Method for predicting temperature field of long-span space steel structure based on spatio-temporal graph neural network
By using a spatiotemporal graph neural network-based approach, a graph model was established and combined with an attention mechanism to solve the efficiency and accuracy problems of temperature field prediction for large-span spatial steel structures. This enabled rapid and accurate temperature field prediction, supporting safety assessments during construction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TIANJIN UNIV
- Filing Date
- 2026-03-03
- Publication Date
- 2026-06-09
AI Technical Summary
Existing technologies are insufficient to efficiently and accurately predict the spatiotemporal temperature distribution of large-span spatial steel structures under solar radiation, resulting in delayed temperature effect warnings during construction. Furthermore, traditional monitoring methods are costly and have limited coverage.
A spatiotemporal graph neural network-based approach is adopted. A three-dimensional geometric model is established and transformed into a graph model. The spatial and temporal features of the nodes are extracted using graph neural networks and recurrent neural networks. The temperature field is predicted by combining an attention mechanism, taking into account factors such as solar radiation and ambient temperature.
It enables accurate and rapid prediction of the temperature field of large-span spatial steel structures, reduces monitoring costs, improves the real-time performance and accuracy of predictions, and supports safety assessments during construction.
Smart Images
Figure CN122174640A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the interdisciplinary field of structural engineering and artificial intelligence, and more specifically, to a method for predicting the temperature field of large-span spatial steel structures based on spatiotemporal graph neural networks. Background Technology
[0002] With the widespread application of large-span steel structures in stadiums, transportation hubs, and other large public buildings, their safety during construction and operation has received increasing attention, highlighting the growing importance of structural temperature field research. Under solar radiation, large-span steel structures are prone to highly non-uniform temperature distribution, leading to significant temperature stress and structural deformation, directly affecting construction accuracy and long-term safety.
[0003] Current research primarily employs the finite element method (FEM) for numerical simulation of structural temperature fields. While this method comprehensively considers geometric and material nonlinearities, it suffers from complex modeling, time-consuming computation, and its accuracy heavily relies on the precision of boundary condition settings. This makes it difficult to meet the demands for real-time, rapid prediction and feedback of temperature fields during construction. In engineering practice, monitoring typically relies on a limited number of temperature sensors. This not only results in high deployment costs and limited coverage but also only acquires localized temperature data, failing to comprehensively reflect the overall temperature distribution of the structure and making it difficult to predict future trends. Consequently, temperature effect early warnings exhibit significant lags.
[0004] Therefore, proposing a method that can efficiently and accurately predict the spatiotemporal temperature distribution of large-span spatial steel structures under solar radiation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] In view of this, this application provides a method for predicting the temperature field of large-span spatial steel structures based on spatiotemporal graph neural networks, which can efficiently and accurately predict the spatiotemporal distribution of temperature in large-span spatial steel structures and solve the problem of difficult temperature monitoring of large-span spatial steel structures.
[0006] A method for predicting the temperature field of large-span spatial steel structures based on spatiotemporal graph neural networks includes: A three-dimensional geometric model of the steel structure to be tested is established. Transient thermal analysis of each node inside the steel structure is performed using the finite element method to obtain the spatial distribution cloud map of the structural temperature and the time series data of the node temperature. The three-dimensional geometric model is transformed into a graph model to reflect the topological features of the steel structure under test, and the node feature matrix and edge feature matrix in the graph model are determined. A temperature field prediction model is constructed based on the graph model. The temperature field prediction model is configured to take the node feature matrix and edge feature matrix as input, extract the spatial dependency features of each node using a graph neural network, extract the time series features representing the temperature changes of the nodes from the spatial dependency features using a recurrent neural network, determine the periodic trend features of the time series features using an attention mechanism, and map the extracted features into predicted node temperature values after fusion. The temperature field prediction model is trained using the spatial distribution cloud map of the structure temperature and the time series data of the node temperature. The temperature field distribution of the steel structure under test is determined by inputting the data from a limited number of monitoring points on the steel structure under test into the trained temperature field prediction model.
