A dynamic graph neural network driven high arch dam spatiotemporal deformation prediction method

Abstract

The application discloses a dynamic graph neural network driven high arch dam space-time deformation prediction method, relates to the technical field of dam operation safety monitoring and management, and comprises the following steps: preprocessing dam body deformation, water level, temperature and other monitoring data, and constructing a prediction data set containing measuring point three-dimensional coordinates and radial deformation; combining the measuring point coordinates and historical deformation to generate a static candidate adjacency matrix, dynamically adjusting the correlation weight according to the window characteristics through cross attention, and obtaining a dynamic adjacency matrix; taking the dynamic adjacency matrix as a constraint, extracting the spatial coupling characteristics of the measuring points by GCN, modeling the long-time sequence deformation law by Mamba, constructing a GCN-Mamba model, iteratively training the optimal offline model according to the prediction error, and reserving the optimal offline model; inputting real-time monitoring data to complete deformation prediction, calculating online loss, fine-tuning local parameters by using a small learning rate, and realizing online adaptive updating of the model. The application adopts the above method, has high prediction accuracy and supports online adaptive updating.

Description

Technical Field

[0001] This invention relates to the field of dam operation safety monitoring and management technology, and in particular to a method for predicting the spatiotemporal deformation of high arch dams driven by a dynamic graph neural network. Background Technology

[0002] Due to their large volume, height, wide range of water level fluctuations, and complex spatiotemporal deformation distribution under load, ultra-high arch dams are extremely complex nonlinear static indeterminate structures. For these types of high arch dams, numerous monitoring points are deployed during engineering projects, resulting in long-term, multi-point, and continuous deformation monitoring data. Deformation monitoring models for high arch dams are an important component of dam safety monitoring; accurately predicting deformation at monitoring points helps identify abnormal trends and assists in safety assessments and operational management.

[0003] Existing methods for predicting the deformation of high arch dams have gradually evolved from traditional statistical models to deep learning models, but they still have shortcomings in multi-measuring-point collaborative prediction and online applications. Early studies were mostly based on water pressure-temperature-aging models and their improved forms to interpret and predict dam deformation. These methods have good engineering interpretability and can reflect the influence of factors such as water level, temperature, and aging on dam deformation. However, most of them still focus on single-measuring-point modeling, which makes it difficult to fully characterize the collaborative deformation relationship between different dam sections, different elevations, and different measuring points of high arch dams.

[0004] With the development of machine learning and deep learning methods, LSTM, GRU, attention mechanisms, combined prediction, and residual correction methods have been widely used for dam deformation prediction. These methods enhance the model's ability to learn nonlinear time dependencies and responses to complex environments, but their main goal is still to improve the time series prediction accuracy of single points or a small number of measurement points. They still do not adequately consider the spatial correlation, response similarity, and structural transmission relationships between multiple measurement points.

[0005] To address the problem that single-point modeling cannot accurately reflect the overall response of the dam, multi-point collaborative prediction has gradually become an important direction in the research of deformation monitoring of high arch dams in recent years. However, existing methods mostly rely on clustering partitions, correlation coefficients, spatial distances, or empirical settings to express the relationships between measurement points. The correlation between measurement points is usually fixed before modeling, making it difficult to adapt to the dynamic changes in spatial correlation under different water level stages, temperature effects, and operating conditions. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting the spatiotemporal deformation of high arch dams driven by dynamic graph neural networks, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, this invention provides a method for predicting the spatiotemporal deformation of high arch dams driven by a dynamic graph neural network, comprising: S1. Collect and preprocess the high arch dam deformation and environmental monitoring data by cleaning outliers and interpolating missing values. Based on the basic principle of the arch dam spatiotemporal deformation tracking and prediction model, determine the relationship between environmental loads such as upstream and downstream water levels and temperatures and the deformation response of the measuring points. Construct a high arch dam deformation tracking and prediction dataset and divide it into training set, validation set and prediction set in a ratio of 8:1:1. S2. Construction of a dynamic-static adjacency matrix: Based on the 3D spatial coordinates of the measurement points in the training set and historical deformation response data, static spatial associations are learned to generate a static candidate adjacency matrix; then, based on the node features of the measurement points in the current input window, the edge association strength is dynamically adjusted through a cross-attention mechanism to output a dynamic adjacency matrix. ; S3. Offline training of the spatiotemporal prediction network: using a dynamic adjacency matrix. To constrain the graph structure, a graph convolutional network (GCN) is used to aggregate the spatial coupling features of measurement points, and a Mamba network is used to capture the long-term temporal dependencies of deformation at a single measurement point, thus building a GCN-Mamba spatiotemporal deformation prediction model. The network parameters are iteratively updated using a loss function constructed based on the error between the predicted and actual deformation values, and the model with the best performance on the validation set is saved as the initial model for the online prediction stage. S4. Online Real-Time Prediction and Incremental Fine-Tuning: Real-time monitoring data collected on-site is input into the trained optimal initial model, and the dynamic adjacency matrix... The deformation prediction results of each measuring point are output under constraints; the online prediction loss is calculated by comparing with the measured deformation values; and the network parameters of the model are incrementally updated using a small learning rate to achieve adaptive updating of the model.

