目前各煤矿根据《煤矿安全规程》的明确规定都已安装水文监测系统，并且积累了大量的水文数据，但是存在数据分析能力不足，利用率低的问题。本文对矿井历史水文数据进行深入分析，建立水文参数的深度学习预测模型，为预防矿井水害发生提供技术支持，对确保煤矿安全生产具有重要意义。

    水文参数分析是为后期预测模型的建立提供依据。矿井监测的各水文参数具有明显的噪声多、非线性等特点，通过自相关图对其平稳性进行判定，自相关系数在很长的延迟期内一直为正，说明其具有典型的非平稳性。另外，通过皮尔逊相关系数和散点图矩阵对多序列水文参数的相关性进行分析，结果表明，不同参数传感器间存在不同程度的相关性，同一参数各监测点的数据也存在相关性。

    矿井水文数据呈现出的非平稳性特征导致传统的单序列预测方法难以达到有效的预测效果。为了提高预测精度，论文提出CEEMD_GRU 模型。首先通过 CEEMD 将水文数据分解为多个平稳的子分量；其次，通过 PACF 确定水文数据的滞后期数，从而确定输入神经元个数；然后通过 GRU 神经网络学习各分量的变化规律并进行预测；最后对各分量预测结果进行融合得到最终的预测值，并与其它五种神经网络模型基于两组数据进行对比实验，测试集均方根误差分别平均降低了36.38%、25.48%。

    单序列预测基于数据的自相关性进行建模，然而，矿下的各水文传感器在空间上也具有相关性，本文利用这种相关性建立GRU_Attention 模型。首先将多维序列输入到GRU 中，实现高层次特征学习；然后将 GRU 的输出作为注意力机制的输入，通过注意力机制挖掘多维输入与输出的关联关系，计算特征权值；最后，将 GRU 层输出与特征权值加权求和对各输入进行增强或削弱，并将得到的特征表达向量输入全连接层计算最终预测值。与其它五种神经网络进行对比实验，测试集均方根误差平均降低了 31.1%。

    论文对矿井水文参数进行了深入分析，构建深度学习预测模型并进行实验验证。结果表明，论文所提模型具有更好的预测效果。可为矿井防治涌水和排水系统设计等方面提供技术支持。

At present, all coal mines have installed hydrological monitoring systems in accordance with the clear provisions of the "Coal Mine Safety Regulations" and have accumulated a large amount of hydrological data, but there are problems with insufficient data analysis capabilities and low utilization rates. This paper conducts an in-depth analysis of mine historical hydrological data, establishes a deep learning prediction model of hydrological parameters, provides technical support for preventing mine water damage, and is of great significance to ensure coal mine safety production.

The hydrological parameter analysis is to provide the basis for the establishment of the later prediction model. The mine monitoring hydrological data has obvious characteristics of noise and non-linearity. The stationarity of the hydrological data is judged by the autocorrelation graph. The autocorrelation coefficient is always positive in a long delay period, indicating that it has typical non-stationarity. In addition, the correlation of multi-sequence hydrological parameters is analyzed by Pearson correlation coefficient and scatter diagram matrix. The results show that there are different degrees of correlation among different parameter sensors, and the data of each monitoring point with the same parameter also have correlation.The non-stationary characteristics of mine hydrological data make it difficult for traditional single-sequence forecasting methods to achieve effective forecasting results. In order to improve the prediction accuracy, the paper proposes the CEEMD_GRU model. First, the hydrological data is decomposed into multiple stable sub-components by CEEMD; secondly, the number of lag periods of the hydrological data is determined by PACF, so as to determine the number of input neurons; and then the change law of each component is learned and predicted by the GRU neural network; Finally, the prediction results of each component are fused to obtain the final prediction value, and compared with the other five neural networks model based on two sets of data, the root mean square error of the test set is reduced by 36.38% and 25.48% on average.

Single-series forecasts are modeled based on the autocorrelation of the data. Then, the different hydrological sensors in the mine are also spatially correlated. This paper uses this correlation to establish the GRU_Attention model. First, input the multi-dimensional sequence into the GRU to achieve high-level feature learning; then use the output of the GRU as the input of the attention mechanism, and use the attention mechanism to mine the relationship between the multi-dimensional input and output, and calculate the feature weight; finally, the GRU layer The weighted summation of output and feature weights enhances or weakens each input to obtain a feature expression vector, which is input to the fully connected layer to calculate the final predicted value. Comparing experiments with five other neural networks, the root mean square error of the test set is reduced by 31.1%.

