Hybrid Deep Learning Approach for Accurate Detection and Multiclass Classification of Broken Conductor Faults in Power Distribution Systems
The identification and classification of broken conductor faults in power distribution systems (PDSs) is challenging because of the nonlinear and complex nature of the unbalanced three-phase currents. Previous approaches to identifying faults fail to resolve the issue of precise fault identification...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2024-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10788708/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The identification and classification of broken conductor faults in power distribution systems (PDSs) is challenging because of the nonlinear and complex nature of the unbalanced three-phase currents. Previous approaches to identifying faults fail to resolve the issue of precise fault identification because of limited data availability and inefficient data modeling approaches. In this study, to identify the broken conductor faults and classify them into several scenarios in three-phase open-circuit distribution networks, a new hybrid method of deep learning based on Complex Value Neural Networks (CVNNs) and Long Short Term Memory (LSTM) is introduced. Our method refines input using the Proportional Topology Optimization (PPTO) algorithm that enhances the signal quality of unbalanced three-phased current signals for fault identification. Moreover, we use an appropriate DenseNet pre-trained model to extract features and estimate the topology state with reliable data representation. The developed CVNN framework was built with TensorFlow’s Sequential API, which provided efficient computing of complex-valued inputs and weights and, thus, the proper handling of the intricate signals often encountered in power systems. The proposed approach was validated on three publically accessible datasets using a five-fold cross-validation technique. Results were realized in the form of Classification Accuracy (98.22% / 99.83% / 98.22%), F-measure (0.8881 / 0.9983 / 0.88), Precision (0.9207 / 0.9983 /0.92), Recall (0.8578 / 0.9983 / 0.85), and Area Under the Curve (AUC) (0.9433 / 0.9917/ 0.9958). It is shown that the proposed method has higher fault detection and classification accuracy compared to three traditional classification approaches, namely, Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), and three state-of-the-art methods: 1) Stockwell transform +SVM, 2) Fast Fourier Transform + SVM, and 3) Hilbert-Huang transform of vibration data and power spectral density + Artificial Neural Network. Overall, the performance of the proposed methodology was satisfactory and can be used in real-time fault identification from PDSs installed in complex load environments and industries. |
|---|---|
| ISSN: | 2169-3536 |