A Physics-Guided Lightweight Multi-Task Learning Framework for Transmission Line Fault Diagnosis

Accurate transmission-line fault diagnosis is critical for reliable power-system operation and rapid restoration. This paper proposes a Squeeze-and-Excitation Lightweight Multi-Task Learning (SE-Lite-MTL) framework for simultaneous fault-type classification, faulted-line identification, and spatial localization. Evaluated on a ۱۱۰-kV transmission-line dataset, the proposed method bypasses traditional statistical dimensionality reduction. Instead, it integrates raw electrical measurements with explicitly constructed physics-guided characteristics, creating a robust feature space further stabilized by targeted data augmentation. The architecture utilizes a compact feature-recalibration mechanism and a shared residual representation, requiring only ۲۴,۸۹۱ trainable parameters. Evaluated over ten independent stratified runs, the SE-Lite-MTL framework achieves a fault-type classification accuracy of ۱۰۰.۰۰%, a faulted-line identification accuracy of ۹۷.۸۵%, and a mean position error (MAE) of ۰.۹۴۸۴%. Compared to the baseline benchmark, the proposed architecture improves line identification and reduces absolute localization error by approximately ۱۷.۵%, while achieving an ۸۰% reduction in computational complexity. With a sub-millisecond inference time, the SE-Lite-MTL establishes a highly efficient accuracy-complexity balance, demonstrating substantial potential for real-time diagnostic deployment in modern digital substations.

کلیدواژه ها:

Transmission-line fault diagnosis ، Multi-task learning ، Physics-guided features ، Squeeze-and-Excitation ، Lightweight neural networks