专家经验引导分层强化学习的配电网灾后协同恢复策略OA

[Objective]Extreme disasters are prone to trigger multi-point faults in distribution networks,and the lack of coordination between network reconfiguration and fault emergency repair will degrade the efficiency of fault restoration.This paper proposes a coordinated post-disaster restoration strategy for distribution networks based on expert experience-guided hierarchical reinforcement learning(HRL).[Methods]Firstly,a decision-making model for post-disaster coordinated restoration of distribution networks is constructed.Taking the minimization of load loss,switch operation cost and maintenance travel cost as the optimization objectives,the model comprehensively considers the operation constraints,network topology constraints and fault repair constraints of distribution networks,as well as the time-varying traffic accessibility constraints of transportation networks.Secondly,targeting the sequential decision-making characteristics of post-disaster restoration,a hierarchical reinforcement learning decision framework consisting of high-level task selection and low-level action execution is established.The high-level layer selects network reconfiguration or repair scheduling tasks according to real-time system states,while the low-level layer generates topology reconfiguration operations,repair task allocation schemes and route adjustment strategies under corresponding constraint conditions.Finally,expert trajectories are adopted to pre-train the policy network,and the parameters of the hierarchical policy are continuously updated through online interactive training.[Results]Random disaster scenario tests on a coupled system of the modified IEEE 33-bus distribution network and the Sioux-Falls traffic network demonstrate that the average cumulative load loss of the proposed method is 5.45 MWh,which is reduced by 22.43%and 29.57%compared with the mixed-integer second-order cone programming(MISOCP)and binary particle swarm optimization(BPSO)algorithms,respectively.The average online solution time of the proposed method is 0.032 7 s,which is significantly lower than that of the comparative methods.[Conclusions]The proposed method realizes the coordinated optimization of post-disaster network reconfiguration and repair scheduling for distribution networks.It exhibits superior performance in reducing cumulative load loss and improving online decision-making efficiency,and can provide auxiliary decision support for the rapid restoration of distribution networks under extreme disaster conditions.