| The power grid guarantees the daily electricity consumption of human beings,and has the characteristics of high dispatching cost and intensive technology.It is of great significance to conduct real-time dispatching of the power grid under different operating conditions to achieve stable power transmission.Faced with the problems of strong interference,high coupling,and multi-time variation in the power transmission process,the grid dispatching method based on expert experience modeling and analysis has limited adaptability to the multi-variable grid topology.Therefore,this topic focuses on resisting sudden power grid disconnection.The optimization problem of power grid dispatching strategy under faults,a power grid system dispatching method based on reinforcement learning is studied,and the research work is mainly carried out in the following three aspects:(1)Construct the method framework of "imitation learning + reinforcement learning",and use expert experience data to reduce the cost of online learning;(2)Design a simulated operation mechanism to rehearse the power grid scheduling process,improve the fault tolerance of the scheduling strategy,and reduce the scheduling cost;(3)Aiming at the problems of large power grid scheduling action space,inaccurate disconnection and reconnection actions,and single reward basis,three power grids improvement methods are proposed: multi-level reduction model,disconnection and reconnection action priority selection mechanism,and complex environment adaptive reward function.Simulation experiments show that the proposed method for real-time dispatching optimization of power grid has good robustness when the power grid is disconnected.Compared with the basic dispatching model method,the success rate of comprehensive dispatching optimization is increased by more than 2 times.The method also achieved the third highest score in the L2 RPN NEURIPS 2020 test of the Codalab open platform(as of May 2022). |
PDF Full Text Request