Bränsleförbrukningsprediktion för bussar : Maskininlärning medvägrelaterade faktorer

Heavy vehicles contribute considerably to greenhouse gas emissions around the world, and fuel consumption leads to significant economic costs. This thesis investigates how the fuel consumption of buses can be modeled from roadway-related features, using different machine learning models. The results can be used for optimizing and analyzing routes for heavy vehicles, including buses. The methods in the thesis require the use of several sources of data, including road network data from multiple sources, fuel consumption data, and GPS data from buses. Extensive preprocessing is performed, including the development of an algorithm for matching a vehicle’s GPS signals to road segments. Several machine learning models are implemented. These are XGBoost and Multi Layer Perceptron networks, as well as the two sequential models: Long Short-Term Memory networks and transformer architectures. The results for the models show that the Long Short-Term Memory model gives the best results when comparing the actual fuel consumption with the predictions of the respective models. However, all models have difficulties generalizing to new, unseen roads. Notably, a simple K-Nearest Neighbors model, when trained and tested within the same geographical region, outperforms the more complex models trained and evaluated across different regions. These results suggest that sequential models can be advantageous for predicting fuel consumption from roadway-related features, but also that a significant challenge lies in the models’ difficulty with generalization.