Prediction of water-conducting fracture zone height under repeated mining in close-distance coal seam groups using a BP neural network

Abstract

Accurate prediction of the height of the water-conducting fracture zone (WCFZ) is essential for assessing hydraulic connection between mining-induced fractures and overlying aquifers in close-distance coal seam groups. Repeated extraction differs from isolated single-seam mining because later disturbance acts on an overburden that has already experienced separation, fracturing, and partial compaction. Cumulative mining thickness and interlayer spacing may therefore exert coupled and strongly nonlinear influences on fracture-zone development. In this multi-project engineering-data study, 30 anonymized cases compiled from close-distance coal-seam mining projects were used for model development, and 25 non-overlapping engineering cases not used in model development were retained for independent validation. Mining depth, cumulative mining thickness, interlayer spacing, overburden hardness coefficient, face length, and seam dip angle were used as inputs to a 6-9-1 BP neural network trained by damped least-squares Levenberg–Marquardt optimization. On the four-case hold-out set, the BP model achieved an MRE of 1.02%, an RMSE of 0.71 m, an MAE of 0.61 m, and an R2 of 0.9987. Repeated five-fold cross-validation yielded an MRE of 4.03% ± 0.93%, while independent validation produced an MRE of 2.67% and an R2 of 0.9834. Ablation, Garson analysis, and the M–D response surface consistently indicated that cumulative mining thickness controls the overall scale of WCFZ development, whereas interlayer spacing regulates repeated-mining superposition. The results establish a transparent framework integrating mechanism-informed variables, nonlinear prediction, model interpretation, stability evaluation, and uncertainty assessment.