Remaining Useful Life Prediction of Lithium-Ion Batteries Using Physics-Informed Inverted Attention Transformer

Abstract

PredictingHow long a battery can keep working before it stops functioning properly is very important for electric vehicles, systems that store energy from renewable sources, and devices that are carried around. This remaining operating time is known as Remaining Useful Life (RUL). Accurate RUL prediction helps prevent sudden battery failures and allows better planning of maintenance. In this work, we propose a transformer-based deep learning framework for predicting the RUL of lithium-ion batteries. Three transformer models are evaluated using two widely used public datasets: the Oxford Battery Degradation Dataset and the NASA Battery Aging Dataset. First, a baseline iTransformer model is used to learn patterns from battery measurements such as voltage, current, temperature, and cycle time. Next, a time-informed iTransformer is introduced, where positional encoding is added so that the model can better understand the order of battery cycles. Finally, a physics-informed iTransformer is proposed by adding a smoothness constraint during training so that the predicted degradation pattern follows realistic battery behavior. Experimental results show that the physics-informed iTransformer provides more stable predictions and lower errors compared to the other models. The model also performs well when tested on a different dataset, which shows good generalization ability. This study demonstrates that combining transformer models with simple physics knowledge can improve battery life prediction and support the development of reliable battery management systems.