Research On Control Strategy Of Fuel Cell Gas Supply System Based On Deep Reinforcement Learning

With the increasingly serious problems of energy shortage and environmental pollution,new energy sources represented by fuel cells have been widely used.Among them,the proton exchange membrane fuel cell(PEMFC)is broadly applied in many fields represented by the electric vehicle industry because of the advantages of higher output power density and mass power,non-pollution,rapid start-stop,and high reliability.The fuel cell engine system uses a stack and auxiliary subsystems to convert the chemical energy of the hydrogen fuel and oxidizer into electrical output to power the vehicle.In an in-vehicle environment,the frequently changing power demand puts forward high requirements for the powerability of fuel cells.Whether the output power of the fuel cell system can meet the demand mainly depends on whether the reactants can be adequately supplied,which is determined by the control effect of the gas supply system.For nonlinear fuel cell gas supply system with high complexity and large time lag,the current commonly used control strategies have poor dynamic performance,which makes it difficult to achieve effective real-time supply of both gases,leading to the reduction of output performance and even damaging the cell structure to produce safety hazards.To address these problems,considering that deep reinforcement learning algorithms can achieve real-time tracking control of nonlinear systems,a control strategy based on deep reinforcement learning is proposed,optimized and tested in this dissertation.The main research work and contributions are as follows:(1)For the multi-parameter and multi-variable fuel cell gas supply system,a mathematical model of the fuel cell gas supply system is established based on existing literatures and actual fuel cell engine systems.Considering the nonlinearity and high complexity of the system,a control-oriented simulation model of the fuel cell gas supply system is constructed with the help of parameter fitting and other methods,and the electrical characteristics of the stack and the output characteristics of some key elementary components are analyzed.(2)A control strategy based on deep reinforcement learning is proposed for the flow characteristics of fuel cell gas supply system,with the oxygen excess rate and hydrogen excess rate as the control indexes,air compressor and hydrogen circulation pump as the control objects,and the desired control index value obtained from offline simulation as the control target.In addition,the pressure difference between the two poles of the stack is controlled by separate follower control to reduce the control complexity.Meanwhile,considering the coupling between the two gases,two control strategies,individual and coordinated,are designed.On this basis,several typical deep reinforcement learning algorithms are trained offline and tested online.The comparison experimental results show that the overall dynamic control effect of the joint deep reinforcement learning controller based on deep deterministic policy gradient is better.(3)To further improve the steady-state performance of the controller,a fusion strategy is used to optimize the deep reinforcement learning controller.The optimized control strategy is compared with feedforward control,fuzzy PID control and traditional deep reinforcement learning algorithm in simulation tests.The experimental results show that the control strategy can obtain good steady-state performance improvement on the basis of securing the dynamic response speed of the system.(4)Based on the hardware-in-loop devices and controllers,a hardware-in-loop test system is designed by using a simulation model of the fuel cell gas supply system,and is used to perform hardware-in-loop testing of the control strategy proposed in this dissertation.It is further verified that the control strategy is feasible and effective,with good dynamic and static control effects,and can be applied to the practical fuel cell engine system.