Accurately estimating the state of charge (SOC) of lithium-ion batteries by the battery management system (BMS) is crucial for efficient energy management and power distribution control in electric vehicles (EVs). By far, data-driven methods have been extensively used in estimating the SOC of lithium-ion batteries to efficiently operate EVs with limited resources. However, the learning ability of these methods needs further enhancement. Therefore, this paper aims to analyze the SOC performance of an extended-input long short-term memory (ELSTM) model. The inputs of the ELSTM are improved with identified battery parameters based on an adaptive multi-timescale identification method with frequency difference decomposition. Second, an adaptive singular value decomposition-transformed unscented Kalman filter (ASVDUKF) is proposed to integrate all unknown variables and guarantee the positive definiteness of the error covariance matrix into a vector to form a multi-fusion model that robustly outputs the denoised and optimized SOCs based on the estimations of the ELSTM. The SOC results demonstrate that the ELSTM outperforms the LSTM, which uses conventional input data. Following that, the ELSTM-ASVDUKF model ensures high accuracy and stability with optimal mean absolute error, root-mean-square error, and mean absolute percentage error of 0.0939%, 0.1074%, and 0.1257%, respectively. The proposed model is validated using various full-cycle current rate and temperature data from two different batteries, establishing it as a promising solution for future development.