New uncertainty-aware AI framework could improve fuel cell degradation forecasting

Researchers have developed an uncertainty quantification-based framework for predicting degradation trends in proton exchange membrane fuel cells, aiming to make fuel-cell prognosis more reliable under realistic operating conditions. Instead of offering only a single predicted value, the new method provides both point estimates and interval estimates with probability density information, a step that could make degradation forecasting more useful for real control and maintenance decisions in fuel-cell vehicles and stations.

Accurately predicting degradation in proton exchange membrane fuel cells, or PEMFCs, is important because fuel-cell performance does not decline in a perfectly smooth or fully predictable way. In practical systems, degradation can be influenced by load variation, environmental conditions, noise in measured data, and uncertainty inside the prediction model itself. If those uncertainties are ignored, operators may get a forecast that looks precise but does not offer much confidence about how trustworthy it really is. That is a serious limitation for applications such as vehicle control, station management, maintenance scheduling, and performance optimization, all of which benefit from knowing not only what degradation trend is expected, but also how uncertain that forecast may be.

The authors of the new study argue that many existing prediction methods still fall into exactly that trap. They can provide point estimates, but they do not adequately consider measurement errors from experimental environments or the inherent cognitive uncertainty of the model. As a result, the output may be numerically convenient but operationally incomplete. A more useful prediction framework, especially for engineering systems that evolve over time, would ideally account for uncertainty directly and express degradation in a way that better reflects the actual reliability of the forecast.

To address this need, the researchers developed a deep-learning prediction framework that combines a bidirectional gated recurrent unit, or BiGRU, model with a truncated Bayes by backpropagation through time algorithm, referred to in the paper as TB. According to the article, the TB algorithm reconstructs fixed model parameters into probability density distributions, effectively transforming the model's output from a simple point estimate into an interval estimate accompanied by probability density information. In practical terms, this means the framework is designed not just to predict degradation, but to describe the uncertainty around that prediction in a mathematically explicit way.

That shift is important because degradation forecasting in fuel cells is not purely a pattern-recognition problem. It is a decision-support problem. If a controller, operator, or system planner is going to rely on a prediction, it helps to know whether the model is confident or uncertain, stable or sensitive, and robust or easily disturbed by noise. The paper therefore frames uncertainty quantification not as a secondary add-on, but as part of the core prediction task. It also introduces four evaluation indicators intended to characterize both forecasting accuracy and uncertainty quantification capability, helping provide a more informative assessment of model quality than conventional error metrics alone.

The reported performance improvements are notable. Under dynamic conditions, the proposed TB-BiGRU model improved mean absolute error and root mean square error by 37.28% and 36.09%, respectively, compared with TB-GRU. The study also states that TB-BiGRU significantly outperformed both TB-GRU and Bayesian GRU in uncertainty quantification capability. These gains matter because they suggest the model is not simply adding uncertainty intervals to an otherwise unchanged predictor. Instead, it is improving the underlying forecasting quality while also strengthening the credibility and interpretability of the result.

The framework was further evaluated under different working conditions and noise levels, and according to the paper, TB-BiGRU remained superior to seven comparison models in prediction accuracy while showing better noise resistance and stability. This is particularly important for PEMFC applications, where operating conditions can change frequently and sensor data may not always be clean. A model that performs well only under narrow or low-noise test settings may have limited field value. By contrast, a model that maintains strong accuracy and uncertainty characterization under varying conditions is much better aligned with real engineering needs.

Taken together, the study suggests that the future of PEMFC degradation prediction may depend not only on making models more accurate, but on making them more informative. More work will still be needed to validate the approach in broader real-world datasets and operational systems. Even so, the framework offers a strong example of how uncertainty-aware deep learning could improve prognosis for hydrogen energy technologies. For fuel-cell vehicles and stations, where control and maintenance decisions are often made under imperfect information, predictions that include both expected trends and quantified confidence could be far more valuable than point estimates alone.

Reference

Author:

Bingxin Guo ^a, Changjun Xie ^{a b c}, Wenchao Zhu ^{a b}, Yang Yang ^{a b}, Hao Li ^a, Yang Li ^d, Hangyu Wu ^a

Title of original paper:

Uncertainty quantification-based framework for predicting degradation trends of proton exchange membrane fuel cell

Article link:

https://www.sciencedirect.com/science/article/pii/S2773153725000477

Journal:

Green Energy and Intelligent Transportation

DOI:

10.1016/j.geits.2025.100297

Affiliations:

a School of Automation, Wuhan University of Technology, Wuhan 430070, China

b Hubei Key Laboratory of Advanced Technology for Automotive Components, Wuhan University of Technology, Wuhan 430070, China

c Modern Industry College of Artificial Intelligence and New Energy Vehicles, Wuhan University of Technology, Wuhan 430070, China

d Department of Electrical Engineering, Gothenburg 41296, Sweden