为解决我国陕北浅埋煤层开采过程中顶板含水层涌(突)水问题，基于“三图-双预测”理论，将机器学习算法应用到顶板含水层涌（突）水危险性预测中，以提升预测准确性，为陕北矿企的煤层顶板含水层水害防治工作提供有力依据，本文的主要研究内容和成果如下：

(1)通过分析浅埋煤层顶板含水层涌（突）水机理，得出煤层顶板含水层主要水源为风化基岩水，其发生涌（突）水的危险性主要由含水层风化基岩富水性及导水裂隙带发育高度共同决定，并分析风化基岩富水性和导水裂隙带发育高度的主要影响因素。

(2) 含水层风化基岩富水性预测模型的构建。首先基于方差过滤和交叉递归特征消除算法对含水层富水性影响因素进行特征选择，并利用XGBoost模型构造高级特征。然后基于stacking策略融合SoftMax和随机森林，构建含水层风化基岩富水性预测模型。最后使用钻孔实测数据集与其他模型进行对比实验，经过五折交叉验证，结果表明：融合模型的预测准确率提高了2.8%。

(3)导水裂隙带高度预测模型的构建。将导水裂隙带发育过程看成时间序列问题，使用一维卷积神经网络(CNN)对导水裂隙带数据进行特征提取，然后使用提取到的特征训练长短时记忆网络(LSTM)，同时利用导水裂隙带数据训练LightGBM模型，基于预测误差倒数调整LSTM和LightGBM模型预测结果的权重，从而得出导水裂隙带发育高度预测值。实验结果表明：相比于其他模型，本模型的平均绝对误差MAPE和均方根误差RMSE降低了0.41和0.0822，验证了本模型具有较高的准确性。

(4)模型应用。使用本文构建的预测模型，对相邻矿区中未进行抽水钻孔实验的区域，进行顶板含水层风化基岩富水性及导水裂隙带高度预测，得到富水性分区图及顶板冒裂（导水裂隙带）安全性分区图。基于ArcGIS软件叠加两图得到含水层风化基岩涌（突）水危险性分区图，为浅埋煤层开采防治水工作提供科学指导。

In order to solve the problem of gushing (sudden) water in the roof aquifer during the mining of shallow buried coal seams in northern Shaanxi Province, machine learning algorithms are applied to predict the risk of gushing (sudden) water in the roof aquifer based on the theory of "three maps - double prediction" to improve the prediction accuracy and provide a strong basis for the prevention and control of water hazards in the roof aquifer of coal seams in northern Shaanxi mining enterprises, the main research contents and results of this paper are as follows.

 (1) through the analysis of shallow buried coal seam roof aquifer gushing (sudden) water mechanism, it is concluded that the main water source of coal seam roof aquifer is weathered bedrock water, and its risk of gushing (sudden) water is mainly determined by the water-richness of the weathered bedrock of the aquifer and the height of water-conducting fissure zone development together, and analyze the main influencing factors of the water-richness of the weathered bedrock and the height of water-conducting fissure zone development.

 (2) Construction of water-richness prediction model for weathered bedrock of aquifers. Firstly, we select features based on variance filtering and cross-recursive feature elimination algorithm for aquifer water-richness influencing factors, and construct advanced features using XGBoost model. Then, the water-richness prediction model of the weathered bedrock of the aquifer is constructed by fusing SoftMax and random forest based on stacking strategy. Finally, a comparison experiment was conducted with other models using the borehole measured dataset, and after a five-fold cross-validation, the results showed that the prediction accuracy of the fusion model was improved by 2.8%.

(3) Construction of a hydraulic fracture zone height prediction model. The hydraulic fissure zone development process is viewed as a time series problem, and the features are extracted from the hydraulic fissure zone data using a one-dimensional convolutional neural network (CNN), and then the extracted features are used to train a long and short term memory network (LSTM), while the LightGBM model is trained using the hydraulic fissure zone data, and the prediction results of the LSTM and LightGBM models are adjusted based on the inverse of the prediction error The weights of the predicted results of the LSTM and LightGBM models are adjusted based on the inverse of the prediction error to obtain the predicted height of the hydraulic fracture zone development. The experimental results show that the Mean Absolute  Percentage Error (MAPE) and Root Mean Square Error (RMSE) of this model are reduced by 0.41 and 0.0822 compared with other models, which verifies the high accuracy of this model.

(4) Model application. Using the prediction model constructed in this paper, the water-richness of the weathered bedrock of the top slab and the height of the hydraulic fracture zone are predicted in the adjacent mine area where no pumping borehole experiments are conducted, and the water-richness zoning map and the safety zoning map of the top slab fracture (hydraulic fracture zone) are obtained. Based on ArcGIS software overlaying the two maps, we obtained the water risk zoning map of the weathered bedrock of the water-bearing layer, which provides scientific guidance for water control work in shallow buried coal seam mining.

[1] 苗彦平,蔚波,姬中奎,等.侏罗纪浅埋煤层开采典型水害模式及分区[J].煤矿安全,2021,52(08):75-82.

