Objective The eastern and northern regions of China exhibit thick unconsolidated layers and thin bedrocks, with water-sand inrush disasters occurring frequently. Therefore, accurately predicting the risk of water-sand inrushes from coal seam roofs holds great significance for safe coal mining. However, the water-sand inrushes in these regions exhibit complex disaster-causing mechanisms, which involve the coupling effects of multiple factors. Accordingly, the real-time prediction of the water-sand inrush risk in the field poses challenges including high risk and cost. These issues lead to difficult data acquisition and severely insufficient samples, limiting the accuracy and performance of traditional prediction models. Therefore, there is an urgent need to explore effective prediction methods suitable for a small sample size.

Methods Based on a review and analysis of the measured field data and historical cases of mining faces adjacent to unconsolidated layers, this study determined 11 factors influencing water-sand inrushes (e.g., the thickness of aquifers at the unconsolidated-layer bottom and bedrock thickness) and constructed the original sample dataset. Subsequently, the intrinsic relationships and correlations among various influencing factors were discovered using Spearman correlation. A risk prediction model for water-sand inrushes, termed CTGAN-DPSO-RF, was developed based on conditional tabular generative adversarial networks (CTGAN), detecting particle swarm optimization (DPSO) algorithm, and random forest (RF) algorithm. The quality of data synthesized using CTGAN was explored. Furthermore, the effectiveness of the CTGAN-DPSO-RF model was validated through comparison with the DPSO-SVM and DPSO-XGBoost models, along with two engineering cases.

Results and Conclusions Among the 11 influencing factors, the caving zone height exhibited the strongest correlation with mining height (correlation coefficient: 0.93), while the water pressure of aquifers at the unconsolidated-layer bottom presented the weakest correlation with the height of the hydraulically conductive fracture zone. The data synthesized using CTGAN highly resembled the original data, with a comprehensive quality score reaching up to 85.03%. The DPSO algorithm yielded an optimal fitness of 0.9265 after the hyperparameter tuning, outperforming the particle swarm optimization (PSO) algorithm. The CTGAN-DPSO-RF model yielded accuracy (A_c), weighted precision (P_w), weighted recall (R_w), and weighted F1-score (F1_w) consistently reaching 1.0 on the test set, outperforming its counterparts. The risk prediction results of two mining faces derived using the proposed model were consistent with actual mining conditions. By synthesizing high-quality data to expand the sample dataset and optimizing hyperparameters, the proposed model effectively overcomes the low accuracy and poor performance of traditional models under a small sample size, providing a new method for predicting the risk of water-sand inrushes from coal seam roofs under conditions of thick unconsolidated layers and thin bedrocks.