Physics-constrained transfer learning: Open-circuit voltage curve reconstruction and degradation mode estimation of lithium-ion batteries

Abstract

Open-circuit voltage (OCV) updates are the key to accurate state of charge (SOC) estimates over lifetime. Degradation modes (DM) are directly coupled to OCV estimation. They offer a more detailed analysis of the battery’s state of health (SOH) and yield optimized usage strategy, and with that, a prolonged lifetime. In this study two data-driven models are coupled with physics-based models and compared in regards of their OCV and DM estimation accuracy: Two temporal convolutional — long short-term memory neural networks (TCN-LSTM) are trained from synthetic NCA-graphite battery data for OCV curve estimation (model 1) and alignment parameter estimation (model 2). Both models are fine-tuned with varying amounts of experimental NMC-graphite battery data during the transfer learning (TL) step. In the subsequent physics-constraining part the DMs are derived via optimization (model 1), i.e., fitting the OCV with half cell open-circuit potentials, or directly via mathematical equations (model 2). Both models prove that fine-tuning data from one aging path suffices, if it includes the maximum appearing DMs of the target domain. For these use cases both models maintain OCV mean absolute errors (MAEs), DM MAEs and SOH mean absolute percentage errors (MAPEs) under 10 mV, 3.10 % and 1.98 %, respectively. The model 2 has less computational complexity and reaches slightly better results but requires labeled target data including alignment parameters for its application. This study shows that synthetic data is eligible for TL, even for varying cell chemistries, and that the mechanistic model helps to physically constrain the output.