Data-driven full life-cycle state parameter assessment of Li-ion batteries

Abstract

<p indent="0mm">The battery management system (BMS) is a vital link between the power battery, the onboard system, and the engine. Adjusting the output power depending on the current driving situation of the vehicle and the surrounding environment ensures the lithium-ion battery’s safe, stable, and effective operation. It is essential to optimize the charge and discharge rate of the cell and prolong its lifetime. The battery management system has three key estimation parameters—State of charge (SOC), state of health (SOH), and remaining useful life (RUL)—That have key reference values for the battery’s current use. As the above state characteristics cannot be computed directly from the battery, a broad and highly predictive approach must be developed to estimate the condition parameters. This review analyzes the development of the application of battery state parameter estimation, explains the model concepts, compares the differences between models, and explores illustrative typical applications. Patterns are categorized into traditional models and data-driven machine learning models in this review. The traditional model’s electrochemical model provides a highly accurate description of the internal behavior of the battery through the battery reaction mechanism and explains the charge transfer process between electrodes from a chemical theory perspective. However, the electrochemical model is not suited for online state parameter estimation due to its maximum complexity and high computational stress among the standard models. The equivalent circuit model, which employs circuit components like resistors and capacitors to describe the dynamic features of the battery, is based on a simpler idea than the complicated partial differential equations that are used to describe the battery’s fundamental electrochemical operations. However, because of the low state parameter estimation precision, it is not compatible with emerging battery management systems with high state parameter estimation accuracy. To enhance the performance of the model, the fusion model integrates the features of different models. Though, the application of fusion models for battery management systems needs more research due to the inclusion of models with additional sources of error and the challenge of creating failure criteria for the fused models. The commonly used machine learning algorithms, such as a k-nearest neighbor, support vector machine, random forest, and artificial neural network models, are briefly introduced in this review. It also provides a systematic overview of the model training process from data pre-processing to model evaluation. This review also offers a comprehensive overview of the whole model training process, from data pre-processing through algorithm selection to model performance assessment, and it lists the variables to be considered when choosing an algorithm. Finally, this review discusses the development of data-driven machine learning models for state parameter estimation and examines the estimated performance of models created by the techniques employed in representative papers. The research suggests that long short-term memory network-based models have better estimation performance for estimating battery condition parameters. The data-driven machine learning model is not ideal and still requires improvement since the model that the machine learning algorithm learned is a black box model lacks interpretation of the results and places strict constraints on the source and quality of data. On this basis, it is proposed that existing or more sophisticated machine learning models are enhanced and their estimation accuracy and stability are evaluated. In order to achieve joint estimation models with high accuracy and scalability, machine learning models for multi-tasking also need to be developed. Furthermore, to overcome the unpredictability of machine learning for the internal aging of batteries, it is also advised to create fusion models using a combination of electrochemical models and computer algorithms. To increase the amount of data available to researchers and aid in the advancement of artificial intelligence in the field of state parameter estimation, it is also advised to integrate experimental data, standardize data storage standards, and create databases.