Random Forest Ensemble-Based Predictions of On-Road Vehicular Emissions and Fuel Consumption in Developing Urban Areas

Abstract

The transportation sector is one of the primary sources of air pollutants in megacities. Strict regulations of newly added vehicles to the local market require precise prediction models of their fuel consumption (FC) and emission rates (ERs). Simple empirical and complex analytical models are widely used in the literature, but they are limited due to their low prediction accuracy and high computational costs. The public literature shows a significant lack of machine learning applications related to onboard vehicular emissions under real-world driving conditions due to the immense costs of required measurements, especially in developing countries. This work introduces random forest (RF) ensemble models, for the urban areas of Greater Cairo, a metropolitan city in Egypt, based on large datasets of precise measurements using 87 representative passenger cars and 10 typical driving routes. Five RF models are developed for predicting FC, as well as CO2, CO, NOx, and hydrocarbon (HC) ERs. The results demonstrate the reliability of RF models in predicting the first four variables, with up to 97% of the data variance being explained. Only the HC model is found less reliable due to the diversity of considered vehicle models. The relative influences of different model inputs are demonstrated. The FC is the most influential input (relative importance of >23%) for CO2, CO, and NOx predictions, followed by the engine speed and the vehicle category. Finally, it is demonstrated that the prediction accuracy of all models can be further improved by up to 97.8% by limiting the training dataset to a single-vehicle category.