Under non-uniform distributed layers, the vertical difference used in GRAPES (Global/Regional Assimilation and Prediction System) model can only achieved one order accuracy. A second order scheme was designed and introduced into GRAPES model for the process of vertical discretization. Ideal test with 1-D profile showed the new scheme could improve the accuracy of difference computation, while the improvement was not so markedly compared to the error caused by difference process itself. Density flow ideal test was conducted to verify the correctness and stability of new scheme in GRAPES model. A statistical evaluation of medium-range with second order scheme showed an improvement of forecast skill in large-scale fields, especially for the forecast after 120 hours. In addition, the second order scheme was tested with a real-case experiment for the extreme rainfall at South China, which again indicated improvement for the forecast of precipitation after 48 hours.