State of charge estimation for lithium-ion battery using Transformer with immersion and invariance adaptive observer

Accurate state-of-charge (SOC) estimation lays the foundation for lithium-ion batteries' long-life and safe services. This paper exploits a new machine-learning method and an adaptive observer to estimate the battery's SOC. First, a Transformer neural-network is employed to predict the SOC with the sequence of current, voltage, and temperature data as inputs. Second, an innovative immersion and invariance (I&I) adaptive observer is applied to reduce the oscillations of the Transformer's prediction. The lead of the Transformer network lies in that it has an overview of the entire input sequence and obtains richer information than other conventional neural networks. Besides, the I&I adaptive observer is competent for correcting possible learning fluctuations and ensuring that the battery parameter estimation error is confined within an invariant manifold. The proposed methods are validated with experimental data. The results demonstrate their higher SOC estimation accuracy than a popular baseline method.