Prediction of surface settlement around subway foundation pits based on spatiotemporal characteristics and deep learning models

Abstract

To obtain more accurate surface settlement prediction results around subway foundation pits, a novel spatiotemporal deep hybrid prediction model (STdeep model) that includes both the main and spatiotemporal blocks is proposed. The main block is constructed using the convolutional long short-term memory neural network, and the fundamental patterns of the surface settlement data from a network are extracted using the main block. To obtain more detailed spatiotemporal characteristics of the surface settlement data from a network, a spatiotemporal block is established. The convolutional neural network and the gated recurrent unit neural network are introduced into the spatiotemporal block to obtain the dynamic spatial and temporal characteristics of surface settlement data. To deeply integrate the main and spatiotemporal blocks, a dual-stream structure is designed. In addition, two cases are designed to test the STdeep model. The experiments indicate that the spatiotemporal relationships among the surface settlement data obtained from different monitoring points in a network are dynamic during the process of subway foundation pit excavation, and the STdeep model exhibits the best performance and excellent stability.