Accurate Remaining Useful Life (RUL) estimation of lithium-ion batteries is essential for ensuring the reliability and safety of modern energy systems. However, data scarcity, complex operating conditions, and model uncertainty remain major challenges. This paper presents a neural network-based framework that integrates transfer learning and Monte Carlo (MC) dropout to enhance prediction accuracy, generalization, and reliability. Five state-of-the-art neural architectures are benchmarked using two public datasets and subsequently transferred on the smaller University of Michigan Battery Lab (UMBL) dataset. Experimental results show that transfer learning works effectively only when models are fine-tunned properly. Furthermore, MC dropout effectively quantifies predictive uncertainty, providing confidence intervals that prevent overconfident maintenance decisions. Among all models, the Gated Recurrent Unit (GRU)-based network achieved the best trade-off between predictive accuracy and uncertainty calibration. The proposed framework demonstrates a practical and scalable approach for reliable battery prognostics, supporting safer and more efficient battery management systems.
Cross-Dataset Battery Rul Prediction Using Multi-Source Pretraining and Uncertainty Estimation
Ding, Siwen;Martiri, Luca;Cristaldi, Loredana;Faifer, Marco
2026-01-01
Abstract
Accurate Remaining Useful Life (RUL) estimation of lithium-ion batteries is essential for ensuring the reliability and safety of modern energy systems. However, data scarcity, complex operating conditions, and model uncertainty remain major challenges. This paper presents a neural network-based framework that integrates transfer learning and Monte Carlo (MC) dropout to enhance prediction accuracy, generalization, and reliability. Five state-of-the-art neural architectures are benchmarked using two public datasets and subsequently transferred on the smaller University of Michigan Battery Lab (UMBL) dataset. Experimental results show that transfer learning works effectively only when models are fine-tunned properly. Furthermore, MC dropout effectively quantifies predictive uncertainty, providing confidence intervals that prevent overconfident maintenance decisions. Among all models, the Gated Recurrent Unit (GRU)-based network achieved the best trade-off between predictive accuracy and uncertainty calibration. The proposed framework demonstrates a practical and scalable approach for reliable battery prognostics, supporting safer and more efficient battery management systems.| File | Dimensione | Formato | |
|---|---|---|---|
|
Cross-Dataset_Battery_Rul_Prediction_Using_Multi-Source_Pretraining_and_Uncertainty_Estimation.pdf
Accesso riservato
:
Publisher’s version
Dimensione
884.8 kB
Formato
Adobe PDF
|
884.8 kB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.



