Battery electric modeling is a central aspect to improve the battery development process as well as to monitor battery system behavior. Besides conventional physical models, machine learning methods show great potential to learn this task using in-vehicle data. However, the performance of data-driven approaches differs significantly depending on their application and utilized data set. Hence, a comparison among these methods is required beforehand to select the optimal candidate for a given task.In this work, we address this problem and evaluate the strengths and weaknesses of a wide range of possible machine learning approaches for battery electric modeling. In a comprehensive study, various conventional regression methods and neural networks are analyzed. Each method is trained and optimized based on a large and qualitative data set of automotive driving profiles. In order to account for the influence of time-dependent battery processes, both low pass filters and sliding window approaches are investigated.As a result, neural networks are found to be superior compared to conventional regression methods in terms of accuracy and model complexity. In particular, Feedforward and Convolutional Neural Networks provide the smallest average error deviations of around 0.16%, which corresponds to an RMSE of 5.57mV on battery cell level. With automotive time series data as focus, neural networks additionally benefit from their ability to learn continuously. This key capability keeps the battery models updated at low computational costs and accounts for changing electrical behavior as the battery ages during operation.

中文翻译：

---

基于机器学习的电池电建模方法综合研究

电池电气建模是改进电池开发过程以及监控电池系统行为的核心方面。除了传统的物理模型，机器学习方法显示出使用车载数据学习这项任务的巨大潜力。然而，数据驱动方法的性能因它们的应用和使用的数据集而异。因此，需要事先比较这些方法以选择给定任务的最佳候选者。在这项工作中，我们解决了这个问题，并评估了各种可能的电池电动建模机器学习方法的优缺点。在一项综合研究中，分析了各种传统的回归方法和神经网络。每种方法都基于汽车驾驶档案的大量定性数据集进行训练和优化。为了考虑时间相关电池过程的影响，研究了低通滤波器和滑动窗口方法。结果，发现神经网络在准确性和模型复杂性方面优于传统的回归方法。特别是，前馈和卷积神经网络提供了大约 0.16% 的最小平均误差偏差，这对应于电池单元级别的 5.57mV 的 RMSE。以汽车时间序列数据为重点，神经网络还受益于其持续学习的能力。