An interpretable hybrid LSTM–transformer framework with temporal attention fusion for lithium-ion battery state-of-charge estimation

State-of-charge estimation is essential for lithium-ion battery management, influencing electric vehicle safety, efficiency, and lifespan. Recurrent neural networks are promising, but single network models often fail to capture both short-term fluctuations and long-term dependencies, resulting in a less accurate estimate under varying operating conditions. This paper proposes a parallel hybrid deep learning framework combining Long Short-Term Memory (LSTM) and Transformer architectures with a Temporal Attention Fusion (TAF) mechanism to overcome such limitations. The proposed model is trained on a mixed dataset of three different drive cycles, namely steady state, urban and highway driving conditions and evaluated separately on two new drive cycles with different aggressiveness and duration. The results show a good generalization ability across the different drive cycles, as the model has a high estimation accuracy for both unseen test cycles. Interpretability analysis shows that the TAF mechanism is dynamically adaptive to operating conditions, depending more on LSTM for short-term temporal variations and augmenting the contribution of Transformer for long-term contextual dependencies. The robustness tests with Gaussian noise at different signal-to-noise ratios show good tolerance to moderate measurement disturbances with almost negligible degradation of the metrics. The complexity analysis shows an efficient trade-off between accuracy and computation cost, which is suitable for practical implementations of battery management systems.