Rapid (practical) methodology for creation of fuel cell systems models with scalable complexity

摘要

In order to study various aspects of fuel cell systems, like a fuel cell propulsion system for transportation, several challenges arise: in actual real-world operation, as opposed to benchmark tests, the system is subject to a variety of non-stationary and environmental nuisance factors that are hard to monitor and control; investigating the system's behavior at the limits of its ranges while avoiding any adverse effects; due to sensor capabilities and costs, not every relevant variable can be monitored with sufficiently high temporal resolution. For these reasons, simulation tools are playing a crucial role in the analysis of these system aspects. The first step is therefore to create a mathematical representation of the system (a model) which can then be embedded into a simulation environment. To this end, a methodology is needed for the rapid creation of the mathematical representation of a system which is capable of overcoming the hurdles of dynamic and transient variables. Usually, knowledge-based modeling a system this complex takes several years to accomplish and still does not take nuisance factors into account. In contrast, the approach presented here can be finished within a fraction of that time. We propose to employ black-box adaptive modeling; the key issue in here, selecting an appropriate set of input features, can be solved by either applying iterative wrapper methods, or by making use of the automatic relevance detection technique that has been developed earlier within the framework of Bayesian neural networks. These procedures allow to easily scale the complexity of models in order to accommodate different constraints in terms of modeling effort, sensor availability and cost, and required model accuracy. Our approach can as well be used for the development of diagnostic models for on- and off-board diagnostics.