Particle filter based on partial accurate observation for high dimensional stochastic dynamic system

Abstract

With the improvement of computer performance, particle filter as a method of state estimation, is widely used for nonlinear non-Gaussian dynamic systems. In order to reduce the partical degeneracy in the state estimation process of high-dimensional stochastic dynamic systems which commonly occurs in the conventional particle filter, an improved particle filter method based on adding more accurate direct observations of some dimensions of the system state is proposed in this paper. By arranging a plurality of sensors in the stochastic dynamic system which are able to directly and precisely measure some dimensions of the state, lots of accurate state information can be obtained. Combined the information with the inverse process of the system measurement, the particles move towards the high observation likelihood region, so that the effectiveness and accuracy of resampling process can be greatly enhenced. In the case study, a stochastic dynamic system of near-surface wind farm is established by the theory of computational fluid mechanics. Numerical results show that the proposed PF method is able to enhance the accuracy of the estimation for high dimensional stochastic dynamic system.