Data-driven combustion modeling of homogeneous charge compression ignition processes for real-time use in model predictive control



Patrick Schaber

Research Associate


+49 241 80 48151



A growing global demand for mobility currently leads to 20% of annual global CO2 emissions. In addition, pollutant emissions in metropolitan areas continue to rise. New demands of an increasingly conscious population are also leading to a major need for regenerative and low-emission drive alternatives. These challenges are met by a combustion process that has hardly been exploited to date in the form of low-temperature combustion (NTV). NTV offers great potential for increasing efficiency and reducing pollutants within the engine. This makes it a promising process for resolving the conflicting goals of simultaneously reducing greenhouse gas and pollutant emissions.

Subject is the development of a combustion model of the NTV. Given the complexity and large amount of data, physical modeling is not feasible in real time using current technology. As an alternative to computation time intensive physical modeling, the NTV will be modeled using artificial neural networks (aNN). Using aNN, abstract modeling of the relationships between input and output variables can be achieved. To model changes due to cycle couplings with greater optimality, network architectures such as recurrent neural networks and convolutional neural networks are used. These classes of aNN are able to interpret patterns and temporal changes in the best way possible.

Objective of this project is to contribute to a stabilization of the overall process in the cylinder. For this purpose, an already existing physical charge exchange model will be extended in order to be able to fully model the overall process. The cylinder state at the beginning of combustion is provided by the charge exchange model. Subsequently, the combustion model predicts the combustion and feeds the cylinder state back into the charge exchange model.

To control the combustion model, it is necessary to integrate the model into a model predictive control (MPC) with central multivariable control. In the MPC, a so-called online process optimization takes place. Optimization problems are solved in real time and a high computational effort is required. Here, a control unit is necessary to ensure a real-time capable computation.

Results of the work can later provide important insights into the controllability of the NTV for research group FOR2401. In subsequent steps, the results could contribute to a technical use of the advantages of low-temperature combustion. Possible follow-up projects regarding the combination of NTV and e-fuels also promise further advantages of this combustion process.



RWTH Aachen