[0008] Preferably, the high arch dam deformation tracking and prediction dataset in step S1 includes upstream water level, time factor, temperature factor, spatial three-dimensional coordinates, and radial deformation automatically monitored at each measuring point of the high arch dam.

[0009] Preferably, step S2 contains the following: S21. Use GRU and MLP to extract the historical response embedding information and spatial location embedding information of each measuring point, respectively. S22. The connection strength between any two measurement points is directly learned through the dual-channel Pairwise Edge Scorer and the static candidate adjacency matrix is ​​obtained by adaptive Top-K sparsification. S23. Based on the construction of a static candidate adjacency matrix, a cross-attention mechanism is used to dynamically adjust the association strength of candidate edges according to the monitoring status of the current window, resulting in a dynamic adjacency matrix. .

[0010] Preferably, step S21 specifically includes: For the i-th measurement point, its historical deformation sequence during the training phase is as follows: The historical deformation sequence is encoded using GRU, compressing the long-term deformation evolution characteristics of the measurement points into historical responses, ultimately yielding the historical response embedding information: ; Spatial coordinates of the measuring point Input coordinate encoding MLP to obtain spatial location embedding information: .

[0011] Preferably, step S22 includes: Measuring points , The GRU historical response embedding information and MLP spatial location embedding information are fed into the dual-channel PairwiseEdge Scorer to directly learn the connection strength between the two measurement points. The measurement points are obtained by fusing the dual-channel scores. and measuring points Static correlation strength between ,satisfy: ; in, ; This represents the Sigmoid activation function, which maps the input to the (0,1) interval and is used to generate weight coefficients; Indicates the characteristics of sensor readings , The calculation yields the representative measurement point. Correlation information between sensor measurements; Indicates the measurement point Information relating to spatial locations; Top-K sparsity is applied to the static correlation strength for each measurement point. i After removing self-connects, select the connection with the highest strength. K Each measurement point is selected as a candidate neighbor: ; And finally, the static candidate adjacency matrix is ​​obtained: .

[0012] Preferably, step S23 specifically includes: The cross-attention mechanism dynamically adjusts the association strength of candidate edges based on the current monitoring status in the sliding window: ; ; in, This represents the dynamic nodes obtained from the current sliding window; For dynamic monitoring status query; , These are static prior keys and values; The node structure representation obtained from static prior information; , , These are the learnable parameter matrices; It is a static set of candidate neighbors; For the implicit vector dimension; These are learnable parameters used to adaptively adjust the influence of static prior scores on dynamic graph updates.

[0013] Preferably, step S3 specifically includes: Let the node feature matrix at a certain time step be... The dynamic adjacency matrix is To preserve node information, we first add self-joins to the dynamic adjacency matrix, and then normalize the dynamic adjacency matrix. ; ; in, It is the identity matrix; for The degree matrix; This is the normalized dynamic adjacency matrix; The dynamic adjacency matrix after adding self-connections; Based on the normalized dynamic adjacency matrix, the graph convolutional layer is represented as: ; in, For the first l Layer node representation, For learnable weight matrix, It is a non-linear activation function; After graph convolution processing, a node temporal representation with fused spatial correlation information is obtained. The temporal data is then rearranged according to the measurement point dimension to obtain an independent temporal sequence for each measurement point. Enter the Mamba module; For the input sequence Mamba generates a time step, input mapping, and output mapping based on the current input: ; ; in, This represents the discretization step size of the current time step; and This represents the relevant mapping parameters of the input; Indicates A learnable nonlinear mapping network as input; After the selectivity parameters are determined, the continuous state is discretized as follows: ; ; No. k The selective states at each time step are recursively derived as follows: ; ; in, This represents the hidden state at the current time step. Here, D represents the Mamba output representation; D represents the input mapping matrix in the Mamba selective state-space model, which maps the input features. The output is directly mapped and used as the direct input to the output branch; Mamba then receives the spatially augmented sequence after dynamic graph convolution: ; in, The output of the last time step is taken as the temporal state representation of the current window, and the future deformation prediction value is obtained through the prediction layer: ; in, This represents the output characteristics of the Mamba module at the last time step within this time window.

