【Objective】To achieve safe and efficient autonomous driving of high-speed trains, this paper proposes a compensated model predictive control (MPC) algorithm based on recursive KalmanNet (RKN).【Method】 The method adopts RKN, which combines data-driven and physical modeling, for dynamics modeling, and employs dual gated recurrent units (GRU) to directly learn the Kalman gain and covariance matrices, enabling efficient and accurate state prediction. Furthermore, a proportional compensation module (PCM) is designed to correct the MPC output, while the RKN covariance update is used to adaptively adjust the compensation gain, ensuring the control input approaches the optimum. The stability of the algorithm is proved via Lyapunov theory, and its effectiveness is validated on a hardware-in-the-loop simulation platform. Experimental results.【Result】show that, compared with MPC based on the extended autoregressive integrated moving average model, tracking performance and comfort improve by 64% and 65.8%, respectively; compared with LSTM-based MPC, the improvements are 80.4% and 42.5%, respectively, confirming the superiority of the proposed algorithm.