BROWSE
Detail
Journal Article
Detailed Architectural Design of a Multi-Head Self-Attention Model for Lithium-Ion Battery Capacity Forecasting
- Authors
- Juyeon Park, Gyeong Ho Lee, Jangkyum Kim, Yoon-Sik Yoo, Il-Woo Lee
- Issue Date
- 2025-03
- Citation
- IEEE Access, v.13, pp.48212-48225
- ISSN
- 2169-3536
- Publisher
- Institute of Electrical and Electronics Engineers Inc.
- Language
- English
- Type
- Journal Article
- Abstract
- As the adoption of lithium-ion batteries increases, concerns over safety incidents and replacement costs have become increasingly pressing. Accurate battery state prediction is thereby vital for reducing costs and ensuring safety. This paper introduces a novel system based on a transformer encoder architecture, leveraging Multi-Head Self-Attention (MHSA) layers to enhance battery capacity prediction. To address variability in battery data collection, we implement robust preprocessing techniques and a sliding window method to standardize data input. Positional encoding is applied to embed sequence order information at the input stage, while residual connections and layer normalization between MHSA layers optimize the learning process. Final predictions are generated through a fully connected neural network (FNN). Experimental results on the NASA dataset, obtained from repeated charge-discharge cycles, demonstrate that the proposed model mostly achieves superior predictive accuracy across multiple cell datasets, as measured by MAPE and RMSE. The proposed model showed enhanced performance, achieving a MAPE of 0.9918% and an RMSE of 0.02542 in experiments on specific cells. Furthermore, compared to other models in most experiments, it showed a performance improvement of at least 41% in MAPE and 29% in RMSE. Notably, the model maintains excellent performance while reducing complexity. Additionally, to prevent overestimation of overall performance due to high accuracy in only the early stages, we conducted interval-based performance evaluations, confirming that the proposed model consistently provides accurate predictions with minimal variance across different stages compared to other models.
- KSP Keywords
- Architectural Design, Attention model, Battery state, Data Collection, Different stages, Early stages, Enhanced performance, High accuracy, Ion batteries, Learning Process, Lithium-ion batteries(LIBs)
This work is distributed under the term of Creative Commons License (CCL)
(CC BY)
(CC BY)
