This work proposes a composite anti-disturbance particle filtering (PF) approach for non-Gaussian state estimation problem in the presence of additive disturbance (AD) with switched dynamics. The presence of AD will lead to sample deviation in the stage of particle generation, which degrades both efficiency and robustness of the standard PF algorithm. To overcome this problem, we propose a composite filter structure where an IMM disturbance predictor is adopted in the inner layer for real-time AD prediction and the PF is used in the outer layer to deal with the non-Gaussian PDF. The predicted value of AD will be exploited in the PF to compensate for the sample deviation, hence the particles generated by the PF are more likely to fall into the important regions of the state space, which implies higher sampling efficiency and enhanced robustness. A numerical example is adopted to show that, compared with the brute-force PF, the proposed approach is more robust to the AD.