Support vector machine based emissions modeling using particle swarm optimization for homogeneous charge compression...

Area of Science:

• Internal combustion engine research
• Combustion science
• Artificial intelligence in engineering

Background:

• Internal combustion engines face pressure to improve efficiency and reduce emissions.
• Homogeneous Charge Compression Ignition (HCCI) is a promising efficient combustion strategy.
• Predicting HCCI engine emissions is complex and challenging for conventional models.

Purpose of the Study:

• Develop a simple, accurate model for predicting steady-state emissions of a single-cylinder HCCI engine.
• Explore feature selection methods for optimizing machine learning models.
• Compare the performance of Support Vector Machine (SVM) models against Artificial Neural Networks (ANNs) for HCCI emission prediction.

Main Methods:

• Utilized a metaheuristic optimization-based Support Vector Machine (SVM) for emission prediction.
• Investigated five different feature sets, including up to seven engine inputs, for SVM model development.
• Compared linear and non-linear SVM models with an Artificial Neural Network (ANN) model.

Main Results:

• The best SVM model, using 26 features (linear, squared, and cross-correlated inputs), achieved R-squared values between 0.72 and 0.95.
• CO and CO2 emission predictions showed the best model fits.
• SVM models demonstrated greater robustness to feature selection and avoidance of local minimums compared to ANNs, especially with limited data.

Conclusions:

• The developed machine learning-based HCCI emission models provide accurate predictions.
• The feature selection approach optimizes model accuracy while minimizing computational costs.
• SVM models offer a more consistent prediction performance than ANNs when training data is limited.