PiST-Net: a physics-informed spatiotemporal framework for low-voltage distribution grid topology inference

Abstract
Accurate topology identification is pivotal for the effective management of modern low-voltage distribution networks. Existing methods often exhibit vulnerability to measurement noise, limited adaptability to dynamic reconfigurations, and high computational latency. To address these challenges, we propose PiST-Net, a novel physics-informed spatiotemporal framework that integrates Bidirectional Long Short-Term Memory (BiLSTM) and Transformer architectures. Specifically, the BiLSTM module extracts fine-grained local temporal features from nodal voltage sequences, while a Transformer encoder-decoder—augmented with physics-informed masking and symmetry constraints—reconstructs the global network connectivity. Two high-fidelity simulation datasets, PiST-Net_DS1 (fixed 16 nodes) and PiST-Net_DS2 (variable 2–16 nodes), were generated via MATLAB Simulink to validate the framework. Experimental results demonstrate that PiST-Net achieves superior global identification accuracies of 99.77% and 98.54% on DS1 and DS2, respectively, significantly outperforming traditional approaches and models relying solely on recurrent neural networks. By offering a robust balance between high accuracy and real-time inference speed, PiST-Net provides a scalable and physically consistent solution for fault localization and intelligent control in complex, next-generation smart grids.