基于并行粒子群优化算法的变分资料同化

为了提升粒子群优化算法(PSO)应用到实际大气和海洋模式的资料同化时的计算时效性,针对一种最新提出的基于方向的粒子群优化算法(PSOBD),使用消息传递接口(MPI)和整体同步并行(BSP)计算模型,设计PSOBD的并行算法,较好地实现了BSP结构下PSOBD的全局通信操作。为检验并行化后的PSOBD的计算时效,将其用于潜水方程的资料同化并与基于串行的PSOBD的同化数值试验结果比较。大量的孪生同化试验结果的统计分析显示,并行化后的PSOBD与串行PSOBD一样能产生高质量的同化结果,且计算时效相对串行PSOBD提高了13倍以上。这一结果为PSOBD算法用于实际模式的四维变分资料同化(4D-Var)提供了依据。

Variational data assimilation based on parallel particle swarm optimization

When particle swarm optimizations(PSO) are applied to real atmospheric and oceanic four dimensional data assimilations (4D-Var), some obstacles need to be overcome, and one of them is the long computing time. To solve this problem, a new variant particle swarm optimizer based on directions (PSOBD) was presented recently, and efficient parallel strategies were explored and attempt made to design parallel PSOBD algorithm with Message Passing Interface(MPI) and Bulk Synchronous Parallel(BSP) computing model which should not affect the performance of the serial PSOBD. In the test of the parallel strategy, numerous twin experiments of the 4D-Var with the SWE model were conducted, and the numerical results statistically analyzed. The results show that the parallel PSOBD produces the same high quality assimilation results as the serial PSOBD and the traditional adjoint methods, and that the assimilation time by parallel PSOBD is only 1/13 of the one cost by the serial PSOBD in SWE model.These encouraging results provide the basis for the application of the PSOBD to 4D-Var in real atmospheric or oceanic models.