Research On Operation And Control Strategy Of Distribution Network Based On Deep Reinforcement Learning

Actively promoting the clean and low-carbon transformation of energy and accelerating the construction of a new power system are important means for our country to achieve the "dual carbon" goals.As the final link of the power system,the distribution network connects the transmission network and the power load.In the face of the new situation where the composition of sources and loads in the distribution network is rapidly changing and the number of measurement devices is gradually increasing,how to reasonably utilize various new and old control resources such as distributed power sources,static reactive power compensators,and on-load tap changers has become an urgent problem in dealing with the increasing power fluctuation in the distribution network operation and management.In recent years,deep reinforcement learning(DRL)methods have gradually shown their potential in the power field due to their advantages in continuous control decision-making and complex mapping relationship learning.However,there are still some issues worth exploring.Therefore,studying the application of deep reinforcement learning in distribution network operation and management has great engineering significance and application value.This thesis aims to explore a more suitable DRL-based control model based on a distribution network simulation environment.Firstly,in order to establish the subsequent control models and generate operational decisions,the study focuses on data-driven power flow solution and particle swarm optimization methods to accelerate the training process of the DRL control model and calibrate its operational effectiveness.Brief introductions to relevant neural network knowledge and the DRL-based process for optimization and operational control of distribution networks are provided,laying a theoretical foundation for future research.Next,to address the optimization and operational control issues in distribution networks,a DRL control model for distribution network operation and management is investigated,which relies solely on controllable devices(photovoltaic inverters and static reactive power compensators)for control.After introducing the control problem and the control target,DRLbased distribution network operation and management models based on DDPG,TD3,and SAC are explored,showcasing the design processes of reward functions,networks,and parameters.A comparative analysis of the three methods is conducted through experiments in three different scenarios,demonstrating the effectiveness of the DRL control model.Lastly,to fully utilize controllable resources for operational control of distribution networks,a DRL control model is developed that incorporates both discrete and continuous devices as control means(including photovoltaic inverters,static reactive power compensators,and on-load tap changers).The handling method for discrete OLTC actions is presented,and a mixed-action(discrete-continuous action)control model for distribution network devices based on SAC is designed.Experimental validation is conducted,and the results confirm the effectiveness of the research method.The content of this thesis has certain reference significance and engineering value for the design of DRL-based control models for distribution networks.