为了预防采空区安全事故的发生，提高采空区稳定性预测的准确性，选取了影响采空区稳定性的11个主要因素作为特征值，建立了以Random Frost、Adaboost、ExtraTrees和LightGBM为初级学习器，Logistic-Regression为次级学习器的Stacking模型，对采空区稳定性进行预测。将从实际矿山获取的 60 组数据拆分成训练集与测试集对Stacking 模型进行训练学习，同时对比单模型与 Stacking 模型的预测结果。使用F1值与AUC值对模型进行评价，Stacking模型F1值为0.967，AUC值为0.97，远高于Random Frost等传统单一机器学习模型。对于未放入模型训练的测试集数据，Stacking模型的预测准确度也优于Random Frost等传统单一机器学习模型。结果表明：Stacking 模型相比单模型的机器学习方法能够更加精准有效地预测采空区的稳定性等级。

In the process of mining，the mining technology mainly based on the open field method will leave a large number of goaves after mining，which has great safety risks.Therefore，the prediction of the stability of the goaf is particularly important.In recent years，with the advent of the era of big data，machine learning and deep learning technology has been introduced into the mining industry.More and more scholars begin to use the prediction method of machine learning to study the stability of goaf.It is found that every single machine term model has its scope of use and scene，when facing a new goaf problem，we can not immediately give the most effective model to adapt to it.At this time，the stacking method derived from integrated learning is born.This article is mainly divided into three parts.The first part mainly introduces the Stacking method，explains the main steps of the method and the choice of the learner.Stacking is generally a two-layer structure.The first layer is composed of various elementary learners.Input the original data set to train them.The second layer uses the output of these primary learners as a new data set and new features into the secondary learner （meta-learner） for training，to complete the stack fusion of the model.The Stacking model requires that the correlation between the primary learners should be as small as possible，and the performance difference between the primary learners should not be too large.The second-level learner should choose a simpler learner.The second part is the construction of Stacking model.This paper selected eleven factors affecting the stability of mined-out area，they are the rock mass structure （X ₁），geological structure（X ₂），groundwater （X ₃），goaf engineering layout （X ₄），mining perturbation（X ₅），adjacent cavity （X ₆），the volume of the goaf （X ₇），the exposed area of goaf roof （X ₈），buried depth （X ₉），goaf span mined-out area ratio （X ₁₀），rock compressive strength（X ₁₁）.By using the eleven dimensional data and the classification of goaf stability as training data，we put them into the stacking model with random frost，AdaBoost，extratrees and lightgbm as primary learners and logistic regression as secondary learners for training.Another part of the training data was used as test data to compare the performance of stacking model and single machine learning model.The experimental results show that the F1 and AUC of stacking model are 0.967 and 0.97 respectively.Far higher than all other single machine learning models.The stacking model is better than other single machine learning models and show stronger generalization ability.Therefore，as the third part，it is concluded that the stacking model can predict the stability of goaf better than the single machine learning model.

http://www.goldsci.ac.cn/article/2020/1005-2518/1005-2518-2020-28-6-894.shtml