Accelerating real-time topology identification: A lightweight method via ranking-consistent knowledge distillation

Real-time and accurate topology identification (TI) is crucial for the safe and efficient operation of modern distribution networks, yet it is hindered by a fundamental trade-off since high-accuracy methods are often too computationally intensive for real-time applications or deployment on resource-constrained edge devices. To resolve this conflict, we propose a novel knowledge distillation method named Ranking-Consistent Knowledge Distillation (RCKD), which focuses on transferring the relational ranking of connection confidences from a teacher to a student model. The RCKD first trains a high-performance graph neural network as a teacher model to accurately learn the network’s topological characteristics. Subsequently, this topological knowledge is distilled into a highly lightweight multi-layer perceptron student model, significantly accelerating inference speed. The core innovation of RCKD lies in its unique distillation process. Unlike conventional methods that force the student to mimic the teacher’s exact output probabilities, RCKD compels the student to learn the relative ranking of connection confidence across all potential network edges. By preserving this relational knowledge, the RCKD effectively transfers the teacher’s topological reasoning without needing explicit graph structures or costly message-passing operations during inference. Experiments on the IEEE 33-bus and 123-bus systems show that RCKD reduces inference latency by more than an order of magnitude, and the resulting distilled model achieves superior accuracy under a properly selected teacher model, even surpassing its teacher, while also demonstrating strong robustness across various practical scenarios. Additional analyses on model complexity and teacher architecture further confirm the deployment potential and effectiveness of the proposed method for real-time resource-constrained applications. • Proposing a novel knowledge distillation method for real-time topology identification. • Transferring relative connection rankings to a lightweight student model. • Graph-free inference achieves over 10times speedup while maintaining high accuracy. • Demonstrating high robustness against noise, missing data, and DER integration.