Attention-Based Multimodal Transformer-LSTM Fusion Networks for Enhanced Lithium-ion Battery State-of-Charge Prediction with Aging Pattern Consideration

Abstract

Lithium-ion batteries play a pivotal role in electric vehicles (EVs) and energy storage systems, where accurate State-of-Charge (SoC) prediction is essential for ensuring the efficiency and safety of battery management systems (BMS). This study conducts an in-depth analysis of the temperature factors that most significantly influence battery remaining capacity prediction, with a particular focus on accurately predicting battery SoC variations under extreme temperature conditions ranging from - 30 degrees C to 80 degrees C. The research methodology employs a multimodal neural network that com-bines Transformer architecture, which demonstrates superior performance in processing static characteristic data, with Long Short-Term Memory (LSTM) networks, which exhibit exceptional capabilities in time-series data processing. The model effectively integrates battery temperature performance data with NASA aging datasets through an attention-based fusion approach, enabling efficient information exchange between heterogeneous data modalities. Experimental results demonstrate that the proposed model achieves an MAE of 0.2438 +/- 0.016 on the overall test set across - 30 degrees C to 80 degrees C. Additionally, the model attained an RMSE of 0.3156 +/- 0.020 and an R & sup2; of 0.9501 +/- 0.009, representing a 24.6% improve-ment over baseline LSTM models, and an additional 14% improvement attributable to the attention mechanism when compared to simple concatenation-based fusion. Furthermore, the attention-based fusion approach demonstrates marked performance improvements compared to simple combination methods. These results demonstrate potential applicability to electric vehicles and energy storage systems through comprehensive laboratory validation under diverse temperature conditions (-30 degrees C to 80 degrees C), though field testing is required for deployment verification.