To tackle the challenges of low accuracy in estimating the vertical offset distance of vibration sources in underground power optical cables and the difficulty in adapting to complex propagation media, we propose Spatio-Temporal Orthogonal Transformer (STOTFormer) based on phase-sensitive Optical Time Domain Reflectometry ( \(\phi \) -OTDR). The method adopts Patch Embedding modules and a dual-branch Transformer architecture to separately extract temporal and spatial features from vibration signals, and integrates a Multi-Scale Spatio-Temporal Fusion (MSSTF) with a Cross-Orthogonality Loss to enable precise modeling of nonlinear propagation patterns. In real buried experiments with vertical distances ranging from 0 to 5 m, STOTFormer outperforms physical models based on TDOA, MUSIC, and FDOA, machine learning models (SVM, Random Forest, k-Nearest Neighbor, and Adaboost) and deep learning models based on CNN models (CNN-SVM, CNN-TDOA), and Transformer models (Swin-T, ViT). For the three vibration events (digging, falling objects, hammer), STOTFormer achieves the highest localization accuracy within ±0.5 m, ±0.3 m, and ±0.1 m thresholds, reaching 99.72%, 98.60%, and 79.78%, respectively, demonstrating excellent consistency across distances and strong adaptability to complex underground environments.