To reduce fuel costs and the environmental impacts of the transportation sector, efficient fuel management is crucial. Traditionally, fuel consumption in the transportation sector has been predicted using empirical and simple analytical models, which has contributed to further research into how machine learning models can improve these predictions. This study investigates and compares how well the machine learning models Linear regression, Random Forest and LightGBM can predict fuel consumption, as well as which driving style-related parameters and external factors mainly affects the predicted consumption. 

In the study, the quantitative approach and the experimental research method are central, where machine learning models are trained and then evaluated. The data the models are trained on is collected from two cars, and to increase the possibility of the training process returning the most optimal results, the data is pre-processed. After the machine learning models are trained, they are evaluated using the evaluation metrics R2-score, Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). Finally, the most prominent parameters related to fuel consumption are evaluated and identified, using a SHAP (Shapley Additive exPlanations) analysis. 

The results of the study show that LightGBM predicted fuel consumption with the highest accuracy of all the models, followed by Random Forest with a slightly reduced accuracy. The linear regression model performed significantly worse, mainly due to its inability to discover complex and non-linear relationships. Furthermore, the SHAP analysis shows that the variables EngineCoolantTemperature and VehicleSpeedInstantaneous were assigned the highest weight in the predictions of all models, meaning that the engine temperature and the instantaneous speed of a vehicle have the highest impact on the predicted fuel consumption out of the factors included in this study.