A Multi-Scale Hybrid Neural Network with Attention and Intelligent Optimization for High-Precision State of Charge Estimation of Lithium-Ion Batteries
TL;DR
A hybrid neural network optimized by an improved sparrow search algorithm for SOC estimation, combining temporal convolutional networks for local feature extraction and bidirectional long short‑term memory networks for long‑term dependency capture, promising wide battery management system applications.
Abstract
With the rapid development of the new‑energy lithium‑battery industry, precise state‑of‑charge (SOC) estimation is vital for lithium‑ion battery system safety and performance. Traditional empirical or single‑model methods fail to adapt to SOC variations amid complex operating conditions and long service cycles. This paper puts forward a hybrid neural network optimized by an improved sparrow search algorithm for SOC estimation. Combining temporal convolutional networks for local feature extraction and bidirectional long short‑term memory networks for long‑term dependency capture, the model integrates short‑ and long‑term information. A self‑attention mechanism weights key time steps to strengthen feature extraction, while the improved algorithm optimizes hyperparameters to avoid manual tuning drawbacks. Experiments under multiple temperatures and working conditions show the method’s high accuracy, robustness and generalization, with maximum mean absolute error kept at 0.8%–1.2%, promising wide battery management system applications.