Abstract: In this study, the perturbation forecast model GRAPES_PF appropriate for the implementation of four dimensional variational data assimilation (4D-Var) system has been developed based on the regional numerical weather prediction model GRAPES. GRAPES_PF involves a set of linear perturbation forecast equations including momentum, thermodynamic, moisture and continuity, which are derived from the non-hydrostatic primitive equations used in GRAPES on a terrain-following vertical coordinate framework. A semi-implicit, semi-Lagrangian and two time-level integration scheme is applied to the linear equations. Spatial discretization is performed on the Arakawa staggered C-grid in the horizontal and the Charney-Phillips grid in the vertical. The Helmholtz equation that only contains perturbation Exner pressure at future time step of integration is obtained by eliminating other variables in the linear perturbation equations. Similar to the nonlinear model, the generalized conjugate residual (GCR) method is used to solve the perturbation Helmholtz equation. A numerical experiment has been designed to evaluate the GRAPES_PF model by applying an initial perturbation of mesoscale high pressure centered at model domain and predicting its evolution with time. The same initial perturbation of high pressure is also added to nonlinear model so that the evolution of the perturbation can be traced as truth for verification. We then verify the perturbations predicted by the linear GRAPES_PF model against those of the nonlinear GRAPES model. Results show that the initial pressure perturbation induces a fast-moving-outbound internal inertial gravity wave through the well-known geostrophic adaptation process. The linear GRAPES_PF model produces results similar to that of nonlinear GRAPES model with high accuracy: the initial pressure perturbation subsequently induces increments in the fields such as horizontal winds, vertical velocity, potential temperature and water vapor, which are almost identical to those of the nonlinear model. The main conclusion is that the perturbation forecast model GRAPES-PF, as a reasonable linear version of the nonlinear GRAPES model, can offer a good scientific base for the 4D-Var data assimilation system to be developed in the future.