A method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN
By integrating GCN and RNN into a spatiotemporal deformation prediction model for rockfill dams, the spatial collaborative deformation patterns among measuring points are captured and long-term temporal characteristics are extracted. This solves the shortcomings of temporal dependence and spatial correlation in the deformation prediction of rockfill dams, and achieves high-precision deformation trend prediction.
Patent Information
- Application Number
- CN202611190121.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-06
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies have failed to effectively handle temporal dependencies and spatial correlations in the deformation prediction of rockfill dams, resulting in insufficient prediction accuracy and safety, especially in scenarios involving long-term time series and multi-point collaborative deformation.
A spatiotemporal deformation prediction model for rockfill dams was constructed by fusing graph convolutional networks (GCN) and recurrent neural networks (RNN). The spatial collaborative deformation patterns between measurement points were captured by the graph adjacency matrix, and long short-term memory networks (LSTM) were used to mine the long-term temporal features of the deformation sequence. The model was then trained and validated using a hybrid loss function.
It significantly improves the prediction accuracy of the overall deformation trend of rockfill dams, achieving an average relative error of less than 4.2%, and overcomes the shortcomings of traditional methods in predicting the coordinated deformation of multiple measuring points and long-term time-dependent deformation.
Smart Images

Figure CN123116124A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of water conservancy engineering technology, and more specifically, to a method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN. Background Technology
[0002] Rockfill dams, as an important dam type in water conservancy and hydropower projects, are widely used due to their good terrain adaptability and economic efficiency. However, rockfill dams experience significant spatiotemporal deformation during both construction and operation. Accurately predicting their deformation development trend has become a core technical problem in the field of dam safety monitoring. Existing deformation prediction methods still have certain shortcomings in handling time-series dependencies and spatial correlations, affecting prediction accuracy and the reliability of safety assessments. Therefore, conducting research on accurate spatiotemporal deformation prediction of rockfill dams has become an important technical direction in the field of dam health diagnosis.
[0003] The patent (CN111811468A) establishes a physical causal model and calculates the settlement deformation, entanglement deformation and creep deformation of the dam body respectively, which can provide certain predictive analysis for the settlement deformation of high core rockfill dams.
[0004] The patent (CN117272872A) extracts the trend terms of data information with the same water level and temperature, separates the mixed periodic components, and establishes water level component model, temperature component model and time-effect component model, which can predict deformation to a certain extent and analyze the influence of each component factor.
[0005] In summary, existing technologies mostly employ physical mechanics models or statistical regression analysis for deformation prediction, which presents the following problems: Methods based on traditional physical mechanics models rely on empirical formulas and simplifying assumptions, failing to fully consider the temporal characteristics and spatial propagation patterns of rockfill dam deformation. Calculation parameters often need to be determined based on engineering experience, easily leading to a decline in long-term prediction accuracy and difficulty in accurately depicting the nonlinear evolution of rockfill dam deformation. Methods based on statistical regression analysis use component separation techniques, which can analyze the influence of each component factor to some extent, but their ability to capture spatial collaborative deformation patterns between monitoring points is limited, failing to effectively handle drift issues in deformation data, and their extraction of long-term time-series dependency features is not deep enough, resulting in significant uncertainty in prediction results. All of the above methods lack effective spatiotemporal information fusion mechanisms, and their prediction efficiency and data utilization need improvement when dealing with complex scenarios involving multi-point collaborative deformation and long-term time-series dependencies.
