Cases Guided Reinforcement Learning Control for Stripe Flatness in Quenching System

This paper addresses the optimization of strip steel flatness during the quenching process, which is significantly influenced by cooling speed and uniformity. Due to the nonlinearity and coupling of multiple variables in the quenching process, the plate shape control system relies on models and existing knowledge when performing online optimization. Considering the computational time cost due to the low utilization of experience by widely used online optimization algorithms for rolling and quenching processes, this paper proposes a reinforcement learning-based architecture for quenching plate shape control policy. This strategy is designed to adjust parameters in quenching equipment for effective plate shape control. A data-driven broad learning system incorporating a random vector functional link model is utilized to predict flatness. The predicted flatness serves as a basis for reward calculation, enabling the TD3 agent to refine its learning process. To enhance convergence efficiency, the agent is guided by manufacturing cases during training. The experimental results on the same production samples demonstrated that the proposed method was effective in increasing the percentage of good board shapes by 6.04 % and reducing the percentage of poor board shapes by 0.59 %. Meanwhile, compared with PSO, the proportion of good plate shape is 0.34 % higher with this method.