Tracing identification model of unbalanced load in distribution network based on graph attention network and feature extraction

• Proposed DGAT framework enables unbalanced load traceability in distribution networks efficiently. • Integrates dynamic gaph construction, joint frequency-spatiotemporal filtering, and lightweight deployment. • Achieves 95.7 % fault location accuracy, outperforming state-of-the-art by 8.1 % reliably. • Edge deployment attains 28FPS processing and low 9.3 W power for real-time diagnostics. The growing integration of renewable energy exacerbates load imbalance in distribution networks, creating complex challenges for real-time identification due to dynamic topologies, harmonic interference, and edge deployment constraints. To address this, we propose an innovative Dynamic Graph Attention Network (DGAT) framework. Its core innovations include an Electrical Sensitivity Coefficient-driven dynamic graph construction that jointly encodes topological and temporal electrical states, a Harmonic Suppression Module for frequency-domain noise filtering, and a physics-informed lightweight strategy guided by Electrical Centrality for edge deployment. Evaluated on the IEEE 123-node system, our model achieves a 95.7% localization accuracy, which outperforms benchmarks by 8.1% to 23.4%, and attains a 200 ms response time at 28 FPS with 9.3 W power in simulated edge environments. These results demonstrate DGAT’s potential as an effective solution for real-time, edge-based unbalanced load identification in smart distribution networks.