Abstract
Accurate early prediction of Li-ion battery aging facilitates new product optimization and application management. Here, a joint modeling scheme is proposed. It is dedicated to decoupling cell-to-cell variability and cycle-by-cycle nonlinear aging in Li-ion batteries, enabling accurate cycle life and capacity trajectory predictions. An aging end-point early predictor is established based on the multiple output Gaussian process (MOGP). It enables better performance by sharing knowledge across outputs, including cycle life and life capacity. For degradation trajectory prediction, an aging process trajectory predictor is constructed based on the ‘prompt-learning’ paradigm neural network, and the life capacity from MOGP is used as the prompt input. Moreover, a Bayesian search strategy is proposed to obtain optimal feature combinations, and feature alignment is used to narrow the gap between source and target domains. The proposed scheme is developed on the source domain with 41 cells and tested on three special target domains with 128 cells. As a result, the predicted error of cycle life can be within 7.87%, and that of the capacity trajectory can be within 6.53%. The prediction results prove its high adaptability and effectiveness. The achievements will promote high-dimensional parameter design and optimization, accelerating battery development and reducing testing costs.
[Display omitted]
•Accurate prediction is achieved for both LIB cycle life and aging trajectory.•An aging end-point early predictor is built using the multi-output Gaussian process.•A prompt-learning paradigm joint model is used for degradation trajectory prediction.•Optimization strategies are proposed for feature combination and alignment.•Effectiveness is demonstrated on 169 commercial cells for 80 protocols.