Data-driven Model with Ensemble Learning Predicting Thermal Runaway of Real Working Condition Vehicles

Abstract

Battery failure is a big obstacle that should be tackled for new energy vehicles, and thermal runaway is one of the principal threats, causing vehicle fire and leading to casualties. So, it is urgent and vital to developing an algorithm that can predict if and when the thermal runaway will happen and then send alerts to passengers. Nevertheless, it is hard to make a precise prediction because the causing factors of thermal runaway are complicated and comprehensive, and it can not only be triggered from inside the power battery but also from the external force. We aim to make more accurate predictions as much as possible; thus, we construct a combined machine learning algorithm that is highly accurate and flexible to predict the probability of lithium battery thermal runaway that happens in real life. By considering voltage and temperature, abnormal current, single battery consistency, and overcharge risk factor separately, we build a stacked model consisting of five sub-models linked with grid-search chosen hyperparameters.