Diesel Engine Modeling and Optimization for Emission Reduction

Liu, Y. Diesel Engine Modeling and Optimization for Emission Reduction. University of Wisconsin-Madison, 2005.

This work investigates the benefits of using multiple injection strategies with up to 5 pulses/cycle for optimizing HSDI diesel combustion and emissions. An improved multi-step phenomenological soot model, enabled by its high computational efficiency, was adopted for the computational optimizations using the KIVA-3V code. This model was validated on heavy-duty and HSDI diesel engines with single and split injections over wide operating ranges. Excellent agreements were seen with the measured data for engine-out soot quantity, soot volume fraction, particle size, and number density.

The models were applied using an optimization framework, which consists of tools of searching, evaluation, and data analysis. A new local search technique called the Non-Gradient Step-Controlled (NGSC) algorithm was developed that places no requirement on response surface continuity/differentiability. A global optimization technique, called Path-Tracking Global NGSC (PTGN) algorithm was proposed. The PTGN solves the “exploration vs. exploitation” problem by systematically combining the local (NGSC) and global search techniques. By comparing to Genetic Algorithm—based optimization on the same HSDI split injection problem, the PTGN algorithm exhibited significantly higher search efficiency and accuracy, in addition to stronger global convergence.

A non-parametric regression method, called the Component Selection and Smoothing Operator (COSSO), was revised and coded to serve as the data analysis tool. The COSSO method was applied on the data obtained from the engine optimizations. Models correlating engine responses (e.g., merit function value, emissions, BSFC) and engine operating parameters were constructed through the data-driven processes by extracting information from the data samples. Though facing the challenges of undesigned, un-evenly-distributed, and small size data samples with regard to the control factors’ high-dimensionality, complex correlations between responses and factors, and complicated interactions between control factors, cross validations showed satisfying agreements between the sample data and model predictions. Based on the well-constructed models, the importance of the control factors could be ranked according to their significance in the models, and interactions between control factors and their influence on responses could be quantitatively assessed. Moreover, the solutions can be diversified to meet various criteria. The potential of using the COSSO method for assisting optimal parameter search was also explored.