Scene-GCN: a time-series prediction method in complex monitoring environments through spatial–temporal knowledge graph (ST-KG)

作者：Weihao Li, Yufeng He, Lizhi Tao, Jifa Chen, Xing Wang, Yong Ge, Hui Lin · 发表于：International Journal of Geographical Information Systems · 年份：2025 · DOI：10.1080/13658816.2025.2479183 · 被引用次数：5 · 研究领域：Traffic Prediction and Management Techniques、Time Series Analysis and Forecasting、Anomaly Detection Techniques and Applications

Current forecasting methods for complex geographic scenes often fail to adequately account for both intrinsic factors and external environmental variables, which limits their predictive accuracy. This study proposes Scene-GCN, a Graph Convolutional Network (GCN)-based model that utilizes a spatio-temporal knowledge graph (ST-KG) to integrate static and dynamic environmental factors with the scene’s spatio-temporal characteristics to predict trends. First, variations at each monitoring target point are abstracted as geo-processes, and instantaneous monitoring values are represented as geographic states, forming a coupled “process-state” ST-KG model tailored for monitoring networks in complex geographic scenes. Environmental knowledge, both dynamic and static, is then embedded within the nodes of the ST-KG, capturing historical and real-time contextual influences through feature computation. Next, a temporal graph convolutional network (T-GCN) cell is developed to quantify the relationships between spatio-temporal features, environmental knowledge, and target predictions. Experimental evaluation on a real-world time-series dataset for monitoring surface deformation in a tailings dam shows that our method outperforms baseline models. Ablation studies further show that the integration of the ST-KG enhances forecasting performance, reducing root mean square error (RMSE) by 49.2% compared to a T-GCN model without ST-KG. Additionally, leveraging a graph neural network framework impr...