[0007] One possible implementation involves establishing a three-dimensional geometric model of the steel structure to be tested, including: Identify the nodes and components of the steel structure to be tested; Assign coordinates, equivalent surface area, convective heat transfer coefficient, and ambient temperature attributes to the nodes; Assign the component the attributes of length, cross-sectional area, surface area, and solar radiation. A three-dimensional geometric model of the steel structure under test is constructed based on the nodes and components with assigned attributes.
[0008] One possible implementation involves using the finite element method to perform transient thermal analysis on each node inside the steel structure under test, obtaining a spatial temperature distribution cloud map of the structure and time series data of node temperatures, including: Simulate the steel structure component under test and determine the equivalent heat generation rate based on the component properties corresponding to the steel structure under test; The heat dissipation rate of a node is determined by its equivalent surface area, convective heat transfer coefficient, and ambient temperature. A transient thermal analysis finite element model of the steel structure under test is established. The three-dimensional geometric model is divided into elements and assigned material thermal property parameters, including density, specific heat capacity and thermal conductivity. The equivalent heat generation rate of the component is applied to the corresponding component unit as a bulk heat source, and the heat dissipation rate of the node is applied to the node as an equivalent convective heat transfer boundary condition, and the ambient temperature and initial temperature field are set. Set the total calculation time and time step of the transient thermal analysis, solve the nodal temperature distribution of the structure at each time step, and obtain the temperature calculation results in the whole time domain. Map the node temperature results at any time step to the three-dimensional geometric model to generate a spatial distribution cloud map of the structural temperature at the corresponding time. The temperature values of each node at each time step are extracted in chronological order and stored according to the node number to form node temperature time series data.
[0009] One possible implementation involves transforming the three-dimensional geometric model into a graphical model that reflects the topological features of the steel structure under test, including: Extract all structural nodes from the three-dimensional geometric model to form a node set for the graph model; The three-dimensional spatial coordinates, equivalent surface area, convective heat transfer coefficient, and ambient temperature at the corresponding time of each structural node are used as the node feature matrix of the graph model. Extract the structural components that connect each structural node in the three-dimensional geometric model to form the edge set of the graph model; The edge feature matrix of the graph model is constructed based on the length, cross-sectional area, surface area, and solar radiation of each structural component.
[0010] In one possible implementation, the temperature field prediction model takes the node feature matrix and edge feature matrix as input and uses a graph neural network to extract the spatial dependency features of each node, including: For each structural node in the graph model, determine the neighboring nodes that are directly connected to the structural node through components; Obtain the node features of the neighboring nodes and the edge features connecting the neighboring nodes and the structural nodes; Information aggregation is performed on the node features and the edge features; Spatial feature update is performed on the aggregated features using an activation function; After multi-layer graph convolution, the spatial embedding vector of each structural node is obtained, which is used to represent the spatial dependency features of the structural node.
[0011] One possible implementation involves using a recurrent neural network to extract time-series features characterizing node temperature changes from the spatially dependent features, including: The spatial dependency features are input into the gated loop unit. The update gate inside the gated loop unit controls the degree to which the state of the previous moment is maintained, and the reset gate inside the gated loop unit controls the degree to which the state of the previous moment affects the current input. Based on the adjustment results of the update gate and the reset gate, calculate the hidden state output at the current moment; By utilizing the update process of the hidden state output, the time series features of node temperature changes over time are extracted.
[0012] One possible implementation involves using an attention mechanism to determine the periodic trend characteristics of the time series features, including: Obtain the time series features output by the recurrent neural network; By introducing a query, key, and value matrix, attention weights are obtained by calculating the correlation between different time steps; The attention weights are used to weight and aggregate the time series features to generate a comprehensive feature vector in the time dimension, which is used to characterize the global periodic change trend of the temperature field.
