Based on the conjugate code technique,a meso scale adjoint model is constructed,and the lateral boundary conditions of the model is discussed.Then assimilation and prediction experiments are performed by adopting the adjoint and Derber's assimilation schemes,and using the quasi real data and the descent algorithm of conjugate gradient method in assimilation.The results show that the both schemes achieve fine assimilation effect with Derber's scheme better than conjugate one.It is suggested that the variational assimilation is an effective skill to improve the prediction accuracy.