Comparing Deep Neural Networks for 1D Multi-Label Power Quality Disturbance Classification: ResNet, MobileNet, and DenseNet

Abstract

Power quality (PQ) analysis has become increasingly challenging due to the widespread usage of power electronic devices and the variability of operational events in electric grids integrating renewable energy resources. Deep neural networks (DNNs) have emerged as the prevalent approach for identifying PQ disturbances (PQDs). The 1D DNN-based disturbance classification architectures, in contrast to 2D models, operate directly on original waveform measurements, offering sequential views to capture irregular PQDs. However, there remains a research gap on the impact of architectural choices of DNN frameworks in this context, as well as on understanding their robustness to noise and adequacy for multi-output identification tasks. This work delves into three advanced DNN architectures, namely ResNet, MobileNet, and DenseNet, to assess their suitability for 1D PQD classification tasks. The performance of different classification models is compared using synthetic datasets following the IEEE 1159-2019 standard. The results demonstrate that ResNet outperforms MobileNet and DenseNet in terms of average accuracy and noise tolerance.