To address the degraded control precision in model predictive control caused by time-varying aerodynamic parameters of hypersonic vehicles, this paper proposes an online model predictive control method incorporating a forgetting factor extreme learning machine (FF-ELM). A dual-mode framework is established, combining a BP neural network baseline model with an ELM-based online residual compensator, to achieve real-time identification of dynamic system parameters. The baseline model provides prior aerodynamic knowledge, while the online compensator dynamically corrects model prediction deviations through a recursive weight updating mechanism with forgetting factors. Simulation results demonstrate that compared to the recursive least squares method, the proposed approach reduces the mean square error of aerodynamic coefficient prediction by 89.3% and 78.9% in sinusoidal and spiral trajectory tasks, respectively. Trajectory tracking accuracy improves by 30%, with the average single-step optimization time controlled at 25 ms. This method offers a novel technical pathway for real-time precision control of complex dynamic systems.)