[0006] To address the shortcomings of the existing technologies, this invention proposes a spatiotemporal deformation prediction model for rockfill dams that integrates GCN and RNN. The model aims to capture the spatial collaborative deformation patterns between measurement points through graph convolutional networks and to mine the long-term temporal evolution characteristics of deformation sequences using recurrent neural networks, thereby significantly improving the prediction accuracy of the overall deformation trend of the dam. Summary of the Invention
[0007] The main objective of this invention is to disclose a method for predicting the spatiotemporal deformation of rockfill dams that integrates GCN and RNN, comprising: Constructing a spatial correlation graph model of multiple measuring points for a rockfill dam includes: collecting raw displacement data and environmental parameters from multiple measuring points; preprocessing the raw data; determining the spatial adjacency relationships between measuring points; calculating edge weights based on cross-sectional spacing, spatial distance, and distance attenuation parameters between measuring points; constructing a graph adjacency matrix based on the spatial adjacency relationships and edge weights; and initializing node features. The graph adjacency matrix is a weighted adjacency matrix of the dam's topology. , where n is the number of measurement points, and each element A in the matrix represents the connection weight between measurement points; the node features include: three-dimensional displacement components and corresponding monitoring timestamps; Specifically, the raw data includes: displacement data of key measuring points at the dam crest, dam waist, river valley bottom, and bank slope connection sections; the environmental parameters include: dam operating water level, air temperature, and reservoir water temperature. Furthermore, the preprocessing of the raw data includes: outlier detection and interpolation repair of the raw data, elimination of measurement noise using the Kalman filter algorithm, and standardization of the data dimensions of different measurement points using Z-Score. Furthermore, when determining the spatial adjacency relationship between the measuring points: when two measuring points are located on the same monitoring section, they are determined to be adjacent; when two measuring points are located on adjacent monitoring sections, they are determined to be cross-section proximity; when the distance between two measuring points is less than a preset threshold, they are determined to be adjacent. Specifically, the calculation of edge weights refers to the calculation based on the cross-sectional spacing, spatial distance, and distance attenuation parameters between measuring points. The specific formula for calculating the edge weights is as follows: ,when Otherwise, it is 0; in, For measuring points and Spatial distance between them For distance attenuation parameters, This is a preset distance threshold.
[0008] Based on the graph adjacency matrix and node features, a graph convolutional network is used to extract spatial cooperative deformation features between measurement points. The feature expression is enhanced by multi-order spatial propagation and spatial feature attention weighting to generate spatial features. Specifically, the multi-order spatial propagation refers to enhancing the capture of the spatial propagation patterns of rockfill dam deformation from the dam crest to downstream and from the main body to the bank slope through propagation paths of different orders. The specific multi-order spatial propagation calculation formula is as follows: ; in, For the first First-order propagation coefficient, For the maximum propagation order, This is the initial node feature matrix; The spatial feature attention refers to automatically learning the correlation strength between different measurement points to enhance the ability to capture key deformation propagation paths. The specific formula for calculating spatial feature attention is as follows: ; ; in, and They are nodes and eigenvectors, Let i be the set of neighboring nodes. This is the weight matrix. For attention vectors, Let be the transpose of vector a. Sum the exponential scores of all neighboring nodes of node i; Furthermore, the extraction of the long-term time-series dependency features includes: capturing the time-cumulative effect of rockfill dam deformation and aging conditions; the attention mechanism includes: focusing on the historical moment that contributes the most to the current prediction and adaptively assigning weights to different historical moments.
[0009] Based on the aforementioned spatial features, a recurrent neural network is used to mine the long-term temporal dependency characteristics of the spatial features. Temporal modeling is performed through a long short-term memory network, and long-term dependency features are extracted using a temporal decay coefficient and an attention mechanism to generate temporal features. The spatial and temporal features are fused together, and the displacement prediction results of each measuring point of the rockfill dam are output. Specifically, the fusion of spatial features and temporal features includes: projecting the spatial features onto a dimensional space with the same dimensions as the temporal features using a projection matrix, and then fusing them using a residual connection method; the projection matrix is... , bias is .
[0010] The graph convolutional network and recurrent neural network are trained and their accuracy verified by combining loss functions. Specifically, the combined loss function is composed of a weighted average absolute error loss term and an average deviation loss term, and the loss function is specifically defined as follows: ; ; ; in, For the sample size, The true value of the i-th sample. The predicted value of the i-th sample; and The weighting coefficient has a range of values. ∈0.5,1.0, ∈0.1,0.5; This ensures that the squared error between the predicted and actual values is minimized. This enhances the robustness of anomaly prediction; Specifically, the accuracy verification includes: using a five-fold time series cross-validation method to evaluate the stability and generalization ability of graph convolutional networks and recurrent neural networks, calculating the mean relative error δ, root mean square error RMSE and coefficient of determination R2, and introducing cross-year extrapolation test to verify the prediction generalization ability of graph convolutional networks and recurrent neural networks beyond the training period.
