Research On Data-Driven Volt-VAR Control Strategy Of Active Distribution Network

With the proposal of the national "carbon peaking and carbon neutrality" goal and the development of renewable energy technologies,the penetration rate of distributed power sources in the distribution network,mainly composed of new energy sources,is gradually increasing.The proportion of new types of loads represented by electric vehicles on the demand side has also significantly increased.The uncertainty of the power source and load,as well as the diversity of controllable resources,make the distribution network exhibit the characteristics of diverse states,flexible structures,and complex operation control,posing significant challenges to the economic,safe,and reliable operation of the distribution network.The thesis is based on traditional reactive power optimization of the distribution network,relying on digital technology in the power system and using data empowerment as the basic idea.It focuses on the voltage control problem of the distribution network under new source-load access and combines the multiple controllable elements of the distribution network to improve the voltage quality,reduce network losses,and enhance the economic and security operation of the distribution network.The main research content is as follows:(1)To address the problem of voltage fluctuation in the distribution network caused by the uncertainty of new energy sources and loads,and the difficulty of obtaining accurate model parameters of the distribution network,a multi-intelligent body reactive voltage control method for the distribution network considering the participation of new energy inverters is proposed.Based on reinforcement learning theory,the reactive voltage optimization problem is transformed into a Markov game problem and solved by a deep deterministic policy gradient algorithm.The method is data-driven and does not rely on accurate tide modeling and source-load prediction of the distribution network,and makes decisions based only on real-time state information observations of the distribution network.At the same time,it combines the framework of centralized training and decentralized execution of multi-intelligent reinforcement learning,and the regulation unit decision is based only on grid state information acquisition and neural network feedforward operation,which has better real-time performance;(2)To fully exploit and utilize grid data,a sample-efficient data-driven active-reactive power coordination control strategy for the distribution network is proposed.The active power regulation ability of electric vehicle charging stations is taken into consideration.Firstly,based on a travel chain prediction method,the charging power curve of electric vehicles is obtained.Then,combined with reactive power regulation devices such as new energy inverters,a real-time optimization Markov game process for active-reactive power coordination in the distribution network is established.Secondly,a reinforcement learning experience enhancement method is proposed,and a new multi-agent deep reinforcement learning algorithm is developed,which effectively addresses the low sample efficiency and high sample demand challenges of general reinforcement learning methods.The algorithm combines attention mechanisms to improve control performance and enhance the practicality of data-driven methods for voltage optimization problems.(3)To address the problems of high dimensionality and lack of real-time performance of centralized regulation and control problems in distribution networks,a reactive voltage in-place control framework for distribution networks based on the cloud-side cooperative architecture is proposed,and a partitioned two-stage reactive voltage optimization model is established under this framework.First,a multi-scenario hierarchical clustering distribution grid partitioning method based on voltage sensitivity is proposed to cope with the challenge of source-load uncertainty on static partitioning accuracy and to realize decoupling of voltage relationships within each voltage region as the basis for partitioning in-place control;then,a mathematical model-based approach is used to solve the slow time-scale day-ahead regulation problem and to complete real-time intra-day dispatching based on a multi-intelligent reinforcement learning method.The method uses edge computing device as the main body of regional control,and each sub-region only needs to independently observe the state information in the corresponding region during decision-making,without global information acquisition and cloud communication,which effectively reduces the action and observation dimension of the intelligent body and has better real-time performance.(4)To address the problems of lack of interaction of zone-in-place control to cope with structural state changes of distribution networks,insufficient global optimality and dependence on zoning accuracy,a distributed real-time control method of voltage in distribution networks based on graph reinforcement learning method is proposed;firstly,a distributed voltage zone autonomous control framework of distribution networks is proposed in conjunction with the edge-to-edge cooperative architecture,and a partially observable Markov decision process is established.The process belongs to a partially observable problem in the field of reinforcement learning and involves weak communication of data sharing between adjacent regions,which is difficult to solve by reinforcement learning methods based on centralized training and decentralized decision making.Therefore,a multi-intelligent reinforcement learning method based on decentralized training,decentralized decision-making framework with hierarchical graphical recurrent networks is proposed.The method allows neighboring regional intelligences to share state observation data and empowers each intelligence to make distributed decisions.Each region can achieve regional voltage control and optimize the global network loss of the distribution network based on its own state data and communication with neighboring regions,realizing regional distributed autonomy without complex communication.