Prediction of hazardous gaseous emissions from a gasoline engine during cold starts using machine learning methods

Internal combustion engines generate higher exhaust emissions of hazardous gases during the initial minutes after engine start. Experimental data from a state-of-the-art turbo-charged 3-cylinder, 999 cc gasoline engine are used to predict cold start emissions using two Machine Learning (ML) models: a Multilayer Perceptron (MLP) which is a fully connected neural network and an Encoder-Decoder Recurrent Neural Network (ED-RNN). Engine parameters and various temperatures are used as input for the models and NOx (Nitrogen Oxides), CO (Carbon monoxide) and unburned hydrocarbon (UHC) emissions are predicted. The dataset includes time series recordings from the Worldwide harmonized Light-duty vehicles Test Cycle (WLTC) and four Real Diving Emissions (RDE) cycles at ambient and initial engine temperatures ranging from -20 °C to +23 °C. In total, 21 cases are considered, consisting of eight different ambient temperatures and five distinct driving cycles. Each case consists of a sequence of 2500 samples taken at 5 Hz. The training process utilized seven input variables and three output variables (emissions). Two validation scenarios were defined. The first scenario assessed the ability of the models to predict emissions at ambient temperatures not included in the training process. The second, more challenging scenario, tested the ability of the models to predict emissions for unseen driving cycles, but at temperature levels that were included in the training process. Both models predicted the validation cases with reasonable accuracy in the first scenario. However, in the second scenario, the MLP model failed to predict the data accurately, while the ED-RNN model delivered significantly better results, thus demonstrating greater robustness. The low inference CPU-time (Central Processing Unit) of the ED-RNN model makes it suitable for real-time prediction and emission control.