Multi-step prediction of battery state of health based on self-supervised pre-training and transfer learning using the xPatch model

Battery state of health (SOH) is a critical parameter in battery management systems, as its accurate prediction ensures safe operation and efficient energy management. However, batteries experience complex nonlinear aging processes, and variations in materials or usage conditions lead to diverse degradation modes, limiting prediction accuracy. Moreover, single-step SOH prediction often fails to meet multi-scenario energy management requirements. To address these challenges, this paper proposes a multi-step SOH prediction method for lithium-ion batteries (LIBs) based on self-supervised transfer learning with the xPatch model. The xPatch model combines convolutional neural networks (CNNs) and multi-layer perceptrons (MLPs) as its backbone, while a dual-stream mechanism captures long-term degradation trends and cyclic characteristics, balancing accuracy and computational efficiency. To adapt to different degradation patterns, a self-supervised learning (SSL) strategy integrating reconstruction, denoising, and prediction is designed. This framework not only reconstructs degradation behaviors but also enhances robustness against noise, improving generalization across LIB types and conditions. The proposed method is validated on three LIB aging datasets. Experimental results show an average SOH R2 of 0.964, while transfer learning improved R2 by 2.1 % compared to direct training, confirming the effectiveness of the approach in achieving accurate and robust SOH forecasting.