Coalbed methane (CBM) development involves multiple interacting physical fields, and different coupling schemes can lead to distinctly different production behaviors. A thermo-hydro-mechanical model accounting for gas–water two-phase flow and matrix dynamic diffusion (TP-D-THM) is developed and validated, achieving an error rate below 10%. By embedding the numerically estimated reservoir physical parameters of the Qinshui Basin into the numerical model, multi-field couplings during CBM production, the evolution of physical parameters, and the depth-dependent effects on production characteristics were revealed. The main findings are as follows: The inhibitory effect of water on CBM recovery consistently exceeds the promoting effect of temperature. As burial depth expands, the inhibitory effect first diminishes, then intensifies, ranging from 19.73% to 28.41%, while the thermal promotion effect exhibits a monotonically increasing trend, fluctuating between 8.55% and 16.33% and stabilizing below 1000 m. Temperature and burial depth do not alter the trend in gas production rate. For equilibrium permeability, reproducing a decrease–increase–decrease rate pattern requires explicit inclusion of water and matrix-fracture mass exchange terms, which can explain why different scholars obtained varying gas production rate trends using the THM model. Matrix adsorption-induced strain is the primary control on permeability evolution, and temperature amplifies the magnitude of permeability change. The critical depth essentially reflects the statistical characteristics of reservoir petrophysical properties. A dimensionless critical depth criterion has been proposed, which comprehensively considers reservoir pressure, permeability, and a fractional coverage index. For burial depths ranging from 650 to 1350 m, the TP-D-THM model can be simplified to the gas-mechanical model accounts for matrix dynamic diffusion (D-HM) with an error below 5%, indicating that thermal and water effects nearly cancel each other.

To optimize the support performance of rock bolts in high-stress environments, this study employs the ANSYS (Version 2022 R2) finite element numerical simulation method to systematically investigate the influence of bolt geometrical parameters (rib spacing, rib height, and bolt diameter) on the stress state of the anchoring system. A bolt–resin–sleeve model was established to analyze Mises equivalent stress distribution and peaks under a 150 kN pull-out load. The simulation results indicate that a rib spacing of 36 mm effectively promotes the diffusion of pre-stress into deeper regions, with peak stress in the bolt rod and resin ring increasing by 34.42% and 61.64%, respectively, compared to a spacing of 12 mm. Further increase in rib spacing provides limited enhancement in peak stress. A rib height of 1.0 mm achieves optimal system performance without excessively compromising the interfacial stress level. Increasing the diameter to 22 mm raised peak stress in the bolt, sleeve, and resin by 14.19%, 30.48%, and 50.77%, respectively, compared to 18 mm, balancing load capacity and material use efficiently. The optimal parameter set (36 mm spacing, 1.0 mm height, and 22 mm diameter) was validated in a field trial in Zhongmacun Mine’s 3903 East Transportation Bottom Drainage Roadway. Monitoring recorded maximum roof subsidence of 102.9 mm, stabilizing within 25 days (daily deformation < 0.2 mm), confirming the excellent performance of the bolt support system with this parameter combination in high-stress roadways. This study provides a theoretical basis and engineering reference for the optimal design of high-performance rock bolts.

The height of water-conducting fracture zone (WCFZ) critically impacts coal mine safety and water hazard control. This study develops a Bayesian Optimization Algorithm-optimized Multilayer Perceptron (BOA-MLP) model to predict WCFZ height in three-soft coal seams in Henan Province. The Analytic Hierarchy Process-Entropy Weight Method integration identified mining height, face length, and burial depth as primary factors. Trained on 32 field-measured WCFZ samples, BOA intelligently optimized hyperparameters (hidden layer neurons, learning rate, batch size). The optimal configuration (64 neurons in the first hidden layer, 32 neurons in the second hidden layer, learning rate of 0.001, batch size of 32) yielded test-set RMSE = 1.98 m, MAE = 2.23 m, MAPE = 2.67%, R²= 0.973—outperforming manual MLP with 15.0–28.5% error reductions and a 6.06% R² improvement. Statistical analysis using the Wilcoxon signed-rank test confirmed these improvements are significant (p = 0.027) with large effect size (Cohen’s d = 0.83). Comprehensive uncertainty and sensitivity analyses demonstrated the model’s robustness, with 84.4% of predictions falling within 95% confidence intervals and burial depth identified as the most influential parameter. Field validation at Panel 15,030 in Yaoling Mine confirmed the accuracy of BOA-MLP: predicted 28.1 m vs. measured 27.3 m (2.9% error), surpassing empirical formulas (18.3%) and manual MLP (7.3%). These results demonstrate that integrating Bayesian optimization substantially enhances the stability, generalizability, and predictive accuracy of the MLP model. The proposed BOA-MLP approach thus offers a robust and reliable methodology for forecasting WCFZ heights in three-soft coal seams, with significant implications for mine water management and safety planning.