Remaining useful life of power transformers using efficient surrogates and deep learning techniques
Abstract
The assessment of remaining useful life (RUL) in power transformers (PTs) is critical, given their essential role in the reliable distribution of electrical energy. According to international standards, the RUL is primarily influenced by the temperature of the hotspot (HST) located near the insulating paper that separates the coils from the mineral oil. Degradation of this insulating paper can lead to catastrophic failures, as its breakdown results in direct contact between the coil and the oil, potentially triggering explosions and abrupt transformer malfunctions. To solve this problem, this study presents a novel approach that addresses the limitations of traditionally simplified modeling frameworks by using a reduced model based on sparse Proper Generalized Decomposition (sPGD) of a three-dimensional high-fidelity convective model that provides predictions of the HST behavior in real time. This surrogate model utilizes extensive historical operational data including ambient temperature and transformer power consumption collected over several years to estimate the transformer’s time to failure. The proposed methodology demonstrates an impressive ability to evaluate the RUL of power transformers in less than 2 seconds by simulating the HST over 50 years, while faithfully reproducing the predictions derived from multiphysics and high-fidelity models. Furthermore, this study proposes a data-driven framework based on a Neural ODE for estimating the degradation function of the insulation paper of power transformers, relying solely on the known time of failure and historical operational data. The generated dataset allows the validation of this methodology, which enables the accurate estimation of both the loss of life of a transformer during its use cycle, as well as the estimation of the degradation function of the insulating paper. The framework provides a robust and generalizable approach for real-time health assessment and life cycle management of critical assets.
Related
- Efficient Deep Gaussian Process Surrogates via Mini-Batch Latent MAP Estimation
- Macrocanonical Generator Networks: Data-Efficient Neural Surrogates for Amortized Physics Simulation
- Deep-Learning Solvers and Surrogates for Infinity and p-Laplace Problems
- PaECTER: Patent-level representation learning using citation-informed transformers
- Learning on-the-fly quasi-stationarity diagnostics from low-dimensional surrogates
BibTeX
@article{achour2026remaining,
title = {{Remaining useful life of power transformers using efficient surrogates and deep learning techniques}},
author = {Lila Achour and Sebastian Rodriguez and Paul-Henri Langlois and Fikri Hafid and Chady Ghnatios and Xavier Kestelyn and Francisco Chinesta},
journal = {Advanced Modeling and Simulation in Engineering Sciences},
year = {2026},
month = jun,
doi = {10.1186/s40323-026-00329-6},
url = {https://hal.science/hal-05848708v1},
note = {CC BY-NC-ND},
}
RIS
TY - JOUR
TI - Remaining useful life of power transformers using efficient surrogates and deep learning techniques
AU - Lila Achour
AU - Sebastian Rodriguez
AU - Paul-Henri Langlois
AU - Fikri Hafid
AU - Chady Ghnatios
AU - Xavier Kestelyn
AU - Francisco Chinesta
PY - 2026
DA - 2026/06/13
JO - Advanced Modeling and Simulation in Engineering Sciences
DO - 10.1186/s40323-026-00329-6
UR - https://hal.science/hal-05848708v1
AB - The assessment of remaining useful life (RUL) in power transformers (PTs) is critical, given their essential role in the reliable distribution of electrical energy. According to international standards, the RUL is primarily influenced by the temperature of the hotspot (HST) located near the insulating paper that separates the coils from the mineral oil. Degradation of this insulating paper can lead to catastrophic failures, as its breakdown results in direct contact between the coil and the oil, potentially triggering explosions and abrupt transformer malfunctions. To solve this problem, this study presents a novel approach that addresses the limitations of traditionally simplified modeling frameworks by using a reduced model based on sparse Proper Generalized Decomposition (sPGD) of a three-dimensional high-fidelity convective model that provides predictions of the HST behavior in real time. This surrogate model utilizes extensive historical operational data including ambient temperature and transformer power consumption collected over several years to estimate the transformer’s time to failure. The proposed methodology demonstrates an impressive ability to evaluate the RUL of power transformers in less than 2 seconds by simulating the HST over 50 years, while faithfully reproducing the predictions derived from multiphysics and high-fidelity models. Furthermore, this study proposes a data-driven framework based on a Neural ODE for estimating the degradation function of the insulation paper of power transformers, relying solely on the known time of failure and historical operational data. The generated dataset allows the validation of this methodology, which enables the accurate estimation of both the loss of life of a transformer during its use cycle, as well as the estimation of the degradation function of the insulating paper. The framework provides a robust and generalizable approach for real-time health assessment and life cycle management of critical assets.
SN - 2213-7467
ER -