[2] 武强,许珂,张维.再论煤层顶板涌(突)水危险性预测评价的“三图-双预测法”[J].煤炭学报,2016,41(06):1341-1347.

[3] 武强,黄晓玲,董东林,等.评价煤层顶板涌(突)水条件的“三图-双预测法”[J].煤炭学报,2000(01):62-67.

[4] 任晓波,吴瑞芳.基于煤层顶板充水含水层水压的“修正三图-双预测法”研究[J].煤炭工程,2016,48(S2):131-133.

[5] 侯恩科,龙天文,刘庆利,等.基于ArcGIS的洛河组含水层富水性评价[J].煤田地质与勘探,2019,47(02):151-156.

[6] 王汉斌. 基于ArcGIS的首旺煤矿突水水害危险性评价[D].中国地质大学(北京),2017.

[7] 张玉军,宋业杰,樊振丽,等.鄂尔多斯盆地侏罗系煤田保水开采技术与应用[J].煤炭科学技术,2021,49(04):159-168.

[8] 侯恩科,纪卓辰,车晓阳,等.基于改进AHP和熵权法耦合的风化基岩富水性预测方法[J].煤炭学报,2019,44(10):3164-3173.

[9] 杨帆.张家峁井田二盘区风化基岩水文地质特征[J].煤炭技术,2021,40(09):64-66.

[10] 侯恩科,童仁剑,王苏健,等.陕北侏罗纪煤田风化基岩富水性Fisher模型预测方法[J].煤炭学报,2016,41(09):2312-2318.

[11] 曾一凡,李哲,宫厚建,等.顶板风化基岩含水层富水特征与涌(突)水危险性预测[J].煤炭工程,2018,50(02):100-104.

[12] 康红普,徐刚,王彪谋,等.我国煤炭开采与岩层控制技术发展40a及展望[J].采矿与岩层控制工程学报,2019,1(02):7-39.

[13] 郭文兵,赵高博,白二虎.煤矿高强度长壁开采覆岩破坏充分采动及其判据[J].煤炭学报,2020,45(11):3657-3666.

[14] 孙超,薄景山,孙有为.采空区沉陷研究历史及现状[J].防灾科技学院学报,2008,10(04):128-131.

[15] 张雁. 防止煤层顶板水溃入矿井的预警系统研究[D].煤炭科学研究总院,2008.

[16] 管祯. 煤矿采空区“三带”的探测与应用[D].中国矿业大学,2020.

[17] 崔永青,孙亮.马堡煤矿采动覆岩“三带”划分数值模拟研究[J].煤矿现代化,2019(02):92-94+97.

[18] 高延法.岩移“四带”模型与动态位移反分析[J].煤炭学报,1996(01):51-56.

[19] 左文喆,王斌海,程紫华,等.矿井涌水量预测方法综述[J].化工矿物与加工,2016,45(09):71-74.

[20] Orekhov V V, Khokhotva S N. HYDROGEOLOGICAL MODEL OF THE TERRITORY OF KOWSAR HYDRAULIC PROJECT[C]. Proceedings of Moscow State University of Civil Engineering/Vestnik MGSU, 2015 (3): 59-68.

[21] Gupta C P,Thangarajan M,Gurunadha R.Evolution of regional hydrogeologic setep of a hard rock aquifer through R-C analog model[J]. Ground Water,1985, 31(9): 154-157.

[22] 侯恩科,纪卓辰,车晓阳,等.基于改进AHP和熵权法耦合的风化基岩富水性预测方法[J].煤炭学报,2019,44(10):3164-3173.

[23] 刘伟涛,周华强.水体下开采导水裂隙带发育高度模拟与实测[J].煤炭技术,2016,35(10):206-208.

[24] 侯恩科,童仁剑,王苏健,等.陕北侏罗纪煤田风化基岩富水性Fisher模型预测方法[J].煤炭学报,2016,41(09):2312-2318.

[25] 彭睿娥.Bayes判别法在厚松散层薄基岩条件下煤层顶板涌水等级预测中的应用[J].内蒙古煤炭经济,2019(08):81-104.

[26] 徐星,李喆,曾珠.煤层底板突水危险性AHP-Fisher判别分析模型[J].煤炭技术,2018,37(07):153-155.

[27] 董丽丽,费城,张翔,等.基于LSTM神经网络的煤矿突水预测[J].煤田地质与勘探,2019,47(02):137-143.

[28] 刘天泉.“三下一上”采煤技术的现状及展望[J].煤炭科学技术,1995(01):5-7+62.

[29] 侯恩科,张萌,孙学阳,等.浅埋煤层开采覆岩破坏与导水裂隙带发育高度研究[J].煤炭工程,2021,53(11):102-107.

[30] 车晓阳,侯恩科,孙学阳,等.沟谷区浅埋煤层覆岩破坏特征及地面裂缝发育规律[J].西安科技大学学报,2021,41(01):104-111+186.

[31] 侯恩科,冯栋,谢晓深,等.浅埋煤层沟道采动裂缝发育特征及治理方法[J].煤炭学报,2021,46(04):1297-1308.

[32] 马雄德,王苏健,蒋泽泉,等.神南矿区采煤导水裂隙带高度预测[J].西安科技大学学报,2016,36(05):664-668.

