Silicon Content Control for Blast Furnace: Model-free DDQN-ICM Method

Silicon content control is critical to ensure the quality of molten iron and the stable operation of the blast furnace. The complex reactions and large uncertainty of the blast furnace make it hard to control. Several predictive control-based methods were already studied in Silicon content control. However, the requirements of prediction model design and short predictive horizon limit the generalization and long-term optimization of the methods. To tackle these issues, we propose a deep double Q-learning (DDQN) and intrinsic curiosity module (ICM) based Silicon content control method. First, the deep Q-learning framework employed in our approach eliminates the need for a prediction model and supports long-range simulations, thereby enhancing algorithm generalization and long-term optimization capabilities. Second, DDQN and ICM help to avoid the over-optimistic of action taking and improve exploration efficiency to tackle the large uncertainty of the blast furnace. Third, to ensure operational safety, we have designed a regulation strategy that incorporates ε-greedy exploration and ICM, mitigating the risks posed by exploration actions. Finally, we validate the effectiveness of our method using on-site data, demonstrating that DDQN-ICM achieves superior performance in terms of tracking accuracy and disturbance rejection.