Model sensitivity of DPFTransformer.

Deep foundation pit settlement prediction based on machine learning is widely used for ensuring the safety of construction, but previous studies are limited to not fully considering the spatial correlation between monitoring points. This paper proposes a transformer-based deep learning method that considers both the spatial and temporal correlations among excavation monitoring points. The proposed method creates a dataset that collects all excavation monitoring points into a vector to consider all spatial correlations among monitoring points. The deep learning method is based on the transformer, which can handle the temporal correlations and spatial correlations. To verify the model’s accuracy, it was compared with an LSTM network and an RNN-LSTM hybrid model that only considers temporal correlations without considering spatial correlations, and quantitatively compared with previous research results. Experimental results show that the proposed method can predict excavation deformations more accurately. The main conclusions are that the spatial correlation and the transformer-based method are significant factors in excavation deformation prediction, leading to more accurate prediction results.

基于机器学习的深基坑沉降预测技术已广泛应用于保障工程施工安全，但以往研究的局限在于未能充分考量监测点间的空间相关性。本文提出一种基于Transformer的深度学习方法，可同时覆盖基坑监测点间的空间相关性与时间相关性。该方法将全部基坑监测点整合为单向量形式，以此纳入所有监测点间的空间关联信息，进而构建对应数据集。该深度学习架构以Transformer为核心，能够同时处理时间相关性与空间相关性。为验证模型的预测精度，本文将所提方法与仅考量时间相关性、未考虑空间相关性的长短期记忆网络（LSTM）及RNN-LSTM混合模型进行对比，并与既往研究成果开展定量比对。实验结果表明，所提方法可更为精准地预测基坑变形。核心结论显示，空间相关性考量与基于Transformer的方法是提升基坑变形预测精度的关键影响因素，可有效获得更准确的预测结果。