A hybrid machine learning pipeline combining temporal convolutional networks with recursive feature elimination can reliably track voltage degradation in proton exchange membrane fuel cells under familiar conditions, but its predictive capability degrades when transferred across operating regimes without recalibration. Evaluating the pipeline across standardized durability datasets, researchers demonstrated test mean absolute errors of 0.0452 volts on the IEEE PHM dataset and 0.0478 volts on the FC Lab benchmark. While the predictive framework maintained steady performance under synthetic input disturbances, direct migration across different load conditions and separate hardware test benches increased forecasting errors by more than fifty percent. These results outline clear boundaries for purely data-driven fuel cell prognosis, establishing that low historical errors alone do not guarantee operational generalization across diverse working environments.
Proton exchange membrane fuel cells serve as zero-emission electrochemical power sources for transportation and stationary infrastructure, yet commercial deployment remains complicated by inevitable functional decay. Over prolonged operation, catalyst activity diminishes, membrane electrode assemblies degrade, and operational fluctuations cause gradual voltage decline. Forecasting this loss of performance and calculating the remaining useful life of the fuel cell stack is essential for scheduling maintenance and avoiding unexpected system failures. Although physics-based models capture underlying transport phenomena and electrochemical kinetics, their governing parameters prove exceptionally challenging to recalibrate as operating conditions fluctuate and internal components age. Conversely, standard data-driven algorithms often struggle when redundant operational measurements mask subtle decay patterns or when training procedures inadvertently introduce statistical information leakage from future data.
To confront these operational challenges, the investigative team assessed an end-to-end prognostic workflow structured around four sequential phases: leakage-controlled preprocessing, train-only feature selection, temporal neural modeling, and global hyperparameter optimization. The complete sequence was designed to test reliability rather than claim an entirely novel neural architecture. Denoising began with a five-point median filter applied to raw operating telemetry to remove spurious transient spikes and measurement noise. Normalization parameters were strictly fitted using only the chronologically isolated training split of the data, ensuring that validation and test records remained unseen during scaling. By enforcing this chronological boundary, the evaluation avoided optimistic estimation biases that frequently arise when normalization statistics or feature rankings are computed across an entire continuous operating record prior to model evaluation.
Feature refinement was conducted using random forest recursive feature elimination, which systematically pruned uninformative sensor channels based on impurity reduction within decision trees. For the IEEE PHM dataset, which initially provided seven operational streams comprising total stack voltage, hydrogen inlet temperature, air outlet temperature, hydrogen inlet pressure, air outlet pressure, air inlet flow rate, and hydrogen outlet flow rate, the recursive algorithm systematically eliminated redundant variables until identifying a compact subset of three core features: stack voltage, hydrogen inlet temperature, and hydrogen outlet flow rate. On the parallel FC Lab dataset, the algorithm converged on stack voltage, hydrogen outlet temperature, and hydrogen inlet flow rate. Five-fold cross-validation confirmed that pruning secondary pressures and temperatures prevented overfitting while retaining the essential dynamic context needed to forecast stack voltage trajectories.
Temporal sequence modeling was executed using an improved temporal convolutional network featuring stacked residual blocks and dilated causal convolutions. Causal convolutional layers ensured that output predictions depended strictly on preceding observations without leaking future temporal information into the present calculation. By incorporating a dilation factor of two alongside kernel sizes of seven or nine, the network expanded its receptive field exponentially, enabling it to span hundreds of historical time steps without pooling or sacrificing temporal resolution. Skip connections within each residual block provided identity mapping paths that alleviated vanishing gradient problems across deeper layers. The final extracted temporal representations were processed through a four-layer fully connected head with node dimensions progressing from 48 to 32, 16, and 1, outputting multi-step voltage forecasts across prediction horizons spanning from 5 to 24 steps from a fixed 24-step input history.
Structural hyperparameters were tuned using particle swarm optimization with a population of 20 particles iterated over 50 generations, minimizing mean squared error on the validation split. The optimization algorithm searched a defined space encompassing kernel numbers between 32 and 128, kernel sizes between 3 and 9, residual block counts from 1 to 3, and dropout regularization rates ranging from 0.1 to 0.3. The converged configuration for the primary IEEE PHM dataset selected 64 convolutional filters, a kernel size of 7, a single residual block, and a dropout rate of 0.20. Learning rate, network dilation, batch size, and the 24-step historical input window were held invariant across all models to ensure an equitable, controlled comparison against baseline architectures.
