Estimating Lithium-Ion Battery Health Status: A Temporal Deep Learning Approach With Uncertainty Representation

Precise estimation of the state of health (SOH) of lithium-ion batteries is essential for the reliable and safe functioning of energy-storage systems. However, existing data-driven methods seldom account for the uncertainties involved in the battery degradation process, which diminishes the reliability of SOH estimation models. To address this issue, this article proposes a novel temporal deep learning approach with uncertainty representation. In this methodology, temporal convolutional networks (TCNs) are employed for front-end feature extraction, effectively capturing temporal patterns within battery data. The Transformer architecture is utilized for SOH prediction, leveraging its capability to model complex dependencies in sequential data. To enhance predictive performance, the Transformer decoder is replaced by bidirectional long short-term memory quantile regression (BiLSTMQR), allowing for the estimation of SOH values at various quantiles, providing a more detailed understanding of the battery’s health status. Finally, kernel density estimation (KDE) is applied to transform these quantile estimates into probability density functions, offering a probabilistic interpretation of the SOH values. Experimental findings indicate that the proposed framework significantly improves the precision and reliability of SOH estimation when compared to other intelligent methods, making it a promising solution for real-time monitoring and management of lithium-ion batteries.