From Local Patterns to Global Context: A Multimodal Deep Learning Approach for Complex Power Quality Disturbance Recognition

ABSTRACT The increasing penetration of renewable energy sources introduces complex and mixed power quality disturbances (PQDs) that challenge traditional diagnostic approaches. To address the limitations of convolutional neural networks (CNN) in capturing long‐range temporal dependencies, this study proposes an intelligent classification framework integrating fast Fourier transform, one‐dimensional CNN, lightweight Bidirectional Encoder Representations from Transformers (LBERT1d) and a signal‐based cross‐attention (SCA) mechanism. The proposed MCNN1d‐LBERT1d‐SCA framework leverages multimodal time–frequency feature fusion, combining local and global representations to enhance the recognition of concurrent and nonstationary PQDs. A synthetic dataset following IEEE Std. 1159 was constructed encompassing 25 disturbance types with multiple signal‐to‐noise ratios, to ensure robustness and generalisation. Experimental results demonstrate acceptable performance, achieving an average accuracy of 99.30% on the synthetic dataset and maintaining better reliability under noise conditions down to 20 dB. Validation using real‐world IEEE PES data and MATLAB/Simulink simulations yielded accuracies of 95.88% and 97.47%, respectively, confirming the model's strong adaptability and real‐time capability. These results indicate that the proposed hybrid deep learning framework offers a practical and scalable solution for intelligent PQD monitoring contributing to the reliability and stability of modern power systems.