Accurate prediction of diaphragm wall deformation is critical for ensuring the safety of deep foundation pit excavation. Diaphragm wall deformation exhibits significant temporal and spatial characteristics, and traditional prediction methods often face problems in stability and accuracy when handling multi-step prediction tasks. This paper proposes a dynamic updating model for diaphragm wall deformation based on a fully convolutional network (FCN), featuring an innovative design for spatiotemporal feature extraction and multi-step dynamic prediction of complex deformation data. The proposed model effectively preserves spatial information through its fully convolutional architecture and enhances feature transmission by incorporating skip connection strategies, thereby improving stability and accuracy in multi-step predictions. The model’s effectiveness was validated using measured data from a foundation pit project of a metro station on Nanjing metro line 11. The results demonstrate that the model has significant advantages in capturing diaphragm wall deformation trends and identifying maximum deformation locations, providing a novel insight for the prediction of complex time-series data.