Early detection of thermal runaway in lithium-ion batteries under extreme conditions using phase change materials

• Simulation of an externally heated lithium-ion battery. • Building a multiple multivariate time-series dataset. • Interpretation of the dataset using SHapley Additive exPlanations. • Early detection of thermal runaway in lithium-ion batteries with long short-term memory model. In the current study, a machine-learning model for the early detection of thermal runaway in lithium-ion batteries is improved. A nickel manganese cobalt battery is modeled and validated using a multi-scale multi-domain approach. Following model validation, the cell is coated with phase change material and thermal runaway is trigßgered by an external heat source. In the simulation phase, 144 thermal runaway data are obtained. The voltage, current, phase change material temperature, and battery temperature data are recorded in time-series. After the preparation of the data set, a long short-term memory model is built to predict the thermal runaway at an early stage. Once the prediction model is built, the trade-off relationship between the prediction performance of the model and the training time is investigated in more detail. As a result, it was found that the thermal runaway onset time could be predicted with an error of 5.33 s using the first 40 s of battery operation data in training and after 70 s of model evaluation. Increasing the training time to 120 s decreased the thermal runaway onset time prediction error to 2.67 s.