This paper combines the probabilistic rough set with the multi-granulation rough set in the multi-granulation space. Then the models of optimistic multi-granulation probabilistic rough set and pessimistic multi-granulation probabilistic rough set are established, respectively. Some properties of these models are investigated. This paper analyzes the knownledge classification accuracy of multi-granulation rough set with probability distributions theory, the approximate accuracies are increased when comparing with Qian's multi-granulation rough set. Finally, rationality and feasibility of the theory is verified by an example.