Standard LSTM performance metrics.

Partial discharge (PD) detection in power transformers is critical for preventing insulation failures in modern power grids, yet remains challenging due to signal complexity and environmental noise. Existing methods struggle with accurate PD classification under strong electromagnetic interference and varying load conditions. This study proposes a novel hybrid Acoustic-VMD and CNN-LSTM model featuring: (1) sample entropy-optimized variational mode decomposition (automatically determining modes and penalty factor), (2) parallel 1D-CNN (5 layers and bidirectional LSTM (2 layers, 256 units) branches, and (3) hierarchical attention mechanisms (8 heads) for dynamic feature fusion. Experimental results demonstrate superior performance with 96.2% classification accuracy for multi-source defects (38% improvement over wavelet methods), 5.8mm mean absolute localization error (53% better than TDOA), and consistent 4.2° angular accuracy under high noise, while maintaining practical 0.8s processing time. The research conclusively establishes that synergistic integration of adaptive signal processing and attention-based deep learning significantly advances PD diagnostics, achieving both computational efficiency and robust performance in complex operational environments.