坚硬覆岩综放开采导水裂隙带高度机器学习预测研究OA

Accurately predicting the height of the water-flowing fractured zone(WFFZ)is of great significance for the rational design of waterproof coal pillars when mining under loose aquifers.To address the prediction of WFFZ height in fully mechanized caving mining under hard roof conditions,this study systematically collected 38 sets of field-measured data from 31 hard roof working faces nationwide and proposed a Voting ensemble machine learning prediction model tailored for small sample data.Four machine learning algorithms—k-Nearest Neighbors(KNN),Random Forest(RF),Gradient Boosting Decision Tree(GBDT),and XGBoost-were selected as base learners.Fea-ture engineering that incorporated empirical formulas as prior knowledge was developed;the Optuna framework was utilised for hyperparameter tuning;and a"soft voting"strategy was employed to construct the Voting ensemble mo-del.This approach effectively addresses the stability and accuracy issues inherent in prediction models dealing with small sample sizes and has been applied to on-site engineering practice.The research results indicate that the Voting ensemble model achieves high prediction precision,outperforming both the standard"Guide"formulas and single machine learning models,whilst also clarifying the model's error distribution range under varying degrees of geolo-gical complexity.Engineering applications demonstrate that the predicted values align well with field-measured data.These research findings provide a theoretical basis for the design of waterproof coal pillars in fully mechanized top-coal caving mining under hard roofs.