[0014] Therefore, the above-mentioned method for predicting the spatiotemporal deformation of high arch dams driven by a dynamic graph neural network has the following beneficial effects: (1) For the problem of deformation prediction at multiple measuring points of high arch dams, a dynamic graph GCN-Mamba model is proposed to realize the joint modeling of spatial correlation, dynamic relationship and long time series features.

[0015] (2) Design a static adjacency matrix learning and dynamic adjacency generation mechanism to avoid relying entirely on manual setting of measurement point connection relationships and improve the model's ability to express complex associations of dam measurement points.

[0016] (3) A prediction process combining offline training and online updating was constructed and compared and verified with LSTM, GRU, Mamba, GRU-GCN and other models on multi-point monitoring data of Xiaowan Arch Dam. This provides a new approach for intelligent and accurate prediction of dam deformation response during operation.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the construction of a dynamic graph neural network-driven method for predicting the spatiotemporal deformation of high arch dams according to the present invention; Figure 2 This is a time variation diagram of data from various monitoring points of the dam according to an embodiment of the present invention; wherein, (a) the time variation of monitoring point data for monitoring line 09 is shown; (b) the time variation of monitoring point data for monitoring line 15 is shown; (c) the time variation of monitoring point data for monitoring line 19 is shown; (d) the time variation of monitoring point data for monitoring line 22 is shown; (e) the time variation of monitoring point data for monitoring line 25 is shown; (f) the time variation of monitoring point data for monitoring line 29 is shown; and (g) the time variation of monitoring point data for monitoring line 35 is shown. Figure 3 This is a static association prior diagram of an embodiment of the present invention; wherein, (a) is a schematic diagram of the detection point layout and structural skeleton; and (b) is a schematic diagram of the optimized static prior topology. Figure 4 This is a dynamic weight heatmap of a dynamic matrix according to an embodiment of the present invention; wherein, (a) is a spatial heatmap of node importance; and (b) is a schematic diagram of the comprehensive weight between key points. Figure 5 The following are comparison charts and residual charts of online model prediction fitting for measurement point A22-PL-01, as well as the online model prediction residual chart for measurement point A22-PL-01. (a) is the comparison chart of online model prediction fitting for measurement point A22-PL-02; (b) is the comparison chart of online model prediction residual for measurement point A22-PL-03; (c) is the comparison chart of online model prediction fitting for measurement point A22-PL-03; (d) is the comparison chart of online model prediction residual for measurement point A22-PL-05; (e) is the comparison chart of online model prediction fitting for measurement point A22-PL-05; (f) is the comparison chart of online model prediction residual for measurement point A22-PL-05. Detailed Implementation

[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] Example Please see Figure 1 This invention provides a method for predicting the spatiotemporal deformation of high arch dams driven by a dynamic graph neural network, comprising: S1. Collect and preprocess the monitoring data of dam deformation and environmental quantities by cleaning outliers and interpolating missing values, construct a dataset of dam deformation and its explanatory variables, and divide the dataset into training set, validation set and prediction set in a ratio of 8:1:1.

[0022] S2. Construct a static adjacency matrix using the 3D spatial coordinates of the training set and the deformation response data of each measurement point in the training set; then, under the static adjacency matrix, dynamically generate an adjacency matrix based on the node features of the current input window. A t .

[0023] S3. Based on the dynamic adjacency matrix, GCN performs spatial fusion of each measuring point, and Mamba models the deformation time dependency of each measuring point. A spatiotemporal deformation tracking and prediction model for high arch dams based on GCN-Mamba is constructed. The model parameters are updated by using the error between the predicted value and the true value as the loss function, and the optimal model is saved as the initial model for the online stage.

[0024] S4. Online phase prediction will use real-time received monitoring data as input, and... A t The prediction results are output using the GCN-Mamba model, and the online loss is calculated simultaneously. Some parameters are incrementally updated using a small learning rate.

