Multi-time Scale State Estimation Of Lithium-ion Batteries Using Data Driven Method

Lithium-ion batteries have become the main power source of electric vehicles due to their unique features,including their high energy density,long cycle life,high efficiency,and environmental-friendly performance.The effective management of lithium-ion batteries has become a key technology in the development of electric vehicles.It is significant to estimate the state and to evaluate the performance of lithium-ion batteries via multi-time scale at different usage stages in order to ensure the safety and efficiency of lithium-ion battery applications.Aiming at addressing the issues in battery application,detailed studies have been conducted from the following aspects:To estimate battery state of charge(SOC)over micro time scale,efforts have been made as follows.Three steps in the processmg of SOC online estimation have been analyzed:open circuit voltage(OCV)-SOC mapping,battery modeling and algorithm implementation.Two common tests,low-current OCV test and incremental OCV test,for observing battery open circuit voltage performance are compared.The temperature dependency of battery OCV-SOC relationship is investigated and off-line data-driven method is used to identify the parameters of the first-order RC model.The unscented Kalman filter is implemented to enhance model-based SOC online estimation.Two estimators are evaluated in terms of accuracy and robustness via statistical measures based on which a conclusion can be drawn that the incremental OCV test is more suitable for an approximation of battery OCV-SOC and can provide a more accurate estimated SOC.To estimate battery state of power(SOP)over micro time scale,efforts have been made as follows.The experimental data with regard to complex conditions are analyzed statistically.Both parametric and nonparametric models are established in conditions of temperature,SOC,and cell resistance to estimate battery SOP using data-driven approaches.Two proposed models are compared via a list of statistics and the corresponding scopes of application are discussed.Two models can be utilized together to achieve a higher accuracy of SOP estimation and provide the basis for the power allocation strategy of battery management system.To predict battery remaining useful life(RUL)over macro time scale,efforts have been made as follows.The degradation trends of battery resistance and capacity have been studied.Battery capacity is modeled in terms of cycle number via data fitting.A Bayesian approach is used to update the model parameters periodically and the RUL with corresponding probability density function is thus predicted.The prediction results are updated over time as more data become available,which leads to an increase in prognostic accuracy.This approach reduces the dependency on battery's life modeling result in laboratory environments and thus reduces pre-test costs.To identify failed batteries in the battery pack at one time over macro time scale,efforts have been made as follows.The incremental capacity analysis(ICA)technique is utilized to analyze the charging process of battery pack.Features are extracted from the partial IC curve and a shrinkage method called elastic net is used for variable selection.Based on the inconsistency of selected variables among batteries in one pack,a classification model using linear discriminant analysis(LDA)is proposed to identify failed batteries.It has achieved high classification accuracy and its input variables can be obtained online which is definitely attractive for practical applications.To improve the battery echelon use technology,efforts have been made as follows.Firstly,capacity of retired batteries should be estimated before being secondary used.A batch of lithium-ion batteries retired from 2008 Olympic passenger buses is tested.Based on the experimental data,the distributions of resistance and capacity are analyzed and the correlation between capacity and resistance is studied.A fast estimation model of battery capacity is established based on the correlation using support vector machine.Genetic algorithm is used to optimize the model parameters and help to achieve accurate estimation.The proposed approach has addressed the issues of time-consuming tests and abundance facilities,and thus has improved the economic benefits of battery secondary use.Secondly,the performance of retired batteries should be evaluated on the target of matching battery packs.Parameters,such as capacity,resistance,and the rate of self-discharge,have been analyzed.An evaluation method for battery performance based on Delphi theory and grey relational grade analysis is proposed for the purpose of matching battery packs.This method overcomes the one-sidedness of using single index to evaluate battery performance,and thus provides a reliable theoretical basis for battery secondary use.The data-driven SOC and SOP estimation approaches over micro time scale,the data-driven remaining useful life prediction method,the data-driven failed battery identification model,and the data-driven battery echelon use techniques over macro time scale are proposed in this thesis which solves the practical problems in different stages of battery application.These data-driven estimation techniques for batteries are helpful and useful to improve the safety and economy of the battery management,which possesses important applicable value in practical engineering.