Identification and machine learning prediction of knee-point and knee-onset in capacity degradation curves of lithium-ion cells
Paula Fermín-Cueto,Euan McTurk,Michael Allerhand,Encarni Medina-Lopez,Miguel F. Anjos,Joel Sylvester,Gonçalo dos Reis
DOI: https://doi.org/10.1016/j.egyai.2020.100006
2020-08-01
Energy and AI
Abstract:High-performance batteries greatly benefit from accurate, early predictions of future capacity loss, to advance the management of the battery and sustain desirable application-specific performance characteristics for as long as possible. Li-ion cells exhibit a slow capacity degradation up to a knee-point, after which the degradation accelerates rapidly until the cell’s End-of-Life. Using capacity degradation data, we propose a robust method to identify the knee-point within capacity fade curves. In a new approach to knee research, we propose the concept ‘knee-onset’, marking the beginning of the nonlinear degradation, and provide a simple and robust identification mechanism for it. We link cycle life, knee-point and knee-onset, where predicting/identifying one promptly reveals the others. On data featuring continuous high C-rate cycling (1C–8C), we show that, on average, the knee-point occurs at 95% capacity under these conditions and the knee-onset at 97.1% capacity, with knee and its onset on average 108 cycles apart. After the critical identification step, we employ machine learning (ML) techniques for early prediction of the knee-point and knee-onset. Our models predict knee-point and knee-onset quantitatively with 9.4% error using only information from the first 50 cycles of the cells’ life. Our models use the knee-point predictions to classify the cells’ expected cycle lives as short, medium or long with 88–90% accuracy using only information from the first 3–5 cycles. Our accuracy levels are on par with existing literature for End-of-Life prediction (requiring information from 100-cycles), nonetheless, we address the more complex problem of knee prediction. All estimations are enriched with confidence/credibility metrics. The uncertainty regarding the ML model’s estimations is quantified through prediction intervals. These yield risk-criteria insurers and manufacturers of energy storage applications can use for battery warranties. Our classification model provides a tool for cell manufacturers to speed up the validation of cell production techniques.