Abstract:

In this paper, a new hybrid data assimilation method is designed based on particle filter and 3Dvar. The new method generates an ensemble deviation with background error information through an optimal estimation of particle filter, thus providing flow-dependent background error covariance for 3Dvar. Particle degeneracy has always been the main obstacle of particle filtering in data assimilation field. In order to make the hybrid method work better, an improved residual resampling method is proposed to solve the problem of particle degeneracy. By sampling particles in the normal distribution, the lack of particle diversity caused by degeneracy is solved. Data assimilation experiments were tested on the ideal lorenz-63 model. The results show that the new method is better than the ETKF-3Dvar method when the model error is large, and as the model error increases, the new method is also better than the traditional data assimilation method. In the comparison experiment with hierarchical resampling and general residual resampling, the improved residual resampling method can ensure the stability of the assimilation results within a given time window, while the other two methods have a large deviation in the assimilation results.

Key words: hybrid data assimilation, particle filter, 3Dvar, residual resampling