Multi-scale temporal convolution attention network for state-of-charge estimation in Li-ion batteries

Accurate estimation of the State of Charge (SOC) in lithium-ion batteries is critical for the safety, reliability, and efficiency of energy storage systems and electric vehicles (EVs). The traditional model-based SOC estimation approaches have their limitations when it comes to the issues of battery aging, dynamic load variations, and temperature fluctuations, while the existing deep learning methodologies mostly struggle to capture the short-term transients and long-term SOC evolution at the same time. To address these issues, the present study proposes a Multi-Scale Temporal Convolution Attention Network (MSTCAN), a fusion of multi-scale temporal convolutions for hierarchical feature extraction and an attention-based technique to highlight the most significant signals from voltage, current, and temperature measurements. The data set of Tesla Model-3 2170 Li-ion battery is prepared for processing by the utilization of data cleaning, min-max normalization, and cycle-based segmentation, resulting in enhanced feature quality and temporal representation improvements. An experiment was carried out where the different techniques were compared, and among them, MSTCAN achieving 1.25% Root Mean Square Error (RMSE), 0.97% Mean Absolute Error (MAE), 3.12% max error, and the coefficient of determination (R²) value of 0.998, thus surpassing the traditional and deep learning methods as well. The study's findings affirm both the model's robustness and its reliability, thus providing an attractive solution that could effectively upgrade the performance and even the efficiency of battery management system (BMS) in practical electric vehicle applications.