为提升大直径泥水平衡盾构在复杂地层中掘进的安全稳定性，提出一种融合图注意力网络（graph attention networks，GAT）、长短期记忆网络（long short-term memory，LSTM）、多目标海洋捕食者算法（multi-objective marine predators algorithm，MOMPA）与优劣解距离法（technique for order preference by similarity to ideal solution，TOPSIS）的混合智能优化框架，实现大直径泥水平衡盾构安全稳定性多目标的预测与优化。为实现盾构施工参数的实时调控，提供倾覆力矩、俯仰角与滚动角的多目标动态协同优化途径。首先，基于武汉轨道交通12号线工程大直径泥水平衡盾构隧道施工数据，构建涵盖12项关键施工参数与3项安全稳定性指标的数据库，并通过数据清洗、归一化与序列重构完成预处理； 然后，开发GAT-LSTM时空混合预测模型，建立施工参数与安全指标之间的非线性映射关系，作为MOMPA算法的适应度函数； 在此基础上，集成MOMPA与TOPSIS方法，实现多目标帕累托解集的生成与最优施工参数的实时选取，并通过历史优化值滚动更新的策略，实现动态在线优化。最后通过实例验证表明： 1）GAT-LSTM模型对倾覆力矩、俯仰角与滚动角的预测精度较高，测试集预测精度R2分别达到0.919、0.923与0.976； 2）所提GAT-LSTM-MOMPA-TOPSIS混合优化方法能显著降低各安全指标值，平均优化幅度为18.0%，其中倾覆力矩、俯仰角与滚动角分别降低19.4%、7.9%与26.6%； 3）与传统静态优化方法相比，所提动态优化策略的总体性能提升约11%。
To improve the safety and stability of large-diameter slurry shield tunneling in complex geological conditions, a hybrid intelligent optimization framework that integrates a graph attention network (GAT), long short-term memory network (LSTM), multi-objective marine predators algorithm (MOMPA), and technique for order preference by similarity to ideal solution (TOPSIS) is proposed. The framework enables the prediction and online multiobjective optimization of key shield stability indicators. A multiobjective dynamic coordination strategy was developed for optimizing the overturning moment, pitch angle, and roll angle, thereby supporting the real-time adjustment of shield operating parameters. First, construction data from a large-diameter slurry shield tunnel project in Wuhan were used to establish a database comprising 12 key construction parameters and three safety and stability indicators. The data were processed through cleaning, normalization, and sequence reconstruction. A GAT-LSTM spatiotemporal prediction model was then developed to capture the nonlinear relationships between construction parameters and safety indicators. The model outputs were subsequently used as objective functions in the MOMPA optimization process. MOMPA and TOPSIS were combined to generate a Pareto solution set and identify the preferred set of construction parameters. A rolling-update strategy based on previously optimized parameters was introduced to support online dynamic optimization. Case study results show that (1) the GAT-LSTM model provides accurate predictions of the overturning moment, pitch angle, and roll angle, with test-set R2 values of 0.919, 0.923, and 0.976, respectively; (2) the proposed GAT-LSTM-MOMPA-TOPSIS framework reduces the three safety and stability indicators by an average of 18.0%, with reductions of 19.4% in overturning moment, 7.9% in pitch angle, and 26.6% in roll angle; and (3) compared with conventional static optimization methods, the proposed dynamic strategy improves the reported overall performance by approximately 11%.
