Abstract: Aiming at the problems that a single-structure deep learning model struggles to fully capture the complex information in the charging and discharging process of lithium-ion batteries, and that hyperparameters relying on empirical settings limit the prediction accuracy and stability, an intelligent estimation method integrating two-dimensional grid long short-term memory (2-GLSTM) and cuckoo search algorithm (CSA) was proposed for state of charge (SOC) estimation of lithium-ion batteries. Firstly, a 2-GLSTM deep network was constructed. By capturing the temporal dynamic characteristics of battery current and voltage, the coupling information between the temporal output and the original input was further fused to form a dual-dimensional feature extraction mechanism, thereby realizing the comprehensive extraction of multi-dimensional key information in the charging and discharging process. Secondly, CSA was introduced to perform adaptive optimization using the validation set loss as the fitness function to determine the optimal hyperparameter combination. Finally, the model was trained using data from dynamic stress test (DST) condition and verified under the federal urban driving schedule (FUDS) condition. The model performance was comprehensively evaluated through ablation experiments and comparative tests with other temporal models. The results show that the proposed method achieves a root mean square error (RMSE) of 0.36% and a mean absolute error (MAE) of 0.32% for SOC estimation under FUDS condition; compared with the conventional long short-term memory network, the MAE is reduced by 59.5%. Through the synergistic effect of dual-dimensional feature extraction and hyperparameter adaptive optimization, the proposed method effectively improves the accuracy and stability of SOC estimation under complex dynamic conditions, providing a new technical path for the intelligent development of battery management systems.