为了提高综采工作面资源采出率，实现工作面的高产高效，越来越多的矿区选择发展超长工作面。而工作面倾向长度增加导致综采工作面在开采过程中顶板来压显著增强，来压步距明显缩短等现象，给超长工作面顶板管理与控制带来较大难题。因此，分析超长工作面顶板来压影响因素及特征，并借助深度学习算法实现顶板来压精准预测，对保障超长工作面安全高效开采具有重要意义。

本文以王庄煤矿3505超长综采工作面为工程背景，采用理论分析、数值模拟、现场实测、深度学习等方式，建立了预测模型及模型优化算法，实现了超长综采工作面的顶板来压预测，开发了超长工作面顶板来压预测系统，并以3505超长综采工作面的实例作为验证。主要研究成果包括：

（1）采用理论分析的方法，分析了超长工作面走向与倾向方向上的来压特征，发现在工作面走向方向，超长工作面的初次来压步距与周期来压步距相对普通工作面偏小，工作面倾向方向，工作面顶板挠度呈W型三峰值形状，工作面支架工作阻力表现为中部相对较低，两侧相对较高。分析了地质条件与开采工艺对超长工作面顶板来压的影响，通过灰色关联度理论选择了工作面长度、煤层倾角、埋深、顶板条件、直接顶厚度、基本顶厚度、推进速度、采高八个影响因素作为预测模型的输入参数。

（2）使用数值模拟的方法，分析在不同的工作面长度、不同推进速度、不同采高条件下覆岩应力及位移变化规律。发现在工作面走向方向上，最大支承应力与工作面长度、推进速度呈正相关关系，与采高呈负相关关系。最大垂直位移与工作面长度、采高呈正相关关系，与推进速度呈负相关关系。工作面长度增加，初次来压步距、周期来压步距降低。推进速度越快，最大支承应力峰值发生位置更加靠近煤壁；在倾向方向上，工作面长度大于300m时，垂向应力呈现为工作面中部较小，工作面两侧较大，与倾向长度小于300m的工作面相反。不同工作面长度的最大垂直位移都发生在工作面的中部位置。

（3）在收集到的样本数据中，以9：1划分训练集与测试集，建立了以深度神经网络为共享层，支持向量回归为特定任务层的多任务学习模型。选择了单任务模型BP神经网络及SVR模型作为对比试验，以MSE与R²为评价指标。通过多次验证分析发现：多任务学习预测模型对初次来压强度、初次来压步距、周期来压强度、周期来压步距的MSE值分别为0.1101、0.1221、0.1205、0.1223，R²分别为0.9126、0.9066、0.9091、0.9036，相比于两个单任务模型而言MSE值最小而R²最大。表明多任务学习模型在顶板来压的预测中优于单任务模型。

（4）使用核主成分分析法，将灰色关联度确定的8个输入因素优化为5个模型输入参数，使用遗传算法对深度神经网络进行优化，选择海鸥寻优算法对支持向量回归做超参数寻优。使用优化后的多任务学习预测模型对初次来压强度、初次来压步距、周期来压强度、周期来压步距预测，其结果MSE值分别为0.0219、0.0221、0.0201、0.0191，R²分别为0.9896、0.9835、0.9901、0.9843。相比未优化的多任务学习模型，预测精度及拟合度都有显著提升。

（5）基于优化后的多任务学习模型，使用Django框架与Vue框架，建立了超长工作面顶板来压预测系统，预测系统包括注册登录模块、预测模块及用户管理模块。以王庄煤矿3505超长综采工作面的实测数据验证超长工作面顶板来压预测系统。结果表明：初次来压强度、初次来压步距、周期来压强度、周期来压步距的均方误差分别为2.8%、3.0%、9.9%、7.2%。预测模型误差均小于10%。因此，优化后的多任务学习模型预测精准，模型在未训练过的数据上表现良好，模型可移植性较强。

In order to improve the resource recovery rate of fully mechanized mining face and realize the high yield and high efficiency of working face, more and more mining areas choose to develop super long working face. The increase of the inclined length of the working face leads to the phenomenon that the roof pressure of the fully mechanized mining face is significantly enhanced and the pressure step is significantly shortened during the mining process, which brings great difficulties to the roof management and control of the super-long working face. Therefore, it is of great significance to analyze the influencing factors and characteristics of roof pressure in super-long working face, and to realize the accurate prediction of roof pressure with the help of deep learning algorithm, so as to ensure the safe and efficient mining of super-long working face.

