Tolerated outlier prediction method of excavation damaged zone thickness of drift based on interpretable SOA-QRF ensemble learning

Drift excavation induces excavation damaged zones (EDZ) due to stress redistribution, impacting drift stability and rock deformation support. Predicting EDZ thickness is crucial, but traditional machine learning models are susceptible to potential outliers in dataset. Directly eliminating outliers, however, impacts training effectiveness. This study introduces an EDZ thickness prediction model utilising quantile loss and random forest (RF) optimised by the seagull optimisation algorithm (SOA), enabling median regression with tolerated outlier performance. 209 sets of data sets containing 34 mine borehole data were used to establish the prediction model. Evaluation using <i>R</i>², explained variance score (EVS), mean absolute error (MAE), and mean square error (MSE) demonstrates the superior accuracy of the proposed SOA-QRF model compared to traditional models. Based on the discussion on the treatment of outliers, the outcomes indicate that the SOA-QRF model is more suitable for the dataset with outliers as well as being able to effectuate tolerated outlier prediction. Additionally, three interpretation methods were utilised to explain the SOA-QRF model and enhance the transparency of the model’s prediction process and facilitating the analysis of dispatcher regulation.