Lightweight State of Charge Estimation Model Based on Multi-Timescale Feature Enhancement for Lithium-ion Batteries

Abstract

Accurate state of charge estimation for lithium-ion batteries is severely challenged by multi-timescale coupling and wide temperature fluctuations. To resolve these issues, this paper introduces a feature-enhanced echo state network optimized via an improved sparrow search algorithm. A physics-inspired dual-reservoir architecture is developed to structurally decouple battery dynamics, where a fast reservoir tracks short-term transient polarizations and a parallel slow reservoir preserves long-term energy integration memory. To handle the non-convex hyperparameter space, the baseline algorithm is updated through chaotic Tent map initialization and an adaptive spiral search strategy to escape local optima. Crucially, by executing calibration exclusively offline and circumventing iterative backpropagation, the framework achieves exceptional lightweight efficiency. Extensive validation across dynamic profiles from -5°C to 35°C demonstrates that the model consistently limits the mean absolute error below 2% and the mean squared error below 0.05%, while processing tens of thousands of data points in approximately 1 second. This establishes a highly robust and computationally efficient solution for real-time embedded energy management systems.