A Meta-learning Based Multimodal Neural Network for Multistep Ahead Battery Thermal Runaway Forecasting

An effective forecast method to trigger Thermal Runaway (TR) warning in an early stage is essential for monitoring battery safety. In this article, we propose a novel data-driven approach to perform multistep ahead forecast accurately for battery TR state at cell-level. We formulate this forecasting task as an imbalance data classification task and propose meta thermal runaway forecasting neural network (Meta-TRFNN) to solve it. Essentially, we exploit high-dimensional thermal images along with low-dimensional temperature and voltage data to capture a more representative thermal profile. Moreover, we adapt a meta-learning framework to handle the data deficiency problem. We evaluate Meta-TRFNN on simulated samples and also explore its applicability in the real world with real samples. Although this classification task is highly imbalanced, Meta-TRFNN is still proven effective with limited historical information. Our further comparison experiments not only demonstrate the forecasting ability of Meta-TRFNN, but also validate the benefit of involving high-dimensional thermal images and the efficacy of meta-learning framework.