[33] 姚建明,夏斐.神府矿区柠条塔井田(南翼)水文地质条件及开采方法研究[J].中国煤田地质,2007(01):33-35+59.

[34] 胡小娟,李文平,曹丁涛,等.综采导水裂隙带多因素影响指标研究与高度预计[J].煤炭学报,2012,37(04):613-620.

[35] 李振华,许延春,李龙飞,等.基于BP神经网络的导水裂隙带高度预测[J].采矿与安全工程学报,2015,32(06):905-910.

[36] 张宏伟,朱志洁,霍丙杰,等.基于改进的FOA-SVM导水裂隙带高度预测研究[J].中国安全科学学报,2013,23(10):9-14.

[37] 刘学生,张明,宁建国,等.近浅埋煤层导水裂隙带发育高度影响因素的数值模拟[J].山东科技大学学报(自然科学版),2012,31(05):31-36.

[38] 杨艳国,王军,于永江.河下多煤层安全开采顺序对导水裂隙带高度的影响[J].煤炭学报,2015,40(S1):27-32.

[39] 闫朝波. 张家峁煤矿煤层顶板涌（突）水危险性分区预测研究[D].西安科技大学,2013.

[40] 姬中奎. 张家峁井田烧变岩与水库水力联系及帷幕截流技术研究[D].西安科技大学,2018.

[41] 高利军.顶板风化基岩含水层富水性评价[J].煤炭技术,2019,38(10):105-107.

[42] 曾一凡,李哲,宫厚建,等.顶板风化基岩含水层富水特征与涌(突)水危险性预测[J].煤炭工程,2018,50(02):100-104.

[43] 李东东. 红柳林煤矿2~(-2)煤顶板涌（突）水危险性评价[D].西安科技大学,2017.

[44] 程磊,罗辉,李辉,等.近年来煤矿采动覆岩导水裂隙带的发育高度的研究进展[J].科学技术与工程,2022,22(01):28-38.

[45] 侯恩科,张萌,孙学阳,等.浅埋煤层开采覆岩破坏与导水裂隙带发育高度研究[J].煤炭工程,2021,53(11):102-107.

[46] 王子升.浅埋厚煤层工作面导水裂隙带发育高度研究[J].煤炭工程,2021,53(09):66-70.

[47] 武强,许珂,张维.再论煤层顶板涌(突)水危险性预测评价的“三图-双预测法”[J].煤炭学报,2016,41(06):1341-1347.

[48] 武强,黄晓玲,董东林,等.评价煤层顶板涌(突)水条件的“三图-双预测法”[J].煤炭学报,2000(01):62-67.

[49] Wang S, Dai Y, Shen J, et al. Research on expansion and classification of imbalanced data based on SMOTE algorithm[J]. Scientific reports, 2021, 11(1): 1-11.

[50] Li S, Huang S, Zhou Y. Toxic behaviour detection based on improved SMOTE algorithm and bi-LSTM network[J]. International Journal of Intelligent Internet of Things Computing, 2020, 1(2): 114-128.

[51] Kumari R, Nigam A, Pushkar S. An efficient combination of quadruple biomarkers in binary classification using ensemble machine learning technique for early onset of Alzheimer disease[J]. Neural Computing and Applications, 2022: 1-20.

[52] Lin S, Zheng H, Han B, et al. Comparative performance of eight ensemble learning approaches for the development of models of slope stability prediction[J]. Acta Geotechnica, 2022: 1-26.

[53] Chen T, Guestrin C. Xgboost: A scalable tree boosting system[C]//Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. 2016: 785-794.

[54] Breiman L. Random forests[J]. Machine learning, 2001, 45(1): 5-32.

[55] 曹义亲,谢舒慧.基于网格搜索的特定类别图像去噪算法[J].计算机科学,2020,47(11):168-173.

[56] Pelikan M, Goldberg D E, Cantú-Paz E. BOA: The Bayesian optimization algorithm[C]//Proceedings of the genetic and evolutionary computation conference GECCO-99. 1999, 1: 525-532.

[57] Shin H C, Roth H R, Gao M, et al. Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning[J]. IEEE transactions on medical imaging, 2016, 35(5): 1285-1298.

[58] Selvin S, Vinayakumar R, Gopalakrishnan E A, et al. Stock price prediction using LSTM, RNN and CNN-sliding window model[C]//2017 international conference on advances in computing, communications and informatics (icacci). IEEE, 2017: 1643-1647.

[59] Yu Y, Si X, Hu C, et al. A review of recurrent neural networks: LSTM cells and network architectures[J]. Neural computation, 2019, 31(7): 1235-1270.

[60] Al Daoud E. Comparison between XGBoost, LightGBM and CatBoost using a home credit dataset[J]. International Journal of Computer and Information Engineering, 2019, 13(1): 6-10.

[61] Zhang J, Mucs D, Norinder U, et al. LightGBM: An effective and scalable algorithm for prediction of chemical toxicity–application to the Tox21 and mutagenicity data sets[J]. Journal of chemical information and modeling, 2019, 59(10): 4150-4158.