Transformer autoencoder framework for estimating core temperature of lithium-ion battery from pulse discharge dynamics

This study presents a non-invasive, data-driven approach for estimating the core temperature of lithium-ion batteries using a transformer-based autoencoder model. Pulse discharge data were collected from a Panasonic NCR18650B cell at three different C-rates (0.5C, 1C, and 2C) to train and evaluate the model. The transformer autoencoder leverages its ability to capture long-range temporal dependencies, efficiently reconstructing multivariate battery signals while compressing them into latent representations that serve as proxies for core temperature. Additionally, two-dimensional visualisation using Principal Component Analysis and t-distributed Stochastic Neighbor Embedding of the latent space revealed well-separated and structured clusters, further confirming the model's ability to encode relevant thermal dynamics effectively. These latent features were used as inputs to a random forest regressor, which was trained to predict temperature and validated against a physical-based thermal model. The proposed method achieved high accuracy across all discharge rates, physics-based model, and drive cycle analysis, with R 2 scores of >0.99 outperforming previously reported studies. These results demonstrate the transformer autoencoder's superior ability to extract meaningful temporally structured representation and its robustness in dynamic operating conditions. • Transformer autoencoder infers core temperature from standard battery signals. • Latent space captures key thermal features for predicting core temperature of Li-ion. • Achieves high accuracy across pulse discharge rates for potential real-time BMS use. • Error analysis highlights the model's reliability and low prediction variance.