Abstract: The ensemble Kalman filter (EnKF) is able to obtain the flowdependent background error covariance based on the statistics of the ensemble samples so that it is becoming a research focus in the current data assimilation field. In this paper, a practical GRAPES ensemble Kalman filter data assimilation system is established and its tentative experiment is carried out. In view of the difficulty of assimilating real observations, the observations are organized into batches that assimilated sequentially in this paper. The Schur product is employed in the horizontal and vertical direction for filtering the error correlation noise and alleviating the matrix singularity problem. The ensemble Kalman filter system consistent with the GRAPES model vertical coordinate and forecast variables has been also established. The generation of ensemble samples considers the spacial correlation of model variables and the correlation among the model variables. The initial perturbation field can be obtained by perturbing the control variable in the 3DVar system. Through the ideal and actual observation experiments, the EnKF system has been verified. Furthermore, the experiment results of assimilating regional radiosonds show that the reasonable analysis of the GRAPES EnKF is obtained and the GRAPES EnKF system can be practically applied to the operation forecast. The 12 h forecast experiment using the GRAPES EnKF analysis indicates that the harmony of the GRAPES EnKF system is not as good as the GRAPES 3DVar assimilation system.