Mine ventilation network (MVN) regulation faces severe challenges: strong variable coupling, high search dimensionality, and the inherent conflict between energy conservation and safety constraints. To address these issues, we propose a novel airflow optimization framework integrating a Minimum Influence Spanning Tree (MIST), sensitivity attenuation boundaries, and an Improved Dung Beetle Optimizer (IDBO). Initially, high-influence co-tree chords are strategically extracted via MIST to compress the mathematical optimization dimensionality. Subsequently, effective ventilation resistance search intervals are bounded using sensitivity attenuation, preventing the algorithm from performing invalid searches in high-resistance regions. Furthermore, the standard DBO is enhanced via Fuchs chaotic initialization, Golden Sine and Lens Imaging collaborative learning, and differential mutation to minimize system power consumption. A 46-branch MVN case study validates the approach, identifying an 8-dimensional control combination as the absolute minimum requirement for full compliance. Compared to state-of-the-art baselines (DBO, SSA, WOA, DE), IDBO achieved the lowest power consumption. Post-optimization, the airflow constraint satisfaction rate improved from 89.13% to 100%, and total system power decreased by 11.87% (from 185.03 kW to 163.08 kW). Ultimately, this method robustly achieves Ventilation on Demand (VoD), providing a reliable computational tool for intelligent underground mining.

This study develops an interpretable point-and-interval prediction method for mine ventilation-parameter identification using limited field samples. Guided by ventilation-resistance theory, eight engineering attributes—roadway length, width, height, cross-sectional area, wetted perimeter, equivalent diameter, support type, and cross-sectional shape—were selected, and a prediction dataset was constructed from 260 field samples. The extended Random Feature Model (xRFM) establishes a nonlinear mapping, and the proposed Evolution-Path Guided Ivy Algorithm (EPG-IVY) optimizes hyperparameters within the training data. Prediction intervals are then constructed from out-of-fold residuals using Adaptive Bandwidth Kernel Density Estimation (ABKDE), while SHAP is used to analyze the model’s dependence on engineering features. With a fixed 7:3 split, EPG-IVY-xRFM achieved R² = 0.9754, RMSE = 0.001600, and MAE = 0.001162. Repeated nested five-fold cross-validation yielded R² = 0.9719 ± 0.0084, RMSE = 0.001675 ± 0.000267, and MAE = 0.001225 ± 0.000182, with lower RMSE and MAE than the optimized XGBoost and CatBoost baselines. External validation on 28 independent samples from Balasu Coal Mine yielded R² = 0.890, RMSE = 0.007072, and MAE = 0.003873. The empirical coverage of the 90% ABKDE prediction interval was 0.8846, and its Wilson 95% confidence interval included the nominal coverage level. SHAP results identified support type, roadway width, and equivalent diameter as the main contributing factors in the current model. These results provide a data-driven framework for friction resistance coefficient identification with point prediction, uncertainty estimation, and feature attribution under limited field samples.