安全与环境学报

随着我国煤炭资源开采逐渐向深部延伸，深部煤岩体在高地应力、高温、高渗透压及时间效应叠加的复杂环境下，围岩蠕变问题日益突出，严重威胁矿井的开采安全和生产效率。为有效预测含水砂岩的蠕变行为，提出了一种融合卷积神经网络(Convolutional Neural Networks, CNN)、门控循环单元(Gated Recurrent Unit, GRU)和注意力(Attention)机制的CGA深度学习模型。模型结合CNN的空间特征提取能力、GRU的时间序列建模能力及Attention的动态权重分配能力，提升了对非线性、长时间依赖关系的捕捉能力。利用实测数据对CGA模型中的优化算法、卷积层数、卷积核数量、GRU层数和GRU层神经元数量进行了训练和确定。应用CGA模型对含水砂岩蠕变行为进行了预测，并与实测数据进行了对比。结果表明，与CNN、反向传播神经网络(Back Propagation Neural Network, BPNN)和CNN-GRU模型相比，CGA模型的平均绝对百分比误差(MAPE)分别降低了25.00%、18.93%和12.00%,平均绝对误差(MAE)分别降低了17.84%、13.77%和4.86%,均方误差(MSE)分别降低了26.04%、15.35%和6.02%,均方根误差(RMSE)分别降低了13.99%、8.01%和3.12%,CGA模型的R²达到了0.981 163,表明CGA模型具有更好的非线性拟合能力。利用CGA模型有助于掌握巷道围岩的长期变形行为，为围岩控制方案设计提供基础依据。

To effectively predict the creep behavior of water-bearing sandstone, we propose a deep learning model called CNN-GRU-Attention(CGA), which integrates Convolutional Neural Networks(CNN), Gated Recurrent Units(GRU), and an Attention mechanism. This model leverages the spatial feature extraction capabilities of CNN, the time-series modeling abilities of GRU, and the dynamic weight allocation of the Attention mechanism to enhance the capture of nonlinear relationships and long-term dependencies. The CNN component efficiently extracts local features and spatial dependencies from the data through convolutional and pooling layers, overcoming the limitations of traditional methods in capturing nonlinear features. The GRU effectively models long-term dependencies in sequential data through its update and reset gate mechanisms, addressing the limitations of CNN in time-series modeling. Meanwhile, the Attention mechanism dynamically adjusts the importance of hidden states at different time steps using a trainable weight matrix. Its fundamental principle involves generating normalized weights by calculating attention scores for each time step, which emphasizes creep characteristics during critical phases and resolves the issue of GRU assigning equal weights to all time-step information in time-series processing. The optimization algorithm, number of convolutional layers, number of convolutional kernels, number of GRU layers, and number of neurons in the GRU layer of the CGA model were determined through training with measured data. The final structure of the CGA model includes 3 convolutional layers with 16, 32, and 64 kernels, respectively, utilizing ReLU as the activation function; 3 GRU layers, each containing 64 neurons; and an Attention layer with a dimension of 50. The CGA model was employed to predict the creep behavior of sandstone at water contents of 0, 1.6%, 2.4%, and 3.2% under a pressure of 20 MPa, with the results compared against the measured data. The results demonstrate that the prediction curve of the CGA model closely aligns with the measured values, exhibiting superior evaluation metrics: MAPE of 1.317 323%, MAE of 0.018 417, MSE of 0.008 521, RMSE of 0.092 300, and R² of 0.981 163, all surpassing those of the CNN, Back Propagation Neural Network(BPNN), and CNN-GRU models. Specifically, the CGA model shows a reduction in MAPE by 25.00%, 18.93%, and 12.00% compared with the CNN, BPNN, and CNN-GRU models, respectively. For MAE, it reduces by 17.84%, 13.77%, and 4.86%. In terms of MSE, the reductions are 26.04%, 15.35%, and 6.02%. In terms of RMSE, the CGA model achieves reductions of 13.99%, 8.01%, and 3.12%, while the R² value increases to 0.981 163, indicating that the CGA model exhibits superior nonlinear fitting capabilities. Employing the CGA model enhances the understanding of the long-term deformation behavior of roadway surrounding rock, providing a foundation for the design of surrounding rock control schemes. This advancement is significant for promoting the safe and efficient development of underground mining.