Data-driven early warning strategy for thermal runaway propagation in Lithium-ion battery modules with variable state of charge.

• A multi-mode and multi-task data-driven method is proposed to predict thermal runaway propagation in battery modules. • A thermal runaway propagation simulation model is developed to supplement the thermal runaway experimental data. • The proposed thermal propagation prediction model is validated in 18,650 batteries with different chemical compositions. • The proposed warning strategy is effective for different state of charge and different model window lengths. Thermal runaway (TR) propagation is triggered in a battery pack by abnormalities such as a cell fire or explosion, which leads to severe consequences. Predicting the TR propagation is challenging due to the complex, high non-linearity, and uncertain disturbances of TR. This paper establishes an electro-thermal coupling simulation model of TR propagation to supplement experimental data and public datasets for model training and verification. Then, a data-driven fusion model named Multi-Mode and Multi-Task Thermal Propagation Forecasting Neural Network (MMTPFNN) is established quantitative advance multi-step prediction of TR propagation in Li-ion battery modules, and a temperature-based TR propagation grading warning strategy is proposed. The TR propagation is mainly influenced by the thermal characteristics of surrounding batteries, and the temperature distribution in the entire battery module is of great significance to the prediction of TR propagation. Herein, the model is presented by using the thermal image and the discrete operating data of cells. Furthermore, because TR is a small probability event, obtaining the thermal image of the battery module requires additional system memory and computational resources. A switching strategy of the prediction model is established to improve the applicability of the model with the temperature threshold of 60 °C. When the battery is in a safe temperature range (below 60 °C), the long short-term memory (LSTM) model is run to predict the battery temperature. Once the battery temperature is detected above 60 °C, the thermal image is captured, and the MMTPFNN model is run to predict the TR propagation. In the validation section, different network structures are discussed, and different time resolutions and different window settings of the MMTPFNN are compared. Finally, the early warning strategy with three alert levels is introduced, and the effectiveness of the warning strategy with different window settings and initial SoCs is further discussed. [ABSTRACT FROM AUTHOR]