Based on the engineering background of 3505 super-long fully-mechanized mining face in Wangzhuang Coal Mine, this paper establishes a prediction model and a model optimization algorithm by means of theoretical analysis, numerical simulation, field measurement and deep learning, and realizes the prediction of roof weighting in super-long fully-mechanized mining face. The roof weighting prediction system of super-long working face is developed and verified by the example of 3505 super-long fully-mechanized mining face. The main research results include :

(1)By using the method of theoretical analysis, the characteristics of the pressure in the strike direction and the tendency direction of the super-long working face are analyzed. It is found that in the strike direction of the working face, the initial pressure step distance and the periodic pressure step distance of the super-long working face are smaller than those of the ordinary working face. In the tendency direction of the working face, the roof deflection of the working face is W-shaped three-peak shape, and the working resistance of the working face support is relatively low in the middle and relatively high on both sides. The influence of geological conditions and mining technology on the roof pressure of super-long working face is analyzed. Using the grey correlation theory, eight influencing factors of working face length, coal seam dip angle, buried depth, roof condition, direct roof thickness, basic roof thickness, advancing speed and mining height are selected as the input parameters of the prediction model.

(2)Using the method of numerical simulation, the variation law of stress and displacement of overburden rock under different working face length, different advancing speed and different mining height is analyzed. It is found that in the strike direction of the working face, the maximum bearing stress is positively correlated with the length of the working face and the advancing speed, and negatively correlated with the mining height. The maximum vertical displacement is positively correlated with the length of the working face and the mining height, and negatively correlated with the advancing speed. With the increase of working face length, the first weighting interval decreases. The faster the advancing speed, the peak value of the maximum bearing stress occurs closer to the coal wall ; in the tendency direction, when the length of the working face is more than 300 m, the vertical stress is smaller in the middle of the working face and larger on both sides of the working face, which is opposite to the working face with a tendency length of less than 300 m. The maximum vertical displacement of different working face lengths occurs in the middle of the working face. With the increase of propulsion speed, the maximum vertical stress increases. The vertical stress contour shows an approximate bimodal shape. With the increase of mining height, the maximum bearing stress decreases and the influence range of the maximum bearing stress increases. The vertical stress distribution field of goaf in working face with different mining heights shows the phenomenon of low abutment pressure in the middle and high abutment pressure on both sides of the middle. The maximum vertical displacement is positively correlated with the mining height.

(3)In the collected sample data, the training set and the test set are divided by 9:1, and a multi-task learning model with deep neural network as the sharing layer and support vector regression as the specific task layer is established. The single task model BP neural network and SVR model are selected as the comparison test, and MSE and R² are used as the evaluation indexes. Through multiple verification analysis, it is found that the MSE values of the multi-task learning prediction model for the initial weighting intensity, the initial weighting step, the periodic weighting intensity, and the periodic weighting step are 0.1101,0.1221,0.1205, and 0.1223, respectively. The R2 is 0.9126, 0.9066, 0.9091, and 0.9036, respectively. Compared with the two single-task models, the MSE value is the smallest and the R² is the largest. It is shown that the multi-task learning model is superior to the single-task model in the prediction of roof pressure.

(4)The kernel principal component analysis method is used to optimize the eight input factors determined by the grey correlation degree into five model input parameters. The genetic algorithm is used to optimize the deep neural network, and the seagull optimization algorithm is selected to optimize the support vector regression. The optimized multi-task learning prediction model is used to predict the initial pressure strength, the initial pressure step, the periodic pressure strength, and the periodic pressure step. The MSE values are 0.0219,0.0221,0.0201, and 0.0191, respectively, and the R² is 0.9896,0.9835,0.9901, and 0.9843, respectively. Compared with the unoptimized multi-task learning model, the prediction accuracy and fitting degree are significantly improved.

