Variable elimination algorithm was proposed for inference using Bayesian networks. In this paper, we explore further on simplifying structures of Bayesian networks to reduce computational complexity. We propose the concepts of omissible node and replaceable node, and prove that we could delete the omissible nodes and replace replaceable nodes and their ancestors without affecting inference results using Bayesian networks. In many cases, the network can be simplified by our proposed methods, and therefore, the computational efficiency could be improved in average.