[0011] According to this invention, by constructing a spatial correlation graph model of multiple measuring points for a rockfill dam, the spatial correlation of measuring point displacement data is explicitly modeled as a graph structure, providing a topological foundation for spatial feature extraction by graph convolutional networks. Employing a three-layer cascaded architecture, multi-order spatial propagation, and attention weighting mechanism of graph convolutional networks, spatial collaborative deformation features are extracted from the displacement data of multiple measuring points. This automatically captures the spatial propagation patterns of deformation from the dam crest to downstream and from the main body to the bank slope, overcoming the shortcomings of traditional methods in quantifying the spatial correlation information of multiple measuring points. Using the LSTM gating mechanism, temporal decay, and attention mechanism of recurrent neural networks, the long-term temporal dependence characteristics and time accumulation effect of deformation sequences are fully explored, effectively solving the problem of decreased prediction accuracy over long time series. By organically fusing spatial and temporal features through a spatiotemporal feature fusion network and jointly training with a hybrid loss function, the prediction accuracy index meets the requirement of an average relative error δ≤4.2%, achieving high-precision prediction of the overall deformation trend of the rockfill dam. Attached Figure Description
[0012] Figure 1 This is a schematic diagram of a method for predicting the spatiotemporal deformation of a rockfill dam by integrating GCN and RNN, according to an embodiment of the present invention. Figure 2 This is a visualization of the spatial correlation map of multiple measurement points of a rockfill dam, based on a method for predicting spatiotemporal deformation of rockfill dams that integrates GCN and RNN, according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the monitoring area and observation point distribution of a spatiotemporal deformation prediction method for rockfill dams that integrates GCN and RNN, provided by an embodiment of the present invention. Figure 4This is a schematic diagram of the spatiotemporal deformation prediction accuracy verification of a spatiotemporal deformation prediction method for rockfill dams that integrates GCN and RNN according to an embodiment of the present invention. Figure 5 This is a schematic diagram of the prediction error distribution of a spatiotemporal deformation prediction method for rockfill dams that integrates GCN and RNN, provided by an embodiment of the present invention. Detailed Implementation
[0013] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.
[0014] like Figure 1 As shown, this invention discloses a method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN, comprising the following steps: Step S100: Construct a spatial correlation graph model of multiple measuring points for a rockfill dam, including: collecting raw data of displacement of multiple measuring points and environmental parameters, preprocessing the raw data, determining the spatial adjacency relationship between measuring points, calculating edge weights based on the cross-sectional spacing, spatial distance and distance attenuation parameters between measuring points, constructing a graph adjacency matrix based on the spatial adjacency relationship and edge weights and completing the initialization of node features. like Figure 2 As shown, a high-precision displacement monitoring system deployed on the surface of the rockfill dam is used to collect displacement data of key parts of the dam; environmental parameters such as the dam's operating water level, air temperature, and reservoir water temperature are collected simultaneously to construct a set of benchmark parameters for displacement-environment correlation analysis. The principle for arranging measuring points is to cover typical deformation-sensitive areas such as the dam crest, dam waist, river valley bottom, and bank slope connection sections. The sampling frequency is set to once per hour, and the storage period is no less than 365 days for continuous monitoring data. For example, a rockfill dam project with a concrete panel is selected. The dam is 156m high and the crest is 432m long. 15 monitoring sections are arranged along the dam axis. 3 to 5 measuring points are arranged on each section, for a total of 59 displacement measuring points. The measuring points are numbered in the order of the sections as V1, V2, ..., V59. Each measuring point is a node in the map.
[0015] In this implementation step, outlier detection and interpolation repair are performed on the original displacement data, Kalman filtering algorithm is used to eliminate measurement noise, and Z-Score standardization is used to unify the data dimensions of different measurement points.
[0016] The specific steps for determining spatial adjacency relationships and calculating edge weights are as follows: Step 1: Based on the structural characteristics and deformation propagation law of the rockfill dam, determine the spatial adjacency relationship between the measuring points; Step 2: The criteria for determining spatial adjacency are as follows: When two measuring points are located on the same monitoring section, they are determined to be adjacent, and the initial weight of the edge is calculated based on the distance between the sections. When two measuring points are located on adjacent monitoring sections, they are determined to be adjacent across sections, and the initial weight of the edge is calculated based on the distance between the upstream and downstream sections. When the distance between two measuring points is less than a preset threshold At that time, it was also determined to be a neighboring relationship. Specifically, the formula for calculating edge weights is: ,when Otherwise, it is 0; in, For measuring points and Spatial distance between them For distance attenuation parameters, This is a preset distance threshold.
[0017] The specific steps for constructing the graph adjacency matrix and initializing node features are as follows: Step 1: Construct a graph adjacency matrix based on spatial adjacency relationships. For those The graph of each measurement point, and its adjacency matrix. Defined as: ; Step 2: To ensure the stability of information transfer during graph convolution operations, self-loops are added to the adjacency matrix, resulting in an adjacency matrix with self-loops. ; Step 3: Encode the historical displacement time series data of each measuring point into a feature vector, set the feature dimension to 32 dimensions, and initialize the node features; wherein, the node features include: three-dimensional displacement components and the corresponding monitoring time.
