The regrinding process of hematite is a unique beneficiation process for low-grade hematite, where the control of the ore feed pressure is a key process. To ensure precise control of the ore feed pressure, this paper combines conventional PI control with signal compensation technology and reinforcement learning to propose an intelligent PI control method. Firstly, describing the controlled object as a low-order linear model with frequent changes in unknown nonlinear dynamic systems, the compensation signal is designed using precisely calculated unknown nonlinear terms from the previous moment and closed-loop operating data, superimposed on the output of the PI controller to improve the dynamic performance of the control system. Then, reinforcement learning and real-time operational data of the closed-loop system are used to iteratively determine the optimal parameters for the PI controller and compensator, thereby optimizing the operation of the closed-loop system. Finally, through comparative experiments with existing methods and physical experiments using real industrial process data, the experimental results demonstrate the superiority of the proposed method.