Research On Topology Identification Technology Of Distribution System Based On Advanced Measurement Infrastructure

With the construction and development of the new generation of power systems,the structure and operation of the distribution system have become more complex and changeable,bringing new challenges to the distribution network dispatching,safety analysis,and refined management.Accurate network topology is an indispensable input for power system load flow analysis,and it is also the prerequisite and basis for many other advanced operation management functions such as three-phase imbalance management,network loss reduction,and fault location identification.This dissertation proposes a novel topology identification method by deeply mining the data obtained from the advanced metering infrastructure(AMI).Firstly,the Markov Random Field(MRF)model is innovatively introduced to find the nodal correlations of voltage magnitudes for the distribution system topology identification.Unlike previous work to identify the topology by manually and individually analyzing the dependency between any neighboring nodes,MRF directly maps the overall relationship for the total distribution system,which avoids error propagation.And a revised maximum likelihood method is firstly proposed to solve the MRF model.The pseudo-log-likelihood method is proposed by replacing the global partition function with the local partition function.In this way,the computational complexity becomes linearly related to the node and sample number,rather than an exponential increase,which efficiently avoids dimension explosion.Besides,the L2 regularization method is proposed to employ regularization to penalize complex models.It reduces the negative impact of inessential parameters by adjusting their value,and thus reduces the tendency to over-fit the data.Secondly,the maximum correlation based iterative screening algorithms are proposed for medium-voltage and low-voltage distribution systems respectively.It generates a node-to-node adjacency matrix from the correlation matrix by iteratively selecting two nodes with the largest correlation,constructing nodal connections,generating the adjacency matrix,and finding the optimal radial topology of the distribution system.Finally,the topology identification model is implemented on edge computing units.The incremental learning model is proposed by continuously updating and putting data in terminals to dynamically train the topology generation model.The parallel programming model is proposed to divide the training model into sub-problems to efficiently implement the algorithm on multiple terminals.