A certain UUV exhibits a high degree of coupling between its states and control variables, making it difficult for classical reduced-order methods to be applied. Meanwhile, considering that the control variables must satisfy the physical constraints of the UUV, a hierarchical model predictive control method for the UUV is proposed. Firstly, through state transformation, the UUV controlled system is decomposed step by step into an intermediate system and a terminal system. Between them, the intermediate system is directly driven by the control variables, and a quasi-infinite horizon model predictive control method is employed to design the control law, ensuring that the control variables meet the physical constraints of the UUV. The terminal system is driven by the states of the intermediate system and directly reveals the dynamic characteristics of the UUV's states. By designing the form of the state transformation, the stability of the terminal system is guaranteed. Ultimately, by integrating the intermediate and terminal systems, stable tracking of the desired states of the UUV is achieved. To further enhance the robustness of the proposed method, an LSTM neural network is employed to observe and estimate disturbances. The numerical simulation results demonstrate that the proposed hierarchical model predictive control method for UUV can effectively control the UUV to track the desired states. Meanwhile, the designed LSTM-based disturbance observer is capable of accurately predicting disturbances. Compared with the hierarchical model predictive control method without disturbance observation, the LSTM-based hierarchical model predictive control approach significantly enhances the smoothness of the control process.