Online joint estimation of lithium electronic main states based on WGAN-informer hybrid model

Ensuring the safe and stable operation of lithium batteries requires accurate estimation of State of Health (SOH) and State of Charge (SOC). However, practical engineering faces challenges such as complex electrochemical mechanisms, loss of key parameter measurements, and intricate interactions between state quantities. To address these issues, this paper proposes a joint estimation framework for SOH and SOC across multiple time scales. Leveraging the generative adversarial network structure, Wasserstein Generative Adversarial Network (WGAN), and an enhanced cascade model of the long-series prediction model, Informer, the framework enhances quantitative measurement data quality and achieves precise SOH prediction over an extended time scale. Additionally, it employs a modified Informer encoder structure for SOC estimation at a larger time scale and enhances multi-task model training with a loss function based on mean square error uncertainty. Experimental results demonstrate that the WGAN structure effectively restores high-precision data under significant loss, reducing errors in subsequent estimation models. The Informer structure outperforms other algorithms across diverse time scales, yielding root-mean-square errors of 0.021 for SOH prediction and 0.015 for SOC estimation, showcasing significant accuracy improvements. •The employed deep learning model adeptly accomplishes accurate simultaneous estimation of the lithium battery state.•The informer structure within the joint model enables extended temporal span for state estimation.•The WGAN structure aids accurate estimation of missing battery data, boosting overall model performance.