| Summary: | Accurate prediction and forecasting of soil mass deformation in deep excavation pits are pivotal for risk monitoring and safety assessment. Nonetheless, the complex underlying dynamics inherent in field sensing measurements pose challenges to the forecasting endeavor. In light of these challenges, the present study leverages recent strides in deep learning and introduces a spatiotemporal learning framework tailored to forecast soil mass deformation marked by resilient temporal interconnections and spatial associations. This study focuses on developing a Multi-Point Recurrent Convolutional Neural Network (RCNN) model for predicting sensor-based temporal patterns. This model integrates data feature fusion to extract spatiotemporal latent features from the dataset, thereby constructing a surrogate model for forecasting soil mass deformation. The proposed methodology is deployed to forecast strain responses in a deep excavation pit using a dataset spanning over five months. A comparative analysis is conducted, contrasting the performance of the proposed approach with that of a conventional temporal-only network. The analysis reveals that the prediction errors generated by the Multi-Point RCNN are predominantly concentrated within the range of 10% for all sensors, with a high-confidence interval (CI) of 96%, compared to the RCNN model (82%) and the LSTM model (79%). The compelling outcomes underscore the efficacy of the Multi-Point RCNN approach as a promising, dependable, and computationally efficient method for accurately predicting soil mass deformation in deep excavation pits, grounded in data-driven principles.
|
|---|