Incorporating uncertainty and reliability for battery temperature prediction using machine learning methods

Abstract

Temperature prediction of lithium-ion batteries is essential to prevent aging and degradation of batteries while ensuring safe and reliable operation. Use of simple machine learning methods for battery temperature prediction given current and voltage inputs is challenging due to battery cycling-induced aging. Since these methods do not generalize to the data outside the training domain, reliability issue arises in battery thermal management. As it is tough to train the battery temperature prediction models with all the possible data the battery is expected to see during real-world usage, uncertainty-aware models are needed. To address this, instead of point prediction, range prediction is performed using the conformal prediction method. The proposed method provides a band of temperature predictions for the input of current, voltage and operational time. Results show high prediction accuracy with seventy-nine percent of actual temperature measurements falling 100% within the narrow-predicted band. The remaining data points are seen within 0.4% of the predicted temperature bounds. The conformal method outperforms point prediction methods showing over 70% improvement in temperature prediction accuracy for pulsed and random walk battery cycling profiles, proving to be a precise and reliable battery temperature prediction tool under aging.