Zinc roasting temperature field control with CFD model and reinforcement learning

Abstract

As the most critical variable in the zinc roasting process, the roasting temperature has been heavily researched for stable control through its average value. However, relying solely on the average temperature cannot convey the entire temperature field information necessary to achieve the optimal production state. To address this, This paper initially proposes a control scheme for the zinc roasting temperature field. First, a computational fluid dynamics (CFD) temperature field model was established through the mechanism of the roasting process. The influence of the feeding position on the temperature field was incorporated into the mechanism model, which provided the basis for the subsequent real-time control. Second, a convolutional Q-learning network (CQLN) is proposed to learn the mapping from state and action to Q value. CQLN can fully mine the spatial information of the temperature field. Then, the feed rate and feed location are adjusted in real-time to obtain the optimal roasting temperature field. Finally, extensive comparative experiments were conducted. Experimental results show that control performance of the proposed method is better than that of the comparison methods, with more uniform temperature distribution and smaller steady-state error.