Low voltage topology mapping through network discovery events applied to AI-based digital twins

Accurate knowledge of Low-Voltage (LV) distribution topology is critical for reliable operation, advanced monitoring and large-scale integration of distributed energy resources (DERs). In practice, topology records in GIS/NIS are frequently incomplete or outdated, while field verification remains costly. This paper presents a smart-meter (SM)–driven methodology for LV topology mapping that combines a data-trained surrogate model with physics-inspired sensitivity analysis. A feedforward Deep Neural Network (DNN), trained on historical SM measurements spanning diverse operating conditions (including DER-driven net generation), is used as a model-free digital twin to emulate customer-to-voltage relationships. Virtual Network Discovery Events (NDEs) are then generated by applying controlled perturbations within the surrogate to obtain voltage-response signatures that support topology inference without physical intervention. Phase groups are identified through dimensionality reduction and hierarchical clustering, and customer connectivity is inferred from the similarity structure of the resulting voltage-response signatures. The method is applied independently per feeder, enabling scalable execution across multi-feeder LV networks. Validation on six real feeders from an urban Spanish network demonstrates accurate voltage emulation and high-fidelity phase and topology reconstruction using only existing SM infrastructure. • Network Discovery Events enable adaptive low-voltage topology mapping. • Hybrid AI–physics approach enhances low-voltage Digital Twins. • Validated on real Spanish distribution network with high precision.