A new parallel PQDs classification method based on an optimized NLM and an improved DRSN-TCN model

Abstract

Considering the multidimensional characteristics of single and composite Power Quality Disturbances (PQDs), this article proposes a new parallel PQDs classification method based on an optimized NLM and an improved DRSN-TCN model. The method effectively solves the problems of high computational complexity, overfitting, and gradient explosion in existing serial method. Meanwhile, it could restrain the interference of noise on the inherent characteristics of PQDs in actual power grids, effectively extracts PQDs features, and thus improves classification accuracy, with strong noise robustness and generalization ability. Firstly, an optimized non-local means (NLM) denoising method is employed to process the noisy PQDs signals. The Greater cane rat algorithm (GCRA) is utilized to adaptively determine the optimal parameters for NLM. By estimating and performing weighted averaging for each sample point in the noisy signal, the method effectively preserves signal detail features, thereby achieving an accurate reconstruction of the original signal. Secondly, to overcome the poor anti-noise capability of the Deep Residual Shrinkage Network (DRSN) model, a new threshold function is proposed to replace the original soft threshold function, enhancing its anti-noise interference capability; to address the issue of the Temporal Convolutional Network (TCN) model's complex structure leading to prolonged training times, a reverse TCN structure is proposed. This structure decreases the receptive field layer by layer as the network depth increases, reducing training parameters and improving training efficiency. Finally, the high-dimensional PQDs features extracted from DRSN and TCN are fed into a feature fusion module for classification. To verify the effectiveness of this method, a parallel classification model is built based on the PyTorch framework, and simulation and comparative experiments on single and composite PQDs are conducted. The results show that the proposed method can effectively classify PQDs, under no noise, 40 dB, 30 dB, and 20 dB noise conditions, the classification accuracies for 16 types of single and composite PQDs are 99.34 %, 98.89 %, 98.12 %, and 96.28 %, respectively, demonstrating the model's strong noise robustness and generalization capability. To further validate the superiority of this method, comparative experiments were conducted among the proposed model and other models such as EWT+SVM, CNN, CNN-LSTM, Transformer, TCN, and DRSN. The results indicate that the improved DRSN-TCN model converges smoothly without oscillation and achieves better classification accuracy. Therefore, the proposed model demonstrates certain advantages over other models.