[0013] One possible implementation involves fusing the extracted features and mapping them to predicted node temperatures, including: The spatial dependency features, time series features, and periodic trend features are fused at the vector level using a vector concatenation method to obtain a comprehensive feature vector; A linear transformation is performed on the comprehensive feature vector to map it to the predicted temperature value of the corresponding structural node.
[0014] One possible implementation involves training the temperature field prediction model using the spatial distribution cloud map of the structure's temperature and the time series data of the node's temperature, including: Using the node temperature time series data as supervision labels, the node feature matrix, edge feature matrix and environmental parameters at the corresponding time are used as inputs to the temperature field prediction model, and the network parameters are optimized by minimizing the loss function of temperature prediction error.
[0015] One possible implementation involves training the temperature field prediction model, which further includes: Based on the Bayesian optimization method, the key hyperparameters of the temperature field prediction model are automatically searched and combined for optimization; wherein, the key hyperparameters include the number of hidden units, the number of network layers, the dimension of hidden states in the GRU, the number of attention heads in the self-attention mechanism, the learning rate, the L2 regularization coefficient, and the batch size.
[0016] Compared with the prior art, the technical solution provided in this application has the following beneficial effects: This application transforms the physical topology and connection relationships of large-span spatial steel structures into a graph structure model, and uses a spatiotemporal graph neural network as the core algorithm to achieve accurate modeling of the spatial correlation characteristics of complex structures. By aggregating neighbor node information through a graph convolutional network and learning the heat transfer correlation between nodes, a recurrent neural network is used to capture the dynamic law of temperature field evolution over time, and a self-attention mechanism is introduced to identify long-term trends such as solar radiation, ambient temperature, and component shading. By comprehensively considering the influence of various complex time-varying factors such as solar radiation, ambient temperature, and component shading, the application achieves accurate and rapid prediction of the spatiotemporal distribution law of the entire temperature field. Attached Figure Description
[0017] Figure 1 The flowchart shows a method for predicting the temperature field of a large-span spatial steel structure based on a spatiotemporal graph neural network, which is provided in Embodiment 1 of this application.
[0018] Figure 2This is a schematic diagram of the spatiotemporal graph neural network architecture provided in Embodiment 1 of this application.
[0019] Figure 3 This is a flowchart of a method for extracting spatial dependency features between structural nodes using a graph neural network, as provided in Embodiment 1 of this application.
[0020] Figure 4 This is a schematic diagram of the model prediction results provided in Embodiment 2 of this application. Detailed Implementation
[0021] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0022] Example 1 See Figure 1 This is a flowchart of a method for predicting the temperature field of a large-span spatial steel structure based on a spatiotemporal graph neural network, provided in Embodiment 1 of this application. Figure 1 As shown, the specific implementation steps of the above method include: Step 101: Establish a three-dimensional geometric model of the steel structure to be tested, and perform transient thermal analysis on each node inside the steel structure to be tested using the finite element method to obtain the spatial distribution cloud map of the structural temperature and the time series data of the node temperature.
[0023] In this embodiment, the aforementioned three-dimensional geometric model is constructed by determining the node and component information of the steel structure to be tested and configuring corresponding attributes for each node and component. Node attributes include coordinates, equivalent surface area, convective heat transfer coefficient, and ambient temperature. Component attributes include length, cross-sectional area, surface area, and solar radiation.
[0024] Using the aforementioned three-dimensional geometric model as the structural thermal analysis model, transient thermal analysis is performed on each node inside the steel structure under test. Specifically, this application uses the ANSYS transient thermal analysis module and LINK33 elements to simulate structural components, calculating the equivalent heat generation rate through solar radiation and component properties. The heat dissipation rate is calculated through the equivalent surface area of nodes, convective heat transfer coefficient, and ambient temperature. The structural temperature is obtained by combining the heat generation of components and the heat dissipation of nodes, thus yielding the aforementioned spatial distribution cloud map of structural temperature and time series data of node temperature.
