Evaluation of Different Machine Learning Tools in End-to-End Prediction of Vehicle Fuel Consumption in California

Accurate measurement of fuel consumption and related CO₂ emissions are of great importance for stakeholders of the transportation systems. In this study, intelligent machine learning models were trained based on comprehensive field testing in California, to rectify the accuracy of fuel consumption measurements. In addition to conventional regression model, more complex model representations such as neural network and ensemble method were incorporated in the learning phase. To achieve optimal performance, the corresponding hyperparameters were tuned to the dataset. To avoid overfitting, first, the dataset was split into training and validation sets. Then, procedures such as early-stopping and regularization were applied to the learners to keep the model representation as simple as possible. The learned parameters were validated by tracking the accuracy of the predictions on the validation set. Finally, the performance of the candidate models was compared. The results reveal higher accuracy achieved by the ensemble method with additional generalization. The accuracy trends assured that the learned models are not prone to overfitting. The ensemble model can be used as a powerful tool in decision-making, planning, and management of transportation systems.