A lightweight Transformer-LSTM Method for Long-Sequence Battery State-of-Health Estimation

Modern Electric Vehicles (EVs) and grid-level energy storage systems rely on accurate lithium-ion battery capacity forecasting to ensure operational safety. However, conventional recurrent neural network (RNN) approaches often struggle to capture long-range dependencies, resulting in suboptimal predictions. Accurately predicting battery health over long-sequence remains a key challenge for developing advanced Battery Management Systems (BMS) and improving Remaining Useful Life (RUL) estimation. In this paper, a lightweight Transformer-
LSTM (TF-LSTM) framework is developed. Compared to the traditional TF-LSTM, it integrates multi-head self-attention and positional encoding with LSTM’s gating mechanism, while simultaneously leveraging Automatic Mixed Precision (AMP)
to further reduce computational overhead. The performance of the proposed model has been validated through numerical simulations and leave-one-out cross-validation on four cells from the CALCE Li-ion Battery Aging Dataset. The experimental results indicate that the proposed approach achieves a 39.2% lower relative error (RE) compared to conventional RNNs for extended sequence lengths. Therefore, it demonstrates the enhanced generalization of the model, making it well-suited for large-scale EVs battery monitoring and industrial energy storage prognostics requiring precise long-horizon forecasting.