The current surge in the need for Li-ion batteries to power electric vehicles has also translated in a need for more advanced models that can predict their behavior, but also quantify the uncertainty in their predictions, given the amount of variables involved and the varying operating conditions. This manuscript proposes a new Bayesian physics-informed recurrent neural network, where the battery discharge curve is described using the Nernst and Butler–Volmer equations, the activity correction term within such equations is modeled with two multilayer perceptrons, and approximate Bayesian computation by subset-simulation is used to train the weights, bias and the physical parameters representing the maximum charge available and the internal resistance. The challenges found during the adaptation and implementation of the Bayesian training algorithm to the recurrent physics-informed cell are described, along with the approaches proposed to overcome them. The performance of the Bayesian hybrid model presented in this paper has also been evaluated using data from NASA Ames Prognostics Data Repository, and the results show comparable accuracy to the standard approach with backpropagation, and a flexible and realistic quantification of the uncertainty. Furthermore, the uncertainty related to the physical parameters of the hybrid model can be evaluated in semi-isolation of the weights and bias of the MLPs, providing a sensitivity tool to assess the relative importance between different parameters.

当前对锂离子电池为电动汽车提供动力的需求激增，也意味着需要更先进的模型来预测其行为，同时考虑到所涉及的变量数量和不同的运行情况，还可以量化其预测的不确定性。状况。本手稿提出了一种新的基于贝叶斯物理学的循环神经网络，其中使用 Nernst 和 Butler-Volmer 方程描述电池放电曲线，此类方程中的活动校正项用两个多层感知器建模，并通过子集近似贝叶斯计算模拟用于训练权重、偏置以及表示最大可用电荷和内阻的物理参数。描述了贝叶斯训练算法对循环物理信息单元的适应和实施过程中发现的挑战，以及提出的克服这些挑战的方法。本文提出的贝叶斯混合模型的性能还使用 NASA 艾姆斯预测数据存储库的数据进行了评估，结果显示与反向传播的标准方法相当的准确性，以及对不确定性的灵活而现实的量化。此外，与混合模型的物理参数相关的不确定性可以在 MLP 的权重和偏差的半隔离中进行评估，从而提供了评估不同参数之间的相对重要性的灵敏度工具。