Yongzhi Zhang

Lithium-Ion Battery Pack State of Charge and State of Energy Estimation Algorithms Using a Hardware-in-the-Loop Validation

An adaptive H infinity filter approach is proposed to estimate the multistates including state of charge (SOC) and state of energy (SOE) for a lithium-ion battery pack. In the proposed approach, the covariance matching technique is used to adaptively update the covariance of system and observation noises and the recursive least square method is used to identify the battery model parameters in real time. The hardware-in-the-loop (HIL) platform for battery charge/discharge is set up to evaluate the accuracy and robustness of the SOC and the SOE estimation and compare the proposed approach with the multistate estimators using an extended Kalman filter and an H infinity filter. The experimental results indicate that the adaptive H infinity filter-based estimator is able to estimate the battery states in real time with the highest accuracy among the three filters.

A Double-Scale, Particle-Filtering, Energy State Prediction Algorithm for Lithium-Ion Batteries

In order for the battery management system (BMS) in an electric vehicle to function properly, accurate and robust indication of the energy state of the lithium-ion batteries is necessary. This robustness requires that the energy state can be estimated accurately even when the working conditions of batteries change dramatically. This paper implements battery remaining available energy prediction and state-of-charge (SOC) estimation against testing temperature uncertainties, as well as inaccurate initial SOC values. A double-scale particle filtering method has been developed to estimate or predict the system state and parameters on two different time scales. The developed method considers the slow time-varying characteristics of the battery parameter set and the quick time-varying characteristics of the battery state set. In order to select the preferred battery model, the Akaike information criterion (AIC) is used to make a tradeoff between the model prediction accuracy and complexity. To validate the developed double-scale particle filtering method, two different kinds of lithium-ion batteries were tested at three temperatures. The experimental results show that, with 20% initial SOC deviation, the maximum remaining available energy prediction and SOC estimation errors are both within 2%, even when the wrong temperature is indicated. In this case, the developed double-scale particle filtering method is expected to be robust in practice.

Evaluation of the model-based state-of-charge estimation methods for lithium-ion batteries

To achieve accurate battery SoC, the Gaussian is applied to construct battery model. It is able to simulate the time-variable, nonlinear characteristics of battery. To adaptively adjust the Gaussian battery model parameter set and order, a novel online four-step model parameter identification and order selection method is proposed. To further evaluate the Gaussian battery model estimation accuracy, another two kinds of representative battery models including the combined model and Thevenin model are built as comparisons. Results based on three kinds of Kalman filters show that the maximum SoC estimation error of each case is within 2% and the Gaussian model has the best accuracy for voltage prediction as well as SoC estimation.

A LSTM-RNN method for the lithuim-ion battery remaining useful life prediction

Prognostics and health management (PHM) can ensure that a battery system is working safely and reliably. Remaining useful life (RUL) prediction, as one main approach of PHM, provides early warning of failures that can be used to determine the necessary maintenance and replacement of batteries in advance. The existing RUL prediction techniques for lithium-ion batteries are inefficient to learn the long-term dependencies of aging characteristics with the degradation evolution. This paper investigates deep-learning-enabled battery RUL prediction. The long short-term memory (LSTM) recurrent neural network (RNN) is employed to learn the capacity degradation trajectories of lithium-ion batteries. The LSTM RNN is adaptively optimized using the resilient mean square back-propagation method. The developed LSTM RNN is able to capture the underlying long-term dependencies among the degraded capacities such that an explicitly capacity-oriented RUL predictor is constructed. Experimental data from one lithium-ion battery cell is deployed for model construction and verification. This is the first known application of deep learning theory to battery RUL predictions.