Lithium-ion Battery High-fidelity Modeling,parameter Identification,and Lifetime Prognostics For Electric Vehicles

Developing electric vehicles is an important approach to ensuring energy security and transitioning to a low-carbon economy.It is also a national strategy to realize independent innovation and promote technological breakthroughs.Lithium-ion batteries,with high energy density,low self-discharge rate,and long cycle life,are the main energy storage devices in electric vehicles.The performance of lithium-ion batteries has direct effects on the power and security of electric vehicles.To guarantee the safe and efficient operation of batteries in demanding driving conditions,effective battery management is required.However,lithium-ion battery is a complex electrochemical system,and its internal states are time-varying and unmeasurable.Furthermore,the performance of lithium-ion batteries will deteriorate with usage,which brings potential safety hazards.Aiming at real-world applications,this thesis investigates high-fidelity electrochemicalthermal coupled modeling,non-destructive parameter identification,and battery lifetime prognostics.The main work of this thesis can be summarized as follows:Firstly,the setup of battery testing platform and simulation systems are introduced.Different tests can be conducted based on the battery testing platform,which provides data for model and algorithm validation.The simulation system provides software support for battery modeling and codes debugging,and it also provides hardware support to run models and algorithms.Secondly,high-fidelity electrochemical-thermal coupled modeling is studied.Based on the pseudo-2-dimensional(P2D)battery model,a reformulated P2 D model with high accuracy and low complexity is to obtained through mathematical reformulation.The whole reformulation process avoids introducing physical simplification,and therefore the reformulated P2 D model is more accurate than the physically simplified P2 D model.Moreover,to increase the applicability of the model under different temperatures,an electrochemical-thermal coupled model which has high accuracy under different working conditions is proposed.Thirdly,since the accuracy of the electrochemical-thermal coupled model is largely dependent on the accuracy of the parameters,a non-destructive parameter identification method is proposed.For a highly non-linear system with a number of parameters such as the P2D-based electrochemical-thermal coupled model,sensitivity analysis is carried out to determine parameters that need to be identified.Then,by finding the suitable identification conditions for these parameters,the identification efficiency can be improved.To avoid divergence problems,feasible initial guess values are generated by deep learning algorithms.Genetic algorithm and the filtering method are combined to identify the sensitive parameters,which improves the identification accuracy.Additionally,the electrochemical-thermal coupled models with different complexity is compared systematically.Under different working conditions and hardware conditions,simplified and non-simplified electrochemical-thermal coupled models have their advantages and disadvantages.Through systematic comparison,we propose the concept of model generalization.That is,the ability of the model to maintain accuracy under different working conditions using the same set of parameters.Taking the state-of-charge(SOC)estimation as an example,the feasibility of using electrochemical-thermal coupled models in real-world applications is investigated.Both the simplified and non-simplified electrochemical-thermal coupled models are used for SOC estimation on a desktop computer and an embedded system.Finally,a novel hybrid physics-based and data-driven approach is proposed for battery lifetime prediction.Based on the electrochemical model and measured voltage data,a hybrid feature that incorporates both aging mechanisms and data information of a battery is extracted.Using data augmentation techniques,the training accuracy of the deep learning algorithm is increased.Then,the historical battery degradation data are classified using health-related features and the clustering method.A non-iterative prediction method is built based on sequence-to-sequence deep neural networks.The proposed battery lifetime prediction method provides accurate prediction results under various battery aging conditions and can be used in early cycles.