Universal Battery Degradation Forecasting Driven by Foundation Model Across Diverse Chemistries and Conditions

TL;DR

The paper addresses the challenge of forecasting battery capacity fade across diverse chemistries, formats, and operating conditions by aggregating a large, heterogeneous corpus from 20 public aging datasets (1,704 cells and 3.96 million cycles). It proposes a Time-Series Foundation Model (TSFM) framework built on the Timer backbone, with LoRA-based parameter-efficient adaptation and physics-guided contrastive representation learning to capture shared degradation patterns while remaining transferable to unseen regimes. The authors demonstrate that a single unified model achieves competitive or superior accuracy on seen datasets and strong zero-shot generalization to unseen chemistries and formats, aided by interpretability analyses (LIME) that reveal physically meaningful attribution patterns. The results highlight the potential for scalable, cross-platform battery health forecasting in real-world BMS applications and point to future work integrating field data and broader deployment in practical workflows.

Abstract

Accurate forecasting of battery capacity fade is essential for the safety, reliability, and long-term efficiency of energy storage systems. However, the strong heterogeneity across cell chemistries, form factors, and operating conditions makes it difficult to build a single model that generalizes beyond its training domain. This work proposes a unified capacity forecasting framework that maintains robust performance across diverse chemistries and usage scenarios. We curate 20 public aging datasets into a large-scale corpus covering 1,704 cells and 3,961,195 charge-discharge cycle segments, spanning temperatures from $-5\,^{\circ}\mathrm{C}$ to $45\,^{\circ}\mathrm{C}$, multiple C-rates, and application-oriented profiles such as fast charging and partial cycling. On this corpus, we adopt a Time-Series Foundation Model (TSFM) backbone and apply parameter-efficient Low-Rank Adaptation (LoRA) together with physics-guided contrastive representation learning to capture shared degradation patterns. Experiments on both seen and deliberately held-out unseen datasets show that a single unified model achieves competitive or superior accuracy compared with strong per-dataset baselines, while retaining stable performance on chemistries, capacity scales, and operating conditions excluded from training. These results demonstrate the potential of TSFM-based architectures as a scalable and transferable solution for capacity degradation forecasting in real battery management systems.

Figures (4)

• Figure 1: The flowchart of the large-model-driven battery capacity forecasting.(a) Unlike conventional, task-specific capacity degradation models, our approach constructs a rich, standardized corpus of degradation trajectories and employs a single time-series foundation model to forecast capacity under previously unseen conditions. (b) In the physics-informed training strategy, the endogenous variable (capacity) and exogenous variables (operating conditions) are treated differently: charge–discharge features are incorporated as auxiliary physical embeddings during training to exploit domain knowledge, while the model remains independent of these potentially missing physical inputs at inference time.
• Figure 2: Dateset overview.(a) Overview of the unified battery degradation corpus constructed in this study. The integrated dataset spans a broad range of electrochemical systems and application scenarios; NMC-based cells account for the vast majority of samples, while emerging sodium-ion, zinc-ion, and mixed-chemistry batteries occupy only a small fraction of the corpus. (b) Representative statistical analysis of DFs with respect to battery chemistry. The panel shows 3D scatter plots, 2D projections, and the corresponding MI and ANOVA effect sizes ($\eta^2$) between DFs and chemistry labels. (c) Correlation-coefficient heat map and scatter plots between physical features and capacity (x-axis: physical features; y-axis: capacity). Specifically, we compute ten physical descriptors, including charging time, the mean and variance of current and voltage, voltage kurtosis, and related statistics.
• Figure 3: The illustrations of capacity forecasting results.(a) Capacity forecasting performance of the proposed method on unseen datasets, together with the MAPE differences relative to competing baselines, showing that our approach maintains superior accuracy even under previously unobserved conditions. (b) Capacity forecasting results on seen datasets, where the model exhibits robust performance across a wide spectrum of battery chemistries, form factors, and operating scenarios. Using a single time-series foundation model, our method matches or even surpasses multiple supervised state-of-the-art time-series forecasting baselines. (c) Distribution of per-cell average errors and Friedman ranking on unseen datasets. (d) Distribution of per-cell average errors and Friedman ranking on seen datasets. The combined box--violin plots visually highlight the advantage of our approach, and the Friedman rankings show that the proposed single foundation model never falls outside the top three and attains the best rank on many datasets.
• Figure 4: LIME-based interpretation of degradation forecasting.(a) Comparison of LIME importance maps for the proposed method and competing baselines on unseen datasets. (b) Comparison of LIME importance maps for the proposed method and competing baselines on seen datasets. (c) Relationship between explanation discrepancy and forecasting accuracy: scatter plot of LIME distribution distance versus $\log(\mathrm{MAPE})$ for all methods, linking differences in attribution patterns to model performance.