This paper deals with the gravity prediction problem of a novel 6-DOF hybrid robot. Note that the gravity load of the actuated joint is large and significantly varies with the configuration, seriously affecting the accuracy and stability of the robot. This paper proposes a gravity prediction strategy that takes the robot rigid body dynamics model as prior knowledge and combines it with a small amount of experimental data. Comparing with the model-driven methods, this method ensures the accuracy and generalizability of the prediction results, while outperforms typical Learning-from-Scratch (LFS) networks in terms of data efficiency. Sufficient experiments have been conducted on a prototype machine, and the results showed an average reduction of 39.98% in prediction error.