Estimating the State of charge (SOC) a crucial aspect of the keys of Lithium-Ion Battery BMS Management. At present, data-driven calculations are predominantly used to determine SOC levels. One of the learning methods used is a recurrent neural network (RNN) which performs very well. We propose a bidirectional long-term memory neural network (BiLSTM) for SOC estimation. This BiLSTM also undergoes optimization with several variations used. In this paper, we compare the variations between BiLSTM and BiLSTM-AM. Using public data that has been published, we propose 4 data that we process using separate BiLSTM and BiLSTM-AM neural networks for 4 batteries. The SOC estimation obtained from each battery will be assessed, with an R2 value 0.99, enabling the model to produce more accurate and better results. © 2025 IEEE.