Research On Multi-Microgrid Economice Dispactching Based On Hybrid Heuristic Deep Reinforcement Learning Algorithm

Large-scale distributed energy sources directly connected with the power system will reduce the operational reliability of large power grids.Microgrids are autonomous systems that integrate distributed energy sources.The orderly connection of microgrids to the power grid can improve the flexibility of the power system.Interconnecting geographically adjacent microgrids to form a multi-microgrid system can achieve energy complementarity between microgrids,improve energy supply reliability,and reduce the impact of uncertainties in microgrids on the upper-level energy supply system.The economic dispatch of microgrid and multi-microgrid system directly affects the economy of system operation.Increased awareness of data protection among stakeholders poses challenges to traditional centralized control architectures.Hence,a new dispatching architecture needs to be investigated to achieve decentralized autonomy of microgrids and optimal coordinated operation of a multi-microgrid system.The research content and innovations of the paper are as follows.(1)Based on the mathematical model of combined cooling,heating,and power microgrid,this paper establishes a centralized-decentralized two-level dispatching architecture for combined cooling,heating,and power multimicrogrid system.The proposed centralized-decentralized architecture enables decentralized autonomy of each microgrid and optimal coordinated operation of a multi-microgrid system,avoiding the limitations of the traditional centralized optimization architecture that requires access to global information.The simulation results verify that the double-layer deep reinforcement learning algorithm has higher search accuracy and search speed than the traditional heuristic algorithm.(2)The coupling of different types of energy sources increases the difficulty of dispatching and controlling the combined cooling,heating and power microgrid.The classical mathematical programming methods have the advantages of high solution efficiency and reliable convergence,but these methods are not easy to solve nonlinear and nonconvex problems directly.Traditional heuristic algorithms do not depend on the mathematical form of the optimization objective,but the computational time of such algorithms is long.Based on the above analysis,a double-layer deep reinforcement learning algorithm is proposed in this study,where the upper layer is utilized to solve the initial operation strategy of controllable units for multiple time periods,and the lower layer is employed to perform the adjustment of the operation strategy of controllable units for single time periods.(3)This study combines a double-layer deep reinforcement learning algorithm with a dual population evolutionary game genetic algorithm to form a hybrid heuristic deep reinforcement learning algorithm.The hybrid heuristic deep reinforcement learning algorithm is applied to solve the centralized-decentralized two-level dispatching architecture proposed in the previous section.The simulation results show that the optimal coordinated operation of combined cooling,heating and power multi-microgrid system can effectively reduce the operation cost of multi-microgrid system and the load power fluctuation at the point of common coupling.Compared with the traditional centralized heuristic algorithms,the hybrid heuristic deep reinforcement learning algorithm improves the efficiency of model solving and the economic efficiency of multi-microgrid systems.Then,this study reduces the impact of forecast information deviations of a multi-microgrid system on the upper-level energy supply system by adjusting the strategy in the intra-day rolling horizon dispatch stage to reduce the load power fluctuation at the point of common coupling.Finally,this study applies the centralized-decentralized two-level dispatching architecture with the hybrid heuristic deep reinforcement learning algorithm to a multi-microgrid system considering power flow constraints,and verifies through simulation experiments the conclusion that the proposed method can reduce the total operating cost of the multi-microgrid system while accessing only local information and ensuring that the node voltage and line power do not cross the limits.