A data-driven method for predicting thermal runaway propagation of battery modules considering uncertain conditions

Thermal Runaway Propagation (TRP) of lithium-ion battery packs has serious hazards. However, the TRP prediction is challenging because of the substantial uncertainty and hard-to-acquire data. To solve this problem, a fuzzy system and multi-task CNN-LSTM method are proposed to predict TRP multiple steps ahead. The TRP dataset is constructed by 25 sets of experiments and 130 sets of simulations. The uncertain SoC, charging and discharging conditions, and thermal runaway (TR) trigger points are considered in both experiments and simulations. Then, the fuzzy system is introduced to reason about the TR probability of the battery and optimized by a sparrow search algorithm (SSA). A multi-task CNN-LSTM model is proposed to extract fuzzy and physical information by employing a convolutional neural network (CNN) and multiple long short-term memory (LSTM) neural networks, respectively, and output the temperature of multiple cells simultaneously. Finally, the models are evaluated in the simulation and experimental validation sets with different window lengths and time resolutions. The results show that the fuzzy information significantly improves the prediction accuracy of the method, with a coefficient of determination (R2) of 98.48% for the 3s prediction horizon and 97.27% for the 18s prediction horizon in the experimental validation set.