Identification of composite power quality disturbances based on dual-modal parallel feature extraction

The identification of composite power quality disturbances is challenging due to the diversity of disturbance categories and the limited feature representation available from single-channel signals. To address these challenges, a dual-modal spatiotemporal feature fusion approach based on a bidirectional cyclic dual residual network (DCRF-Net) is proposed. The framework adopts a time-frequency-image dual-stream architecture. The temporal processing branch employs stacked one-dimensional cyclic dual residual blocks (1D-CDRB) to progressively extract multi-scale temporal features, which are further enhanced by a BiGRU module to capture dynamic temporal dependencies. The image branch integrates a two-dimensional cyclic dual residual block (2D-CDRB) with a triple-directional spatial-channel attention (TriSCA) mechanism. The two feature modalities are then fused to construct highly discriminative dual-modal spatiotemporal representations. Validation is conducted on a simulated dataset comprising 29 composite disturbance categories. Within the 20 dB to 50 dB signal-to-noise ratio range, the proposed method demonstrates superior accuracy compared with single-modal techniques. The DCRF-Net achieves an average accuracy improvement of 2.6% and 1.7% over the ResNet18 and DenseNet-SC models, respectively, while reducing network parameters by 90% and 97%.