为提高盾构模型预测性能，以武汉长江隧道大直径泥水盾构隧道工程为背景和数据来源，基于图卷积网络（GCN）和长短期记忆神经网络（LSTM）构建掌子面压力预测模型，利用贝叶斯优化算法（BO）对预测模型的关键超参数进行寻优。采用SHAP方法对预测模型进行全局解释，计算每个输入参数对预测目标的Shapley值，提高模型的解释度和透明度。研究结果表明： 1）所提出的BO-GCN-LSTM方法在各历史时间步长下均具有较高的精度，拟合优度（R2）平均值为0.943，均方根误差（ERMSE）平均值为0.245，平均绝对误差（EMAE）平均值为0.173，平均绝对百分比误差（EMAPE）平均值为1.183%。2）在历史时间步长t-1—t-10中，时间步长t-3的R2、ERMSE、EMAE、EMAPE分别为0.953、0.233、0.159、1.151%，运行速率为1.7次/s，表现出最佳整体预测性能。3）通过SHAP方法进行全局解释，可以确定对研究目标影响较大的参数为气垫舱压力、进出排浆压力和刀盘挤压力差，为大直径泥水盾构隧道掌子面压力管控提供有价值的决策依据。基于BO-GCN-LSTM深度学习模型可以有效预测隧道掌子面压力，有助于盾构驾驶员做出合理的参数调整。
The authors conduct a case
analysis of a large-diameter
slurry shield tunnel crossing the Yangtze river in Wuhan, China. The Bayesian
optimization algorithm (BO) is employed to optimize key hyperparameters in the
prediction model for tunnel face pressure, which is based on graph
convolutional networks (GCN) and long shortterm memory
(LSTM) neural networks. Furthermore, the Shapley additive explanations (SHAP)
method is applied to globally interpret the prediction model and calculate the
Shapley value of each input parameter for the prediction target, enhancing the
model′s interpretability and transparency. The research
findings are as follows: (1) The proposed BOGCN-LSTM method demonstrates high accuracy across all historical time
steps, achieving an average goodness of fit (R2), root
mean square error (ERMSE), mean absolute error (EMAE), and mean absolute percentage error (EMAPE)of 0.943, 0.245, 0.173, and 1.183%, respectively. (2) Among
historical time steps t-1 to t-10, the metrics at time step t-3—R2 of
0.953, ERMSE of 0.233, EMAEof
0.159, and EMAPE of 1.151%—show the best overall predictive performance, with a computational
running rate of 1.7 times per second. (3) The global interpretation results
using the SHAP method indicate that the air cushion chamber pressure, inlet and
outlet slurry pressures, and cutterhead squeezing pressure difference
significantly influence the research objectives, offering valuable decision-making insights for controlling tunnel face pressure in large-diameter slurry shield operations. The BO-GCN-LSTM deep learning model effectively predicts tunnel face pressure,
assisting shield tunneling operators in making informed parameter adjustments.
