A study based on a data-driven machine learning approach
Examinator: Jonas Fredriksson, Inst för elektroteknik
The market for battery electric vehicles (BEVs) is growing steadily because of the environmental benefits, compared to conventional vehicles. The limited range and long charging times are however large issues for BEVs. To diminish the issue of range-anxiety among drivers, the energy consumption prediction needs to be accurate when setting a destination in the satellite navigation.
This thesis aims to develop machine learning approaches that can predict the energy consumption for BEVs, split into propulsive and auxiliary consumption. The data used will consist of vehicle collected data from Volvo Cars and geographical data from a navigation supplier. Since a machine learning model’s performance relies on its inputs, a major part of the thesis will be spent on choosing the correct parameters and modifying the data to create better predictions.
Linear regression, multi-layer perceptron, recurrent neural network, and gradient boosting were all machine learning models that were compared in the study. After the study it can be concluded that, for the prediction of the propulsive energy consumption, the choice of model did not matter significantly. The performance
of the auxiliary prediction was very similar for all models except the RNN, which showed worse results. It was also concluded that a more reasonable approach to show the predicted energy consumption for a trip, is to show a span of possible consumption instead of an absolute value. The main reason being that there is always a difference between the predicted and actual speed, acceleration, and total time for a trip.