposition, inter-domain communication, data transposition, parallel I/O etc. we develop supporting kits to implement the techniques above, which employs HDF5 format to manage data and I/O efficiently.Parallel implementation techniques Based on the detail researches on the theory and implementation of variational data assimilation algorithm, we design and implement the parallel computing of the global spectral model and its tangle-line/adjoint model, which are the important components of 4D-var system.. Parallization of the unified 3D-var system get better performance than WRF 3DVAR on 32-CPU platform and the speedup reaches 14.3, the global spectral model with resolution of TL399 can reaches 80.53 in speedup when paralleling on 112-CPU computer platform.