Abstract: In the GRAPES_GFS (Global Regional Assimilation and PrEdiction System, Global Forecast System), a one-dimensional reference profile based on isothermal atmospheric structure is used. Previous research studies have introduced a three-dimensional reference profile into the dynamic core of the GRAPES_GFS, and a series of benchmark experiments have been carried out to verify the correctness and accuracy of the new method. The present study mainly focuses on real-data prediction experiments, comparing and analyzing the advantages and disadvantages of different three-dimensional reference state setting methods. The climatic average method is then used to carry out four-dimensional variational cycle forecasting experiments for two summer months. The results show that after using the three-dimensional reference profile, the comprehensive prediction performance of the model has been improved, the biases in the predicted tropospheric height and temperature have been reduced, and the severe loss of mass in long-term integration process has also been significantly alleviated. In addition, by comparing the kinetic energy spectra, it can be seen that the energy dissipation of the model at upper levels is significantly smaller after using the three-dimensional reference atmosphere, and the changes of the energy spectra are more consistent with observations.