Transformer-customer relationship identification based on deep Gaussian mixture model in low-voltage distribution system

Abstract

Accurate transformer-customer relationship is critical for better operation and management of low-voltage distribution system. It is of high cost to establish and check the profiles of transformer-customer relationship by manual maintenance, which is often mistakenly archived too. This paper proposes a novel automatic identification method based on a deep Gaussian mixture model with voltage data. The improved model is built with feature extraction network, clustering network and splitting and merging network, which can adjust the number of transformers adaptively and dynamically. It can also identify outliers of the data and effectively solve the uncertainty and imbalance of samples. Finally, the proposed method is tested in a practical power supply district which has 2752 customers with 9 transformers. Five different test cases are set to verify its effectiveness. The method can dynamically adjust to the actual number of transformer under different initial clustering numbers. The experimental results demonstrate that our model achieves an accuracy of up to 98.8% in profiling unlabeled users, while achieving an error correction accuracy of 99% in the case of imbalanced users in the transformer area. To further validate the superiority of our model, we conducted robustness experiments under different user scales and data loss scenarios. The results consistently outperform current state-of-the-art algorithms.