Data-Driven State-of-Charge Estimation for Lithium-Ion Batteries Using Variational Inference

The nonlinear and time-varying characteristics of internal electronic circuits in battery systems pose challenges to the practical application of traditional state-of-charge (SOC) estimation methods. To address these challenges, we first propose a novel variational inference neural network (VINN), which combines deep neural networks with variational inference to enable accurate SOC estimation solely based on measured current, voltage, and temperature. Specifically, the proposed VINN employs an encoder to infer a latent SOC distribution from measurement inputs, and a decoder is used to reconstruct the inputs for computing the evidence lower bound (ELBO), which guides the network optimization. Meanwhile, a physics-informed loss is introduced by using Coulomb counting to derive the prior of SOC, ensuring consistency with the battery’s circuit behavior. Finally, experiments on realistic Panasonic 18650PF dataset demonstrate that VINN achieves superior estimation accuracy and generalization performance under both supervised and unsupervised settings compared to state-of-the-art methods.