⚡ electrical engineering

Physics-Guided Top-Oil Temperature Prediction for Operational Thermal-State Assessment of Oil-Immersed Power Transformers

This paper proposes TIR-MSTLA, a physics-guided framework that combines a first-order heat-balance model with adaptive residual learning and multi-scale attention mechanisms to significantly improve the accuracy of short-term top-oil temperature predictions for oil-immersed power transformers under varying field conditions.

Original authors: Chengjie Wang, Xiaozhou Fan, Bowen Liu, Yidi Liu, Bowen Xue

Published 2026-09-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Chengjie Wang, Xiaozhou Fan, Bowen Liu, Yidi Liu, Bowen Xue

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Inside the massive, humming vaults of electrical substations, oil-immersed power transformers stand as the silent guardians of the grid. These are not merely metal boxes; they are complex thermal engines that convert electricity while generating immense heat. To keep them from overheating and failing, engineers must constantly monitor the temperature of the oil at the very top of the tank. This "top-oil temperature" is the primary clue that tells operators how hard the transformer is working and how much more load it can safely handle before its internal insulation begins to degrade. If the oil gets too hot, the transformer ages rapidly, risking a catastrophic failure that could leave entire regions in the dark. For decades, the standard way to predict this temperature has been to rely on fixed mathematical rules derived from the manufacturer's nameplate. These rules assume the transformer behaves like a perfect, unchanging machine, reacting to load changes in a predictable, steady rhythm.

However, real-world transformers are far from perfect machines. They age, their cooling systems change efficiency with the seasons, and the flow of oil inside them shifts in ways that fixed rules cannot capture. When the weather turns cold or the load spikes unexpectedly, these standard models often fall behind, predicting a temperature that is either too high or too low compared to reality. This lag is dangerous because it can lead to unnecessary power restrictions or, worse, a failure to warn operators of an impending thermal crisis. The core challenge for scientists is to create a prediction method that respects the fundamental laws of physics—so it remains reliable—but is flexible enough to learn from the specific, messy behavior of each individual transformer in the field.

A team of researchers from North China Electric Power University has developed a new approach to solve this problem, creating a system that blends the stability of physics with the adaptability of modern data learning. Their method, tested on real transformers operating at 110 kilovolts and 220 kilovolts, starts by building a solid physical baseline. This baseline acts like a trusted reference point, calculating what the temperature should be based on the load, the outside air temperature, and the transformer's original design specifications. It uses a simple, first-order heat balance concept: the oil heats up as electricity flows through it and cools down as it loses heat to the environment, a process that takes time. This physical model ensures the prediction never drifts into impossible territory, keeping the results grounded in reality.

But the researchers knew that this physical baseline alone was not enough to handle the quirks of field operation. To bridge the gap between the ideal model and the real machine, they added a layer of intelligent learning that acts as a correction mechanism. This system does not try to guess the temperature from scratch; instead, it learns the specific "residual" errors—the small, consistent differences between what the physical model predicts and what the sensors actually measure. It pays close attention to how the transformer reacts over different time scales, from the immediate jolt of a load change to the slow, creeping heat accumulation over days. By analyzing these patterns, the system can adjust its prediction in real-time, compensating for things like a radiator that is slightly less efficient than the design manual suggests or a local oil flow that has shifted due to aging.

The results of this hybrid approach were striking when tested against one year of continuous field data from two different transformers. The new method proved to be significantly more accurate than the traditional fixed-parameter models used by industry standards, as well as more reliable than several advanced data-driven algorithms that rely solely on historical patterns. On the 110 kilovolt transformer, the new system reduced the average prediction error by 77 percent compared to the standard model, bringing the error down to a mere 0.145 degrees Celsius. On the larger 220 kilovolt unit, it reduced the error by 53 percent, achieving an average error of 0.178 degrees Celsius. Perhaps most importantly for safety, the researchers found that the prediction error never exceeded two degrees Celsius across any continuous period of operation, a level of stability that the older models simply could not maintain, especially during seasonal shifts or sudden load changes.

This level of precision has direct and practical consequences for how power grids are managed. Because the prediction is so accurate, operators can trust the numbers when deciding how much electricity a transformer can safely carry. The system provides a narrow, reliable band of uncertainty around the predicted temperature, allowing engineers to calculate a "load margin" with confidence. If the predicted temperature is close to a safety limit, the small error margin means they know exactly how close they are to the danger zone, rather than guessing based on a wide, uncertain range. This precision also improves the assessment of insulation aging. Since the rate at which transformer insulation degrades is extremely sensitive to temperature, even a small error in prediction can lead to a large miscalculation of the transformer's remaining life. By reducing the prediction error from the degree level to the sub-degree level, the new method tightens the uncertainty in aging calculations, ensuring that transformers are neither retired prematurely nor pushed beyond their safe limits.

The study demonstrates that the future of power grid monitoring lies not in choosing between rigid physical laws and flexible data learning, but in combining them. By anchoring the prediction in the fundamental physics of heat transfer and then using data to correct for the unique imperfections of each machine, the researchers have created a tool that is both scientifically sound and practically robust. This approach offers a way to keep the grid running more efficiently and safely, ensuring that the silent guardians of the electrical network are monitored with a clarity that was previously out of reach. The work confirms that when physics guides the learning process, the result is a system that understands the machine not just as a collection of data points, but as a living, breathing piece of infrastructure that responds to the world around it.

