The retrospective time integration scheme presented on the principle of the self-memory of the atmosphere is applied to the mesoscale grid model MM5, constructing a mesoscale self-memorial model SMM5, and then the shortrange prediction experiments of torrential rain are performed in this paper. Results show that in comparison with MM5 the prediction accuracy of SMM5 is obviously improved due to its utilization of multiple time level past observations,and the precipitation area and intensity predicted by SMM5 are closer to observational fields than those by MM5.