Prediction of Solid Oxide Fuel Cell Performance using Artificial Neural NetworkDUAL CONFERENCE IN INNOVATION AND AUTOMATION (2017)
NASA’s Johnson Space Center has recently begun efforts to eventually integrate air-independent Solid Oxide Fuel Cell (SOFC) systems, with landers that can be propelled by LOX-CH4, for long duration missions. Using landers that utilize such propellants, provides the opportunity to use SOFCs as a power option, especially since they are able to process methane into a reactant through fuel reformation. Various lead-up activities, such as hardware testing and computational modelling, have been initiated to assist with this developmental effort.
One modeling approach, currently being explored to predict SOFC behavior, involves the usage of artificial neural networks (ANN). Since SOFC performance characteristics are inherently complex, ANN can account for such nonlinear characteristics of multiple input and output variables to better predict its performance. There are only a handful of studies on the development of dynamic ANN to predict the transient behavior of SOFCs, and none of them account for inlet variables at both electrodes, in the way that this modeling work investigates. The work to be presented involves prediction of the dynamic performance of a SOFC stack while operating conditions change including ramp changes in current and step changes in flow rates. In this work, six input variables considered consist of load current, anode and cathode inlet temperatures, pressures and flow rates, which have not been explored together in this manner before. The output variable that represents the performance is stack voltage. The SOFC performances of the dynamic ANN models of two different structures (time delay only, and time delay and NARX) are judged based on accuracy and consistency. To develop an optimal ANN model, the number of hidden layer neurons is varied from 1 to 25. The model responses (> 10 hidden neurons) were demonstrated to predict the experimental data very well. The coefficient of determinations of all models ranged from 0.960 to 0.995 while their MSE ranged from 0.01 to 2. The results how promise of ANN modeling approaches for real-time prediction of transient behavior and use in control design to assist in the system integration of SOFC into the lander.
- Fuel Cell,
- Neural Network
Publication DateFall October 28, 2017
Citation InformationM. A. Rafe Biswas and Kamwana N Mwara. "Prediction of Solid Oxide Fuel Cell Performance using Artificial Neural Network" DUAL CONFERENCE IN INNOVATION AND AUTOMATION (2017)
Available at: http://works.bepress.com/mohammad-biswas/22/