基于图深度学习的桥梁群温湿度时空演化预测方法

The environmental effects on urban bridge groups exhibit strong spatial-temporal correlations. To accurately predict spatial-temporal distribution of temperature and humidity across bridge groups，a spatial-temporal graph selective state space model was proposed，which primarily consisted of three modules：multi-granularity data fusion module，spatial-temporal graph convolution module and graph selective state space module. The multi-granularity data fusion module was designed to fully extract and utilize recent，medium，and long-term features embedded in the monitoring data，thereby ensuring prediction accuracy across multiple temporal granularities；the spatial-temporal graph convolution module simultaneously captured spatial and temporal data features during the prediction of the dynamic evolution of temperature and humidity within bridge group system；the graph selective state space module incorporated a selection mechanism to more accurately represent the spatial-temporal evolutionary characteristics of the data，which allowed for the dynamic adjustment of model learning patterns while simultaneously optimizing computational efficiency. The proposed model was validated using environmental temperature and humidity datasets from urban regional meteorological stations and bridge groups. Comparative experiments were conducted against other methods based on graph neural networks and state space models. Mean absolute error （MAE），root mean square error（RMSE），and mean absolute percentage error（MAPE）were adopted as quantitative metrics to evaluate prediction performance. The results indicate that，compared with baseline models such as DSTAGNN and SpoT-Mamba，the proposed method achieves average prediction performance improvements of 6. 74% in mean absolute error（MAE），7. 10% in root mean square error（RMSE），and 5. 97% in mean absolute percentage error（MAPE）.