Research On Optimal Design Of Power System Stabilizer Based On Reinforcement Learning

Power system is a large-scale nonlinear system with numerous energy sources and complicated connections.In the studies of it,the power system stability control is always a key issue.Power system stabilizer is a widely applied and maturely developed technique,which has been a basic method to enhance the power system stability.With the development of grid interconnection and the addition of new energy,it is getting more and more complicated about power system operation,which poses a difficult problem of the security and transient stability control of power system.It is necessary to find new methods to design new types of power system stabilizer of better performance.Reinforcement learning is a very suitable method for optimization and control problems.Reinforcement learning can find the global optimal value and get rid of the dependence on accurate physical model.The learning strategy is adaptive and scalable.In addition,a variety of control theories can be applied to the design of power system stabilizer,which can give full play to the unique advantages of their respective control theories.Therefore,this paper discusses the optimization design of power system stabilizer by uniting control and reinforcement learning.The main content of this paper includes:(1)Introduce the research status of power system stabilizer design,which elaborates the application of control theory and reinforcement learning.(2)Based on synergetic control and reinforcement learning,a model-free decentralized coordinated power system stabilizer is proposed.,The manifold is constructed,according to the synergetic control.And design it as an optimal control problem.The agent is trained by reinforcement learning.All the input signals are local signals,which can achieve decentralized coordinated control and avoid wide area communication problems.The designed PSS does not rely on the model,and has good robustness.(3)Based on model free adaptive control and reinforcement learning,a wide area PSS design method is introduced.Use the improved MFAC,and employ reinforcement learning to optimize the MFAC adaptive law.This method does not need the system model for both on-line and off-line design.The trained agent can realize the adaptive of the optimal controller parameters,and improve the initial value setting problem of the original MFAC controller,which can deal with the variation of oscillation characteristics in time-varying power system.