MLP networks applied to State of Charge estimation in electric vehicle simulation with normed driving cycle

Kawe Monteiro de Souza; José Rodolfo Galvão; Fernanda Cristina Corrêa; Jorge Augusto Pessatto Mondadori; Paulo Broniera Junior; Maria Bernadete de Morais França

doi:10.20906/SBAI-SBSE-2023/3881

Kawe Monteiro de Souza Post-Graduation in Electrical Engineering, Universidade Estadual de Londrina, PR
José Rodolfo Galvão Graduate Program in Electrical Engineering, Universidade Tecnológica Federal do Paraná, Campus Ponta Grossa, PR
Fernanda Cristina Corrêa Graduate Program in Electrical Engineering, Universidade Tecnológica Federal do Paraná, Campus Ponta Grossa, PR
Jorge Augusto Pessatto Mondadori Senai Institute of Information Technology and Communication (ISTIC), Laboratory LAPSEE-PIM, Londrina, PR
Paulo Broniera Junior Senai Institute of Information Technology and Communication (ISTIC), Laboratory LAPSEE-PIM, Londrina, PR
Maria Bernadete de Morais França Post-Graduation in Electrical Engineering, Universidade Estadual de Londrina, PR

DOI: https://doi.org/10.20906/SBAI-SBSE-2023/3881

Keywords: State of Charge, Battery Pack Simulation, Neural Networks, Electric Vehicle SOC, Multi Layer Perceptron

Abstract

Monitoring the battery pack is key to ensuring the safe operation of electric vehicles (EVs). This process is done by the Battery Management System (BMS), which uses, for example, variables such as State of Charge (SOC) to determine the current charge and health of the battery pack. Thus, this paper presents the application of a multilayer perceptron network (MLP) for estimating the state of charge of a battery pack using a simulation of an electric vehicle in the UDDS normative cycle and compares the result with two traditional methods: Kalman Unscented Filter (UKF) and Coulomb Counting (CC) also implemented in this work. The performance of the estimators were evaluated using the statistical metrics of Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). Although CC showed superior performance, with an RMSE of 0.1086, MLP was considered the most appropriate solution due to factors intrinsic to the methodology, ranking second with an RMSE of 0.36517.