A power quality disturbance classification method based on dynamic multi-scale time-frequency feature complementary fusion

To address the challenges of feature extraction under noisy conditions and the difficulty of classifying compound Power Quality Disturbances (PQDs) due to complex time-frequency interactions, this paper proposes a novel PQDs classification method based on dynamically integrated multi-scale time-frequency feature fusion. First, a dynamic multi-scale module is constructed to extract features in both time and frequency domains. This module comprises multi-scale atrous convolutions and dynamic perception units. The former captures multi-scale time-frequency features, while the latter enhances the interaction between local and global information through adaptive weighting, thereby improving adaptability to various types of PQDs. Next, a dual-branch cross time-frequency feature complementary fusion module is introduced. Through feature alignment, information redistribution, and weighted fusion, this module achieves complementary enhancement of time-frequency information, significantly improving the discrimination of compound disturbances. Finally, a fully connected classification layer is employed to accurately classify the disturbance types. Experiments conducted on a 29-class PQDs dataset demonstrate that the proposed model exhibits superior performance in noise robustness, compound disturbance identification, and generalization in practical scenarios, making it well-suited for power quality disturbance recognition systems.