Machine Learning for Emission Modeling of Fossil-fueled and Hydrogen-fueled Internal Combustion Engines

• Author / Creator

  Shahpouri, Saeid

• Development of fast and accurate emission models for engine-out and tailpipe of internal combustion engines (ICEs) using machine learning (ML) and hybrid methods are the focus of this thesis. The application is on medium and heavy-duty vehicles powered by both fossil fuels and alternative fuels like hydrogen. This thesis is structured in three interconnected phases that sequentially build upon each other to establish steady-state and transient emission models for fossil-fueled and hydrogen-fueled engines.
  The first phase proposes a new approach for physics-based combustion modeling
  of low-carbon fuels, by replacing traditional thermo-kinetic combustion mechanisms
  with computationally efficient ML models for laminar flame speed (LFS) calculation.
  LFS is a crucial input for physics-based pre-mixed combustion models. The LFS models are developed by creating a large dataset of LFS through combustion mechanism simulations and training ML methods using this dataset. The results are models that can predict LFS for blends of low carbon fuels including hydrogen, methanol, and ammonia over wide range of temperature, pressure, exhaust gas recirculation (EGR) and air-fuel equivalence ratio that represents combustion conditions in ICEs. These models are hundreds of thousands times faster than traditional thermo-kinetic models in LFS calculation.
  To optimize steady-state operation of an engine it is important to model and
  predict engine-out emissions. The second phase of this research focuses on developing steady-state emission models for both fossil-fueled and hydrogen-fueled engines. Physics-based, black-box (BB), and gray-box (GB) emission models are created and compared against each other to predict emissions from a compression ignition (CI) diesel engine. Subsequently, steady-state BB and GB NOx and soot emissions models for the hydrogen-diesel engine are developed. The resulting BB and GB models can predict NOx and soot emissions for both fossil-fueled and hydrogen-fueled engine with R2 between 0.95 and 0.99 considering the trade-off between accuracy and computational cost.
  The third phase focuses on transient emission modeling. First, transient emission
  models of the dual-fuel hydrogen-diesel engine are developed using the experimental transient data of engine-out emissions. Training models using steady-state data and directly training models with transient data are investigated. Classic ML algorithms and deep-learning (DL) time-series networks are utilized for GB and BB transient emission modeling. The models developed using transient data and times series network showed the best performance, predicting transient NOx emissions with R2 between 0.96 and 0.97. Next, transient tailpipe emissions from a heavy-duty truck are investigated using real-driving on-road emission data from a Class 8 truck, from over 10000 Km of driving. This data is employed to develop various time-series DL algorithms with different input feature sets, complexity levels, and training dataset sizes. The best developed model can predict tailpipe real driving instantaneous NOx emissions with R2 of 0.91 and cumulative NOx emissions with less than 2% error.
  The computationally efficient BB emission models developed in this study can predict over 10000 cases per second which makes them suitable for engine and aftertreatment system model-based control. On the other hand, the GB models provide higher accuracy but require more computational power, which makes them suitable for diagnostics, calibration, and hardware-in-the-loop (HIL) setups for performance optimization and tailpipe emission reduction without requiring expensive experimental tests. These techniques can contribute to mitigating emissions from the transportation sector.

• Subjects / Keywords

  • Machine Learning
  • Emission Modeling
  • Hydrogen
  • Internal Combustion Engines
  • Deep Learning

• Graduation date

  Fall 2023

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/r3-rqfz-wh55

• License

  This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.