Abstract:
In order to solve the prediction problem of the roof settlement in the unsupported period of roadway excavation in the deep roadway engineering of coal mines, the support vector machine (SVM) tool of artificial intelligence was introduced. Combined with the sparrow search optimization algorithm (SSA), a prediction model of the roof settlement during its unsupported period based on SSA-SVM was proposed. Taking the displacement data during the roof’s unsupported period in the deep underground roadway excavation process of the Great Wall No.5 Mine in Inner Mongolia as a sample, six influencing factors including uniaxial compressive strength (UCS), rock integrity (RQD), ground stress, roadway width-span ratio, non-support roof time and artificial mining were selected. The comprehensive influence weights of the data were summarized through applicability, correlation and classification consistency evaluation. The accuracy of ten-fold cross-validation was used as the fitness function to train and test the SSA-SVM prediction model with different population numbers. The optimal parameter model of population number was selected by error correlation coefficient (RMSE, MAPE,
R2), ROC curve, AUC ± Std, running time and standard deviation rate
η. The model was applied to the 1902S return airway to predict the settlement of the roadway during the roof’s unsupported period, and compared with the actual mine pressure monitoring data of the roadway. The results show that the prediction performance of SSA-SVM model is better when the population number is 90. The RMSE of training samples is
0.0165, MAPE is 22.54%, and
R2 is
0.8295. The RMSE of the test sample is
0.0156, MAPE is 22.37%, and
R2 is
0.8490. The realism AUC reaches the maximum of
0.8467, and the standard deviation Std is the minimum of
0.0115. The shortest running time is
8.7239 s and the standard deviation rate is maintained at 0.12%. In the field application of the 1902S return airway, there is no large deviation between the predicted value and the actual value, which is maintained in the range of linear fitting
y = 0.90
x and
y = 1.10
x. Both the error correlation coefficient and AUC ± Std satisfy the requirements of test accuracy. The prediction effect of this model can provide an important guidance for subsequent support design and reinforcement of support measures.