Long Short-Term Memory Networks for the Prediction of Fuel Cell Voltage and Efficiency

EasyChair Preprint 9126

Abstract

Fuel cells are once again experiencing an upswing, as they are a possibility for climate-friendly mobility. Nevertheless, the aim is to operate them at the highest efficiency, which is highly dependent on the operating condition. In order to operate fuel cells at the highest possible efficiency continuously, fast-calculating and reliable predictions are essential.

One approach to provide these predictions is artificial neural networks (ANN), which are significantly faster compared to phenomenological models. In this work, recurrent neural networks (RNN) are trained with dynamic data of a proton exchange membrane fuel cell (PEMFC). Due to the different time scales of the processes that occur during the operation of the fuel cell, the latest operating state is not sufficient for a precise prediction. Since, for example, the absorption and release of water takes place slowly, earlier states are required in order to consider these processes. Therefore, the choice fell on RNN with long short-term memory cells (LSTM), which are trained using time series of various dynamic operating cycles. Thus, all time scales are regarded in one combined modelt hat offers fast prediction.

Keyphrases: Fast Running Model, LSTM, fuel cell

@booklet{EasyChair:9126,
  author    = {Martin Angerbauer and Michael Grill and André Casal Kulzer},
  title     = {Long Short-Term Memory Networks for the Prediction of Fuel Cell Voltage and Efficiency},
  howpublished = {EasyChair Preprint 9126},
  year      = {EasyChair, 2022}}