Development and assessment of a random-forest model for real-time prediction of diesel engine-out particulate matter emissions

The containment of air pollution resulting from transportation sector is worldwide considered as one of the most important targets to be still achieved. The emissions of particulate matter (PM) assume a relevant role when Diesel engine-based vehicles are concerned. In the present paper, an innovative model for the prediction of engine-out emitted PM mass has been developed and assessed on two different Diesel engines. A semi-empirical approach is defined as reference model and initially validated through the ISO/IEC Guide 98-3:2008 procedure and its potentials of the being coupled to a predictive combustion model are demonstrated. Then, relying on the findings of the semi-empirical approach, a Random Forest (RF) algorithm has been thoroughly analysed and selected as a promising solution for real-time testing with on-line computed variables. Furthermore, an automatic feature selection and calibration procedure of the algorithm hyper-parameters has been developed. Very interesting performances were recorded as the reference model prediction accuracies were reproduced and comparable results were obtained when only ECU-measured variables were considered. The presented Random Forest model can be intended to be part of a pollution-oriented real-time powertrain control strategy that could on accurate and repeatable PM estimations.