[0025] Specifically, this application establishes a transient thermal analysis finite element model of the steel structure under test. The three-dimensional geometric model is divided into elements and assigned material thermal property parameters, including density, specific heat capacity, and thermal conductivity. The equivalent heat generation rate of the components is applied as a volume heat source to the corresponding component elements. The heat dissipation rate of the nodes is applied as a convection heat transfer boundary condition to the nodes, and the ambient temperature and initial temperature field are set. The total calculation time and time step of the transient thermal analysis are set, and the nodal temperature distribution of the structure at each time step is solved step by step to obtain the temperature calculation results in the full time domain. The nodal temperature results of any time step are mapped to the three-dimensional geometric model to generate a spatial distribution cloud map of the structural temperature at the corresponding time. The temperature values of each node at each time step are extracted in chronological order and stored according to the node number to form nodal temperature time series data.
[0026] Step 102: Transform the above three-dimensional geometric model into a graph model that reflects the topological features of the steel structure to be tested, and determine the node feature matrix and edge feature matrix in the graph model.
[0027] Specifically, the above graphical model consists of four elements, which can be formally represented as follows: .
[0028] In the formula, V This represents the set of nodes, including all structural nodes in the large-span spatial steel structure, including top chord nodes, bottom chord nodes, web member nodes, and support nodes. Each node corresponds to a component connection point in the structure. The spatial coordinates of each node are obtained through finite element geometric modeling. In this embodiment, the total number of nodes is denoted as... N The node numbers are arranged in geometric coordinate order.
[0029] The above E Let be the set of edges, representing the connections between nodes formed by components. Each edge... Connect two nodes and Each edge corresponds to an actual component in the structure. The edge set includes all components. The edge set is stored in the form of an adjacency matrix. and If there are component connections, then ,otherwise The total number of edges is denoted as The edge numbering should be consistent with the component numbering.
[0030] The above X This is the node feature matrix. Representative dimension A numerical matrix is used to describe the physical and environmental parameters of each node. Each node's feature vector contains... d Each feature term includes three-dimensional spatial coordinates ( xi , y i , z i ), equivalent surface area S n convective heat transfer coefficient h c Ambient temperature at the corresponding time T a .
[0031] E f This is the edge feature matrix. The node feature matrix is also shown. E f Representative dimension is A numerical matrix is used to record the physical properties of each edge (i.e., component). Each edge feature vector contains... k Each feature term includes length. l ij Cross-sectional area AE ij Surface area SE ij Solar radiation HE ij .
[0032] Step 103: Construct a temperature field prediction model based on the above graphical model.
[0033] In this embodiment, a temperature field prediction model is established based on graph neural networks, recurrent neural networks, and an attention mechanism. The aforementioned temperature field prediction model is configured as a spatiotemporal graph neural network architecture, such as... Figure 2 As shown in the image.
[0034] Specifically, the temperature field prediction model includes a spatial feature extraction module, a temporal feature extraction module, a periodic trend capture module, a feature fusion module, and an output mapping module.
[0035] The aforementioned spatial feature extraction module is configured to extract spatial dependency features between structural nodes using a graph neural network, taking the node feature matrix and edge feature matrix as input. This module receives the node feature matrix at each time step. Sum of edge feature matrices And based on the edge set Information is aggregated based on the defined topological relationships.
[0036] Specifically, see Figure 3 This is a flowchart of a method for extracting spatial dependency features between structural nodes using a graph neural network, provided in Embodiment 1 of this application. Figure 3 As shown, the above method specifically includes: Step 201: For each structural node in the graphical model, determine the neighboring nodes that are directly connected to the structural node through components.
[0037] Step 202: Obtain the node features of neighboring nodes and the edge features connecting neighboring nodes and structural nodes.
[0038] Step 203: Aggregate information on node features and edge features.
[0039] Step 204: Update the spatial features of the aggregated features using an activation function.
