EUVIMEDEUVIMED entdecken
Crossref

Multivariate Time-Series Forecasting with Cross-Attention for Long-Horizon Battery Cycling Degradation

Xu Zhang, Florent Magaud, Mahdi Abid, Arnaud Demortiere

ECS Meeting Abstracts · 07.07.2026 · MA2026-01 (6) · 672-672

DOI: 10.1149/ma2026-016672mtgabs ↗

Kurz gefasst

Fachartikel in der Zeitschrift ECS Meeting Abstracts, erschienen am 07.07.2026. Verfasst von Xu Zhang und 3 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

In the context of the accelerating next industrial revolution propelled by modeling, data analytics, AI and digital twins across fields such as energy and transportation, batteries are key technology. The simulation of battery systems enables accurate prediction of electrochemical profiles and forecasting of degradation to optimize performance, lifetime, and safety. Physics-based (mechanistic) models provide interpretability and extrapolation power by encoding conservation laws and degradation mechanisms, enabling virtual testing beyond measured conditions. However, they can be computationally intensive, require difficult-to-identify parameters, and may still miss complex failure pathways or cell-to-cell variability. Conversely, data-driven models have recently advanced through modern AI, including time-series forecasting and generative approaches that learn directly from cycling profiles to emulate voltage/current trajectories and ageing trends at scale. Their limitations are strong dependence on data quality, coverage, and representativeness, with reduced robustness under distribution shift and limited physical transparency. Despite these constraints, combining rich electrochemical datasets with modern AI opens perspectives strong to accelerate characterization, shorten experimental campaigns, reduce cost, and ultimately speed up the development and deployment of safer, longer-lasting battery technologies. Battery performance, lifetime, and safety are governed by strongly coupled effects of temperature, charging protocols, usage patterns, and chemistry, whose reliable characterization still relies on long, costly, and labor-intensive testing under controlled conditions. To address these constraints, our research focuses on electrochemical time-series forecasting using sequence-to-sequence LSTM networks. We propose a deep learning architecture built on stacked LSTM layers, trained on multiple datasets, and augmented with a cross-attention block to capture regime changes and long-range dependencies across multichannel signals, thereby linking local cycle-scale voltage-current profile dynamics with long-term capacity (ageing) evolution. We exploited a large open-source dataset provided by the Toyota laboratory (LFP/Graphite), including measurements of voltage, current and the number of charge-discharge cycles (40 different batteries in 2 years). These data show that voltage and the current clearly have seasonality but as the number of cycles increases, the duration of a seasonality decreases, so we have nonstationary data. The electrochemical data were acquired without a constant time step, so we interpolated the data to have time steps required for time series auto-gressive models. After completing the pre-processing stage, we trained a sequence model comprising two stacked LSTM layers followed by a fully connected output head. To better exploit the information content of the input sequences, we integrated a cross-attention module upstream of the LSTM block, enabling enhanced feature aggregation and redistribution prior to temporal modeling. We conducted a systematic hyperparameter study, exploring input sequence lengths (starting at 90 timesteps to the nominal dataset window and match extending to longer contexts), the depth of the LSTM stack, the number of hidden units per layer, and the number of cross-attention cores, in order to identify the best-performing configuration. Model development proceeded in two stages: we first trained and validated on a single dataset, then scaled training to several dataset files to improve generalization and robustness. On the held-out test set, the resulting model achieved a mean squared error (MSE) of 0.001, demonstrating a strong ability to reproduce both curve deformation and the evolution of cycle duration with ageing. In addition to pointwise error metrics, we used Dynamic Time Warping (DTW) to quantify temporal misalignment between predicted and reference trajectories, providing a dedicated measure of cycle-length drift over longer forecasting horizons. For extended-horizon forecasting, we observe that cycle-length predictions can progressively drift from the reference trajectory, revealing compounding errors in dynamically unstable phases. To counter this, we move to a foundation-style model trained across the full set of 40 batteries, leveraging the diversity of regimes and ageing pathways to improve robustness without over-specifying physics. On the held-out test set, this scaled model reaches MAE = 0.0066 and MSE = 0.0011, while substantially reducing long-horizon drift and preserving high-fidelity reconstruction of local voltage–current profile deformation. Crucially, the SOH trajectory remains closely aligned with ground truth up to 2000 cycles, supporting reliable long-range forecasting for performance and lifetime assessment. Figure 1

Zitieren

APA

Zhang, X., Magaud, F., Abid, M., & Demortiere, A. (2026). Multivariate Time-Series Forecasting with Cross-Attention for Long-Horizon Battery Cycling Degradation. ECS Meeting Abstracts, MA2026-01(6), 672-672. https://doi.org/10.1149/ma2026-016672mtgabs

Vancouver

Zhang X, Magaud F, Abid M, Demortiere A. Multivariate Time-Series Forecasting with Cross-Attention for Long-Horizon Battery Cycling Degradation. ECS Meeting Abstracts. 2026;MA2026-01(6):672-672. doi:10.1149/ma2026-016672mtgabs

BibTeX
@article{Zhang2026,
  author = {Zhang, Xu and Magaud, Florent and Abid, Mahdi and Demortiere, Arnaud},
  title = {Multivariate Time-Series Forecasting with Cross-Attention for Long-Horizon Battery Cycling Degradation},
  journal = {ECS Meeting Abstracts},
  year = {2026},
  volume = {MA2026-01},
  number = {6},
  pages = {672--672},
  doi = {10.1149/ma2026-016672mtgabs}
}
Kostenlos starten →