Home
Scholarly Works
Feedforward and NARX Neural Network Battery State...
Conference

Feedforward and NARX Neural Network Battery State of Charge Estimation with Robustness to Current Sensor Error

Abstract

State of charge (SOC), the remaining usable charge of the battery divided by its nominal capacity, is one of the most important parameters for managing Li-ion battery packs. This work investigates two types of artificial neural network-based SOC estimators: a feedforward neural network (FNN) and a nonlinear autoregressive exogenous model (NARX) network. These networks are trained and tested with battery drive cycle and charging data for a Tesla Model 3 electric vehicle. Measured temperature, along with different combinations of filtered and unfiltered voltage and current, are used as model inputs. The NARX, which benefits from having SOC from the prior time step as an input, is shown to have substantially less error than the FNN, even when there is a significant current sensor offset error which prevents the NARX from simply functioning as a coulomb counter. Overall, the NARX is shown to be accurate for the most difficult highway drive cycles with steep grades and to be robust against large current sensor errors.

Authors

Vieira RN; Kollmeyer P; Naguib M; Emadi A

Volume

00

Pagination

pp. 1-6

Publisher

Institute of Electrical and Electronics Engineers (IEEE)

Publication Date

June 23, 2023

DOI

10.1109/itec55900.2023.10187084

Name of conference

2023 IEEE Transportation Electrification Conference & Expo (ITEC)

Labels

View published work (Non-McMaster Users)

Contact the Experts team