Electric vehicles (EVs) are central to achieving carbon neutrality, with the battery pack acting as the crucial energy storage system. However, applying models designed for single cells directly to battery packs can be problematic because of variations in electrochemical parameters such as capacity and internal resistance, even among cells from the same production batch. These discrepancies can lead to significant errors in the state of charge (SOC) estimation. To address this issue, we propose an algorithm combining the cell mean model (CMM) with a long short-term memory (LSTM) neural network for more accurate SOC estimation in battery packs. By analyzing the differences among individual cells, we identify those with the most pronounced variations and those that reach the cut-off voltage first as representative cells. The CMM is used to summarize the pack's overall characteristics, and an extended Kalman filter (EKF) is employed for preliminary SOC estimation. Finally, the LSTM model refines the SOC estimate by learning complex dynamics between initial SOC values, representative cell data, and the actual pack SOC. Experimental results show that this approach achieves a root mean square error and mean absolute error under 1 %, significantly improving SOC estimation accuracy in dynamic conditions compared to traditional methods.