Explainable AI framework predicts transformer lifespan and load in high-voltage grids

Power transformers are among the most critical and expensive assets in high-voltage transmission networks, and their unexpected failure can trigger service interruptions, economic losses, costly replacements and reduced system reliability. A new study published in the journal Discover Electronics on 8 October 2026 proposes an explainable artificial intelligence framework that brings together two tasks usually studied in isolation: forecasting the electrical load a transformer will face, and estimating how many years of useful life remain in its insulation. The work, led by Khamdan Annas Fakhryza and Efa Yumna Purwono, combines four different machine learning models in a single stacking architecture and then uses SHAP-based explanations to show which physical variables drive each prediction.

The framework targets 150 kV transmission systems and was trained and evaluated on a physics-informed dataset of 26,255 hourly samples spanning three years. Rather than being purely statistical, the synthetic data were generated using established transformer thermal-aging principles: electrical loading influences hotspot temperature, hotspot temperature determines the aging acceleration factor, and cumulative aging determines the remaining useful life. The dataset contains 22 input features covering electrical quantities such as load demand, voltage, current, power factor and reactive power, thermal variables such as oil and winding hotspot temperatures, environmental conditions including ambient temperature and humidity, and operational and aging-related variables such as load ratio, maintenance score, aging acceleration, oil quality index and transformer age.

At the heart of the method is a hybrid stacking ensemble. The outputs of these four models are then fed into a second-level meta-learner, itself based on XGBoost, which learns the optimal nonlinear combination of their predictions. To prevent information leakage, the meta-learner is built from out-of-fold predictions generated through cross-validation, so each prediction comes from data unseen during the corresponding base-model training.

Feature engineering plays a substantial role. The researchers generated lag features at three intervals, the previous hour, the previous day and the previous week, allowing the models to capture both short-term fluctuations and recurring weekly patterns. Rolling means and standard deviations characterise local operating behaviour and degradation trends, while sine and cosine transformations encode hourly and seasonal cycles so that time remains continuous across boundaries. All continuous variables were normalised with Min-Max scaling, and extreme observations were mitigated through statistical clipping and filtering. Before model development, the data were arranged chronologically and split 80 percent for training and 20 percent for testing, and a target-leakage audit separated the variables used directly in the physics-based remaining-life construction from any future-derived quantities.

For transformer remaining useful life prediction, the proposed stacking model achieved the best results among the five models evaluated: a mean absolute error of 1.470 years, a root mean squared error of 2.057 years, a mean absolute percentage error of 6.39 percent and a coefficient of determination of 0.962. Compared with the best individual learner, XGBoost, the stacking model reduced the mean absolute error from 1.670 to 1.470 years and the root mean squared error from 2.240 to 2.057 years, while raising the coefficient of determination from 0.948 to 0.962. The Random Forest outperformed the plain neural network on this task, and the LSTM-Attention model also performed strongly, with a mean absolute error of 1.760 years and a coefficient of determination of 0.941, suggesting that temporal representation contributes meaningfully to lifespan estimation.

Load forecasting told a more nuanced story. For one-step-ahead prediction, the simple Persistence baseline achieved the strongest performance, followed by ARIMA, while the proposed stacking model delivered competitive results among the machine-learning models, with a root mean squared error of 12.78 megawatts and a coefficient of determination of 0.809. The authors are explicit that they do not claim universal forecasting superiority; their stated contribution is the integration of heterogeneous prediction models and explainability into a common predictive-maintenance workflow. To put the forecasting error in context, the simulated load ranged from roughly 42.6 to 213.52 megawatts with a mean of about 110 megawatts, so the 12.78-megawatt error corresponds to approximately 11.6 percent of the mean simulated load. A 150-MVA transformer rating is given as a reference capacity, though the authors note that direct comparison between megawatt load and MVA rating requires consideration of the operating power factor.