[0040] Specifically, for any node The spatial feature update process is defined as follows:
[0041] In the formula, N( i ) is a node The set of adjacent nodes, For the current node at the th l Features of the layer e ij As edge features, ( • ) is the edge information aggregation function. ( • ) is the activation function. Indicates the neighbor node at the th l Characteristics of the layer.
[0042] Step 205: After multi-layer graph convolution, obtain the spatial embedding vector of each structural node. , used to represent the spatial dependency characteristics of structural nodes.
[0043] The aforementioned spatial feature extraction module can automatically learn the thermal conduction correlation between nodes based on the structural topology, thereby enabling the feature extraction of temperature coupling effects between different components.
[0044] The aforementioned time feature extraction module, built on a recurrent neural network (GRU), is used to extract the dynamic features of temperature changes of structural nodes over time and to characterize the dependencies of nodes in the time series dimension.
[0045] Specifically, the node embedding vector sequence obtained by the spatial feature extraction module at multiple time steps Using the above spatial dependency features as input, a recurrent neural network is used to extract time-series features representing node temperature changes from the aforementioned spatial dependency features. In this embodiment, the above spatial dependency features are input to a gated recurrent unit (GRU). The update gate within the GRU controls the degree to which the previous state is maintained, and the reset gate within the GRU controls the degree to which the previous state affects the current input. The GRU-based gating structure enables the memorization and updating of historical temperature features. The state update process of the GRU is defined as follows: ; ; ; ;
[0046] In the formula, For nodes at time steps Spatial eigenvectors. For time step The hidden state output. The gate vector is used to update the state and control the degree to which the previous state is preserved. The reset gate vector is used to control the degree to which the state of the previous time step affects the current input. This is the candidate state vector. and This is the model weight matrix. This is a bias term. This indicates the Hadamard element-wise multiplication operation. This is the Sigmoid activation function.
[0047] Based on the adjustment results of the aforementioned update and reset gates, this application determines the hidden state output at the current moment. Utilizing the update process of the hidden state output, the time-series features of node temperature changes over time are extracted. A gating mechanism is used to achieve dynamic memorization and forgetting of the temperature time series, enabling the model to accurately capture the temperature evolution patterns caused by diurnal variations, seasonal fluctuations, etc.
[0048] The aforementioned periodic trend capture module is configured to use an attention mechanism to determine the periodic trend features of the aforementioned time series features. Specifically, the periodic trend capture module uses the node time feature sequence output by the GRU. As input, feature weighting aggregation is achieved by calculating attention weights at different time steps, as shown below:
[0049] In the formula, These are query, key, and value matrices, respectively. For feature dimensions.
[0050] The weighted output is represented as a comprehensive feature vector in the time dimension. This model is used to describe the global periodic variation trend of the temperature field. Through a self-attention mechanism, the model can identify the diurnal cycle, seasonal variation, and nonlinear time-lag effects of temperature, thereby improving the global stability and accuracy of the prediction.
[0051] The aforementioned feature fusion module is specifically used to concatenate and fuse the extracted features into a vector to obtain a comprehensive feature vector. , is represented as:
[0052] In the formula, " indicates a vector-level concatenation operation. The nodes representing the output of the spatial feature extraction module i At time step t Spatial feature vectors, The node representing the output of the time feature extraction module i At time step t The node time-dependent feature vector, The node representing the output of the periodic trend capture module i Global periodic feature vector.
[0053] This fusion method can achieve a comprehensive expression of nodes in three information dimensions: spatial topology, temporal evolution, and periodic change, while maintaining the independence of various feature spaces, thereby improving the model's ability to fit the complex coupling effects of structural temperature fields.
[0054] The aforementioned output mapping module, built upon a fully connected layer, maps the fused integrated feature vector to the corresponding node's predicted temperature value. This module obtains the node's predicted temperature value by performing a linear transformation on the node's integrated feature vector. Predicted temperature at the next time step , is represented as:
[0055] In the formula, Represents the weight matrix. For bias vectors, For nodes The predicted temperature value.
