Roasting temperature distribution control using multi-agent reinforcement learning

As the initial stage of the zinc smelting process, the roasting process plays an important role in the entire smelting process. The temperature distribution in the roasting process directly determines final product quality, production efficiency, and operation safety. Due to the fact that the roasting temperature often shows nonlinear and spatial-temporally varying dynamics, the process control of such a complex dynamic temperature field is challenging. This paper proposes a new approach employing multi-agent reinforcement learning to facilitate roasting temperature distribution control. First, state, action and reward are defined for the roasting process to formulate the roasting process as a Markov decision process (MDP). Then, to enhance the robustness of the reinforcement learning (RL)-based controller (exploration capability), we propose a multi-agent deep deterministic policy gradient (multi-DDPG) algorithm. The proposed method and the baseline method were run independently for several times via simulation case study. By comparing the overall performance and the worst performance, multi-DDPG showed more stable performance. In the future, it may be possible to achieve temperature distribution control for real roasting processes.