Physics-informed deep learning models for real-time state of health estimation of lithium-ion batteries
Journal of Power Sources, cilt.694, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 694
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.jpowsour.2026.241122
- Dergi Adı: Journal of Power Sources
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Chemical Abstracts Core, Chimica, Compendex, INSPEC, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO)
- Anahtar Kelimeler: DenseNet-121, Joule loss, Lithium-ion batteries, Physics-Informed Deep Learning, State of Health (SOH), Temporal convolutional networks (TCN)
- Ankara Üniversitesi Adresli: Evet
Özet
Accurate State of Health (SoH) prediction is paramount for the reliable operation of Battery Management Systems (BMS) in electric mobility and stationary storage. Since, lithium-ion batteries constitute a complex, non-linear system driven by coupled electro-thermal dynamics, their degradation is intrinsically linked to time-series variables, and physical parameters. Capturing this underlying physics of ageing requires advanced modelling techniques. However, BMS systems have limited resources to overcome estimation requirements. This limitation underscores the need for lightweight models that do not sacrifice accuracy for efficiency. To address this, we propose a Deep Learning Model that relies on real-time measurable signals (voltage, current, and temperature) extracted from merely a 5-minute operational window. To enhance the generalizability of estimation, the fundamental correlation between Joule losses and capacity fade is embedded into Physics-Informed Deep Learning (PI-DL) models, specifically Physics-Informed Temporal Convolutional Networks (PI-TCN) and Physics-Informed Densely Connected Convolutional Networks (PI-DenseNet-121). Experimental results under random-walk (RW) charge–discharge profiles demonstrate that integrating physical laws into deep learning models enables robust SoH estimation. Specifically, the PI-DenseNet-121 model achieved the highest prediction accuracy values with MAE of 0.0444, MSE of 0.0635, and RMSE of 0.0654, while the PI-TCN model recorded best-case MAE, MSE, and RMSE values of 0.0420, 0.0460, and 0.0504, respectively.