In this paper, the mine hydrological parameters are deeply analyzed, and the deep learning prediction model is built and verified by experiments. The results show that the model proposed in this paper has better prediction effect. Provide technical support for mine water inrush prevention and drainage system design.

[1] 李文平,乔伟,李小琴等. 深部矿井水害特征、评价方法与治水勘探方向[J].煤炭学报, 2019,44(08):2437-2448.

[2] 魏丽娜. 煤矿水害类型探讨与防治措施[J].西部探矿工程, 2019,31(03):124-126.

[3] 沈中芹,曾旺,王瑞强等. 基于改进 HFACS-MI 模型的煤矿透水事故致因分析[J].安全 与环境工程, 2020,27(03):178-184.

[4] 董书宁,王皓,周振方.我国煤矿水害防治工作现状及发展趋势[J].劳动保护, 2020,542(08):60-62.

[5] 邢朕国,杜文凤,梁喆,胡进奎.煤矿地下水实时跟踪监测预警系统设计[J].工矿自动化, 2017,43(08):72-75..

[6] 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.

[7] 朱愿福,王长申,李彦周等. 改进的灰色系统理论预测矿井涌水量[J].煤田地质与勘探, 2014,42(04):44-49.

[8] Li Z L, Gao T Y, Guo C, et al. A Gated Recurrent Unit Network Model for Predicting Open Channel Flow in Coal Mines based on Attention Mechanisms[J].IEEE Access, 2020,8(0):119819-119828.

[9] Kaimeng Zhang, Chi Tim Ng, Myung Hwan Na. Real time prediction of irregular periodic time series data[J]. Journal of Forecasting. 2020;39(03):501-511.

[10]王猛,殷博超,张凯歌等. 基于 ARIMA 乘积季节模型的矿井涌水量预测研究[J].煤炭 科学技术, 2017,45(11):199-204.

[11]Daraghmi Yousef-Awwad , Daadoo Motaz. Improved Dynamic Route Guidance based on Holt-Winters-Taylor Method for Traffic Flow Prediction[J].International Journal of Traffic and Transportation Engineering, 2015,04(03):75-83.

[12]Wang J. Prediction model of pulmonary tuberculosis based on gray kernel AR-SVM model[J]. Cluster Computing, 2019,22(02):4383-4387.

[13]田东,韦鑫化,王悦等. 基于 MA-ARIMA-GASVR 的食用菌温室温度预测[J].农业工程 学报,2020,36(03):190-197..

[14]David Adugh Kuhe , Thomas Akwana Obed. An ARMA Model for Short-term Prediction of Hepatitis B Virus Seropositivity among Blood Donors in Lafia-Nigeria[J]. Journal of Scientific Research and Reports, 2019,24(01):1-11. [15]Kang J, Duan Z T, Tang L. Online short-term traffic flow prediction method for LS-SVM[J].Application Research of Computers, 2018.35(10),91-94.

[16]刘航舵. 基于循环神经网络的瓦斯灾害预测预警研究[D].西安科技大学, 2020. [17]郑晓亮,来文豪,薛生. MI 和 SVM 算法在煤与瓦斯突出预测中的应用[J].中国安全科 学学报, 2021,31(01):75-80..

[18]郑方燕,陈鹏霖,石海峰等. 基于BP神经网络的时栅时序预测测量研究[J].仪表技术与 传感器, 2020,00(01):96-99.

[19]钟登华,田耕,关涛等. 基于混沌时序-随机森林回归的堆石坝料加水量预测研究[J].水 力发电学报, 2018,37(08):1-12.

[20]何正义,曾宪华,曲省卫等. 基于集成深度学习的时间序列预测模型[J].山东大学学报 (工学版), 2016,46(06):40-47.

[21]Zhao L L, Wang J J, Hu Y, et al. Conjoint Feature Representation of GO and Protein Sequence for PPI Prediction Based on an Inception RNN Attention Network[J]. Molecular Therapy - Nucleic Acids, 2020,(22):198-208. [22]Bisagno Niccoló , Saltori Cristiano , Zhang B, et al. Embedding group and obstacle information in LSTM networks for human trajectory prediction in crowded scenes[J]. Computer Vision and Image Understanding,2021,203(P0):103-126.

