At the chair of Naturalistic Driving Observation for Energetic Optimisation and Accident Avoidance of the Technische Universität Berlin, currently the project “eHaul” is operated. It is funded by the Federal Ministry for Economic Affairs and Climate Action and deals with the fully automatic exchange of the traction battery of electric trucks, primarily in order to avoid long downtimes of the vehicles when charging the batteries. In order to extend service life and increase the efficiency of the used batteries as well as the utilisation rate of the battery swapping station, an energy demand model for the vehicles is developed. It consists of various sub-models, whereby this presentation will focus the powertrain sub-model. The drivetrain sub-model has been built by using machine learning methods and basic measurement signals from the vehicle as well as secondary data. The motivation is on the one hand a modelling that is independent from type-specific vehicle parameters (exclusive manufacturer knowledge), which are essential when building up physical models. On the other hand, the modelling should be set up "sustainably", whereby the model structure can also be retrained for new vehicles added in the future and only needs to be relearned with a new set of measurement data. The time-consuming search for the appropriate input signals and hyperparameters is then obsolete. This presentation shows the implemented procedure for creating this adaptable drivetrain sub-model. Starting with a study on the selection of a suitable machine learning method, two different approaches for processing the input signals are presented. Subsequently, iterative optimization steps and a validation test of the final model are shown. The drivetrain sub-model for heavy battery electric trucks is represented by a neural network with one hidden layer. Only the variables mass, velocity, acceleration and road gradient are used as input signals. The driving resistance equations for road vehicles show that these variables contribute to the energy demand in non-linear relationships. However, in order to increase the model performance, this physical knowledge was incorporated into a pre-processing of the measurement signals and at the same time the architecture and function of a neural network were utilized. Furthermore, it was shown that the ambient temperature has no significant influence on the energy demand of the drivetrain. The final set of optimal hyperparameters for the neural network was applied to two battery electric trucks (18t panel van and 40t semi-truck) in order to build a vehicle-specific powertrain model based on their measurement data. Despite different data qualities, small deviations of the energy demand for recorded journeys in the low single-digit kWh range could be achieved in the simulation. Due to the high total vehicle mass, heavy commercial vehicles in particular have an increased energy demand, which leads to special challenges for their drivetrain electrification. In this context, high quality data are essential for modelling the drivetrain. Unfortunately, inconsistent data for weight and electric propulsion power demand could be detected on one considered truck. Nevertheless, satisfying results could be achieved. Powertrain and energy demand models for vehicles are a standard problem in primarily industry-driven vehicle development. However, a certain specialty of this presentation is, that it deals with heavy electric trucks, which currently barely start entering the markets. Therefore, there are still few published studies on this special subject. Furthermore, the authors are currently endeavouring to complement the powertrain sub-model with further sub-models for the HVAC system, low-voltage auxiliary consumers, HV battery and the electrical supply of a refrigerated trailer to form an overall vehicles energy demand model. A universal model structure was found which allows to train a simple but sufficiently accurate sub-model with few required measurement signals to predict the energy demand of the powertrain of a heavy battery electric truck. Although there were different data qualities and the method was applied to different vehicle classes, it was able to produce models with satisfying results.

Mr. Martin Gobernatz, Scientific Assistant, Technische Universität Berlin