State of charge estimation for Li-ion batteries using neural network modeling and unscented Kalman filter-based error cancellation

Abstract

Lithium-ion batteries have been widely used as the energy storage systems in personal portable electronics (e.g. cell phones, laptop computers), telecommunication systems, electric vehicles and in various aerospace applications. To prevent the sudden loss of power of battery-powered systems, there are various approaches to estimate and manage the battery's state of charge (SOC). In this paper, an artificial neural network–based battery model is developed to estimate the SOC, based on the measured current and voltage. An unscented Kalman filter is used to reduce the errors in the neural network-based SOC estimation. The method is validated using LiFePO4 battery data collected from the Federal Driving Schedule and dynamical stress testing.

Introduction

Lithium-ion (Li-ion) batteries have attracted attention due to their high energy density and long cycle life compared with other battery chemistries. State of charge (SOC) estimation provides information about when a battery needs to be recharged and allows battery management systems (BMSs) to prolong the battery life by preventing batteries from over-charging or over-discharging.

SOC is defined as the percentage of remaining capacity relative to the maximum capacity of the battery. Many SOC estimation approaches have been developed, among which Coulomb counting [1], [2] is the most popular one. In Coulomb counting, the current is integrated over time to estimate SOC. Although Coulomb counting is easy to implement, the measurement and calculation errors can be accumulated by the integration function, and thus the estimation of SOC tends to drift from the actual values. The voltage-based method [3] estimates the SOC based on a voltage–SOC lookup table. However, voltage-based methods do not work well for Li-ion batteries because of their flat plateau of discharge characteristics. To provide more robust estimates, equivalent circuit models (ECMs) have been proposed for SOC estimation using extended Kalman filters (EKFs) [4], [5], [6], [7], [8], [9] and unscented Kalman filters (UKFs) [10], [11]. Plett [4], [8], [9] developed an enhanced self-correcting model that takes hysteresis effects into consideration to estimate SOC using EKFs. He et al. [7] proposed an improved Thevenin model wherein SOC is estimated using EKFs.

Studies have been conducted to establish data-driven models for battery modeling and SOC estimation that do not require detailed physical knowledge of Li-ion batteries. The commonly used data-driven models include support vector machines [12], [13], [14] and neural networks [15], [16], [17], [18]. For example, Hansen and Wang [12] developed a support vector machine (SVM) based method for SOC estimation. The estimator was validated using US06 dynamical operation data with a root mean square (RMS) error of less than 6%. Anton et al. [13], proposed a state-of-charge estimator using support vector machine. The inputs to SVM were voltage, current and temperature, and output was SOC, with training data and testing data under the same loading condition. Therefore, the generalization ability of the proposed SVM was unknown. Lee et al. [15] developed an SOC estimation approach based on a fuzzy neural network with B-spline membership functions. But this method was only tested using data obtained by constant current discharge, which is different from the real-life loading condition of many battery powered systems like EVs and UAVs. Cheng et al. [16] proposed a SOC estimation method for Ni–MH batteries using a three-layer feed-forward neural network; where the inputs to the neural network (NN) were battery current, temperature, voltage and its first and second derivatives. Though Cheng’s method is promising, the estimation results show high estimation variance, probably due to over-fitting or under-fitting, which is a common problem for data-driven methods. Therefore, researchers have been developing hybrid methods to improve the accuracy of data-driven SOC models. Charkhgard and Farrokhi [19] used an extended Kalman filter (EKF) to infer the SOC based on a radial basis function neural network battery model. However, EKF is only accurate to first order or second order of a nonlinear system in the sense of Taylor expansion. Furthermore, only three inputs were used in Charkhgard and Farrokhi’s model to train the NN, namely, the voltage from the previous sample and the current and SOC of the present sample, which is not sufficient to capture the capacitive effects in the Li-ion battery system dynamics.

LINK 1 - TÌM KIẾM SÁCH/TÀI LIỆU ONLINE (GIÁ ƯU ĐÃI NHẤT)

LINK 2 - TÌM KIẾM SÁCH/TÀI LIỆU ONLINE (GIÁ ƯU ĐÃI NHẤT)

LINK 3 - TÌM KIẾM SÁCH/TÀI LIỆU ONLINE (GIÁ ƯU ĐÃI NHẤT)

LINK 4 - TÌM KIẾM SÁCH/TÀI LIỆU ONLINE (GIÁ ƯU ĐÃI NHẤT)