Mine roof water inflow prediction model for the initial mining disturbance stage based on simulation-driven feature selection and ensemble learning

• Developed W-CGANs-DLBA-XGBoost model for accurate roof water inflow prediction. • Identified key controlling factors through coupled PFC-FiPy simulations. • Enhanced small-sample learning via Wasserstein GAN-based data augmentation. • Optimized parameters using deep reinforcement learning bee algorithm. • Achieved 3.8% prediction error, outperforming traditional methods by 11%—validated model in Wanfu Coal Mine with strong engineering applicability. Water inrush from the roof of coal seams beneath thick unconsolidated layers presents a significant challenge in mine water hazard prevention due to its sudden onset and complex mechanisms. To improve prediction accuracy under static conditions, this study proposes an intelligent water inflow prediction model—W-CGANs–DLBA–XGBoost—by integrating simulation-assisted feature selection and generative data augmentation. Based on coupled Particle Flow Code (PFC) – Finite Volume Partial Differential Equation Python (FiPy) simulations, key controlling factors such as development height of the water-conducting fracture zone, ratio of unconsolidated layer thickness to bedrock, and ratio of burial depth to unconsolidated layer thickness are identified. These, along with conventional geological and hydrogeological variables, form the model input without embedding explicit physical drivers. A Wasserstein Conditional Generative Adversarial Network (W-CGANs) expands the dataset to enhance diversity and distribution coverage. The Deep Q Network-based Bee Algorithm (DLBA) optimizes hyperparameters, while Extreme Gradient Boosting (XGBoost) performs final regression prediction. Application to a thick-unconsolidated-layer mining site in the Wanfu coalfield demonstrates the model’s effectiveness: during the initial mining disturbance phase, it predicts an average water inflow of 314.95 m³/h, with only 3.8% relative error compared to the measured 327.4 m³/h—far better than the 14.8% error of the traditional large-well method. Spatial prediction results closely match actual inflow patterns, successfully identifying high-risk zones. The model exhibits high numerical precision and spatial generalization ability, offering strong physical interpretability and practical value. This provides a reliable approach for early warning and zonal management of water inrush under thick unconsolidated geological conditions.