[0018] Step S110: Based on the graph adjacency matrix and node features, a graph convolutional network is used to extract spatial cooperative deformation features between measurement points. The feature expression is enhanced by multi-order spatial propagation and spatial feature attention weighting to generate spatial features. Specifically, the graph convolutional network adopts a three-layer hierarchical architecture. The first layer has an input dimension of 32 and an output dimension of 64. The second layer has an input dimension of 64 and an output dimension of 128. The third layer has an input dimension of 128 and an output dimension of 64. Each graph convolutional operation is followed by batch normalization, and the ReLU activation function is used to introduce non-linear mapping capabilities.
[0019] In this implementation step, the spatial feature extraction operator is represented as: ; in, Given an adjacency matrix with self-loops, For degree matrix, For the first Hidden layer state, For learnable weight matrix, This is the ReLU activation function.
[0020] Multi-order spatial propagation is used to capture the spatial propagation law of rockfill dam deformation from the dam crest to the downstream and from the main body to the bank slope through propagation paths of different orders. The multi-order propagation operator is expressed as: ; in, For the first First-order propagation coefficient, For the maximum propagation order, The initial node feature matrix, and the multi-order spatial propagation order. Set to 3, propagation coefficients of each order Automatically adjusts based on learnable parameters; For example, multi-order spatial propagation order The value is set to 3, which considers the spatial propagation effects of direct adjacency, first-order proximity, and second-order proximity; the propagation coefficients for each order are... Automatic adjustment via learnable parameters, initial value set to... Multi-stage propagation can effectively capture the spatial propagation pattern of rockfill dam deformation from the dam crest to the downstream and from the main body to the bank slope.
[0021] Spatial feature attention weighting adopts a multi-head attention strategy, assigning differentiated weights to different measurement points, automatically learning the correlation strength between different measurement points, and enhancing the ability to capture key deformation propagation paths; Specifically, the attention score is calculated as follows: ; ; in, and They are nodes and eigenvectors, Let i be the set of neighboring nodes. This is the weight matrix. For attention vectors, Let be the transpose of vector a. This represents a concatenation operation, which joins two vectors end-to-end. LeakyReLU is the activation function, introducing non-linearity and preventing gradient vanishing. exp is the exponential function. Sum the exponential scores of all neighboring nodes of node i. The final attention weight of node i to node j.
[0022] For example, the attention mechanism adopts a multi-head attention strategy, setting up 3 attention heads, with the negative slope of LeakyReLU set to 0.2, and the output dimension of each head is 21-dimensional. Finally, the 63-dimensional spatial weighted feature representation is obtained by concatenating the heads.
[0023] Step S120: Based on the spatial features, a recurrent neural network is used to mine the long-term temporal dependency characteristics of the spatial features, a long short-term memory network is used for temporal modeling, and a temporal decay coefficient and attention mechanism are used to extract long-term dependency features to generate temporal features. Specifically, a long short-term memory network is used for time-series modeling, where the time-series feature update equation is expressed as: ; ; ; ; ; ; in, , , These are the forget gate, input gate, and output gate, respectively. In cellular state, Candidate cell state, Let the input vector be the input vector at the current time. This is the hidden state from the previous moment. Index for the current time step, In hidden state, This represents element-wise multiplication, where tanh is the hyperbolic tangent activation function. , , , This is the weight matrix. , , , For bias vectors, Use the Sigmoid activation function; Specifically, the hidden state dimension of the Long Short-Term Memory network is set to 256 dimensions, and the input is the spatial features output by the graph convolutional network, which are 64 dimensions. The weight matrices of the input gate, forget gate, and output gate are initialized using the Xavier initialization method, and the bias vector is initialized to zero. The sigmoid function is selected as the activation function, which can effectively regulate the flow of information and achieve selective memory and forgetting of information that depends on long-term information.
[0024] The specific formula for extracting long-term time-dependent features is as follows: ; in, This is the time-series decay coefficient. For learnable parameters, For the backtracking window length, This represents the hidden state vector at the τth time step in the past. In the embodiments provided by the present invention, the backtracking window length Setting it to 30 means considering the displacement change trend over the past 30 observation periods; time-series decay coefficient Parameters can be learned by normalizing using the softmax function. The initial value is set to This results in an exponential decline in the contribution of more distant historical periods; the long-term time dependence feature can capture the cumulative effect of rockfill dam deformation over time and the characteristics of aging chemical conditions.