In standardized benchmarking across the primary test splits, the optimized framework achieved superior accuracy relative to conventional recurrent and convolutional models. On the IEEE PHM dataset, the pipeline attained a root mean square error of 0.15, a mean absolute percentage error of 4.61 percent, and a coefficient of determination of 0.95, outperforming recurrent neural networks such as standard long short-term memory networks and gated recurrent units. Statistical comparisons across five independent random initializations demonstrated significant performance differences via analysis of variance and post hoc testing, with Cliff’s delta exceeding 0.474 across all comparisons, indicating a large substantive effect. As multi-step horizons lengthened from 5 to 24 steps, mean absolute error climbed from approximately 0.0421 volts to roughly 0.0655 volts, demonstrating a more gradual accumulation of step error than baseline convolutional networks, which experienced error doubling over identical forecast spans.
To evaluate remaining useful life, stack voltage trajectories were mapped against an end-of-life failure criterion defined as a 5.0 percent reduction from the initial stabilized operational voltage. Monte Carlo dropout with 100 stochastic forward passes generated 95 percent confidence prediction intervals, providing quantified uncertainty bounds alongside point forecasts. On the primary dataset, the pipeline produced a prediction interval coverage probability of 0.9970, an average prediction bias of negative 0.0884, and a composite final score of 0.9901, balancing point accuracy with interval inclusion. On the secondary benchmark, the framework achieved an identical final score of 0.8050 to long short-term memory networks but delivered tighter calibration with an interval coverage probability of 0.9285, verifying that predicted confidence bands effectively tracked the true degradation curve.
Synthetic robustness assessments tested model resiliency against Gaussian white noise up to 20 percent of feature standard deviation, random data loss up to 30 percent, and simulated linear covariate drift of 10 percent. Across these synthetic corruptions, the model maintained coefficients of determination between 0.91 and 0.95 and mean absolute errors between 0.05 and 0.07 volts. In contrast, baseline recurrent and convolutional architectures experienced steeper performance drops under identical noise injections. However, the study emphasized that robustness under simulated mathematical disturbances does not imply invariance to unmodeled physical shifts, noting that synthetic perturbation checks answer distinct analytical questions from true operating regime transitions.
The limitations of data-driven forecasting became evident during physical domain transfer evaluations. Direct migration of a model trained on high-load records to low-load conditions within the same dataset increased mean absolute error by 51.1 percent, rising from 0.045 volts to 0.068 volts while reducing the coefficient of determination to 0.88. Direct cross-dataset transfer from the IEEE PHM platform to the FC Lab system caused an 82.2 percent error increase, reaching a mean absolute error of 0.082 volts. While fine-tuning on labeled target-domain samples partially restored predictive accuracy, reducing errors to 0.049 volts and 0.050 volts respectively, doing so required dedicated target-domain observations that may not be available prior to deployment in commercial systems.
Ablation experiments confirmed that removing feature selection increased prediction error to 5.32 percent on the primary dataset, while replacing the temporal convolutional architecture with gated recurrent units, long short-term memory networks, or conventional convolutional layers increased errors to 5.12 percent, 5.17 percent, and 5.48 percent respectively. Analysis of severe local prediction deviations, such as an observed interval between hours 1250 and 1350 where local error exceeded 0.15 volts, revealed that forecasting failures coincided with rapid load steps and thermal fluctuations that departed from smooth training histories. Because the network incorporates physical operating inputs without embedding explicit electrochemical governing equations, catalyst dissolution dynamics, or water transport constraints, the researchers concluded that future prognostic models must undergo regime-by-regime validation across independently controlled temperature, humidity, stoichiometry, and dynamic cycling before claiming true operational generalization.
Subject of Research: Proton exchange membrane fuel cell degradation and remaining useful life prognosis using machine learning
Article Title: Robustness-oriented RF-RFE-TCN framework for PEMFC degradation and remaining useful life prediction
Article References: Ma, H., Tong, Y., Hung, Y. M., & Wang, X. (2026). Robustness-oriented RF-RFE-TCN framework for PEMFC degradation and remaining useful life prediction. Frontiers in Chemistry, 14, Article 1919324. https://doi.org/10.3389/fchem.2026.1919324
Image Credits: AI Generated
DOI: 10.3389/fchem.2026.1919324
Keywords: proton exchange membrane fuel cells, fuel cell degradation, remaining useful life, temporal convolutional networks, random forest feature elimination, particle swarm optimization, Monte Carlo dropout, prognostics and health management, voltage degradation forecasting, IEEE PHM 2014 dataset, FC Lab durability dataset, cross-regime transfer learning, prediction intervals, dilated causal convolutions, data-driven prognostics, operational telemetry, stack voltage prediction
News Source: Faith Mcneil. (November 19, 2026). Machine Learning Pipeline Accurately Tracks Fuel Cell Degradation but Faces Transfer Limits. Scienmag.



