Thermal runaway of lithium-ion batteries represents a critical safety issue that requires urgent resolution, wherein early identification and warning are considered essential for preventing severe consequences. An early identification and explosion prediction algorithm based on Raman spectroscopic gas analysis was proposed. The method integrated partial least squares regression (PLSR) with a convolutional neural network (CNN), and an attention mechanism was introduced to enhance the focus on and interpretation of key features, thereby achieving accurate detection of characteristic gases associated with thermal runaway, such as C₂H₄, CO, and CH₄. By combining thermal runaway dynamics with the Raman-CNN gas volume fraction prediction algorithm, a model for predicting the time to explosion was constructed. Validation results indicate that the mean absolute error of the model’s predicted explosion time is 24 seconds. It is expected that this warning mechanism can effectively mitigate personnel injuries and property losses resulting from lithium battery thermal runaway.