Cross-attention-enabled long short-term memory for detection and classification of multiple power quality disturbance events

The integration of new generation sources and loads has transformed the power grid into a more flexible and interactive system, though it has also disturbed power quality. Consequently, accurate detection and classification of Power Quality Disturbances (PQDs) are essential for implementing effective corrective actions in smart grids. However, existing methods often struggle to simultaneously capture complex spatio-temporal dependencies and maintain computational efficiency, particularly for combined disturbances where multiple fault patterns overlap. To address this gap, this paper proposes a novel hybrid deep learning framework. The novelty lies in the synergistic integration of a powerful pre-trained Convolutional Neural Network (EfficientNet-B7) for efficient spatial feature extraction, a Bidirectional Long Short-Term Memory (Bi-LSTM) for deep temporal modeling, and a Cross (X)-attention mechanism to adaptively focus on the most discriminative features. A rigorous evaluation confirms the effectiveness of this approach. The model achieves an overall accuracy of 99.62%, precision of 99.7%, recall of 99.51%, and an F1-Score of 99.51% on a clean, balanced dataset. The model's robustness is further demonstrated by its high accuracy on unbalanced data (99.01%) and under challenging 20 dB noise conditions (97.95%). With a rapid average inference time of less than 5 ms, these findings establish that our proposed framework provides a compelling combination of accuracy, robustness, and efficiency, making it a highly promising solution for real-time monitoring in modern smart grid systems.