Neural Network-Based Voltage and Degradation Modeling of a PEM Fuel Cell based CHP System

Kurzfassung

Hydrogen plays an important role in the transition from a fossil fuel based energy system to a renewable energy system, as it offers the flexibility renewable energy sources can-not provide. Combined heat and power systems (CHP) produce heat and power simultaneously resulting in a much more efficient use of the primary energy. A key obstacle, which is in the way of widespread application of hydrogen fuel cells is their degradation. Over time the voltage provided by the fuel cell drops. This phenomenon is not yet completely understood physically and is dependent on many internal parameters, which might be different for every manufactured fuel cell. For the CHP system based on a proton-exchange membrane fuel cell currently under research this thesis explores the possibility of designing a completely data driven model to model the voltage output of the CHP system under dynamic load using neural networks as a vehicle. The network used is trained on solely temporal data to model the voltage output of the CHP system. It is shown that close to no physical data is needed to model the voltage even under dynamic load. Due to the dynamic nature of the load, common architectures typically used for this kind of prediction model – namely recurrent neural networks – cannot be used. Instead simple dense layers are used to approximate the voltage development over time, by feeding the future load profile into the system. Various tests are carried out including a variation of the split between training and testing data as well as the generalizability on an unseen fuel cell stack.