Power-split hybrid powertrains represent one of the most advanced and complex types of powertrain systems. The combination of multiple energy sources and power paths offers great potential but results in complex interactions that require improved strategies for optimal efficiency and emission control. The development and optimization of such operating strategies typically involve algorithms that demand fast computational environments. Traditional high-accuracy numerical simulations of such a complex system are computationally expensive, limiting their applicability for extensive iterative optimizations and real-time applications. This paper introduces a data-based approach designed specifically to address this challenge by efficiently modeling the dynamic behavior of power-split hybrid powertrains using cascaded neural networks. Cascaded neural networks consist of interconnected subnetworks, each specifically trained to represent individual drivetrain components or subsystems. This modular structure allows the networks to cover a larger parameter space within each subnetwork effectively and enables the generation of larger and more targeted training datasets using simpler, separated numerical models. However, this approach is also challenging due to the need for step prediction, potential error accumulation and the requirement for feedback loops. Assumptions are necessary to overcome these limitations and improve the model predictions. The focus is on the possible connection to optimization algorithms with targeting key parameters such as battery state of charge, vehicle velocity, and emissions profiles. By significantly reducing computational demands compared to complex and detailed simulations, the cascaded data-driven models state a promising basis for real-time applications, enabling sophisticated control strategies tailored to the complexities inherent to hybrid electric vehicles. This method accelerates the optimization process, enhances adaptability and scalability of control strategies, and significantly contributes to the development of cleaner, more efficient vehicles.
Growing environmental concerns drive the increasing need for a more climate-friendly mobility and pose a challenge for the development of future powertrains. Hydrogen engines represent a suitable alternative for the heavy-duty segment. However, typical operation includes dynamic conditions and the requirement for high loads that produce the highest NOx emissions. These emissions must be reduced below the legal limits through selective catalytic reduction (SCR). The application of such a control system is time-intensive and requires extensive domain knowledge. We propose that almost human-like control strategies can be achieved for this virtual application with less time and expert knowledge needed by using Deep Reinforcement Learning (DRL). A DRL agent is trained to control the injection of diesel exhaust fluid (DEF) and compared with the performance of a manually tuned controller. The performance is evaluated based on the restrictive emission limits of a possible EURO7-framework and DEF consumption. Applied to a standardized driving cycle (WHTC) and compared with the conventional application, the agent reaches similar emission values with an equally high DEF consumption. The results demonstrate, that the control of an exhaust gas aftertreatment system using DRL is very satisfactory. Furthermore, it is shown, that the methodology can be applied to different engine calibrations and still satisfactory results are achieved. Further work is required to refine the proposed methodology into a fully-fledged tool for application in powertrain development.
To support the transition toward climate-neutral mobility and power generation, internal combustion engines (ICEs) must operate efficiently on renewable, carbon-neutral fuels. Hydrogen, methanol, and ammonia-hydrogen blends are promising candidates due to their favorable production pathways and combustion properties. However, their knock behavior differs significantly from conventional fuels, requiring dedicated simulation tools. This work presents a modeling framework based on quasi-dimensional (QD) engine simulation, including two separate knock prediction models. The first model predicts the knock boundary of a given operating point and combines an auto-ignition model with a knock criterion. The overall methodology was originally developed for gasoline and is here adapted to hydrogen, methanol, and ammonia-hydrogen blends. For this purpose, the relevant fuel properties were incorporated into the auto-ignition model, and a suitable knock criterion was identified that applies to all investigated fuels. The model was validated using experimental data from single-cylinder engine tests. In addition, two entirely new modeling approaches were developed to predict statistical knock values, specifically knock frequency and knock intensity. Each model was calibrated once per fuel and subsequently validated across a wide range of conditions. The results show that the adapted knock boundary model and the new statistical model accurately capture the knock behavior of hydrogen, methanol, and ammonia-hydrogen blends. The methodology enables predictive knock analysis using QD simulation and supports the development of robust, high-efficiency ICEs for future carbon-neutral applications.
