Power quality disturbance classification via a time-frequency feature-fused transformer model with cross-attention mechanism

• Innovative feature extraction method: This study proposes a novel classification framework that combines time-frequency domain analysis with Transformer cross-attention architecture, achieving more efficient feature extraction. • Application of cross-attention mechanism: By introducing the cross-attention mechanism, this method can more effectively capture the correlations and complementarities between different feature branches, significantly improving classification accuracy while maintaining lower model complexity. • High-precision classification results: The method demonstrates high classification accuracy across various types of power quality disturbances, with fewer parameters, making it capable of excellent classification performance even under limited computational resources, which is significant for monitoring and managing power systems. • Enhanced robustness: Through joint time-frequency domain processing, the method exhibits better robustness against noise and power signals under different operating conditions, while also excelling in controlling model complexity. • Potential for real-time applications: Due to its high computational efficiency, the method is suitable for real-time monitoring of power quality disturbances, ensuring grid stability while reducing hardware burden, thus having practical application value. With the increasing penetration of renewable energy in power systems, power quality disturbances (PQDs) exhibit composite patterns and dynamic time-varying characteristics. To address the challenges of dimension redundancy and alignment bias in traditional multimodal feature fusion, this study proposes a novel classification framework that integrates time-frequency domain analysis with a Transformer-cross attention architecture. The core innovation lies in its dual-channel parallel processing design: capturing local temporal features through convolutional neural networks (CNNs), while encoding global dependencies in the frequency domain via a Transformer encoder, followed by dynamic alignment and adaptive cross-domain fusion through a learnable cross-attention mechanism. The proposed method not only demonstrates significantly enhanced robustness in noisy environments but also exhibits clear advantages in terms of computational complexity and cost. Through a comparative analysis using a hardware experimental platform, the model was optimized in resource utilization when compared to mainstream PQD classification methods such as CNN-Transformer and CNN-BiGRU-Attention. This optimization led to a significant reduction in computational requirements, enabling it to operate more efficiently in the complex environments of power systems. This further substantiates its outstanding advantages and practicality under challenging conditions within power grids.