[0056] Based on the output mapping module, the mapping process from the high-dimensional feature space to specific temperature scalars is realized, completing the model's node-level prediction of the structural temperature field. By combining the predicted temperatures of all nodes, the temperature field distribution of the large-span spatial steel structure at the next time step can be obtained.
[0057] Step 104: Train the temperature field prediction model using the above-mentioned structural temperature spatial distribution cloud map and node temperature time series data.
[0058] Step 105: Input the data from a limited number of monitoring points on the steel structure to be tested into the trained temperature field prediction model to determine the temperature field distribution of the steel structure to be tested.
[0059] Compared with the prior art, the technical solution provided in Embodiment 1 of this application has the following beneficial effects: (1) This application transforms the physical topology and connection relationship of a large-span spatial steel structure into a graph structure model and uses a spatiotemporal graph neural network as the core algorithm to achieve accurate modeling of the spatial correlation characteristics of complex structures. By aggregating neighbor node information through a graph convolutional network, the heat transfer correlation between nodes is learned. A recurrent neural network is used to capture the dynamic law of temperature field evolution over time, and a self-attention mechanism is introduced to identify long-term trends such as solar radiation cycle. The influence of various complex time-varying factors such as solar radiation, ambient temperature, and component shading is comprehensively considered, and accurate and rapid prediction of the spatiotemporal distribution law of the temperature field of the entire field is achieved.
[0060] (2) This application realizes an efficient and high-precision temperature field extrapolation method based on spatiotemporal graph neural network. It can quickly invert and predict the overall temperature field distribution of the structure based on real-time monitoring data of limited measurement points and a trained model. It effectively solves the limitations of traditional finite element method, such as low calculation efficiency, difficulty in real-time early warning, high cost due to reliance on dense sensor deployment, and inability to predict future trends. It provides intelligent technical support for structural safety assessment and risk control during construction.
[0061] Example 2 In a specific embodiment, a steel roof structure of a stadium is used as an example. First, 3012 node coordinates and information such as the length, cross-sectional area, and surface area of 8092 components are input into ANSYS. A three-dimensional geometric model of the structure is established using LINK33 elements. The equivalent surface area of the nodes is calculated based on the connection relationship between the nodes and components. Solar radiation on the component surface is calculated based on a clear sky solar radiation model, and the equivalent heat generation rate is calculated in conjunction with the component dimensions. The heat dissipation rate is calculated using the equivalent surface area of the nodes, the convective heat transfer coefficient, and the ambient temperature. A temporal and spatial distribution dataset of the structural temperature field is obtained through transient thermal analysis. The three-dimensional geometric model is then converted into a graphical model using node sets. Edge set Node feature matrix and edge feature matrix The data constitutes the structure of the graph model.
[0062] Subsequently, the graph model and the time-series temperature field data are input into the spatiotemporal graph neural network model constructed in step S3. The model first uses the spatial feature extraction module to perform graph convolution operations on the topological relationships between nodes, obtaining the spatial feature vector of each node at each time step. Then, the temporal feature extraction module performs GRU operations on the time series of node temperatures to extract the dynamic features of temperature changes over time. Next, the periodic trend capture module performs a self-attention mechanism calculation on the historical temperature series to obtain the periodicity and long-term dependence features of temperature changes. Then, the feature fusion module performs vector-level combination of the above spatial, temporal, and periodic features to form a comprehensive node feature vector. Finally, the output mapping module maps this comprehensive feature vector to the predicted node temperature value.
[0063] During the model training phase, the time-series temperature field data obtained from finite element transient analysis is used as supervision labels, and the node feature matrices at corresponding time points are used as supervision labels. Edge feature matrix Using environmental parameters as model inputs, the network parameters are optimized by minimizing the loss function of temperature prediction error.
