Reinforcement Learning based parameter optimization of cascade control loop in the Coke Dry Quenching System

Waste heat boiler is an important part of dry quenching power generation, and its steam drum level cascade three impulse system is characterized by strong perturbation, non-linearity, strong coupling and multiple working conditions. In this paper, a reinforcement learning based parameter optimization of cascade control loop in the coke dry quenching system is proposed for such series control problems. Based on the uncertainty of the state combination between the control parameters of the cascade loop and the liquid level of the steam, this paper uses a BP neural network to model the controller parameters and optimization indexes, uses the NSGA-II algorithm to select the design factors of the system, transforms the PID parameter optimization problem into a combination optimization problem, and finally using the reinforcement learning mechanism gives the optimal control strategy under multiple operating conditions. The simulation experiments are carried out through the data of a 180t/h dry quench coke waste heat boiler steam drum level system, and the results show that compared with similar controllers, the control strategy proposed in this paper can significantly reduce the maximum overshooting amount with a small increment in the rise time.