Parameter Identification for Cooperative SOFC Models on the GPU (Vortrag)
|Parameter Identification for Cooperative SOFC Models on the GPU (Vortrag)|
|Autor||Ekaterina Auer, Andreas Rauh|
|Ort||Szeged, Ungarn (online)|
|vom||13. September 2021 bis 15. September 2021|
Over the last decade, using graphic processing units (GPUs) for scientific computations has given a significant boost to such varied fields as neural networks, bioinformatics, medical imaging, or cryptography. Especially in control engineering, this paradigm can open up possibilities which remained unexplored because of the lack of cheap computing power.
One of such fields is modeling, parameter identification, simulation, and control of solid oxide fuel cells (SOFCs). Models for SOFC temperature are based on partial-differential equations, which are usually discretized wrt. space and time into algebraic equations. A disadvantage of this technique is the lack of flexibility: Only stationary states of SOFC systems can be simulated this way, which is unsuitable for control. Using the same finite volume method without discretization in time, it is possible to arrive at dynamic system models consisting of a set of ordinary differential equations which can be shown to be cooperative. A challenge here is to identify a large number of parameters based on uncertain measured values from the SOFC test rig.
Our special focus is on the property of cooperativity. For a cooperative system with uncertain but bounded parameters, two bracketing systems with crisp parameters can be defined to catch the bulk of uncertainty. (These bounding systems might be coupled with each other if lower and upper interval bounds for the system parameters to be identified appear simultaneously in an equation.) A brute force approach using the GPU would be to partition the parameter search space and evaluate the system over the subintervals in parallel, eliminating the regions inconsistent with available measurements. To avoid a prohibitively large number of system evaluations due to naive interval multi-sectioning schemes, we propose to employ a set of additional simple consistency tests. They represent knowledge from physics such as non-negativity and strict monotonicity of heat capacities and reaction enthalpies that occur multiple times in the dynamic SOFC model (common subexpressions). Such constraints in combination with inequalities reflecting physically meaningful temporal variation rates of the measured SOFC stack temperature help to reduce the number of parameter subintervals. This preprocessing stage is carried out prior to the parallelized evaluation of the system model.
In this contribution, we extend the GPU-based technique to deal with the reaction phase of a dynamic SOFC model. Considering this phase separately from the electrochemically idle heating phase represents a further possibility to cope with the high dimensionality of the parameter space. Finally, we show how using Bernstein polynomials helps to reduce the amount of data for controlled and measured system inputs and outputs.