Real-driving CO2, NOx and fuel consumption estimation using machine learning approaches

Real driving emissions (RDE) testing are gaining attention for monitoring and regulatory purposes because of providing more realistic emission and fuel consumption measurements compared to laboratory tests. This study aims to develop machine learning (ML) based emission and fuel consumption estimation models using real- driving measurement data. A light-duty diesel vehicle equipped with a portable emissions measurement system (PEMS) was driven in an urban test route by 30 participant drivers of disparate backgrounds to obtain a wide variety of data in terms of driving behaviour and traffic conditions. The Pearson correlation coefficient was used to select the input variables among 36 driving behaviours and 6 engine parameters. The CO₂, NO_x and fuel consumption prediction models were developed using linear regression (LR), support vector machine (SVM) and Gaussian process regression (GPR). The results showed that all three models could predict CO₂ with an absolute relative error (ARE) of less than 9%. The GPR model showed the best performance in CO₂ prediction with an R² of 0.74 and ARE of 3.30%. LR model showed the best prediction accuracy for NO_x with an R² of 0.80 and ARE of 8.91%. All three models worked well for fuel consumption prediction, however, GPR showed the best accuracy with an R² of 0.81 and ARE of 3.52%. This method lays a foundation for developing route/region specific emission and fuel consumption models that will help to monitor and reduce the environmental impact and the amount of burned fuel. Moreover, developing models from different driver classes will provide valuable insights into emission-optimal driving behaviour which could be used to train new drivers.