Early Lifetime Prediction of Lithium-ion Batteries Based on Relative Position Matrix and Residual Network

Early lifetime prediction of lithium-ion batteries is critical for timely risk warning and ensuring the safe operation of energy systems. In recent years, deep learning methods with excellent ability in feature extraction, have gained increasing attention in battery prognostics. However, most of them directly stack cycle data as inputs, neglecting to enhance the subtle battery degradation features in early cycles, and suffer from poor model interpretability. Therefore, this paper proposes an early lifetime prediction method for lithium-ion batteries based on relative position matrix and residual network. Furthermore, the model’s decision-making process is visually analyzed using gradient-weighted class activation mapping technology. Experimental results demonstrate that the proposed method can achieve accurate early lifetime prediction, with a root-mean-square-error of less than 126, only using the first 40 cycles data. Additionally, visual analysis indicates that the model focuses more on the dynamic characteristics of battery capacity degradation during discharge.