[0025] The temporal attention mechanism, where the attention weights are calculated using the following formula: ; ; ; in, Index for the current time, For indexing historical moments, For temporary traversal indexes in summing the denominators, For a historic moment The hidden state, This is the feature representation after attention aggregation. and These are the query weight matrix and the key weight matrix, respectively. For attention score weight vectors, The hyperbolic tangent activation function is used. The attention score between the current time t and the historical time i. Here, exp represents the normalized attention weights, and exp is the exponential function.
[0026] Step S130: The spatial features and temporal features are fused and the displacement prediction results of each measuring point of the rockfill dam are output. The specific steps for fusing spatial and temporal features are as follows: Step 1: First, project the spatial features onto a dimensional space with the same dimensions as the temporal features, where the projection matrix... bias ; Step 2: Use residual joins for fusion to maintain the integrity of information flow. The specific formula is as follows: ; in, The spatial features output by the graph convolutional network are 64-dimensional. The temporal features are after attention aggregation, with a dimension of 256. and This is the fusion weight matrix.
[0027] Step S140: The graph convolutional network and recurrent neural network are trained and their accuracy verified by combining loss functions; Specifically, the formula for the combined loss function is as follows: ; ; ; in, For the sample size, The true value of the i-th sample. The predicted value of the i-th sample, and the range of the weight coefficients. ∈0.5,1.0, ∈0.1,0.5, weighting coefficients and Set them to 0.7 and 0.3 respectively. This ensures that the squared error between the predicted and actual values is minimized. This enhances the robustness of anomaly prediction.
[0028] The prediction equation is used to generate and output the prediction results. The specific formula is as follows: ; in, for Predicted displacement value after time step [time] and For output parameters.
[0029] In this implementation step, the Adam optimizer is used for model training, the batch size is set to 32, the number of training epochs is set to 500, and the patience of the early stopping strategy is set to 50.
[0030] The model refers to the fused graph convolutional network (GCN) and recurrent neural network (RNN).
[0031] Specifically, the model accuracy verification and generalization ability evaluation, with the prediction accuracy index meeting the requirement of average relative error δ≤4.2%, involves the following steps: Step 1: Divide the dataset into training set, validation set and test set in chronological order, with a ratio of 8:1:1; Step 2: Use the five-fold time series cross-validation method to evaluate the stability and generalization ability of the model, and calculate the mean relative error δ, root mean square error RMSE and coefficient of determination R2 on the test set; Step 3: Introduce cross-year extrapolation test to verify the model's ability to generalize predictions beyond the training period.
[0032] This invention provides a specific example of a method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN: Example 1: Dam Project: A concrete-faced rockfill dam project is selected, located in the southwest region. The dam is 156m high, 432m long at the top, and 10m wide at the top. The upstream slope ratio is 1:1.4 and the downstream slope ratio is 1:1.5. Measurement point layout: A total of 59 displacement measurement points are arranged, distributed across 15 monitoring sections. One section is arranged every 30m on the dam crest, one section is arranged every 15m at the bottom of the valley, and one section is arranged every 40m on the bank slope connection section. 3 to 5 measurement points are arranged on each section. Displacement range: The maximum displacement change at the dam crest measuring point is 127.4 mm, the maximum displacement change at the dam waist measuring point is 58.6 mm, and the maximum displacement change at the foundation gallery measuring point is 12.3 mm. Water level conditions: The normal operating water level range is 125m to 145m; Temperature conditions: Annual temperature range is -10℃ to 38℃; Monitoring data: From January 2019 to December 2023, a total of approximately 35,000 sets of valid displacement data were collected, with a data integrity rate of over 99% and a time resolution of 1 hour.
[0033] The method of this invention was verified to predict the collaborative deformation of multiple measurement points. With a prediction step size of 1 hour, the average relative error was 2.8%, which is 67.4% lower than the error of the traditional physical and mechanical model, and the spatiotemporal feature fusion effect is significant.
[0034] Example 2: To verify the generalization ability and engineering applicability of the method of the present invention, another dam project built in 2012 was selected as the verification object. The project is located in the Northwest region, with a dam height of 98m, a dam crest length of 285m, an operating water level range of 115m to 135m, and an annual temperature range of -25℃ to 32℃.
