Abstract
Automated phase identification is critical for managing three-phase imbalance in modern low-voltage distribution networks (LVDNs) populated with electric vehicles and microgrids. This paper presents a data-driven approach using Principal Component Analysis (PCA) for noise reduction and DBSCAN for density-based clustering. Validated on real-world smart meter data, the method achieves 97.3% accuracy, offering a scalable alternative to manual on-site inspections.
Introduction
In modern smart cities, the surge of electric vehicle (EV) charging stations, residential microgrids, and solar panels has complicated the management of low-voltage distribution networks. These technologies often lead to three-phase imbalance, where load is unevenly distributed across the grid's three phases. This imbalance results in harmonic distortion, energy waste, and reduced equipment lifespan. To mitigate these issues, grid operators must accurately identify which phase (A, B, or C) each customer is connected to. Traditionally, this requires slow and error-prone manual wire tracing. This paper proposes a "topology-free" algorithm that utilizes existing smart meter voltage data to automate this process.
Methodology
The proposed identification pipeline consists of three primary stages designed to handle the high dimensionality and noise inherent in real-world grid data.
Data Preprocessing and Feature Extraction
Raw voltage time-series data often contains gaps and noise. The researchers employ Lagrange interpolation and low-rank matrix completion to ensure data continuity. Following preprocessing, Principal Component Analysis (PCA) is used to project the data into a lower-dimensional space. While PCA is often used for compression, here it serves as a crucial denoiser, filtering out high-frequency spikes and transient noise that do not represent underlying phase characteristics. By retaining 96% of the variance, the reconstructed voltage curves become significantly smoother, enhancing the "signature" of each phase.
Clustering and Mapping
Rather than using centroid-based methods like K-means, the authors utilize DBSCAN (Density-Based Spatial Clustering of Applications with Noise). DBSCAN identifies clusters based on point density, which is ideal for grid data as it can naturally isolate "noise" points—outlier customers with corrupted or atypical data—rather than forcing them into incorrect phase assignments. Finally, the centroids of these identified clusters are matched against reference voltage vectors from the transformer to assign the final Phase A, B, or C labels.
Experimental Results
The method was validated using a real-world dataset from a secondary substation serving 126 residential customers. Analysis was performed on 112 single-phase consumers with complete 15-minute interval voltage records, using utility archives as ground truth.
Performance Benchmarking
The PCA+DBSCAN approach achieved a 97.3% identification accuracy, correctly labeling 109 out of 112 users. It outperformed other machine learning combinations, including LargeVis+DBSCAN (96.1%) and t-SNE+K-means (94.6%). A key insight was the impact of PCA on feature independence; the average pairwise correlation between features dropped from 0.85 to 0.12, drastically improving the separability of the phases during clustering.
Discussion and Conclusion
While highly accurate, the study noted specific sensitivities in Phase A identification, where some users were either misassigned to Phase C or labeled as noise. Furthermore, the current mapping step still relies on a reference phase vector from the transformer. Future research will focus on refining noise-handling for these edge cases and exploring fully unsupervised mapping techniques that do not require a reference signal. Despite these limitations, the computational efficiency and high fidelity of the PCA+DBSCAN method make it a robust solution for real-time application in smart city power systems.
Summary updated