[0025] This example uses a hyperbolic arch dam, a key reservoir power station located on the middle and lower reaches of the Lancang River in Yunnan Province, my country. The dam is 294.5m high, with a crest arc length of 892.786m. At the time of construction, it was the world's highest arch dam and also the arch dam bearing the greatest water load. The dam crest elevation is 1245m, the normal reservoir water level is 1240m, the dead water level is 1166m, and the flood season limit water level is 1236m. To monitor the horizontal deformation of the dam, 9 vertical lines and 9 inverted vertical lines were installed inside the dam, totaling 53 monitoring points. The time variation of data from each monitoring point is shown below. Figure 2 As shown. The specific process is as follows: For the i Each measurement point has a historical deformation sequence during the training phase. The historical deformation response data is encoded using GRU, compressing the long-term deformation evolution characteristics of the measuring points into historical responses, ultimately yielding the historical response embedding information: ; Spatial coordinates of measuring points Input coordinate encoding MLP to obtain spatial location embedding information: ; After obtaining the two types of node information, a dual-channel Pairwise Edge Score is used to directly learn the connection strength between any two measurement points. and Paired features are constructed for the historical response feature channel and the spatial coordinate feature channel, respectively, and the two scores are fused into the final static connection strength: ; The correlation between all measurement points introduces noise, which negatively impacts the correlation with other measurement points and increases the computational burden. Therefore, Top-K sparsification is applied to the connection strength matrix. For each measurement point... After removing self-connects, select the connection with the highest strength. K Each measurement point is selected as a candidate neighbor: ; The final static prior matrix connection is as follows Figure 3 As shown.

[0026] The static prior graph provides candidate adjacency masks M and prior association scores S. Wherein, M... ij =1 indicates the measuring point With measuring points There are candidate associations. This represents the static prior association strength between the two. During the dynamic graph update process, the model does not freely calculate edge weights between all pairs of points, but only... Attention is calculated and normalized within the candidate edge range to avoid generating noisy connections lacking structural basis, such as... Figure 4 As shown.

[0027] A dynamic graph-constrained GCN-Mamba spatiotemporal prediction network is constructed. The dynamic graph convolution module is used to aggregate spatial features on the measurement point association graph at the current time step, enabling each measurement point to fuse the state information of its candidate neighboring measurement points, thereby characterizing the spatial coupling relationship between multiple measurement points in the dam body. The Mamba module is used to perform temporal modeling on the spatially enhanced node representation sequence.

[0028] For including N The graph structure of measurement points, assuming the node feature matrix at a certain time step is... The adjacency matrix is To preserve node information, self-connections are typically added to the adjacency matrix. ; in, This is the identity matrix. The adjacency matrix is ​​then normalized to avoid instability in the feature scale caused by differences in node degree. ; in, for The degree matrix. Based on the normalized adjacency matrix, the graph convolutional layer can be represented as: ; in, For the first Layer node representation, For learnable weight matrix, This is a nonlinear activation function. The process can be understood as follows: each measurement point aggregates information from related measurement points based on its adjacency weights, and obtains a new node representation through linear transformation and nonlinear mapping.

[0029] After dynamic graph convolution, the model obtains a temporal representation of nodes that incorporates spatial correlation information: ; in, For batch size, For the length of the history window, For the number of measurement points, This is the hidden layer dimension. To model the temporal evolution of each measurement point, this invention will... H Rearranged by measurement point dimension as follows: ; The spatial augmentation representation of each measurement point within the historical window is treated as a time series and input into the Mamba module. Since the monitoring data is a discrete sampling sequence, the continuous system needs to be discretized. For the... The discrete state update at each time step can be written as: ; ; This formula indicates the current state. Not only dependent on the current input It is also affected by the state of the previous moment. Because of its influence, the state-space model can transmit historical information through hidden states, making it suitable for modeling historical cumulative effects and hysteresis responses in continuous monitoring sequences. Unlike traditional state-space models, Mamba introduces a selection mechanism, allowing some state-space parameters to change dynamically with the input.

[0030] For the input sequence Mamba generates a time step, input mapping, and output mapping based on the current input: ; ; in, This represents the discretization step size of the current time step; and This represents the relevant mapping parameters of the input.

