ºìÐÓÊÓƵ

Topic

Lightweight and Explainable Deep Learning Model for EV Battery Voltage Prediction

Department of Electrical and Computer Engineering

Date & location

• Thursday, December 12, 2024

• 10:30 A.M.

• Virtual Defence

Reviewers

Supervisory Committee

• Dr. Daler N. Rakhmatov, Department of Electrical and Computer Engineering, University of Victoria (Supervisor)

• Dr. Thirumarai Ilamparithi Chelvan, Department of Electrical and Computer Engineering, UVic (Member) 

External Examiner

• Dr. Jason Keonhag Lee, Department of Mechanical Engineering, University of Washington 

Chair of Oral Examination

• Dr. Andrew Rowe, Department of Mechanical Engineering, UVic

   

Abstract

Electric vehicles (EVs) play an important role in reducing the greenhouse gas emissions by providing an environment-friendly alternative to the fossil-fuel-based means of transportation. EVs are typically powered by Li-ion battery packs supported by a Battery Management System (BMS). The latter is tasked with monitoring and keeping the battery voltage, current, and temperature within safe operating limits, as well as estimating and improving the battery performance-related parameters, such as the battery state-of-charge and lifespan. In this thesis, we aim to extend the BMS capabilities by enabling battery voltage predictions under a given load profile (i.e., discharge/charge current varying over time). Such predictions are useful for proactive (as opposed to reactive) load management, as they allow a BMS to forecast the battery voltage behaviour under various anticipated load conditions. 

Using a data-driven deep learning (DL) approach, we propose a novel model that generates battery voltage estimates given the battery current, temperature, and consumed charge over time. It has a V-shaped architecture that features two wings to enhance the model explainability. The first wing predicts the steady-state open circuit voltage (OCV) component, based on the consumed battery charge information, while the second wing predicts the transient voltage component, based on the battery current and temperature information. The total number of the model parameters is under 2.6K. 

A well-known experimental dataset was used in this study for training, validation, and testing purposes. This dataset contains measurements taken on a Li-ion battery subjected to various EV driving cycles interleaved with charging cycles. The mean absolute percentage error (between predicted and measured battery voltage values) was under 1%, demonstrating the accuracy of the proposed model.