Robust Reinforcement Learning And Its Application For Decoupling Control Of Gas Collector Pressure Of Coke Oven

In coking process, the stability of gas collector pressure influencesthe coke quality, the life-time of ovens and the producing environment,and its control have a direct impact to the operating condition of thewhole coke oven system. However, the control system of gas collectorpressure is a complicated multivariable, nonlinear, time-varying and bigtime-dealy control object, which have many disturbance factor and thecoupling phenomenon is serious. Therefore, it is difficult to establish amathematical model to reflect working situation of the control system ofgas collector pressure accurately and obtain the desired control effectusing traditional automatic control methods.In order to solve the above problems, the process of coke oven gasgathering and control requirements are analyzed in the paper, and find outthe relevant factors and coupling relationship which influence thepressure. On this basis，a dynamic mechanistic model of coke oven gascollector is established. The key parameters such as butterfly valueopening and gas production are analyzed by model simulation. Thesimulation laid the foundation for the study on control strategy.In the control strategies, the optimal control strategy based onRobust Reinforcement learning algorithm is designed, which based on therelated theory of reinforcement learning and the Robust control theory.Agent through reinforcement learning gain the optimal control strategy,but during the search good strategy learning process, try many possibleperformance is bad strategy, and these strategies may make the system isnot stable. In order to avoid these strategies or action research, the maingoal is to generate stable action the using reinforcement learning optimalcontrol behavior.This method make replace Neural network’s nonlinear andtime-varying as the uncertainty in robust control theory, Using Lyapunovstability theory, Structure suitable Lyapunov-Krasovskii functional，through the robust constraints ensure that the designed system is stable, like this the neural network remain stable when it is learning.Applied tothe coke oven gas collector pressure system, and provide this kind ofstrong coupling, nonlinear and strong disturbance and uncertaindistributed decoupling control for system solutions. MATLAB simulationresults show that the control strategy is effective.