Distributed Thermal Process Modeling of Lithium-Ion Power Battery Based on Limited Knowledge


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
  • Hanxiong LI (Supervisor)
  • Hua DENG (External person) (External Supervisor)
Award date15 Jun 2023


Lithium-ion batteries are the core energy storage components for in key areas such as modern electric vehicles. However, large-scale power batteries' safety performance and cycle life are highly related to their operating temperature. Overheating or severely uneven temperature distribution can cause battery performance degradation or safety incidents. Thus, designing advanced battery thermal management systems with temperature field regulation and local thermal runaway detection is crucial for developing the electric vehicle industry. The effective thermal management strategies require access to accurate thermal models of the battery system.

The battery systems of electric vehicles often operate in variable environments. Due to the lack of real-time update mechanism, the temperature prediction performance of traditional physics-based models may be severely limited. As a result, advanced in-vehicle thermal management systems place higher demands on the reliability of battery thermal models. However, due to unknown internal states and random external disturbances, often only partially known system information can be used for modeling. Especially in some electric vehicle energy storage systems with dense battery elements, the unknown heat flow leads to an unusually complex heat transfer mechanism, making it more difficult to capture the heat transfer dynamics of the battery. Thus, fusing limited mechanistic model knowledge and sampled data to construct a battery temperature field prediction model with both efficiency and reasonable accuracy is not only a engineering challenge with practical value, but also a scientific one.

The key problem addressed in this dissertation is how to develop effective modeling techniques for the thermal processes of lithium-ion power batteries that exhibit spatiotemporal dynamic characteristics, using limited knowledge in the absence of accurate mathematical models. The goal is to achieve precise temperature field predictions. The main research efforts include the following:

First, this dissertation proposes an efficient optimization framework that combines evolutionary algorithms and data-driven surrogate models for multi-scale parameter identification of the battery electrochemical model. The framework pre-screens candidate parameters to significantly reduce the number of times the full-order physics-based model needs to be solved, addressing the computational expense of evaluating a massive number of candidate parameters. Results demonstrate that the proposed method is much more computationally efficient than traditional evolutionary algorithms. Through charge-discharge experiments and idealized physics-based simulations, this study provides a data foundation that reflects the dynamic characteristics of the battery system and establishes the necessary theoretical foundation for subsequent research on thermal process modeling methods.

Given the complex electrochemical reactions inside the battery system, it is difficult to accurately calculate the battery heat generation rate in practice. Thus, this dissertation proposes a novel adaptive reduced-order modeling approach for the case of unknown heat source to solve the battery thermal process modeling problem. First, the spatiotemporal variable of the battery heat equation is decomposed using the spectral method to extract a low-order physics-based model of the dominant system dynamics. Then, a neural network is embedded in the low-order model to adaptively estimate the unknown heat source. Results show that the proposed method has good modeling performance. In particular, the developed adaptive model can predict the temperature field in the continuous spatial domain in real time with limited sensing since the proposed method uses analytic orthogonal eigenfunctions to expand the spatiotemporal variable.

Due to the influence of practical working environments, the uncertainty of battery thermal processes may result in parameter drift and unknown dynamics. Thus, this dissertation proposes a fusion-driven modeling method based on dynamic error compensation. This method first adopts a spatial decomposition strategy to decouple the heat transfer effects between tabs and the battery core, and derives a low-order thermal model for the battery core region using the spectral method. Then, a dynamic error compensation model is constructed to reduce the nonlinear prediction error caused by process/model mismatch. The validation results show that the proposed fusion model can accurately predict the battery temperature distribution and has strong generalization performance, providing new ideas and methods for optimizing the design and real-time control of battery systems.

Given the complex heat transfer mechanisms that are difficult to establish clear dynamic equations, this dissertation proposes two novel purely data-driven methods for modeling battery thermal processes. To meet the rapid response requirements of industrial applications, a spatiotemporal feature adaptation-based incremental learning modeling approach is proposed. This method only requires the latest measurement data for model updates and has outstanding computational efficiency. In addition, to reduce the number of online sensors, a sparse sensing modeling approach based on the reconstruction of temporal dynamics is proposed. The method transfers the spatial distribution features extracted from the offline data to the online stage, and only partial-node information is required for prediction of the full-node temperature field. These two data-driven methods provide novel solutions for modeling complex battery thermal processes, with significant theoretical and practical value.

    Research areas

  • Lithium-ion batteries, Thermal process modeling, Spatiotemporal dynamic system, Temperature field prediction