一种结合各类同化算法的一般方法

A generalized framework to blend different data assimilation models in Bayesian filter

提出了一个将不同滤波方法的估计结果相结合的一般框架,进而得到一个更加稳健的估计结果.首先,将不同方法的预报集合(粒子)结合,并通过相应的统计方法去掉集合中对状态估计贡献非常小的粒子.然后,我们使用处理过的集合样本,利用核密度估计(KDE)得到相应的试验密度(proposal distribution),并从该密度中重新抽样,最后利用贝叶斯滤波方法得到新的估计值.由于不同数据同化方法都有相应的适用范围,新方法的估计结果可以结合各类方法的优点,进而得到一个更加稳健的估计结果。在模拟试验中,我们用集合Kalman滤波(EnKF)与粒子滤波(PF)的集合样本来估计试验分布,从模拟结果可以看出,新方法可以较大的提高原有方法的估计精度,并有效的预防了滤波发散.

A generalized framework to blend different data-assimilation models is proposed in the present work,to obtain a robust estimate of the state.First,ensembles or particles of different data-assimilation methods are collected.A refinement procedure is then used to remove sample contributing too little to the estimate of the state.Second,aproposal distribution is estimated for Bayesian filter from the collected samples by kernel density estimation.Since different data-assimilations are suitable for different situations,the combined method with different data-assimilation results proposed here are more robust and can be applied to more complicated state space models.Judging from the simulation results,the new method improves accuracy of the state estimation in some models.