Deep Learning-Based Ensemble Method for Li-ion Battery State-of-Charge Estimation

Abstract

Lithium-ion batteries (LIB) are vital components of modern electric vehicles and load management in smart grids due to their relatively high energy density, power density, and efficiency. It is important that onboard battery management systems in battery electric vehicles have accurate battery state of charge (SOC) information to gauge the remaining vehicle range and minimize battery degradation through power management. However, owing to highly dynamic vehicle driving habits and the nonlinear nature of SOC relative to other battery parameters such as current, voltage, and temperature, SOC is unable to be measured directly and is difficult to be accurately estimated in real-time. This article proposes a novel Li-ion battery SOC estimation method through a Deep Feedforward Neural Network Multimode Ensemble (DNN-ME). K-means clustering is used to separate the training data into differentiable data subsets, which are then each fed into N deep feedforward neural network (DNN) base learners. The final ensemble output through weighted averaging is less susceptible to error from weight initialization variation than single models, ensuring greater prediction accuracy.