Abstract: In this paper, a new approach for model reduction for discrete-time stochastic systems, which is called regional pole and covariance equivalent realization method, is presented. The major idea of the proposed approach is to construct the reduced-order models which match both the specified regional pole and steady-state covariance value. The resulting reduced-order models will approximate the full-order models in the two important aspects, i. e., transient and steady-state performances. The existence of desired reduced-order models is proved and the analytical expression of these models is also given. Finally, this paper provides an illustrative example.