Research On Distribution Network Reconfiguration Strategy Based On Reinforcement Learning

In recent years,with the rapid development of China's economy,the power load has increased year by year,and the peak load has increased significantly,which has seriously affected the safety and reliability of the distribution network.At the same time,traditional fossil energy is gradually depleted and environmental problems are becoming more and more serious.For these reasons,a large number of renewable distributed generations have been connected to the grid,and the renewable distributed generation grid-connected has changed the power flow and voltage of the traditional distribution network.The structure of the distribution network is becoming more and more complex,and the loss of the network is increasing.Reducing network losses is critical to reducing grid operating costs.Distribution network reconfiguration is the most basic and effective method to optimize the network structure.It is of great significance to balance load,eliminate overload,reduce network loss,improve voltage quality,and improve power supply reliability and economy.Therefore,it is of great significance to study the distribution network reconfiguration with distributed generation access.The main work of this thesis includes:(1)The research background and the concept of distribution network reconfiguration are introduced,the necessity and feasibility of distribution network reconfiguration are analyzed in this thesis.The research status of the reconfiguration model and algorithm,the development and application of deep reinforcement learning are briefly discussed in this thesis.(2)A variable encoding method based on the basic loop is designed.When all tie switches and section switches are closed,the distribution network structure is essentially a connected graph with loops from the point of view of graph theory.Since any reconfiguration result that meets the requirements corresponds to a radial network structure,that is,any network reconfiguration result that meets the requirements corresponds to a spanning tree of the connected graph with loops.Therefore,the second chapter of this thesis introduces a method for solving the number of spanning trees for connected graphs with loops,which provides a feasible solution space for the switching actions during the reconfiguration of the distribution network in Chapter 4.(3)Second-order cone programming method is used to reconfiguration the distribution network.The convex nonlinear power flow equation constraint is transformed into the standard form of second-order cone programming.The basic method is to first introduce slack variables,and use the second-order cone relaxation technique to perform convex relaxation on the inequality constraints with slack variables and quadratic equation constraints to form two Order cone constraint;then replace the variables in the power flow equation with the square term,and propose a model conversion method based on second-order cone relaxation and variable replacement linearization,which converts the original nonlinear equations into linear equations.This strategy successfully converts the original non-convex optimization model into a mixed-integer second-order cone programming problem,which can be solved by convex optimization algorithm package to obtain the global optimal solution.(4)A NoisyNet deep Q-learning reinforcement learning algorithm is used to study the strategy of distribution network reconfiguration.The traditional model-based distribution network reconfiguration method needs to optimize the network structure according to the network parameters.This thesis adopts a method of distribution network reconfiguration based on model-free reinforcement learning.NoisyNet DQN can realize automatic exploration without adjusting the agent's exploration parameters,which speeds up the training process of the agent.Through analysis,the distribution network reconfiguration optimization performance based on the NoisyNet DQN algorithm is better.