Reinforcement-Learning-Based Composite Optimal Control for Looper Hydraulic Servo Systems in Hot Strip Rolling

Abstract

The control scheme of the looper angle plays an essential role in hot rolling, which is directly related to the tension maintenance of the strip. The conventional scheme uses a proportion–integration–differentiation (PID) controller to control the servo valve to drive the hydraulic actuator. For different steel types and production temperature changes, the PID control mainly relies on empirical parameter adjustment, which may bring inaccuracy or inefficiency. The objective of this article is to propose a reinforcement-learning-based looper hydraulic servo optimization control scheme to automatically tune the control gain to optimum. First, the modeling error caused by variable parameters and the influence of external disturbance are considered, and the corresponding control model of the looper system is given. Subsequently, a feedforward controller with a radial basis neural-network-based disturbance observer is used to deal with modeling errors and disturbances. A feedback controller with off-policy reinforcement learning is applied in the meantime. The proposed control scheme can realize uniformly ultimately bounded of the system with Lyapunov theory. Simulation results verify the effectiveness of the proposed method.