Power Quality Disturbances Recognition via Lightweight Deep Learning Framework Integrated With Time-Frequency Analysis Embedded

A novel lightweight time-frequency fusion model is proposed to accurately detect and classify complex power quality disturbances, which is helpful for improving power systems’ security and stability. First, a new Morlet wavelet time-frequency convolutional (MWTconv) layer is introduced to extract discriminative time-frequency information, with wavelet parameters optimized through backpropagation to achieve the adaptive extraction of multi-scale time-frequency features for PQD signals. Second, a novel classification method named MEMNet is proposed, which integrates the MWTconv layer with the primary classification framework, and incorporates an efficient multi-scale attention (EMA) module. MEMNet enables end-to-end training and optimization for PQD signal feature processing and intelligent classification. Finally, experimental validation demonstrates better performance with 99.2% classification accuracy and 6.2ms inference time per sample, outperforming the compared methods. The proposed approach has been rigorously validated through an embedded edge computing platform based on Raspberry Pi 4B, confirming its exceptional robustness and practical applicability in PQD identification.