Machine learning assisted prediction of exhaust gas temperature of a heavy-duty natural gas spark ignition engine

Abstract

Exhaust gas temperature is a key parameter for optimizing engine performance and emissions. Of particular interest is forecasting the exhaust gas temperature in a diesel engine converted to spark ignition natural gas operation, as the combustion process in this engine is significantly different from the one in a conventional engine. The goal was to assess four different machine learning algorithms namely the artificial neural network, random forest, support vector regression, and gradient boosting regression trees, with respect to a physical 1D CFD model and relative to one another, when predicting the exhaust temperature. The spark timing, equivalence ratio, and engine speed were model inputs. First, the artificial neural network predicted the exhaust temperature more accurately than the physical model, because of the complex premixed combustion phenomena inside a conventional diesel chamber. When compared relative to one another, all machine learning models predicted the exhaust gas temperature with acceptable error while also capturing its relationship with the three model inputs. The gradient boosting regression trees predicted the best, but it usually requires high quality noise-free data. The random forest had the least accuracy, but it required the least amount of calibration. The support vector regression had the smallest error, but it required the highest computational resources. The artificial neural network algorithm was the most appropriate, but it required effort in tuning its hyperparameters. Overall, the results showed that well-trained machine learning models can complement more complex physical model while also helping with optimizing the engine performance, emissions, and life.