[0064] To achieve optimal training performance, this embodiment introduces a Bayesian optimization method during training to automatically search for and optimize the model's key hyperparameters. The hyperparameters involved in optimization include the number of hidden units in the graph neural network, the number of network layers, the dimension of the GRU hidden states, the number of attention heads in the self-attention mechanism, the learning rate, the L2 regularization coefficient, and the batch size. Bayesian optimization uses the validation set prediction error as the objective function to search for the optimal combination of hyperparameters that maximizes model performance. Based on this optimal combination, the Adam optimization algorithm is then used to complete model training.
[0065] After the model training is completed, a small amount of temperature data from a few points in the structure is input into the trained model to obtain the temperature inversion results of all nodes in the structure.
[0066] To verify the predictive performance of this method, 48 hours of continuous temperature data from a node on the upper chord of the roof steel structure were selected and compared with the model's predicted temperature. The results are as follows: Figure 4 As shown, the node temperature predicted by this method is highly consistent with the simulation results (R2=0.9989). This result shows that the temperature field prediction method based on spatiotemporal graph neural network proposed in this application can accurately invert the overall temperature field of the structure under limited sensor deployment conditions and achieve effective prediction of the temperature field at future times.
[0067] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for predicting the temperature field of large-span spatial steel structures based on spatiotemporal graph neural networks, characterized in that, include: A three-dimensional geometric model of the steel structure to be tested is established. Transient thermal analysis of each node inside the steel structure is performed using the finite element method to obtain the spatial distribution cloud map of the structural temperature and the time series data of the node temperature. The three-dimensional geometric model is transformed into a graph model to reflect the topological features of the steel structure under test, and the node feature matrix and edge feature matrix in the graph model are determined. A temperature field prediction model is constructed based on the graph model; the temperature field prediction model is configured to use the node feature matrix and edge feature matrix as inputs and extract the spatial dependency features of each node using a graph neural network. A recurrent neural network is used to extract time-series features representing node temperature changes from the spatially dependent features; The periodic trend features of the time series features are determined by using an attention mechanism, and the extracted features are fused and mapped to the node temperature prediction values. The temperature field prediction model is trained using the spatial distribution cloud map of the structure temperature and the time series data of the node temperature. The temperature field distribution of the steel structure under test is determined by inputting the data from a limited number of monitoring points on the steel structure under test into the trained temperature field prediction model.
2. The method for predicting the temperature field of large-span spatial steel structures based on spatiotemporal graph neural networks according to claim 1, characterized in that, Establish a three-dimensional geometric model of the steel structure to be tested, including: Identify the nodes and components of the steel structure to be tested; Assign coordinates, equivalent surface area, convective heat transfer coefficient, and ambient temperature attributes to the nodes; Assign the component the attributes of length, cross-sectional area, surface area, and solar radiation. A three-dimensional geometric model of the steel structure under test is constructed based on the nodes and components with assigned attributes.
3. The method for predicting the temperature field of large-span spatial steel structures based on spatiotemporal graph neural networks according to claim 2, characterized in that, Transient thermal analysis of each node inside the steel structure under test was performed using the finite element method to obtain spatial temperature distribution cloud maps and time series data of node temperatures, including: Simulate the steel structure component under test and determine the equivalent heat generation rate based on the component properties corresponding to the steel structure under test; The heat dissipation rate of a node is determined by its equivalent surface area, convective heat transfer coefficient, and ambient temperature. A transient thermal analysis finite element model of the steel structure under test is established. The three-dimensional geometric model is divided into elements and assigned material thermal property parameters, including density, specific heat capacity and thermal conductivity. The equivalent heat generation rate of the component is applied to the corresponding component unit as a bulk heat source, and the heat dissipation rate of the node is applied to the node as an equivalent convective heat transfer boundary condition, and the ambient temperature and initial temperature field are set. Set the total calculation time and time step of the transient thermal analysis, solve the nodal temperature distribution of the structure at each time step, and obtain the temperature calculation results in the whole time domain. Map the node temperature results at any time step to the three-dimensional geometric model to generate a spatial distribution cloud map of the structural temperature at the corresponding time. The temperature values of each node at each time step are extracted in chronological order and stored according to the node number to form node temperature time series data.
