Real-Time Caustic Ratio Control in Alumina Digestion Using Deep Reinforcement Learning

Due to the lag time in the alumina digestion process (ALDP), traditional control methods that depend on delayed measurements often struggle to maintain the optimal caustic ratio, which is crucial for achieving optimal performance. This paper proposes a novel intelligent operational control method to address this challenge. Our multi-stage framework employs a fuzzy inference system (FIS) to generate an initial setpoint based on process knowledge. Subsequently, a bi-directional gated recurrent unit (BiGRU) network predicts the endpoint caustic ratio, assessing the validity of the initial setpoint. Finally, a deep deterministic policy gradient (DDPG) adjustor refines the setpoint in real-time based on the predicted and target values. Validated with industrial data, this method significantly improves the qualification rate of the caustic ratio compared to manual operation. It demonstrates the effectiveness of DDPG for real-time setpoint adjustment, leading to improved product quality in the ALDP.