Technical Summary: Physics-Guided Top-Oil Temperature Prediction for Operational Thermal-State Assessment

Problem Statement
Top-oil temperature ( $T_o$ ) is a critical operational variable for oil-immersed power transformers, serving as the primary input for loading guides (IEEE Std C57.91, IEC 60076-7), hot-spot estimation, and insulation aging assessments. However, accurate short-term prediction in field operations is challenging due to equipment-specific thermal inertia, varying ambient boundaries, and changing cooling efficiencies. Existing approaches face limitations:

Fixed-parameter thermal models (e.g., standard IEEE/IEC models) often fail to capture field phenomena such as local oil-flow redistribution, radiator-efficiency variations, and season-dependent ambient boundaries because they assume static thermal resistance and capacitance.
Purely data-driven predictors (e.g., LSTM, Transformer) can capture non-linear relationships but may exhibit phase lags, peak deviations, or non-physical fluctuations when physical constraints are weak, particularly under seasonal or load changes.

The objective is to develop a prediction framework that maintains the physical interpretability of standard thermal models while adapting to real-world field deviations to support condition monitoring and loading-margin assessment.

Methodology: The TIR-MSTLA Framework
The authors propose TIR-MSTLA (Thermal-Inertia-Residual Multi-Scale Time-delay Attention), a physics-guided learning framework. The model decomposes the predicted top-oil temperature ( $\hat{T}_{o,t}$ ) into four components:
$\hat{T}_{o,t} = T_{phy,t} + \Delta T_{res,t} + \Delta T_{data,t} + b_t$

Physical Reference ( $T_{phy,t}$ ): A first-order heat-balance recursion serves as the physical backbone, consistent with loading guides. It calculates the trajectory based on load factor, ambient temperature, and nameplate parameters.
Adaptive Thermal Parameters: To address field deviations, the equivalent thermal resistance ( $R$ ) and capacitance ( $C$ ) are modeled as operating-state-dependent functions using correction networks and equipment embeddings (derived from nameplate data). This allows the thermal time constant ( $\tau$ ) to adapt dynamically.
Multi-Scale Time-Delay Attention: Recognizing that thermal response involves short-term, medium-term, and long-term heat accumulation, the model uses three history windows constrained by the thermal time constant. A causal sequence extraction mechanism with attention weights fuses features from these different temporal scales.
Residual Learning and Bias Correction:
- Residual Branch: Learns physically unmodeled deviations (e.g., local flow redistribution) using thermal-process features.
- Dynamic Bias: Corrects systematic offsets dependent on the operating state.
Physics-Guided Training: The model is trained with a composite loss function including prediction error ( $L_{pred}$ ) and a heat-balance residual loss ( $L_{phy}$ ). The latter enforces physical consistency by penalizing violations of the energy balance equation during training, ensuring the model does not learn non-physical dynamics.

Key Contributions

Framework Development: A physics-guided prediction framework that constructs a loading-guide-consistent thermal reference and uses it as a physical backbone, enabling direct comparison with and improvement over fixed-parameter standard models.
Adaptive Compensation: Introduction of adaptive thermal parameters and residual learning to compensate for field-specific deviations (cooling boundary variations, equipment condition differences) that fixed models cannot capture.
Temporal Structure: A thermal-time-constant-guided multi-scale time-delay attention structure designed to capture the multi-time-scale coupling inherent in transformer thermal dynamics.
Field Validation: Comprehensive evaluation using one-year field datasets from 110 kV and 220 kV ONAN transformers, linking empirical error bands to practical engineering applications.

Experimental Results
The method was validated on two datasets (110 kV and 220 kV) spanning approximately 12 months, with a 15-minute prediction horizon.

Accuracy: TIR-MSTLA achieved a Mean Absolute Error (MAE) of 0.1450 °C (110 kV) and 0.1776 °C (220 kV), and RMSE of 0.3577 °C and 0.4862 °C, respectively.
Improvement: Compared to the fixed-parameter IEEE/IEC standard model, the method reduced test MAE by 77.0% (110 kV) and 53.1% (220 kV). It also outperformed data-driven baselines (LSTM, Transformer, PINN).
Stability: All continuous-segment absolute errors remained below 2 °C. The 95th-percentile errors were 0.3525 °C and 0.4326 °C for the two units.
Robustness: Month-level analysis showed that while the fixed-parameter model degraded significantly during seasonal shifts (e.g., RMSE increasing from 0.369 to 1.019 °C for the 220 kV unit), TIR-MSTLA maintained stable, sub-degree accuracy.

Significance and Claims
The paper positions the proposed method not as a replacement for transformer protection or dispatch rules, but as a field-validated input for operational thermal-state assessment.

Uncertainty Reduction: By reducing prediction errors to the sub-degree level, the method narrows the "uncertainty band" around thermal limits. This allows for more precise classification of thermal states (normal vs. warning vs. high-risk) near alarm thresholds.
Loading Margin: The stable error profile supports conservative loading-margin assessments, where the prediction uncertainty is explicitly accounted for in the calculation of available load headroom.
Insulation Aging: The authors demonstrate that the reduced top-oil prediction error significantly tightens the uncertainty in downstream calculations. For a ±0.5 °C error, the induced relative change in the aging acceleration factor ( $F_{AA}$ ) is limited to roughly 12–15%, whereas a 2 °C error (typical of baselines) would propagate to 50–60%.
Compatibility: The framework is designed to be directly compatible with standard hot-spot and loss-of-life calculation chains (IEEE Std C57.91/IEC 60076-7), providing a more accurate intermediate input without altering the underlying physical standards.

The authors conclude that physics-guided top-oil prediction serves as a robust, equipment-level auxiliary input for condition monitoring and risk screening, bridging the gap between theoretical thermal models and complex field realities. Future work is identified as extending the framework to multi-step forecasting and direct hot-spot estimation with sensor validation.

Technical Summary: Physics-Guided Top-Oil Temperature Prediction for Operational Thermal-State Assessment

More like this