[23]闫杨,孙丽珺,朱兰婷. 基于时空相关性的短时交通流量预测方法[J].计算机工程, 2020,46(01):31-37.

[24]侯恩科,龙天文,樊志刚. 解析法预测文家坡煤矿工作面涌水量[J].矿业安全与环保, 2019,46(05):80-84.

[25]王春刚,方刚,刘洋.榆横北区巴拉素井田不同时期矿井涌水量预测[J].煤矿安全, 2020,51(06):212-217.

[26]苗霖田,贺晓浪,张建军等. 矿井涌水量的时间序列分析-水文地质比拟法预测[J].中国 煤炭, 2017,43(03):32-35.

[27]彭辉才,徐卫东,付青等. 贵州绿塘煤矿涌水量预测研究[J].南水北调与水利科技, 2013,11(02):58-61.

[28]施 龙 青 , 赵 云 平 , 王 颖 等 . 基 于 灰 色 理 论 的 矿 井 涌 水 量 预 测 [J]. 煤 炭 技 术 , 2016,35(09):115-118.

[29]贾伦. ARIMA 模型在矿井涌水量预测中的应用[J]. 科学技术创新, 2018,00(27):25-26.

[30]李孝朋,谢道雷,徐万鹏等. 多元回归分析在矿井涌水量预测中的应用[J].煤炭技术, 2016,35(10):189-190.

[31]孟璐. 基于人工神经网络的矿井涌水量预测[J].煤炭与化工, 2020,43(04):65-68. [32]乔美英,程鹏飞,刘震震. 基于 GA-SVM 的矿井涌水量预测[J].煤田地质与勘探, 2017,45(06):117-122.

[33]赵东,臧雪柏,赵宏伟. 基于果蝇优化的随机森林预测方法[J].吉林大学学报(工学版), 2017,47(02):609-614.

[34]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.

[35]彭秀艳,张彪. 基于 EMD-PSO-LSTM 组合模型的船舶运动姿态预测[J].中国惯性技术 学报, 2019,27(04):421-426.

[36]Furlaneto, Dennis Carnelossi, Cavalcanti, et al. Bias effect on predicting market trends with EMD[J]. Expert Systems with Application, 2017.00(82):19-26.

[37]Huang N E, Shen Z, Long S R, et al. The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis[J]. Proceedings Mathematical Physical & Engineering Sciences, 1998, 454(1971):903-995.

[38]赵征,汪向硕. 基于 CEEMD 和改进时间序列模型的超短期风功率多步预测[J]. 太阳 能学报, 2020,41(07):352-358.

[39]Torres M E, Colominas M A, Schlotthauer G, et al. A complete ensemble empirical mode decomposition with adaptive noise[C]. IEEE International Conference on Acoustics, Speech and Signal Processing(ICASSP), IEEE, 2011,4144-4147.

[40]易灵芝,常峰铭,龙谷宗等. 基于进化深度学习短期负荷预测的应用研究[J]. 电力系 统及其自动化学报, 2020,32(03):1-6.

[41]Zhang M H, Wang B Z, Zhou Y T, et al. Prediction of Chaotic Time Series Based on SALR Model with Its Application on Heating Load Prediction[J]. Arabian Journal for Science and Engineering, 2021,19(01):116-122.

[42]Ran X D, Shan Z G, Fang Y F, et al. A Convolution Component-Based Method with Attention Mechanism for Travel-Time Prediction[J] ． Sensors (Basel, Switzerland), 2019,19(09):01-19.

[43]Chen X et al. A Novel Hybrid Model Based on an Improved Seagull Optimization Algorithm for Short-Term Wind Speed Forecasting[J]. Processes, 2021,09(02) : 387-387.

[44]Hochreiter S, Schmidhuber J. Long short-term memory [J]. Neural Computation (S0899-7667), 1997,09(08):1735-1780.

[45]黄林,王电钢,刘萧等. 基于 LSTM 的网络流量预测方法[J].计算机应用研究, 2020,37(S1):264-265+272.

[46]罗奕,向新,王勇. 矿井明渠流量测定方法[J].煤田地质与勘探, 2006,00(01):56-58.

[47]魏健,赵红涛,刘敦楠等. 基于注意力机制的 CNN-LSTM 短期电力负荷预测方法[J].华 北电力大学学报(自然科学版), 2021,48(01):42-47.