[0031] Therefore, the state transition can adaptively adjust according to the current monitoring state. After the selective parameters are determined, the continuous state is discretized as follows: ; ; Therefore, the first k The selective states at each time step are recursively derived as follows: ; ; in, This represents the hidden state at the current time step. This is the output representation for Mamba. Because... , , Will follow input By changing the monitoring status at different time steps, Mamba can selectively retain, update, or suppress historical information, thereby enhancing its ability to express non-stationary deformation sequences and making it more effective in capturing non-stationary changes, stage changes, and long-term dependencies in arch dam deformation sequences.

[0032] Mamba then receives the spatially augmented sequence after dynamic graph convolution: ; in, This invention takes the output of the last time step as the temporal state representation of the current window, and obtains the predicted value of future deformation through the prediction layer: .

[0033] The prediction metrics of the offline model at all measurement points and the metrics of the comparison model are shown in Tables 1 and 2.

[0034] Table 1. Offline prediction metrics of the proposed model at all monitoring points

[0035] The offline comparison model fitting residual plot of survey line A22 is shown below. Figure 5 As shown.

[0036] Table 2 Evaluation Indicators for Comparison Models

[0037] In the online phase, the model uses real-time back-transmitted samples to fine-tune local parameters and adapt to recent data changes. The latest monitoring data received over time is input into the model, and a dynamic adjacency matrix is ​​regenerated under the constraints of the static adjacency matrix. The prediction results are then output using the GCN-Mamba model. Subsequently, the actual monitoring values ​​are back-transmitted, the prediction residuals are calculated, and it is determined whether the current data deviates from the normal deformation evolution pattern. The online loss is calculated, and some parameters are incrementally updated using a small learning rate. Taking the A22 survey line as an example, the online prediction results are shown in Table 3.

[0038] Table 3. Online prediction metrics of the proposed model at all monitoring points

[0039] Therefore, this invention employs a dynamic graph neural network-driven method for predicting the spatiotemporal deformation of high arch dams. It utilizes the spatial coordinates of measuring points and historical response information to learn static spatial relationships. Based on this, a dynamic adjacency matrix characterizes the relationship strength between measuring points over time. Graph convolution is used to extract spatial dependency features, and a Mamba module is introduced to model long-term evolution patterns. Furthermore, an online learning mechanism is combined to enable the model to perform lightweight updates based on newly arriving monitoring data, improving its adaptability in practical prediction scenarios. This invention provides a feasible approach for accurately analyzing the influence of measuring point relationships on the deformation changes of high arch dams during operation and for accurately predicting their future trends.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting the spatiotemporal deformation of high arch dams driven by a dynamic graph neural network, characterized in that, Includes the following steps: S1. Collect and preprocess the high arch dam deformation and environmental monitoring data by cleaning outliers and interpolating missing values. Based on the basic principle of the arch dam spatiotemporal deformation tracking and prediction model, determine the relationship between upstream and downstream water level and temperature environmental load and the deformation response of the measuring points, construct the high arch dam deformation tracking and prediction dataset, and divide it into training set, validation set and prediction set in a ratio of 8:1:

1. S2. Based on the 3D spatial coordinates of the measurement points in the training set and historical deformation response data, learn the static spatial association and generate a static candidate adjacency matrix; then, based on the node features of the measurement points in the current input window, dynamically adjust the edge association strength through a cross-attention mechanism and output a dynamic adjacency matrix. ; S3, using a dynamic adjacency matrix To constrain the graph structure, a graph convolutional network (GCN) is used to aggregate the spatial coupling features of measurement points, and a Mamba network is used to capture the long-term temporal dependencies of deformation at a single measurement point, thus building a GCN-Mamba spatiotemporal deformation prediction model. The network parameters are iteratively updated by constructing a loss function based on the error between the predicted value and the actual deformation value, and the model with the best performance on the validation set is saved as the initial model for the online prediction stage. S4. Input the real-time monitoring data collected on-site into the trained optimal initial model, and then input the data into the dynamic adjacency matrix. The deformation prediction results of each measuring point are output under constraints; the online prediction loss is calculated by comparing with the measured deformation values; and the network parameters of the model are incrementally updated using a small learning rate to achieve adaptive updating of the model.

2. The method for predicting the spatiotemporal deformation of high arch dams driven by a dynamic graph neural network according to claim 1, characterized in that: The high arch dam deformation tracking and prediction dataset in step S1 includes upstream water level, time factor, temperature factor, spatial three-dimensional coordinates, and radial deformation automatically monitored at each measuring point of the high arch dam.