(5)Based on the optimized multi-task learning model, the Django framework and the Vue framework are used to establish a roof weighting prediction system for the super-long working face. The prediction system includes a registration login module, a prediction module, and a user management module. The roof weighting prediction system of super-long working face is verified by the measured data of 3505 super-long fully mechanized working face in Wangzhuang Coal Mine. The results show that the mean square error of the initial pressure, the strength of the periodic pressure and the step distance are 2.8 %, 3.0 %, 9.9 % and 7.2 %, respectively. The error of the prediction model is less than 10 %. Therefore, the optimized multi-task learning model predicts accurately, the model performs well on untrained data, and the model has strong portability.

[1] 王双明,申艳军,宋世杰,刘浪,顾霖骏,魏江波.“双碳”目标下煤炭能源地位变化与绿色低碳开发[J].煤炭学报,2023,48(07):2599-2612.

[2] 国家统计局. 中华人民共和国2023年国民经济和社会发展统计公报[1][N]. 人民日报,2024-03-01(010).

[3] 王国法,张金虎,徐亚军,等.深井厚煤层长工作面支护应力特性及分区协同控制技术[J].煤炭学报,2021,46(03)

[4] 毛兴文. 哈拉沟煤矿450m长壁工作面矿压规律研究[D].西安科技大学,2017.

[5] 钱鸣高,缪协兴.采场上覆岩层结构的形态与受力分析[J].岩石力学与工程学报, 1995 (02):97-106.

[6] 缪协兴,钱鸣高.采场围岩整体结构与砌体梁力学模型[J].矿山压力与顶板管理,1995 (Z1):3-12+197.

[7] 宋振骐.实用矿山压力控制[M].徐州：中国矿业大学出版社，1988.

[8] 缪协兴,钱鸣高.超长综放工作面覆岩关键层破断特征及对采场矿压的影响[J].岩石力学与工程学报,2003,(01):45-47.

[9] 曹胜根,缪协兴.超长综放工作面采场矿山压力控制[J].煤炭学报,2001,(06):621-625

[10] 陈忠辉,谢和平,李全生.长壁工作面采场围岩铰接薄板组力学模型研究[J].煤炭学报,2005,(02):172-176.

[11] 刘畅,刘正和,张俊文等.工作面长度对覆岩空间结构演化及大采高采场矿压规律的影响[J].岩土力学,2018,39(02):691-698.

[12] 付玉平,宋选民,邢平伟.浅埋煤层大采高超长工作面垮落带高度的研究[J].采矿与安全工程学报,2010,27(02):190-194.

[13] 崔树江.超长综放工作面顶板破断规律研究[J].煤炭工程,2010,(02):32-34.

[14] 王家臣, 杨胜利, 杨宝贵,李杨, 王兆会, 杨毅,马焱遥.深井超长工作面基本顶分区破断模型与支架阻力分布特征[J].煤炭学报,2019,44(01):54-63.

[15] 范志忠,付书俊,潘黎明.深部超长孤岛工作面覆岩垮落结构特征研究[J].煤炭工程,2020,52(02):86-90.

[16] 宋选民, 顾铁凤,闫志海.浅埋煤层大采高工作面长度增加对矿压显现的影响规律研究[J].岩石力学与工程学报,2007,(S2):4007-4012.

[17] 邢平伟,宋选民,付玉平,李志军.神东大采高超长工作面矿压显现强度预测研究[J].中国煤炭,2009,35(08):43-47.

[18] 王生彪,黄庆享.浅埋煤层400m超长综采工作面矿压规律研究[J].煤炭科学技术,2018,46(S1):75-80.

[19] 王庆雄,鞠金峰.450m超长综采工作面矿压显现规律研究[J].煤炭科学技术,2014,42(03):125-128.

[20] 张飞,范文胜,孙建岭,徐丽,芭蕾.超长综放工作面的矿压显现规律研究[J].现代矿业,2010,26(05):63-66.

[21] 张智,李金刚.浅埋煤层超长工作面矿压规律研究[J].煤炭科学技术,2021,49(S2):30-33.

[22] 丁国利,鲁喜辉,武少国,刘永强.深埋超长综采工作面矿压规律及支架适应性研究[J].煤炭科学技术,2021,49(03):43-48.

[23] 杨征,李明忠,刘前进,李安宁,薛晨晓.450m超长工作面矿压显现与地表移动变形规律研究[J].煤炭工程,2023,55(02):63-68.

