Application of Reinforcement Learning for Real-Time Optimal Control of the Pellet Induration Process

摘要

Abstract Pellet induration is the process of converting wet iron ore pellets to heat hardened pellets in a high temperature furnace. Handling disturbances in input raw materials to maintain the pellet quality coupled with simultaneous optimization of energy consumption and furnace productivity is an essential requirement of the induration process. It is a challenging control problem that cannot be adequately addressed by traditional Proportional–Integral–Derivative controllers or Model Predictive Controllers. In this paper, we demonstrate a Reinforcement Learning (RL) approach to devise an optimal supervisory control policy for the burner temperatures to optimize the pellet quality in the presence of a stochastic noise in the mean diameter of pellets. The optimal policy learnt by the RL agent is validated for various initial states of an industrial induration furnace and could successfully take the furnace from any initial state to the desired state within two control steps. This work proves that RL can be used to devise effective control policies for real-time control of large-scale complex industrial processes such as induration.