Residual Learning for Real-Time Prediction of Lithium-Ion Battery Degradation in Maritime Applications

Abstract

Accurate prediction of lithium-ion battery degradation is crucial for optimizing performance, extending lifespan, scheduling maintenance as well as safe operation of batteries in maritime energy systems. Traditional physics-based models rely on electrochemical principles to simulate known degradation mechanisms such as Solid-Electrolyte Interface (SEI) formation and Active Material (AM) loss, but they can be computationally intensive, need many electrochemical parameters to be defined, and do not capture effects that are not mathematically modeled. Machine learning (ML) approaches have shown success in predicting battery degradation, however, they need large datasets covering a variety of operational conditions to learn and are agnostic of the underlying mechanisms, as a result of which they lack interpretability and struggle with generalization. This paper applies the residual learning approach for combining physicsbased modeling and ML to predict lithium-ion battery degradation. Reduced-order models are used to describe the SEI layer formation and AM loss at the graphite anode, and combined with the Long-short term memory model. The approach is demonstrated using an open-source degradation dataset for Lithium Ferrous Phosphate cells, and used with a system model to estimate the life of batteries for ship applications. The hybrid approach improves the understandability of the results, is generalizable to batteries with graphite anodes and runs in real-time, thereby enabling real-time monitoring and prognostics for maintenance scheduling, which are of critical importance for maritime applications.