基于MHA-BiGRU的工作面瓦斯浓度预测模型  

A gas concentration prediction model for working faces based on MHA-BiGRU

作　　者：李玉鹏 贾澎涛[2] 刘平 杨鸿宇 杨超社 LI Yupeng;JIA Pengtao;LIU Ping;YANG Hongyu;YANG Chaoshe(Shanmei Group Huangling Jianzhuang Mining Co.,Ltd.,Yan’an 727307,China;College of Computer Science and Technology,Xi’an University of Science and Technology,Xi’an 710054,China)

机构地区：[1]陕西煤业集团黄陵建庄矿业有限公司,陕西延安727307 [2]西安科技大学计算机科学与技术学院,陕西西安710054

出　　处：《中国矿业》2024年第12期180-188,共9页China Mining Magazine

基　　金：国家自然科学基金项目资助(编号:51974236)。

摘　　要：为了准确可靠地评估煤矿的瓦斯安全态势,提高煤矿瓦斯浓度预测模型的精度,提出了基于多头注意力机制(MHA)与双向门控循环单元神经网络(BiGRU)的瓦斯浓度预测模型(MHA-BiGRU)。首先,采用线性插值和滑动窗口方法,对原始的瓦斯相关监测时间序列数据进行预处理和空间重构,为进一步的数据建模奠定基础;其次,鉴于BiGRU模型可以同时捕捉到序列中上下文的信息,用均方误差做损失函数、采用Adam优化算法,构建适用于瓦斯浓度时间序列的优化BiGRU学习模型,并确定在线预测学习的窗口大小和BiGRU模型参数;然后,将多头注意力机制(MHA)嵌入到BiGRU循环结构中,从BiGRU的输出中提取重要特征,多个注意力头可以关注隐藏状态序列的不同部分,捕捉复杂的时序关系和模式,提取的特征再通过循环可以由BiGRU进一步处理,捕捉更深层次的时序依赖;随后,下一层的MHA可以再次优化特征表示;最后,经过全连接层,生成最终的瓦斯浓度预测值。实验结果表明,相较于循环神经网络(RNN)、长短时记忆网络(LSTM)、GRU、BiGRU、MHA-RNN、MHALSTM、iTransformer、Linear等对比模型,MHA-BiGRU模型取得了最好的预测效果,显著提高了煤矿瓦斯安全态势的监管水平。To enhance coal mine gas safety and improve the accuracy of gas concentration prediction,it proposes a novel model,termed MHA-BiGRU,which integrates the Multi-Head Attention mechanism(MHA)with a Bidirectional Gated Recurrent Unit(BiGRU)neural network.The approach begins with the preprocessing and reconstruction of raw gas monitoring time series data using linear interpolation and sliding window techniques,providing a robust foundation for subsequent modeling.The BiGRU model,optimized using the mean square error loss function and the Adam algorithm,effectively captures contextual dependencies within sequential data.Critical parameters,including window size and BiGRU architecture,are meticulously fine-tuned for real-time prediction tasks.The incorporation of the MHA mechanism within the BiGRU network further enhances predictive performance by enabling multiple attention heads to focus on distinct segments of the hidden state sequences,thereby capturing complex temporal patterns.These features are subsequently refined by the BiGRU layers to extract deeper temporal dependencies.The MHA layers are further optimized to enhance feature representation,with the final gas concentration prediction being generated by a fully connected output layer.Empirical results indicate that the MHA-BiGRU model significantly outperforms several baseline models,including RNN,LSTM,GRU,BiGRU,MHA-RNN,MHA-LSTM,iTransformer,and linear models.The superior accuracy achieved by the MHA-BiGRU model underscores its potential to significantly improve coal mine gas concentration predictions,thereby contributing to enhance safety monitoring in mining operations.

关 键 词：瓦斯浓度 多头注意力机制 双向门控循环单元 深度学习 预测 

分 类 号：TD712.5[矿业工程—矿井通风与安全]