⚡ electrical engineering

Online Power-Transformer Hot-Spot Estimation via Spectrum-Aware Physics-Data Hybrid Learning

This study proposes an explainable physics-data hybrid framework that integrates spectrum-aware thermal modeling with conditional C-vine copula residual learning to achieve high-accuracy, probabilistic online estimation of power-transformer hot-spot temperatures and overtemperature risks under complex harmonic and dynamic load conditions.

Original authors: Shichao Cao

Published 2026-09-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shichao Cao

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

Power grids are the invisible arteries of modern life, carrying electricity from wind farms and solar fields to the homes and businesses that depend on them. At the heart of this network sit massive transformers, the heavy-duty machines that step up voltage for long-distance travel and step it down for local use. Inside these steel giants, copper windings carry the current, but as electricity flows, it generates heat. If this heat gets too high, it damages the insulation paper wrapping the wires, shortening the machine's life and risking a sudden failure. For decades, engineers have relied on standard formulas to guess how hot these transformers get, using simple measurements of how much power is flowing through them. However, these old formulas struggle when the power source is unpredictable, such as the fluctuating output of renewable energy, or when the electricity itself is "messy," filled with high-frequency ripples known as harmonics. In a world increasingly powered by wind and sun, knowing the true temperature of a transformer is no longer just a matter of routine maintenance; it is a critical safety issue.

A researcher led by Shichao Cao at Hebei Vocational University of Technology and Engineering set out to solve this problem by creating a smarter way to estimate the hottest spot inside a transformer. Instead of relying on a single, rigid formula or a purely computer-based guess, they built a hybrid system that combines the known laws of physics with real-time data learning. The researcher focused on two specific transformers at a large substation in China that collects power from a 96-megawatt wind farm and an 80-megawatt solar plant. These machines face a unique challenge: the electricity they handle changes rapidly as clouds pass over solar panels or wind speeds shift, and the power electronics used to connect these sources to the grid introduce complex harmonic distortions. The researcher realized that simply measuring the total amount of current was not enough; they needed to understand the specific "shape" of that current, including its high-frequency components, because different frequencies generate heat in different ways.

To tackle this, the researcher first built a physical model that acts like a digital twin of the transformer's cooling system. This model tracks how heat moves from the copper windings into the surrounding oil and then into the air, accounting for the fact that oil takes time to warm up and cool down. Crucially, they modified this model to weigh the heat generated by different parts of the electrical spectrum. They found that two currents with the same total distortion level could produce vastly different amounts of heat depending on which specific frequencies were present. For instance, a current dominated by lower-frequency ripples might heat the machine less than one filled with higher-frequency ripples, even if the total distortion number looked the same. By feeding the actual, detailed spectrum of the current into their physical model, they could calculate a much more accurate baseline temperature than standard methods allowed.

However, even a perfect physical model cannot account for every tiny variation in a real-world machine, such as slight differences in oil flow or sensor placement. To fix the remaining errors, the researcher added a second layer of intelligence using a statistical tool called a C-vine copula. Think of this tool as a highly sensitive tuner that listens to the difference between what the physical model predicts and what the actual sensors measure. It learns how this difference changes based on the current conditions: how fast the load is changing, the outside air temperature, the specific harmonic content of the electricity, and even how the error behaved in the previous few minutes. By analyzing these patterns, the system can adjust its prediction in real time, providing not just a single temperature number, but a range of likely values and a probability of the machine overheating.

The researcher tested their new method in two ways. First, they ran a month-long computer simulation where they could control every variable, including creating scenarios where the total distortion was identical but the frequency mix was different. In these tests, the new method proved its worth by reducing the average error to just 0.636 degrees Celsius, a significant improvement over the standard models which missed the mark by more than 1.3 degrees. More importantly, it correctly identified the hottest moments with far greater precision, avoiding the dangerous tendency of older models to underestimate the temperature during rapid changes.

They then took the system to the field for a fifteen-day trial at the actual substation. During this period, the system monitored the transformers continuously, comparing its estimates against real fiber-optic sensors embedded inside the windings. The results were striking. For the transformer connected to the solar farm, the new method predicted the hot-spot temperature with an average error of only 1.87 degrees Celsius, whereas the standard industrial model was off by more than 8 degrees. When the temperature approached a critical warning limit of 98 degrees Celsius, the new system correctly identified 92.3 percent of the actual overheat events, while the traditional model generated many false alarms or missed the danger entirely. The system also successfully tracked the aging of the insulation, showing that the periods of high heat caused far more wear than isolated, noisy spikes.

This work demonstrates that the future of grid safety lies in blending physical understanding with adaptive data learning. By respecting the complex nature of renewable energy and the specific way different electrical frequencies generate heat, engineers can now monitor their equipment with a level of clarity that was previously impossible. The method does not require expensive new hardware; it simply uses existing data more intelligently, turning a standard transformer into a self-aware asset that can warn operators before a problem becomes a crisis. As the world shifts toward cleaner energy, this kind of precise, real-time insight will be essential for keeping the lights on and the grid safe.

