Forecasting battery capacity for second-life applications using physics-informed recurrent neural networks

Accurately forecasting lithium-ion battery capacity degradation is crucial for optimizing the second-life utilization of these batteries, enabling reliable operation, reduced maintenance costs, and extended life cycle performance. However, achieving consistent forecasting accuracy across cells and over time remains challenging due to significant cell-to-cell variability and substantial changes in real-world usage conditions during the transition from first to second life. In this study, we propose a new physics-informed machine learning method that integrates an aging-aware electrochemical model with a recurrent neural network, creating a physics-informed recurrent neural network (PI-RNN). This hybrid model leverages both physics-based insights and data-driven learning to predict capacity fade under diverse usage conditions, including transitions from first- to second-life applications. We evaluate PI-RNN using two datasets: an open-source NASA dataset comprising 28 lithium cobalt oxide/graphite cells, and a newly collected dataset of 39 commercial lithium iron phosphate/graphite cells, where cells were initially cycled to 80% capacity in their first life before undergoing milder cycling in their second life. While PI-RNN performs comparably to data-driven models in the first-life phase, it demonstrates a clear advantage in second-life forecasting, reducing root mean squared error by approximately 40%–70% compared to baseline models when forecasting periods span the transition from first to second life, even when trained on as few as two cells. Parametric studies highlight the advantages of incorporating physics-based modeling, and uncertainty quantification ensures the reliability of long-term capacity forecasting. In addition, we conducted benchmarking studies to systematically assess the advantages and limitations of the proposed model, thus identifying the scenarios where this approach excels.