An interpretable predictive model for shale gas hazard risk in tunnel excavation based on an improved broad learning system

In complex geological environments, shale gas hazards occur frequently during tunnel construction. The “self-generated and self-contained” nature of its accumulation characteristics makes traditional harmful gas prediction models unsuitable. Therefore, the development of a reliable and high-precision shale gas tunnel risk prediction model is of paramount importance. This study conducts sampling experiments on existing tunnel projects in the northwestern region of Hunan province, China, selecting five key feature parameters: organic carbon abundance, thermal maturity, burial depth, permeability, and porosity. Subsequently, a Broad Learning System (BLS) is utilized as the base classifier, with a residual connection strategy incorporated to construct an improved Broad Learning System (IBLS), significantly enhancing the model’s stability and its ability to capture non-linear relationships. In performance testing, the model achieved an accuracy of 91.67% on an independent test set, indicating good predictive performance under the compiled dataset. Further validation was conducted using a case study from the SGL tunnel project, where the predicted risk levels were consistent with the observed risk levels. Finally, SHapley Additive Explanations (SHAP) were applied to interpret the constructed model, analyzing the marginal contributions of various features to the model’s output across different shale gas risk levels. The results indicate that the proposed IBLS model provides an interpretable auxiliary tool with practical potential for preliminary shale gas tunnel risk prediction.