[24] 刘会利,郭彦科.浅埋中厚煤层超长工作面末采矿压规律研究[J].煤炭工程,2021,53(03):6-10.

[25] 王兆会,唐岳松,李辉,杨印朝,李家龙,王志峰.千米深井超长工作面支架阻力分布特征及影响因素研究[J].采矿与安全工程学报,2023,40(01):1-10.

[26] 赵曦,王涛.三元煤矿超长综放工作面矿压显现规律研究[J].煤炭技术,2020,39(02):39-42.

[27] 刘前进,李安宁,薛晨晓等.中厚煤层千万吨级超长工作面矿压显现与覆岩活动规律研究[J].煤矿安全,2023,54(09):147-155.

[28] 蔺星宇,徐刚,高晓进,张震,刘前进.超长工作面支架工作阻力分布及分区增阻特性研究[J].煤炭科学技术,2023,51(04):11-20.

[29] 杨永康,李建胜,康天合,季春旭.浅埋厚基岩松软顶板综放采场矿压特征工作面长度效应[J].岩土工程学报,2012,34(04):709-716.

[30] 柴敬,高登彦,王国旺,崔亚仲.厚基岩浅埋大采高加长工作面矿压规律研究[J].采矿与安全工程学报,2009,26(04):437-440. .

[31] 黄庆享,黄克军,赵萌烨.浅埋煤层群大采高采场初次来压顶板结构及支架载荷研究[J].采矿与安全工程学报,2018,35(05):940-944.

[32] Tingjiang Tan,Zhen Yang,Feng Chang,Ke Zhao. Prediction of the First Weighting from the Working Face Roof in a Coal Mine Based on a GA-BP Neural Network[J]. Applied Sciences,2019,9(19).

[33] Dong Jianjun et al. Multiple regression method for working face mining pressure prediction based on hydraulic support monitoring dataset[J]. Frontiers in Earth Science, 2023,

[34] Ye Li and Xiaohu Shi. Mine Pressure Prediction Study Based on Fuzzy Cognitive Maps[J]. International Journal of Computational Intelligence and Applications, 2020, 19(03) : 13.

[35] 冯夏庭,王泳嘉,姚建国.煤矿顶板矿压显现实时预报的自适应神经网络方法[J].煤炭学报,1995(05):455-460.

[36] 吴洪词.基于神经网络的采场底板分类与顶板来压预报[J].贵州工学院学报,1996 (04):32-36.

[37] 张丽华,蔡美峰.顶板来压识别与预测的复合小波神经网络方法[J].煤炭科学技术, 2003(07):41-43.

[38] 赵毅鑫,杨志良,马斌杰,等.基于深度学习的大采高工作面矿压预测分析及模型泛化[J].煤炭学报,2020,45(01):54-65.

[39] 程敬义,万志军,PENG Syd S,等.基于海量矿压监测数据的采场支架与顶板状态智能感知技术[J].煤炭学报,2020,45(06):2090-2103.

[40] 左凌云.基于支架工作阻力大数据的工作面区域矿压分析研究[J].煤炭工程,2019, 51(11):60-64.

[41] 巩师鑫,任怀伟,杜毅博,等.基于MRDA-FLPEM集成算法的综采工作面矿压迁移预测[J/OL].煤炭学报：1-11[2021-01-18].

[42] 葛君山.矿压显现观测数据的时间序列分析[J].工矿自动化,2009,35(10):43-45.

[43] 杨硕. 基于PSO-BP神经网络的浅埋煤层工作面顶板矿压预测研究[D].西安科技大学,2010.

[44] 余琼芳,牛冬阳.基于LSTM网络的矿山压力时空混合预测[J/OL].电子科技:1-7[2023-03-05].

[45] 冀汶莉,刘艺欣,柴敬,王斌.基于随机森林的矿压预测方法[J].采矿与岩层控制工程学报,2021,3(03):71-81.

[46] 柴敬,刘义龙,王安义,屈世甲,欧阳一博.基于GRU和XGBoost的矿压显现规律预测[J].工矿自动化,2022,48(01):91-97.

[47] 来兴平,万培烽,单鹏飞等.基于免疫粒子群混合算法优化BP网络的矿压预测方法[J].西安科技大学学报,2023,43(01):1-8.

