基于改进BiLSTM代理模型的基坑分步开挖变形计算方法

Calculation Method for Stepwise Excavation-induced Deformation in Foundation Pits Based on BiLSTM Surrogate Model

摘要:

提出了贝叶斯优化?双向长短期记忆?注意力机制（BO-BiLSTM-Att）代理模型，用于基坑分步开挖变形计算。融合BiLSTM网络和Att，挖掘输入输出的时序映射关系，并通过BO寻找最优超参数，进一步提高计算精度。通过数值算例验证了BO相较于粒子群优化（PSO）、遗传算法（GA）等在超参数优化上的适用性和高效性，并使用由BO得到的超参数训练代理模型。将代理模型在测试集上的计算结果与真实值进行比较，评估该方法在基坑分阶段变形计算中的适用性。同时，对比了相同超参数下BiLSTM网络和BiLSTM-Att在测试集上的计算精度。结果表明，Att可提高代理模型精度。所提代理模型在测试集上的决定系数均值约为0.967，平均绝对误差均值约为0.875 mm。

Abstract:

In this paper， a BO-BiLSTM-Att surrogate model is proposed to efficiently calculate the excavation-induced deformation. The BiLSTM network is integrated with Att to capture temporal mapping relationships between input and output parameters. The Bayesian optimization （BO） is implemented to identify optimal hyperparameters， thereby further enhancing the model’s computational accuracy. Through a numerical case， the applicability and efficiency of BO in hyperparameter optimization are verified by comparison with PSO and GA. Subsequently， the surrogate model is trained by BO-optimized hyperparameters and its effectiveness in staged deformation calculation is demonstrated through the difference between predicted values and true values of test sets. Finally， the computational accuracy between BiLSTM network and BiLSTM-Att under identical hyperparameters is compared. It is shown that Att can improve the model’s computational accuracy. The proposed surrogate model has enhanced the computational accuracy （R²=0.967） and generalization capability （mean average error 0.875 mm） in testing scenarios.