Distributed Reinforcement Learning Algorithm For Dynamic Generation Dispatch Of Automatic Generation Control Based On Virtual Generation Tribe

To adapt the trend in development of the smart Grid from centralized to decentralized,a two-layer decentralized generation command dispatch(GCD)of automatic generation control(AGC)with electric vehicle charging stations based on virtual generation tribe(VGT)is adopted to resolve the curse of dimension emerged in large-scale power systems.Then the first layer of decentralized GCD is undertaken among VGTs,after the chief obtains the generation command of its own VGT,each unit can autonomously calculate the generation command by communicating with its adjacent units of that VGT.And the CTQ is adopted for the first layer of decentralized GCD,while a consensus algorithm is used for the second layer of decentralized GCD to solve the problem of long convergence time effectively caused by the large scale of the AGC units.When taking into account EV charging stations to participate in the second-layer of decentralized GCD,give priority to EV to participate due to its short delay,the remaining power shortage is borne by the traditional AGC units.In particular,all charging stations achieve a consensus on regulation costs,while each traditional unit calculate its own generation command based on the ramp time.Virtual consensus variables are adopted to exchange information among the agents while the actual consensus variables are used to calculate the generation command,which provides more flexibility of the proposed algorithm.Simulations of Hainan power grid prove that the proposed algorithm not only can reduce the regulation costs and enhance the AGC performance,but also effectively achieve an autonomous frequency regulation of electric vehicles and traditional units.When the CTQ is developed to solve the first layer model for dynamic generation dispatch,each VGT collaborates with its adjacent VGTs through exchanging the value-matrix function,then the generation command of each VGT can be calculated autonomously,so that the decentralized and autonomous AGC is achieved.Furthermore,the CTQ can be adopted for fast optimization of dynamic generation dispatch to meet the requirement of AGC control cycle after introducing the transfer learning,which can utilize the effective information of historical optimizing task.Simulations of Guangdong power grid prove that the proposed algorithm can effectively handle the decentralized optimization problem of dynamic generation dispatch for complex large-scale power grid compared with centralized algorithms.Moreover,the CTQ can enhance the AGC performance and reduce the regulation costs.