GNN-LSTM-based full-field temperature prediction of bridges using sparse sensor data and heat-transfer analysis

• Published in: Engineering structures Vol. 342; p. 120922
• Main Authors: Shan, Yushi, Wang, Xiaoyou, Jing, Qiang, Wu, Wanglin, Xia, Yong
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2025
• Subjects: Deep learning; Graph neural network; Heat-transfer analysis; Structural health monitoring; Temperature distribution; Deep learning; Temperature distribution; Heat-transfer analysis; Structural health monitoring; Graph neural network
• ISSN: 0141-0296
• DOI: 10.1016/j.engstruct.2025.120922

Detailed temperature distribution of a bridge is essential to investigate its structural behavior. The conventional physics-based heat-transfer analysis simulates the entire temperature field while with high computational cost and inevitable numerical errors. Data-based relation-mapping models rely on real measurements but are subject to the constraint of limited sensors. This study proposes a novel semi-supervised model integrating graph neural networks (GNNs) and long short-term memory (LSTM) to reconstruct the complete temperature field from sparse measurement data by leveraging spatial neighboring relations and temporal temperature trends. A long-span cable-stayed bridge is used as the testbed. First, a 2D finite element model (FEM) of the steel box girder is established and the heat-transfer analysis is conducted to calculate the detailed hourly temperature distribution. Second, the GNN model is developed, in which the mesh of the FEM informs the graph structure; the calculated temperatures serve as the pseudo labels; and the heat transfer between nodes in the FEM corresponds to message passing between vertices in the graph. Thus, the mesh, data, and essence of the FEM are embedded as physical information for the GNN model. Additionally, LSTM is combined to capture the temporal variation of the sequential temperature. The GNN-LSTM-based model is applied to the cable-stayed bridge to predict the full-field temperature distribution of the girder. The average discrepancy between the predictions and measurements in one year decreases to 0.92 °C, which is significantly reduced compared with the simulation error of 1.98 °C. The proposed model demonstrates its generalization capability by compensating for numerical errors across all four seasons. The discrepancy of the calculated temperature-induced displacements can be reduced by up to 52 % compared with the FEM results. This innovative method effectively narrows the gap between numerical simulation and field measurement, enabling an accurate temperature behavior investigation of long-span bridges. •Graph neural networks and long short-term memory are integrated to predict temperature.•The model can reconstruct the full temperature field from sparse measurements.•Physical information is embedded to improve generalization.•The heat-transfer behavior is approximated by using training data from simulation.•The discrepancy is significantly reduced between simulation and measurement.