Research On Intelligent Control Technology Of Stirred Reactor System Based On Reinforcement Learning

Continuous Stirred Tank Reactor(CSTR)is a device which widely use in petroleum and chemical industries,it has a wide range of applications in chemical,biological and pharmaceutical fields.It has the advantages of simple structure,easy operation and high reaction efficiency,which can realize continuous production and improve production efficiency and product quality.Continuous Stirred Tank Reactor is usually exothermic or heat-absorbing reaction,which requires precise control of the reaction temperature.However,due to the fast flow of reaction materials,it is difficult to control the temperature,which may lead to the reduction of reaction efficiency and the quality of products.When the controller is unable to precisely control the temperature in the kettle,many plants use manual control,but the labor cost and waste resources are increased by this control method,it may cause better results with less effort,and may even produce serious safety hazards and on-site accidents.The reason why the reactor temperature is difficult to control is not only because of the heat absorption and exotherm of the reaction process,but also because of the influence of uncertainties such as environmental influences or external disturbances,the strong nonlinearity,large time lag and strong coupling are showed in the CSTR system.The production process of continuous stirred reactor reaction is taken in the petroleum industry as an example in this paper.Based on the research on CSTR model and control methods by domestic and foreign experts and scholars,the method of mechanism modeling is used,the CSTR model is constructed,and the control strategies is designed around the CSTR system by combining A3 C reinforcement learning(asynchronous advantage actor-critic,A3C),neural network and event-triggered control methods to design control strategies for CSTR systems,and the research is conducted around the modeling and temperature control problems of CSTR systems.The main research of this article is as follows.(1)The mechanism model in the CSTR reaction process is established,and the A3 C reinforcement learning method is selected as the CSTR temperature control method.By analyzing the basic structure,working principle and modeling analysis of CSTR,understanding the working principle and characteristics of CSTR,selecting the key factors of reactor temperature control,and using the mechanistic relationship between various parameters of CSTR(such as feed concentration,feed temperature,coolant temperature,etc.),a mechanistic modeling model building method is proposed.By understanding the development history of reinforcement learning,comparing and analyzing policy-based,value-based,AC,A2 C and A3 C etc.the advantages and disadvantages of methods,the A3 C reinforcement learning method is selected as the CSTR intelligent temperature control method for temperature control of the CSTR system.(2)The CSTR temperature control is implemented using reinforcement learning.Through the detailed introduction of environment model,global network,thread network,interaction mechanism and system temperature control implementation,fully understanding the principle of global network,thread network and interaction mechanism to perform temperature control on CSTR system.The simulation experimental results are showed that the temperature control for CSTR is better implemented with short and stable experimental time by using the A3 C algorithm and it possesses good global performance,but there are still problems in terms of large computation and resource consumption.(3)The resource occupation of the A3 C algorithm is solved by using the eventtriggered approach.First,by understanding the principle and development of eventtriggered control,understanding the event-triggering mechanism,introducing the modules of the event-triggered A3 C algorithm,determining the conditions for event triggering,and designing the control framework of the event-triggered A3 C algorithm based on event triggering;this type of control strategy only samples when certain rules are violated,the global network and thread network are updated only when the event is triggered,thus completing the system is sampled,and in the rest of the time the system does not sample,and the updates of the global and thread networks remain unchanged.Simulation results are showed that the event-driven A3C-based algorithm can greatly reduce the number of system samples,reduce the communication overhead of the system and reduce the computational overhead of the system,thus reduce the resource consumption of the system.