改进双向长短期记忆神经网络的瓦斯涌出量预测  

Enhanced Bi-directional Long Short-Term Memory neural network for gas emission forecasting

作　　者：祁云 白晨浩 代连朋 汪伟 薛凯隆 崔欣超 QI Yun;BAI Chenhao;DAI Lianpeng;WANG Wei;XUE Kailong;CUI Xinchao(School of Mining and Coal,Inner Mongolia University of Science and Technology,Baotou 014010,Inner Mongolia,China;College of Coal Engineering,Shanxi Datong University,Datong 037000,Shanxi,China;Institute of Disaster Rock Mechanics,Liaoning University,Shengyang 110036,Liaoning,China)

机构地区：[1]内蒙古科技大学矿业与煤炭学院,内蒙古包头014010 [2]山西大同大学煤炭工程学院,山西大同037000 [3]辽宁大学灾害岩体力学研究所,辽宁沈阳110036

出　　处：《安全与环境学报》2024年第12期4630-4637,共8页Journal of Safety and Environment

基　　金：国家自然科学基金面上项目(52174188);山西省高等学校科技创新计划项目(2022L448,2022L449);山西大同大学研究生教育创新项目(23CX49)。

摘　　要：为提高瓦斯涌出量预测精度,降低煤矿回采工作面瓦斯涌出超限事故的风险,针对瓦斯涌影响因素众多、难以预测等问题,采用灰狼优化算法(Grey Wolf Optimization,GWO)双向长短期记忆神经网络(Bi-directional Long Short-Term Memory,BiLSTM)的组合模型预测瓦斯涌出量。首先,运用主成分分析法(Principal Components Analysis,PCA)处理瓦斯涌出影响因素,降低数据维度,以减少模型计算时的负担;其次,利用GWO优化BiLSTM模型的学习率(best_lr)、隐藏层层数(best_hd)以及正则化系数(best_l2),可有效避免局部最优解问题,并采用决定系数(R-Square,R^(2))、均方根误差(Root Mean Square Error,RMSE)和平均绝对误差(Mean Absolute Error,MAE)对所建模型预测的结果进行综合评价分析;最后,将该模型应用于内蒙古自治区某矿回采工作面预测瓦斯涌出量。结果显示:PCA GWO BiLSTM组合模型相比于长短期记忆神经网络(Long Short-Term Memory,LSTM)和双向长短期记忆神经网络对应的单一模型,其MAE分别降低20.81%、30.17%,RMSE分别降低0.063、0.142,R^(2)则分别提高了0.023、0.075,表明该模型在复杂因素条件下具有更高的精准度、泛化性和鲁棒性。In the context of coal mine safety management,this paper delves into the factors influencing gas emission exceeding limit accidents in coal mine working faces to enhance the accuracy of gas emission prediction and mitigate associated risks.Subsequently,a combined prediction model is formulated based on the principles of the Grey Wolf Optimization Algorithm(GWO)and the Bi-directional Long Short-Term Memory Neural Network(BiLSTM).Firstly,an analysis of relevant literature identified 13 factors that influence gas outbursts.Principal Component Analysis(PCA)was then applied to process the sample data.This involved normalizing the original sample data and combining it with the component matrix and variance contribution rate.These calculations were used to determine the three main factors that influence gas outbursts.This reduced the dimensionality of the data and alleviated the computational burden on the model.Secondly,the GWO algorithm was employed to optimize the learning rate(best_lr),number of hidden layers(best_hd),and regularization coefficient(best_l2)of the BiLSTM model.This approach helped identify the hyperparameters that are most suitable for the BiLSTM model,effectively mitigating the issue of local optimal solutions.The prediction results of the constructed model are thoroughly evaluated and analyzed using various metrics,including the coefficient of determination(R^(2)),Root Mean Square Error(RMSE),Mean Absolute Error(MAE),and relative error.Finally,the model will be utilized to forecast gas emissions from a mining face in the Inner Mongolia Autonomous Region.The analysis results demonstrate that this model achieves a reduction in MAE of 20.81%and 30.17%respectively,a decrease in RMSE of 0.063 and 0.142 respectively,and an improvement in R^(2)of 0.023 and 0.075 respectively,compared to the single models based on the Long Short-Term Memory Neural Network(LSTM)and Bi-directional Long Short-Term Memory(BiLSTM)neural network.These findings provide evidence that this model exhibits superior accuracy,generaliz

关 键 词：安全工程 瓦斯涌出 灰狼优化算法 双向长短期记忆神经网络 主成分分析法 

分 类 号：X936[环境科学与工程—安全科学]