Open Collections

Data-driven power estimation of electric buses using deep LSTM neural networks and symbolic aggregate approximation filtering

University of British Columbia

This thesis proposes a deep learning framework for estimating the power use of electric vehicle drivetrains, using an artificial recurrent neural network architecture (long short-term memory), and techniques for data curation and analysis. Power samples of battery-electric buses operating in Victoria, B.C. were deployed to train, validate, and test the model. Microtrips were extracted from the dataset by taking out proportions of time spent at standstill. Anomalous trends of the dataset components were detected, marked as outliers, and excluded from the model input using Symbolic Aggregate approXimation (SAX) transformation filtering algorithm. A systematic optimization process is adopted to achieve the best hyperparameter configuration for this estimation task. The resulting fine-tuned deep LSTM model is then created, evaluated, and tested. More effort has been devoted to enhancing the efficacy of the model through investigating alternative model configurations and formulations. Among the several configurations tested, the best power consumption modeling approach yielded an error of about 5% (the RMSE of 0.05747) and a squared correlation coefficient of beyond 90% (R² of 0.92502). [This thesis was revised on October 25, 2021.]

Text

2022-01-31

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/76914

Mechanical Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/