Long Short-Term Memory Networks for Data-Driven Modeling of Proton Exchange Membrane Fuel Cell Systems

Fuel cells (FC) offer a clean and efficient alternative to conventional power sources for automotive powertrain systems. To ensure a fast development of the balance of plant (BOP) and maintain a high system efficiency throughout the operating range, fast simulation models are desirable. Conventional physical simulation models often suffer from great computational effort, limiting their efficiency and area of application. To overcome this weakness a data-driven modeling approach using Long Short-Term Memory (LSTM) networks is used to derive a data-driven digital twin from a physical 0D/1D model of an automotive proton exchange membrane fuel cell (PEMFC) system. First, the physical model is presented and used to generate the training data for the data-driven model. The data generation is followed by a description of the data preprocessing and the training of various LSTM networks on both system and component level. The initialization and progression of the time-series prediction of selected physical quantities is outlined and the accuracy in replicating these quantities is discussed. This work underlines the benefits of data-driven modeling of physical systems using LSTM networks regarding computational effort and accuracy. The derived data-driven digital twins provide a faster processing time than the physical counterpart in the order of several thousand with at the same time minimal deterioration in accuracy. As a result, this methodology facilitates a faster development process and paths the way for optimized control strategies using data-driven ultra-fast running simulation models.