4. The method for predicting the temperature field of large-span spatial steel structures based on spatiotemporal graph neural networks according to claim 1, characterized in that, The three-dimensional geometric model is transformed into a graphical model that reflects the topological features of the steel structure under test, including: Extract all structural nodes from the three-dimensional geometric model to form a node set for the graph model; The three-dimensional spatial coordinates, equivalent surface area, convective heat transfer coefficient, and ambient temperature at the corresponding time of each structural node are used as the node feature matrix of the graph model. Extract the structural components that connect each structural node in the three-dimensional geometric model to form the edge set of the graph model; The edge feature matrix of the graph model is constructed based on the length, cross-sectional area, surface area, and solar radiation of each structural component.
5. The method for predicting the temperature field of large-span spatial steel structures based on spatiotemporal graph neural networks according to claim 1, characterized in that, The temperature field prediction model takes the node feature matrix and edge feature matrix as input and uses a graph neural network to extract the spatial dependency features of each node, including: For each structural node in the graph model, determine the neighboring nodes that are directly connected to the structural node through components; Obtain the node features of the neighboring nodes and the edge features connecting the neighboring nodes and the structural nodes; Information aggregation is performed on the node features and the edge features; Spatial feature update is performed on the aggregated features using an activation function; After multi-layer graph convolution, the spatial embedding vector of each structural node is obtained, which is used to represent the spatial dependency features of the structural node.
6. The method for predicting the temperature field of large-span spatial steel structures based on spatiotemporal graph neural networks according to claim 1, characterized in that, The time-series features characterizing node temperature changes in the spatially dependent features are extracted using a recurrent neural network, including: The spatial dependency features are input into the gated loop unit. The update gate inside the gated loop unit controls the degree to which the state of the previous moment is maintained, and the reset gate inside the gated loop unit controls the degree to which the state of the previous moment affects the current input. Based on the adjustment results of the update gate and the reset gate, calculate the hidden state output at the current moment; By utilizing the update process of the hidden state output, the time series features of node temperature changes over time are extracted.
7. The method for predicting the temperature field of large-span spatial steel structures based on spatiotemporal graph neural networks according to claim 1, characterized in that, Determining the periodic trend characteristics of the time series features using an attention mechanism includes: Obtain the time series features output by the recurrent neural network; By introducing a query, key, and value matrix, attention weights are obtained by calculating the correlation between different time steps; The attention weights are used to weight and aggregate the time series features to generate a comprehensive feature vector in the time dimension, which is used to characterize the global periodic change trend of the temperature field.
8. The method for predicting the temperature field of large-span spatial steel structures based on spatiotemporal graph neural networks according to claim 1, characterized in that, The extracted features are fused and mapped to node temperature prediction values, including: The spatial dependency features, time series features, and periodic trend features are fused at the vector level using a vector concatenation method to obtain a comprehensive feature vector; A linear transformation is performed on the comprehensive feature vector to map it to the predicted temperature value of the corresponding structural node.
9. The method for predicting the temperature field of large-span spatial steel structures based on spatiotemporal graph neural networks according to claim 1, characterized in that, The temperature field prediction model is trained using the spatial distribution cloud map of the structure's temperature and the time series data of the node temperature, including: Using the node temperature time series data as supervision labels, the node feature matrix, edge feature matrix and environmental parameters at the corresponding time are used as inputs to the temperature field prediction model, and the network parameters are optimized by minimizing the loss function of temperature prediction error.
10. The method for predicting the temperature field of large-span spatial steel structures based on spatiotemporal graph neural networks according to claim 1, characterized in that, Training the temperature field prediction model further includes: Based on the Bayesian optimization method, the key hyperparameters of the temperature field prediction model are automatically searched and combined for optimization; wherein, the key hyperparameters include the number of hidden units, the number of network layers, the dimension of hidden states in the GRU, the number of attention heads in the self-attention mechanism, the learning rate, the L2 regularization coefficient, and the batch size.