Comparative Analysis of Learning Approaches for Internal Temperature Estimation of Lithium-Ion Batteries

Accurate estimation of battery temperature is essential for ensuring the safety, reliability, and performance of lithium-ion energy storage systems. Traditional temperature sensors measure only surface temperatures and cannot capture the internal thermal state. Electrochemical impedance spectroscopy (EIS) offers a promising alternative, as impedance responses are highly sensitive to both temperature and electrochemical state. This study investigates neural network regression models for estimating temperature from EIS data combined with the state of charge (SoC). Several input representations were evaluated under identical architectures and training conditions: (i) full impedance spectra, (ii) an autoencoder-derived latent repre sentation of the full spectra, (iii) the imaginary component of the spectra, and (iv) a latent representation learned from the imaginary spectra (Imag-Latent + SoC). Results show that using the imaginary component with SoC achieved the best overall performance, with the lowest error and highest coefficient of deter mination(2 ≈ 0.93in5-foldvalidation). Theautoencoder-basedrepresentation of the full spectra performed worst due to information loss, while the full spec tra remained accurate but required longer training. The Imag-Latent + SoC configuration maintained high accuracy with reduced computational cost. Overall, the findings confirm that the imaginary impedance provides an efficient and accurate representation for data-driven temperature estimation, supporting the feasibility of EIS-based, sensorless temperature monitoring for battery management systems