This paper proposes a data-driven framework for user–feeder topology identification in low-voltage residential power networks using ambient (current and voltage) measurements from smart meters. The framework first prepossesses the raw dataset via wavelet-based denoising, principal component analysis-based dimensionality reduction, and deep learning-based temporal feature extraction. In addition, a deep learning-based anomaly detection approach is also applied. Seven clustering algorithms are adopted for user–feeder relationship identification, and then the results are fused via a result-fusion strategy to enhance the identification accuracy further. Experiments on three real-world residential power networks demonstrate that the proposed approach significantly outperforms the results obtained by a single clustering method and the results obtained by simple voting-based fusion. The proposed approach achieves up to 88% identification accuracy in the considered case studies. Ablation studies are also conducted to validate the importance of each module in the proposed framework.
Guo et al. (Tue,) studied this question.