Abstract: With the rapid development of underground engineering, the demand for deformation monitoring and accurate assessments during foundation pit excavations has been increasing. Existing methods, which are mostly based on time-series models, often neglect the spatial correlations and evolutionary patterns contained in monitoring data, resulting in limited prediction accuracies and insufficient capabilities for long-term trend estimations. To address the aforementioned issues, this study proposes a deformation estimation method that integrates a spatiotemporal pattern analysis with an attention mechanism. By constructing a spatial feature extraction module and an attention-based sequence fusion module, the method proposed dynamically models key spatiotemporal features within monitoring data. A real-world engineering case study demonstrates that the temporal dependency of foundation pit deformation data is relatively weak, with only the latest observation having a significant impact on the prediction results, whereas spatial correlations play a more critical role, providing an important reference for model design. Compared with traditional models, the method proposed shows a superior performance in long-term prediction tasks, achieving stable accuracy improvements of more than 70%, and demonstrating good robustness and engineering applicability.