Data‐Augmentation Spatiotemporal Deep Learning Framework for Robust Prediction of Diaphragm Wall Deformation

Accurate prediction of diaphragm wall deformation is important for deep excavation safety and structural health monitoring (SHM). However, predictive models are often constrained by sparse monitoring data and inadequate learning of spatial dependencies among monitoring points. To address these limitations, a data‐augmentation spatiotemporal prediction (DA‐STP) framework is proposed, in which a deep convolutional generative adversarial network augments sparse monitoring data and an attention‐enhanced convolutional long short‐term memory (AE‐ConvLSTM) model learns spatiotemporal deformation patterns from the augmented dataset. Excavation depth is included as a global input to represent the excavation process. The proposed framework is validated using data from the Wuhu Yangtze River‐Crossing Tunnel, and only measured field data are used for testing. Results show that the optimal 1:3 real‐to‐synthetic ratio reduces the root mean square error by 66.4%, while the AE‐ConvLSTM model achieves long‐horizon prediction with a coefficient of determination ( R 2 ) above 0.95. The proposed framework also outperforms ConvLSTM, LSTM, and gated recurrent unit baselines, reducing root mean square error and mean absolute error by at least 66.5% and 40.8%, respectively. In addition, the framework demonstrates good transferability under similar site conditions, achieving accurate prediction at the receiving shaft without retraining, while limited local fine‐tuning further improves prediction accuracy. The proposed DA‐STP framework, therefore, provides an effective tool for long‐horizon deformation prediction and SHM‐informed early warning in deep excavations, particularly when monitoring data are sparse.