3. The method for predicting the spatiotemporal deformation of high arch dams driven by a dynamic graph neural network according to claim 2, characterized in that, Step S2 is as follows: S21. Use GRU and MLP to extract the historical response embedding information and spatial location embedding information of each measuring point, respectively. S22. The connection strength between any two measurement points is directly learned through the dual-channel Pairwise Edge Scorer and the static candidate adjacency matrix is ​​obtained by adaptive Top-K sparsification. S23. Based on the construction of a static candidate adjacency matrix, a cross-attention mechanism is used to dynamically adjust the association strength of candidate edges according to the monitoring status of the current window, resulting in a dynamic adjacency matrix. .

4. The method for predicting the spatiotemporal deformation of high arch dams driven by a dynamic graph neural network according to claim 3, characterized in that, Step S21 is as follows: For the i Each measurement point has a historical deformation sequence during the training phase. The historical deformation sequence is encoded using GRU, compressing the long-term deformation evolution characteristics of the measurement points into historical responses, ultimately yielding the historical response embedding information: ； Spatial coordinates of the measuring point Input coordinate encoding MLP to obtain spatial location embedding information: 。 5. The method for predicting the spatiotemporal deformation of high arch dams driven by a dynamic graph neural network according to claim 3, characterized in that, Step S22 includes: Measuring points , The GRU historical response embedding information and MLP spatial location embedding information are fed into the dual-channel Pairwise EdgeScorer to directly learn the connection strength between the two measurement points. The measurement points are obtained by fusing the dual-channel scores. and measuring points Static correlation strength between ,satisfy: ； in, ; This represents the Sigmoid activation function, which maps the input to the (0,1) interval and is used to generate weight coefficients; Indicates the characteristics of sensor readings , The calculation yields the representative measurement point. Correlation information between sensor measurements; Indicates the measurement point Information relating to spatial locations; Top-K sparsity is applied to the static correlation strength for each measurement point. i After removing self-connects, select the connection with the highest strength. K Each measurement point is selected as a candidate neighbor: ； And finally, the static candidate adjacency matrix is ​​obtained: 。 6. The method for predicting the spatiotemporal deformation of high arch dams driven by a dynamic graph neural network according to claim 5, characterized in that, The specific content of step S23 is as follows: The cross-attention mechanism dynamically adjusts the association strength of candidate edges based on the current monitoring status in the sliding window: ； ； in, This represents the dynamic nodes obtained from the current sliding window; For dynamic monitoring status query; , These are static prior keys and values; The node structure representation obtained from static prior information; , , These are the learnable parameter matrices; It is a static set of candidate neighbors; For the implicit vector dimension; These are learnable parameters used to adaptively adjust the influence of static prior scores on dynamic graph updates.

7. The method for predicting the spatiotemporal deformation of high arch dams driven by a dynamic graph neural network according to claim 6, characterized in that, Step S3 specifically includes: Let the node feature matrix at a certain time step be... The dynamic adjacency matrix is To preserve node information, we first add self-joins to the dynamic adjacency matrix, and then normalize the dynamic adjacency matrix. ； ； Where I is the identity matrix; for The degree matrix; This is the normalized dynamic adjacency matrix; The dynamic adjacency matrix after adding self-connections; Based on the normalized dynamic adjacency matrix, the graph convolutional layer is represented as: ； in, For the first l Layer node representation, For learnable weight matrix, It is a non-linear activation function; After graph convolution processing, a node temporal representation with fused spatial correlation information is obtained. The temporal data is then rearranged according to the measurement point dimension to obtain an independent temporal sequence for each measurement point. Enter the Mamba module; For the input sequence Mamba generates a time step, input mapping, and output mapping based on the current input: ； ； in, This represents the discretization step size of the current time step; and This represents the relevant mapping parameters of the input; Indicates A learnable nonlinear mapping network as input; After the selectivity parameters are determined, the continuous state is discretized as follows: ； ； No. k The selective states at each time step are recursively derived as follows: ； ； in, This represents the hidden state at the current time step. Here, D represents the Mamba output representation; D represents the input mapping matrix in the Mamba selective state-space model, which maps the input features. The output is directly mapped and used as the direct input to the output branch; Mamba then receives the spatially augmented sequence after dynamic graph convolution: ； in, The output of the last time step is taken as the temporal state representation of the current window, and the future deformation prediction value is obtained through the prediction layer: ； in, This represents the output characteristics of the Mamba module at the last time step within this time window.

Images