Book/Dissertation / PhD Thesis

FZJ-2022-00345

http://join2-wiki.gsi.de/foswiki/pub/Main/Artwork/join2_logo100x88.png Machine Learning in Modeling of the Dynamics of Polymer Electrolyte Fuel Cells

Zou, W. (Corresponding author)RWTH*FZJ*

2021
Forschungszentrum Jülich GmbH Zentralbibliothek, Verlag Jülich
ISBN: 978-3-95806-601-4

Please use a persistent id in citations: http://hdl.handle.net/2128/30164  urn:nbn:de:0001-2022020831

Abstract: Polymer electrolyte fuel cells (PEFCs) are a promising energy conversion technology thatgenerates electricity from hydrogen with low noise, and less or zero emission properties.Phenomena during the fuel cell operation are complex, which are caused by many interrelatedfactors. In addition, the dynamic behaviors of the fuel cells will change due to differentoperating conditions and load changes. A fast response model that can predict the PEFCsdynamic behavior is helpful to implement optimal control to the fuel cell systems obtaining adesired performance.The aim of the thesis is to developing, analyzing and modifying a fuel cell dynamic model,in which a least squares support vector machine (LSSVM) is employed. The efficiency of theLSSVM model is first demonstrated in comparison to experimental data collected from a fuelcell test rig. Analyzing the model’s performance under various fuel cell load changes is carriedout with the help of experimental data collected from our test rig and artificial data generatedby a white-box model that based on the mechanism of the fuel cell systems. Two types ofartificial data are generated: one is idealized artificial data with determined cell voltage andanother one is oscillated artificial data that includes the oscillation on the cell voltage.Various load changes, namely current density changes, are considered in the analysis, andare represented by a combination of two factors called as ramp time and ramp value. Ramp timeis used to show how fast the load is changed and ramp value is used to describe the range ofload change. In addition, considering the data-driven nature of the LSSVM method, samplinginterval of the test rig that determines the frequency of data collection is considered. It is foundthat the performance of the LSSVM model is better when smoother load changes are imposedon the system, so large ramp time and small ramp value are preferable in order to achieve goodmodel accuracy. Moreover, to modeling a high dynamic fuel cell system, a high frequencysampling is suggested to reach a desirable model performance.The thesis defines a working zone for the LSSVM model when predicting the PEFCsdynamic response to sudden load change. Based on the acceptable error to the modeling, a setof workable combinations of sampling interval, ramp time and ramp value can be found. Theworking zone helps to instruct the future application of the LSSVM model when differentoperating load changes are applied.Last but not the least, the LSSVM model is modified in order to improve its modelingperformance when predicting the dynamic behavior of the fuel cell. An online adaptive LSSVMmodel is developed. Determination of initial value of the internal parameters to the LSSVMmodel is optimized by employing a genetic algorithm to search global optimum instead ofmanual search. An adaptive process is carried out to update these internal parameters online.With a suitable starting point of the internal parameters and online updating processes, thisonline adaptive LSSVM model can well deal with complex nonlinear fuel cell systems withfrequent load changes

Note: Dissertation, RWTH Aachen University, 2021

---

Contributing Institute(s):

1. Elektrochemische Verfahrenstechnik (IEK-14)

Research Program(s):

1. 1231 - Electrochemistry for Hydrogen (POF4-123) (POF4-123)

Appears in the scientific report 2021

---

Database coverage:
;

---