State-of-Charge Estimation of Lithium-Ion Battery Integrated in Electrical Vehicle Using a Long Short-Term Memory Network

Improving the accuracy of state-of-charge (SoC) estimation is crucial for electric vehicles (EVs) using Lithium-Ion batteries (LiBs). This helps users reliably predict driving range and optimize the charging process, thereby extending battery life and ensuring safety during use. However, due to temperature, driving mode, and charge-dependent electrochemical nonlinear dynamics, SoC estimation for LiB integrated with EVs remains a significant technical challenge. In particular, SoC estimation in the regions of SoC < 30% and SoC > 80% is often inaccurate due to nonlinearity and sensitivity to battery aging. Accurate estimation in these regions is crucial for making decisions regarding recharging and discharging to prolong battery life and prevent damage. To address this issue, this paper proposes a method for SoC estimation using a Long Short-Term Memory (LSTM) network, which is capable of retaining information on battery characteristics related to changes long term electrochemical parameter changes, such as the number of discharge cycles and the aging effects. The method utilizes practical data from 80,000 samples collected from pure electric vehicle testing under different driving modes, temperatures, and road conditions over a 30-day period. The LSTM network was optimized by adjusting the input data sequence and hidden size to minimize the number of hyperparameters. This makes it suitable for use on low-cost processors with moderate computing power. SoC estimation was evaluated across four SoC test regions: SoC < 30%, SoC > 80%, 30% <= SoC <= 80%, and 0% <= SoC <= 100%. The results were compared with feedforward neural network (FNN) and convolutional neural network (CNN). Despite having a configuration with a hidden size of 96 and a single layer, the LSTM model achieved estimation accuracy with RMSE = 0.0106, MAE = 0.0077, and MAPE = 1.4116%.