|
A CNN–LSTM Based Differential Protection Scheme for Early Detection and Classification of Internal Winding Faults in Power Transformers
|
Ehsan Akbari *   |
| Department of Electrical Engineering, Mazandaran University of Science and Technology, Babol, Iran , e.akbari@ustmb.ac.ir |
|
|
Abstract: (35 Views) |
Power transformers are critical assets in transmission and distribution networks, and internal faults can cause substantial equipment damage and compromise power system stability. The rapid detection of winding faults and their accurate discrimination from transient phenomena remain key challenges in transformer differential protection. This paper proposes an intelligent hybrid model based on a convolutional neural network and long short-term memory network (CNN–LSTM) for the early detection and classification of internal transformer faults. Three-phase differential current signals are used as inputs to the proposed model. The CNN layers extract salient local and spatial features from the current waveforms, while the LSTM layer captures the temporal dependencies embedded in the transient signals. Various operating scenarios are simulated, including normal operating conditions, magnetizing inrush current, and different types of internal faults, namely phase-to-ground, phase-to-phase, and three-phase faults. The evaluation results demonstrate that the proposed hybrid architecture achieves an accuracy of 98.8%, outperforming standalone CNN and LSTM models. With its fast response and high accuracy in discriminating internal faults from magnetizing inrush currents, the proposed model shows strong potential for implementation in next-generation intelligent protective relays. |
|
| Keywords: CNN–LSTM, differential protection, early fault detection, internal fault, long short-term memory, power transformer |
|
|
Full-Text [PDF 1240 kb]
(36 Downloads)
|
Type of Study: Research |
Subject:
Special
Received: 2026/02/1 | Accepted: 2026/04/14 | Published: 2026/05/31
|
|
|
|
|
|
|
| Send email to the article author |
|
|
|
| Add your comments about this article |
|
|
|
