Data-driven state-of-charge estimation of the Panasonic 18650PF Li-ion cell using deep forward neural networks

Abstract

The State-of-Charge (SOC) is a key parameter for the proper functioning of the Battery Management System (BMS) of lithium-ion (Li-ion) batteries, and indicates the amount of charge remaining in the battery. In this work, we present a novel empirical study for the data-driven estimation of the SOC of the Panasonic 18650PF Li-ion cell using Deep Forward Neural Networks (DFNN) and optimization algorithms with adaptive learning rates. Specifically, we model the Urban Dynamometer Driving Schedule (UDDS) drive cycle. Our results suggest that the choice of the optimization algorithm affects the performance of the model and that a DFNN with five hidden layers is the model of optimal capacity when considering 256 units per layer. This optimal DFNN is able to estimate the SOC of the 18650PF Li-ion cell with an error smaller than 0.12% over a 25o C dataset using the Adamax optimization algorithm.