Topology Identification Of Distribution Network Based On Machine Learning

To ensure the safe and stable operation of the distribution network amid the expansion of distributed energy and new load access,real-time monitoring and analysis must be strengthened.Accurate distribution network topology is essential for power system analysis,such as power flow calculation,state estimation,and reactive power optimization.Therefore,identifying the topology of the distribution network quickly and accurately is an urgent and important problem to solve.This article explores intelligent methods for topology identification in distribution networks through the research of "machine learning-based topology identification" with the following main work:1.A two-stage feature selection method is proposed,which first performs collinearity detection and then feature importance filtering.The first stage uses Spearman correlation coefficient and agglomerative hierarchical clustering to select features and remove collinear redundant features.The second stage uses Light GBM to calculate feature importance and remove features with lower importance.The proposed twostage feature selection method adds a collinearity detection step before feature importance analysis,which can effectively reduce the negative impact of collinearity problems on feature importance ranking and model interpretability.2.A Hyper GBM-based intelligent topology identification model for distribution networks is proposed.After two-stage feature selection,the feature data is input into the Hyper GBM model to train individual classifiers of XGBoost,Light GBM,and Cat Boost with different hyperparameters and weight coefficients.The optimal classifier combination is obtained to generate an ensemble model.The effectiveness and superiority of the proposed model are verified through case studies on the IEEE33 node distribution network system,including the possibility of using reduced features for topology identification and the adaptability to different noise levels and missing feature values.Simulation results show that the proposed model can efficiently identify the topology of radial and looped distribution networks with only a small amount of voltage measurement data and has good adaptability to different scenarios.3.A topology identification model for distribution networks based on graph attention networks is proposed.Firstly,the distribution network topology is abstracted as a line graph.Secondly,the edge features(the voltage amplitude difference between adjacent nodes in the distribution network)are extracted based on the node features in the topology graph and used as the corresponding vertex features in the line graph.Then,the preprocessed data is used as the input of multi-head graph attention layer,which aggregates the features between the voltage differences of nodes through weighted aggregation,takes the mean of the vertex features under multi-head attention,and outputs the classified state vector through fully connected layers and softmax layers.Finally,case studies on the IEEE33 node distribution network system show that the topology identification model based on graph attention networks has higher prediction accuracy compared to the intelligent topology identification method based on two-stage feature selection and Hyper GBM.