A GCN‐Transformer Based Spatial–Temporal Data Prediction Method for Location‐Fixed Heterogeneous Monitoring Sensor Network: A Case Study of Dam Seepage Safety Monitoring

ABSTRACT Location‐fixed monitoring sensor networks, which are crucial for geospatial data acquisition and prediction, face significant challenges including the underutilization of data, inadequate modeling of complex spatio‐temporal couplings, and insufficient expression of relationships among heterogeneous sensors. To address these issues, this study shifts the methodological focus from purely algorithmic architectural design to multi‐dimensional spatio‐temporal feature engineering. This study proposes a novel feature representation framework tailored for sensor heterogeneity. At the node level, it extracts multi‐scale temporal dynamics (i.e., trend and seasonality) via STL decomposition. At the edge level, it innovatively constructs a fused graph topology by quantifying physical geographic distance, type compatibility, and historical data‐driven correlations. Furthermore, to effectively utilize these extracted features, the GCN‐Transformer based Spatial–Temporal Data Prediction (GT‐STDP) method is proposed. It integrates Graph Convolutional Networks for spatial dependency learning and a Transformer architecture for long‐range temporal dynamics modeling. Validated using 858,455 monitoring records from the Shanxi Reservoir dam, GT‐STDP significantly outperforms state‐of‐the‐art time‐series models, achieving a prediction accuracy with an R 2 of 0.76 under the full feature configuration. Furthermore, through spatio‐temporal feature representation effectiveness analysis, the optimal feature combination further elevates the R 2 to 0.79. Moreover, the model exhibits strong robustness, maintaining an R 2 of 0.46 even with six consecutive days of missing data, thereby providing a highly reliable analytical framework for infrastructure safety monitoring.