Regression-based oxides of nitrogen predictors for three diesel engine technologies

Area of Science:

• Environmental Science and Engineering
• Automotive Engineering
• Computational Modeling

Background:

• Accurate modeling of diesel engine emissions, particularly oxides of nitrogen (NOx), is essential for vehicle system integration, emissions inventory, and control strategy development.
• Modern diesel engines with advanced technologies (e.g., multiple injection, exhaust gas recirculation, variable geometry turbocharging) exhibit increased transient sensitivity, necessitating sophisticated predictive models.
• Legacy engine models may not adequately capture the complex emission dynamics of newer diesel engine designs.

Purpose of the Study:

• To evaluate the effectiveness of Multivariate Adaptive Regression Splines (MARS) and linear regression (LR) for predicting instantaneous oxides of nitrogen (NOx) emissions from U.S. truck engines.
• To compare the predictive performance of these models across different model-year engines (1992, 1999, 2004), considering transient test procedures.
• To assess the utility of incorporating manifold air temperature (MAT) and manifold air pressure (MAP) into a plug-in model for NOx prediction.

Main Methods:

• Emissions data from 1992, 1999, and 2004 model-year U.S. truck engines were analyzed using both linear regression (with transient terms) and MARS.
• Six input variables (torque, speed, power, and their derivatives) were utilized for the MARS models, with emissions time delay considered for both approaches.
• A plug-in model incorporating MAT and MAP was developed and evaluated for NOx prediction.

Main Results:

• MARS demonstrated strong predictive performance for NOx emissions, with R-squared values of 0.981 (1992), 0.988 (1999), and 0.949 (2004) on a portion of the Federal Test Procedure (FTP).
• Linear regression performed comparably for older engines (R2 = 0.981 for 1992, R2 = 0.988 for 1999) but showed lower accuracy for the 2004 engine (R2 = 0.896).
• The MARS model incorporating MAP and MAT yielded the best overall predictive performance, although performance differences between LR and MARS were not substantial across all conditions.

Conclusions:

• MARS offers a viable and effective approach for modeling instantaneous NOx emissions from diesel engines, particularly for newer, more complex engine designs.
• While MARS generally outperformed LR for the 2004 engine, both methods showed limitations in capturing highly transient emission behaviors.
• The development of plug-in models using parameters like MAP and MAT can enhance NOx prediction accuracy, offering practical applications for engine control and emissions management.