Robust Nonlinear Observer Design with Learning Applied to SOC Estimation in Li-Ion Batteries

Abstract This paper proposes a novel robust nonlinear observer (RNO) with learning capacity (LC) for state of charge estimation in lithium-ion batteries. The observer is designed via a convex optimization formulation that guarantees the input-to-state stability of the estimation error dynamics under exogenous disturbances. An auxiliary correction term, generated by a machine learning scheme, is incorporated to enhance the estimation performance. The learning mechanism employs a feedforward neural network that processes delayed measurements; however, the proposed structure can accommodate more complex learning mechanisms provided that the correction signal remains magnitude-bounded. The proposed scheme is validated through comprehensive simulations and practical experiments, with its performance benchmarked against well-established observers. Both numerical and experimental results demonstrate that the proposed observer outperforms the RNO without LC as well as other well-known observers.