Qualitative Process Knowledge-Assisted Deep Reinforcement Learning Approach for Industrial Process Control

In modern industrial processes, high-quality controllers play a vital role in meeting the demands of the growing number of high-quality products. In this paper, a qualitative process knowledge-assisted deep reinforcement learning approach is proposed for the optimization control of the industrial processes. The core of this approach is to make use of the known qualitative process knowledge to simplify the policy network in deep reinforcement learning (DRL), thereby expediting the learning process of DRL. Initially, an appropriate basic controller is selected according to the characteristics of the process. Subsequently, based on the known qualitative structure information of the basic controller, the basic controller is transformed into a neural network-based form, which replaces the intricate policy network in traditional DRL algorithms. Furthermore, the DRL parameter update mechanism is utilized to adaptively update the network parameters of the basic controller, thus realizing the optimization control of the process. Finally, the feedwater control of the feedwater system in a coalfired power plant is employed to verify the effectiveness of the proposed method. Simulation experiments demonstrate that, in contrast to the traditional DRL method, the proposed method boasts a faster learning speed while attaining acceptable control performance and mitigates the learning complexity of DRL.