Anomaly detection and topology identification of distribution network based on conditional variational autoencoder

In the context of the construction of new power systems, intermittent distributed energy sources such as wind and photovoltaic power, as well as new source-load businesses like electric vehicles and energy storage, are increasingly being integrated into the grid, leading to a more complex distribution network topology. Currently, a significant number of non-automated switches remain in operation, with maintenance work primarily relying on manual topology information checks and updates, which can lead to discrepancies between topology records and actual conditions. Delays in updating or errors in updating the topology of distribution networks can adversely affect the normal operation and stability of the grid. This paper first employs a label-conditioned conditional variational autoencoder (CVAE)-based anomaly detection model to identify and eliminate anomalous samples in distribution-network data. Subsequently, a modified CVAE with an improved task formulation is proposed for topology identification. The topology-identification model is initialized using the parameters learned in the anomaly-detection stage, thereby transferring latent-space knowledge from anomaly screening to topology identification. The validation on a practical dataset demonstrates that the proposed strategy can effectively detect anomalous data and achieve high accuracy in topology identification. For the anomaly detection task, the proposed method attained the best performance of AUC (0.9873), TPR (0.9850), FPR (0.0135) and F1-score (0.9857) among comparative methods. In the topology identification task, it also significantly outperformed comparative approaches, as it achieved an AUC of 0.9865 and an F1-score of 0.9623 on the test dataset.