Abstract:

To address the problems of low accuracy and insufficient adaptability in existing methods for determining the parameters of PIM for predicting surface deformation prediction in goaf areas under thick unconsolidated layers, 36 sets of measured surface movement data from coal mining working faces were selected. The core indicators of mining-geological conditions were screened via Hierarchical Cluster Analysis (HCA), Entropy Weight Method(EWM) and Grey Relational Degree (GRD) analysis. Furthermore, the GRNN model was optimized by integrating K-fold cross-validation with the neighborhood perturbation strategy of SAA, and an SAA-GRNN optimization model was constructed for PIM parameter determination. A case study was conducted using 45 sets of data from coal mining working faces with thick unconsolidated layers in the Jining area. The results show that: seven mining-geological condition indicators can be classified into three categories, and five core input indicators were identified screening, namely mining thickness M, coal seam dip angle α, mining depth H, strike mining degree D₃/H, and unconsolidated layer thickness h. The maximum root-mean-squared error (RMSE) of SAA-GRNN model is no more than 0.190 4, the maximum mean absolute error (MAE) is controlled within 0.133 9, the maximum mean absolute percentage error (MAPE) is 0.153 6, and the overall coefficient of determination (R²) is generally above 0.8. Under the same conditions, the prediction errors are greatly reduced compared with those obtained using Back Propagation (BP) neural network and the conventional GRNN model.

Key words: simulated annealing algorithm (SAA), generalized regression neural network (GRNN), thick unconsolidated layer, probability integral method (PIM), parameter solution