Technical Summary: Online Power-Transformer Hot-Spot Estimation via Spectrum-Aware Physics-Data Hybrid Learning

Problem Statement
The increasing integration of converter-interfaced wind and photovoltaic (PV) generation subjects power transformers to rapid load fluctuations and harmonic-rich currents. Conventional thermal models, while physically interpretable, often exhibit residuals dependent on specific operating conditions, particularly under non-sinusoidal loading. Conversely, purely data-driven predictors may lack reliability when operating outside their training regimes. Furthermore, standard Total Harmonic Distortion (THD) metrics fail to capture the specific frequency-dependent loss mechanisms caused by different harmonic spectra, leading to potential errors in hot-spot temperature estimation and insulation aging assessments. The challenge lies in developing an estimator that maintains physical consistency, accounts for detailed harmonic spectra, and adapts to real-time operational changes without sacrificing interpretability.

Methodology
The study proposes an explainable physics-data hybrid estimator designed for online hot-spot temperature and overtemperature risk assessment. The framework consists of three integrated components:

Spectrum-Aware Physical Model:
- Heat Source Formulation: Instead of relying solely on load factor or THD, the model calculates a normalized heat source using the total branch RMS current and two distinct harmonic loss factors: $F_{HL}$ (winding eddy-current) and $F_{OSL}$ (other stray losses). These factors weight the harmonic spectrum based on frequency-dependent loss exponents (2 for eddy currents, 0.8 for stray losses), allowing the model to distinguish between spectra with identical THD but different harmonic orders.
- Dynamic State Equations: The model employs coupled differential equations for top-oil temperature and winding hot-spot rise. Renewable fluctuations and harmonic effects modify the common heat source, which then propagates through a single thermal network, preserving thermal memory and avoiding the superposition of separate load and harmonic models.
C-Vine Copula Residual Correction:
- To address the residuals between the physical model and fiber-optic measurements, a conditional C-vine copula is employed. This structure learns the complex, non-linear dependence of the residual on five conditioning variables: load factor, harmonic loss factor, ambient temperature, load-rate change, and the lagged residual.
- Unlike a single Frank copula, the C-vine allows different pair-copula families (e.g., Student-t for tail dependence, Gumbel for upper-tail association) to model specific variable relationships. This enables the generation of a median estimate, a 90% prediction interval, and an overtemperature probability.
Online Adaptation and Uncertainty Calibration:
- The system includes an online bias update mechanism using exponential weighting to adjust for persistent shifts in cooling performance or unit-specific behavior.
- A data-quality score ( $q_t$ ) governs the estimation process: if data quality is low, the system reverts to the physical model; if high, it applies residual correction and updates parameters.
- Overtemperature probability is calculated based on a conditional residual distribution relative to a warning threshold (e.g., 98°C), and equivalent aging is computed using the Arrhenius equation.

Key Contributions

Unified Heat-Source Formulation: The paper introduces a method to represent loading and spectral content simultaneously, retaining information discarded by THD alone.
Dynamic State Propagation: It formulates dynamic equations where renewable fluctuations and harmonic effects alter a common heat source before propagating through the thermal network, ensuring thermal history is preserved.
C-Vine Residual Modeling: The application of a C-vine copula to model the residual dependence on multiple operating conditions (load, harmonics, ambient, rate of change, and lagged error) provides a flexible, condition-dependent correction that outperforms single-parameter copulas.
Comprehensive Evaluation: The method is validated through a 30-day chronological simulation and a 15-day field campaign at a 220/500 kV wind-solar-thermal aggregation substation, utilizing causal rolling-origin evaluation to ensure no future data leakage.

Results

Simulation Performance: In a 30-day simulation, the proposed method achieved a Mean Absolute Error (MAE) of 0.636 °C, an RMSE of 0.797 °C, and an $R^2$ of 0.9928. The 90% prediction interval achieved 89.3% coverage. Ablation studies confirmed that replacing the full harmonic spectrum with THD alone increased MAE significantly (to 1.790 °C), and the C-vine structure outperformed a Frank copula variant.
Field Campaign Performance:
- T1 (Top-Oil): The model achieved an MAE of 0.94 °C and RMSE of 1.21 °C, representing a 78.0% reduction in MAE compared to the conventional IEC dynamic thermal model.
- T2 (Winding Hot-Spot): The model achieved an MAE of 1.87 °C and RMSE of 2.34 °C, with an $R^2$ of 0.972. At the 98 °C warning threshold, the recall rate was 92.3% (12 out of 13 samples), significantly outperforming the IEC model which generated numerous false alarms due to positive bias.
Robustness: The method demonstrated stability under missing data (handled via causal forward imputation) and temperature noise, with the physical state recursion maintaining numerical stability during short data gaps.

Significance and Claims
The paper claims that the proposed spectrum-aware physics-data hybrid estimator effectively reduces temperature estimation errors under high harmonic loads, load ramping, and cooling state switching. By distinguishing between harmonic spectra with the same THD, the method provides a more accurate characterization of the heat source. The C-vine correction layer successfully captures the distinct dependence patterns of residuals, offering calibrated uncertainty intervals and traceable overtemperature probabilities. The study concludes that this approach supports temperature estimation, thermal-boundary assessment, and operating event review within a single, traceable workflow, making it suitable for smart-grid asset monitoring in renewable-rich environments. The author notes that while the simulation serves as a controlled mechanism test, the field campaign provides the external operational check, confirming the method's engineering applicability despite limitations such as unmodeled local oil flow and sensor placement variations.

Technical Summary: Online Power-Transformer Hot-Spot Estimation via Spectrum-Aware Physics-Data Hybrid Learning

More like this