Comparison of models for predicting the floor failure depths of coal seams with moderate thicknesses and above under fully mechanized mining

Objective  The failure depth of a coal seam floor represents a critical parameter for assessing the risk of floor water inrushes. With the widespread application of fully mechanized mining with a large mining height, existing empirical equations suffer from insufficient accuracy when used to calculate the mining-induced floor failure depths of coal seams with moderate thicknesses and above.

Methods  Based on the compressive strength of coal seam floors (i.e., hard, moderately hard, and weak floors), 107 sets of data samples of the measured floor failure depths of coal seams with moderate thicknesses and above under fully mechanized mining were classified. Subsequently, models for predicting the floor failure depth were established while considering four factors: mining height, the length of the mining face along its dip direction, coal seam dip angle, and mining depth. The resulting four theoretical models consisted of three linear models (i.e., the quasi-classical empirical equation, the linear support vector regression (SVR) model, and the log-linear mixed model) and one nonlinear model (i.e., the backpropagation (BP) neural network model). Then, the reliability of the four models was compared and verified using two evaluation metrics, i.e., goodness of fit (R²) and mean absolute percentage error (E_MAP), as well as measured data from nine mines.

Results and Conclusions For coal seam floors with three lithologies, the R² values of the four models decreased in the order of the BP neural network model, the log-linear mixed model, the linear SVR model, and the quasi-classical empirical equation model sequentially, suggesting inferior goodness of fit of the latter two models. The log-linear mixed model, the linear SVR model, and the quasi-classical empirical equation model yielded E_MAP values greater than the ideal threshold of 20%. However, the three models showed remarkable improvements compared to the classical empirical equation, which is no longer suited to guiding the prediction of the floor failure depths of coal seams with moderate thicknesses and above under fully mechanized mining due to discrepancies between calculated values and actual data. In contrast, the BP neural network model yielded average relative errors of less than 10%, significantly outperforming other models. By systematically comparing the prediction accuracy of the four theoretical models, this study provides a valuable reference for selecting an appropriate method for calculating the floor failure depths of coal seams with moderate thicknesses and above.