The team subjected the results to several robustness checks. Repeating the experiments with five fixed random seeds gave a load-forecasting root mean squared error of 12.83 plus or minus 0.14 megawatts and, for remaining-life prediction, a mean absolute error of 1.47 plus or minus 0.02 years and a coefficient of determination of 0.962 plus or minus 0.001, indicating the improvements are not artefacts of a particular random initialisation. Paired errors compared with the Wilcoxon signed-rank test produced p-values of 0.018 for mean absolute error, 0.012 for root mean squared error and 0.015 for mean absolute percentage error, all below the 0.05 threshold. An ablation study showed that removing the LSTM-Attention component raised the forecasting root mean squared error to 14.35 megawatts, removing engineered temporal features raised it to 13.64 megawatts, and removing physics-informed features raised it to 13.28 megawatts.

Explainability is where the framework aims to distinguish itself. SHAP analysis, applied to the final stacking ensemble, identified hotspot temperature, aging acceleration and load ratio as the dominant factors influencing transformer health predictions, a ranking consistent with established transformer aging mechanisms in which excessive loading raises hotspot temperature and accelerates insulation degradation. Because those variables are physically related to the synthetic remaining-life construction, the researchers also ran a sensitivity experiment excluding the direct aging-related variables from the predictor set. Performance dropped to a mean absolute error of 1.82 years and a coefficient of determination of 0.944, confirming that these variables carry strong degradation information while showing the model still extracts degradation signals from other operational measurements. SHAP summary, importance and waterfall plots provide both global and per-prediction explanations, which the authors argue helps maintenance engineers recognise early indicators of accelerated degradation and prioritise high-risk transformers based on interpretable evidence rather than black-box outputs.

The framework also translates its predictions into maintenance-oriented outputs. Predicted load and remaining lifespan are jointly used to assign operational risk levels and maintenance recommendations, so transformers showing increasing forecasted loads alongside decreasing remaining lifespan can be flagged as high-risk assets needing enhanced monitoring, while healthier units continue under routine programmes. Uncertainty was quantified with prediction intervals: the 95 percent intervals achieved empirical coverage of 93 to 96 percent across the temporal validation folds, with a mean interval width of roughly 7.5 to 8.5 years, though the authors stress these intervals should be recalibrated with field data before any operational deployment. Computational benchmarks on a MacBook with an Apple M4 processor and 32 gigabytes of memory showed the stacking framework trains in 31.6 minutes and infers in 4.12 milliseconds per sample, a higher training cost than any individual base learner but fast enough for near-real-time prediction under the investigated simulation setting.

The authors are careful about the limits of the evidence. The entire evaluation rests on a physics-informed simulated dataset rather than operational data from utility companies, remaining-life estimation did not incorporate detailed condition-monitoring measurements such as Dissolved Gas Analysis, and only a single 150 kV transmission scenario was considered. Because of this, the reported performance should be interpreted within the scope of the simulation assumptions and not as direct evidence of equivalent performance on field data. Future work, the researchers say, will validate the framework on real multi-site utility datasets, integrate additional condition-monitoring variables, explore transformer-based deep learning architectures, and investigate physics-informed thermal-aging models with real-time monitoring and online learning for adaptive predictive maintenance. If those validations succeed, the approach could offer utilities a transparent, integrated tool for anticipating both the demand placed on their transformers and the aging those demands induce.

Subject of Research: An explainable machine learning framework for transformer load forecasting and remaining lifespan prediction in high-voltage power systems

Article Title: An explainable AI-based predictive maintenance framework for transformer lifespan assessment and load forecasting in high-voltage power systems

Article References: Fakhryza, K. A., & Purwono, E. Y. (2026). An explainable AI-based predictive maintenance framework for transformer lifespan assessment and load forecasting in high-voltage power systems. Discover Electronics, 3(1), Article 145. https://doi.org/10.1007/s44291-026-00294-9

Image Credits: AI Generated

DOI: 10.1007/s44291-026-00294-9

Keywords: transformer, predictive maintenance, load forecasting, machine learning, explainable AI, SHAP, LSTM, XGBoost, high-voltage power systems, remaining useful life, stacking ensemble, insulation aging