Machine learning-based fire early warning model and full-scale experimental validation for energy storage power stations using 314Ah LiFePO4 cells

Energy storage fires are often caused by thermal runaway of lithium batteries. Traditional single-parameter early warning technologies suffer from low accuracy and slow response, making them unsuitable for engineering needs. Therefore, research on rapid and reliable early warning of thermal runaway is of great value. This paper addresses the thermal runaway problem in lithium battery modules. Using a 314Ah lithium iron phosphate battery experimental platform, it collects full-process data on temperature, voltage, gas concentration, and expansion force, clarifying the criteria for determining each stage of thermal runaway: before the pressure relief valve opens, early warning can be achieved through changes in expansion force, characteristic gas concentration, and temperature; during the outbreak stage, the core determining features are a sudden drop in voltage to zero and a peak expansion force. By comparing six machine learning models, including decision trees and random forests, the excellent early warning performance of each model was verified, and the applicable conditions of different models were clarified. LightGBM performs best under single-cell conditions, while the decision tree model shows significant advantages under dual-cell conditions. This paper employs the technique for order preference by similarity to ideal solution (TOPSIS) weighting method to fuse multiple models and construct an integrated early warning model. The overall performance of this model is superior to that of any single sub-model: in the binary warning task, all evaluation metrics exceed 0.92 with a maximum area under the curve (AUC) of 0.9652, and in the multi-class stage-identification task, all metrics exceed 0.93 with a maximum macro-AUC of 0.9625. Combined with SHapley additive exPlanations (SHAP) analysis, it is clarified that expansion force, gas concentration, temperature, and voltage are the core early warning characteristics for different stages of thermal runaway. This study significantly improves the speed and accuracy of lithium battery thermal runaway early warning through machine learning algorithm optimization and model fusion, providing new theoretical basis and practical technical support for the safety control of energy storage batteries.