Mine water inflow (inrush) source identification model based on the RFE-GA-LightGBM model

Accurate identification of mine water inflow sources is essential for water hazard prevention and emergency decision-making in underground coal mining. However, source discrimination is often difficult because different aquifers may exhibit similar hydrochemical characteristics, and mining-induced fractures can enhance hydraulic connection and water–rock interaction. In this study, the 5110 working face of Yushupo Coal Mine was selected as the engineering case, and a mine water source identification model based on recursive feature elimination and genetic algorithm-optimized Light Gradient Boosting Machine, namely RFE-GA-LightGBM, was developed. Piper trilinear diagrams were first used to analyze the hydrochemical types and ion composition characteristics of different aquifers. Field observations showed that the maximum measured height of the water-conducting fractured zone reached 178 m, indicating that roof sandstone aquifers may be hydraulically connected during mining. Recursive feature elimination was then employed to screen key hydrochemical indicators, and Ca 2+ , Na + , SO₄² – , HCO₃ – , and Cl – were selected as input variables for model training. The GA-LightGBM model was compared with default LightGBM, PSO-LightGBM, and Bayesian-LightGBM models to verify the effectiveness of feature selection and hyperparameter optimization. The results show that Cl⁻ is an auxiliary indicator for identifying Lower Shihezi Formation water, whereas SO₄² – is sensitive to coal-measure sandstone aquifers affected by pyrite oxidation and water–rock interaction. Compared with the other models, GA-LightGBM achieved the best overall classification performance, with an accuracy of 96.1%, precision of 89.2%, recall of 89.5%, F1 score of 90.2%, specificity of 98.42%, and AUC of 0.96. The prediction results indicate that the water inflow at the 5110 working face is mainly related to the Taiyuan Formation sandstone aquifer and the K2 aquifer of the Shanxi Formation. This study provides a data-driven and hydrogeochemically interpretable method for rapid mine water source identification under complex mining conditions.