Year: 2018
Type: dissertation
Abstract: Starting 2019, all cars produced by Volvo Cars will be equipped with an electric engine and by 2025 the company's goal is for 50% of their sales to constitute of fully electric vehicles.Due to battery technology limitations, the main difficulty for Volvo Cars' engineers is to develop vehicles with sufficient range.One potential technique for increasing the range of (hybrid) electric vehicles is to predict the future speed-profile of a driver and set vehicle control parameters accordingly.The thesis explores to what extent recurrent neural networks can be used for performing speed-profile predictions.The cell units: RNN, LSTM, and GRU are evaluated as part of two different models (one to many and many to many).The thesis aims to explore recurrent neural networks' ability to predict speedprofiles.The two research questions of the thesis entails: (RQ 1) evaluating multiple types of neurons and model structures and (RQ 2) evaluating the effect of preprocessing of data and tuning of the neural network configurations.The thesis follows the Design Science Research methodology.Proof of concept configurations were implemented, trained, and evaluated on their ability to correctly predict the next 60 samples (seconds) of a vehicle's speed (speed-profile).RQ 1 was trained using only vehicle speed as data feature, and evaluated using ≈ 100 speedprofiles.RQ 2 was trained using both vehicle speed and vehicle battery charge as data features.It was evaluated using ≈ 100 and ≈ 800 speed-profiles.The thesis shows that several neural network configurations are capable of achieving predictions which are better than that of a trivial predictor.The trivial predictor "predicts" that the driver will keep its speed constant for the next 60 seconds.The best configuration developed to answer RQ 1 is a many to many-model with RNN cell units.Its RMSE is 0.77 times that of the trivial predictor.The evaluation of RQ 2 shows that configurations which make use of fewer preprocessing/regularisation parameters achieve lower RMSE.For ≈ 100 speed-profiles, the best configuration used no additional preprocessing or regularisation and achieved an RMSE 0.80 times that of the best configuration from RQ 1.With ≈ 800 speed-profiles, the best configuration made use of smoothening and aligning of the speed-profiles.Its RMSE was 0.62 times that of the best configuration from RQ 1.
Author: Erik Thorsell
Institution: Chalmers University of Technology
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2885728959
