Predicting Degradation of Fuel Cell Membrane Electrode Assemblies from Buses

Harsh Srivastav and Alexandre Bayen

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2025-107
May 16, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-107.pdf

The Alameda Contra Costa Transit District (AC Transit) has provided lifetime data for a limited fleet of their buses that were run on hydrogen fuel cells. Opportunities to study such longitudinal data in clean energy transportation systems are uncommon, particularly over the complete operational lifespan of vehicles. This thesis leverages that dataset to investigate whether early-life performance indicators can be used to accurately predict long-term degradation behavior of fuel cell systems. We have run machine learning models on the predictions of the data to study whether later life degradation behavior can be accurately estimated from earlier time performance. To do so, we have explored using recurrent neural network type models, with variations in both the loss function and the data being trained upon. Long term predictions (over the course of years) are also presented for multiple buses with trends analyzed. These findings have implications for predictive maintenance, fleet management, and the broader deployment of hydrogen-powered transportation systems. Additionally, we discuss the limitations of the modeling approach and suggest future directions for improving predictive accuracy with hybrid modeling strategies and additional contextual data.

Advisor: Alexandre Bayen

\"Edit"; ?>

BibTeX citation:

@mastersthesis{Srivastav:EECS-2025-107,
    Author = {Srivastav, Harsh and Bayen, Alexandre},
    Title = {Predicting Degradation of Fuel Cell Membrane Electrode Assemblies from Buses},
    School = {EECS Department, University of California, Berkeley},
    Year = {2025},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-107.html},
    Number = {UCB/EECS-2025-107},
    Abstract = {The Alameda Contra Costa Transit District (AC Transit) has provided lifetime data for a limited fleet of their buses that were run on hydrogen fuel cells. Opportunities to study such longitudinal data in clean energy transportation systems are uncommon, particularly over the complete operational lifespan of vehicles. This thesis leverages that dataset to investigate whether early-life performance indicators can be used to accurately predict long-term degradation behavior of fuel cell systems. We have run machine learning models on the predictions of the data to study whether later life degradation behavior can be accurately estimated from earlier time performance. To do so, we have explored using recurrent neural network type models, with variations in both the loss function and the data being trained upon. Long term predictions (over the course of years) are also presented for multiple buses with trends analyzed.  These findings have implications for predictive maintenance, fleet management, and the broader deployment of hydrogen-powered transportation systems. Additionally, we discuss the limitations of the modeling approach and suggest future directions for improving predictive accuracy with hybrid modeling strategies and additional contextual data.}
}

EndNote citation:

%0 Thesis
%A Srivastav, Harsh
%A Bayen, Alexandre
%T Predicting Degradation of Fuel Cell Membrane Electrode Assemblies from Buses
%I EECS Department, University of California, Berkeley
%D 2025
%8 May 16
%@ UCB/EECS-2025-107
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-107.html
%F Srivastav:EECS-2025-107