Performance evaluation and interpretation of hybrid models based on light gradient boosting machine to predict coal mining subsidence

Abstract Accurate prediction of coal mining subsidence (CMS) is pivotal for mining design, environmental preservation, safety assurance, and the formulation of preventive measures in mining areas. This study introduces an effective hybrid prediction model for coal mining subsidence based on light gradient boosting machine (LightGBM) and uses Shapley Additive Explanations (SHAP) method to reveal and explain the contribution mechanism and interaction of factors affecting mining subsidence. This study collected a dataset of 163 mining subsidence cases (covering 12 features such as coal seams, rock layers, and mining conditions) to model mining subsidence prediction, and evaluated the performance of the model through multiple performance indicators. The results indicate that the hybrid prediction model developed in this paper demonstrates remarkable performance on the test set, with HGS-LightGBM standing out, achieving the coefficient of determination ( R 2 ) of 0.9589, while the single LightGBM achieves an R 2 of 0.925. Finally, apply model interpretation techniques to analyze the impact of input features on mining subsidence, and explain the prediction principles and decision-making process of the model. The analysis reveals that the thickness of coal seam (m) is the most influential parameter for CMS. Furthermore, a targeted interaction analysis on key parameters was conducted to clarify the impact mechanisms of each influencing factor. In summary, the model established in this study has excellent performance and exhibits significant interpretability and transparency.