[48] 程海星,朱磊,宋立平,刘文涛,徐凯.基于逆向传播神经网络的工作面顶板矿压数据预测[J].煤矿安全,2021,52(05):216-220.

[49] 李慧民,李振雷,何荣军,闫玉彪.基于粒子群算法和BP神经网络的冲击危险性评估[J].采矿与安全工程学报,2014,31(02):203-207+231.

[50] 柴华彬,张俊鹏,严超.基于GA-SVR的采动覆岩导水裂隙带高度预测[J].采矿与安全工程学报,2018,35(02):359-365.

[51] 杨胜利. 基于中厚板理论的坚硬厚顶板破断致灾机制与控制研究[D].中国矿业大学,2019.

[52] 张涛. 采空区层状顶板破损力学特性及控制研究[D].中南大学,2024.

[53] 徐亚军,王国法.液压支架群组支护原理与承载特性[J].岩石力学与工程学报,2017,36(S1):3367-3373.

[54] 吴梵,朱锡,梅志远. 船舶结构力学[M]. 北京:国防工业出版社,2010.

[55] 陈铁云，陈伯真. 船船结构力学[M]. 上海:上海交通大学出版社，1991.

[56] 徐亚军.液压支架顶梁外载作用位置理论研究与应用[J].煤炭科学技术，2015，43(7)：102-106

[57] 赵毅鑫,刘文超,张村,等.近距离煤层蹬空开采围岩应力及裂隙演化规律[J].煤炭学报,2022,47(01):259-273.

[58] 张杰,王斌,白文勇,等.浅埋近距间隔式采空区顶板“双拱桥”结构稳定性研究[J].中国矿业大学学报,2021,50(03):598-605.

[59] 郭纬宇. 巴矿坚硬顶板矿压显现特征及控制研究[D].辽宁工程技术大学,2020.

[60] 常峰.基于GA-BP神经网络的工作面顶板矿压预测模型应用研究[D].中国矿业大学,2019.

[61] 杜旭峰.基于PSO-SVR的浅埋煤层顶板来压规律预测模型研究[D].西安科技大学,2022.

[62] 梁东民,池小楼.工作面推进速度对顶板覆岩活动的影响[J].煤矿安全，2018，49(09):276-279.

[63] 胡进,郜传厚.基于多任务学习的高炉热状态预测[J].中国冶金,2023,33(07):81-90.

[64] 张祥宇. 基于多时段组参模型的PM2.5预测方法研究[D].北京工业大学,2021.

[65] 吴锡娟. 基于深度多任务学习的流域多断面水质预测方法研究[D].西北师范大学,2023.

[66] 邵蓉波,肖立志,廖广志,史燕青,周军,李国军,侯学理.基于多任务学习的测井储层参数预测方法[J].地球物理学报,2022,65(05):1883-1895.

[67] 李沂光,李风铃,宁劲松,卢立娜,段元慧,谷文艳.基于BP神经网络的虾蛄捕捞海域溯源方法[J].渔业现代化,2017,44(05):39-44.

[68] 董建伟. 基于深度学习的船舶主机排烟温度基线模型与预测模型研究[D].大连海事大学,2022.

[69] 侯慧,曾金媛,陈国炎,等.配电开关设备及其控制器可靠性研究综述[J].武汉大学学报(工学版),2018,51(07):627-633.

[70] 闫国华,朱永生.支持向量机回归的参数选择方法[J].计算机工程,2009,35(14):218-220.

[71] 王鑫,吴际,刘超,等.基于LSTM循环神经网络的故障时间序列预测[J].北京航空航天大学学报,2018,44(04):772-784.

[72] 陈帅,黄腾,高大龙.改进BP神经网络模型在索塔变形预测中的应用[J].地理空间信息,2022,20(02):27-32.

[73] 孙仁科,皇甫志宇,陈虎,李仲年,许新征. 神经架构搜索综述[J]. 计算机应用:1-15[2024-04-06]

[74] 杨赟,张丽丽.基于ISOA-SVR模型的短期网络舆情预测[J].计算机工程与设计,2024,45(01):168-176.

[75] 陈杰,任金武,辛亚军.超长综采工作面矿压显现规律研究[J].煤炭技术,2022,41(10):10-14.

[76] 陆娇. 新疆师范大学校园网络用户管理系统的分析与设计[D].吉林大学,2011