[0035] like Figure 3 As shown, the measurement point layout scheme is as follows: a total of 37 displacement measurement points are arranged in this project, distributed across 11 monitoring sections. Compared with Example 1, this project has fewer measurement points and lower spatial coverage density, which can verify the applicability of the model under sparse measurement point conditions.
[0036] Transfer learning validation: Since the training samples for this project are small, the model parameters pre-trained in Example 1 are used as the source domain, and the target domain is adapted through the transfer learning method. The transfer learning strategy adopts feature extraction layer freezing, the parameter freezing rate of the first two graph convolutional layers is set to 65%, and only the last graph convolutional layer and the fully connected layer of the LSTM network are fine-tuned.
[0037] like Figure 4 , Figure 5 As shown, the test results indicate that, under the condition of only 1500 training samples, after transfer learning adaptation, the error statistics after 1 hour show that the maximum absolute error is 0.78 mm, the mean absolute error is 0.23 mm, the root mean square error is 0.31 mm, the standard deviation is 0.25 mm, and the average relative error of the prediction step size after 1 hour is 3.2%, which is a 52.3% reduction in error compared to the baseline model without transfer.
[0038] Error statistics for 24 hours: maximum absolute error: 0.92 mm, mean absolute error: 0.36 mm, root mean square error: 0.45 mm, standard deviation: 0.34 mm, mean relative error for 24-hour prediction step size: 4.0%. The model demonstrates good generalization ability for rockfill dam projects under different geographical locations and climatic conditions.
[0039] According to this invention, by constructing a spatial correlation graph model of multiple measuring points for a rockfill dam, the spatial correlation of measuring point displacement data is explicitly modeled as a graph structure, providing a topological foundation for spatial feature extraction by graph convolutional networks. Employing a three-layer cascaded architecture, multi-order spatial propagation, and attention weighting mechanism of graph convolutional networks, spatial collaborative deformation features are extracted from the displacement data of multiple measuring points. This automatically captures the spatial propagation patterns of deformation from the dam crest to downstream and from the main body to the bank slope, overcoming the shortcomings of traditional methods in quantifying the spatial correlation information of multiple measuring points. Using the LSTM gating mechanism, temporal decay, and attention mechanism of recurrent neural networks, the long-term temporal dependence characteristics and time accumulation effect of deformation sequences are fully explored, effectively solving the problem of decreased prediction accuracy over long time series. By organically fusing spatial and temporal features through a spatiotemporal feature fusion network and jointly training with a hybrid loss function, the prediction accuracy index meets the requirement of an average relative error δ≤4.2%, achieving high-precision prediction of the overall deformation trend of the rockfill dam.
[0040] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.
Claims
1. A method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN, characterized in that, include: Constructing a spatial correlation graph model of multiple measuring points for a rockfill dam includes: collecting raw displacement data and environmental parameters from multiple measuring points; preprocessing the raw data; determining the spatial adjacency relationships between measuring points; calculating edge weights based on cross-sectional spacing, spatial distance, and distance attenuation parameters between measuring points; constructing a graph adjacency matrix based on the spatial adjacency relationships and edge weights; and initializing node features. The graph adjacency matrix is a weighted adjacency matrix of the dam's topology. , where n is the number of measurement points, and each element A in the matrix represents the connection weight between measurement points; the node features include: three-dimensional displacement components and corresponding monitoring timestamps; Based on the graph adjacency matrix and node features, a graph convolutional network is used to extract spatial cooperative deformation features between measurement points. The feature expression is enhanced by multi-order spatial propagation and spatial feature attention weighting to generate spatial features. Based on the aforementioned spatial features, a recurrent neural network is used to mine the long-term temporal dependency characteristics of the spatial features. Temporal modeling is performed through a long short-term memory network, and long-term dependency features are extracted using a temporal decay coefficient and an attention mechanism to generate temporal features. The spatial and temporal features are fused together, and the displacement prediction results of each measuring point of the rockfill dam are output. The graph convolutional network and recurrent neural network are trained and their accuracy verified by combining loss functions.
2. The method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN as described in claim 1, characterized in that, The raw data includes: displacement data of key measuring points at the dam crest, dam waist, river valley bottom, and bank slope connection sections; the environmental parameters include: dam operating water level, air temperature, and reservoir water temperature.
3. The method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN as described in claim 1, characterized in that, The preprocessing of the raw data includes: outlier detection and interpolation repair, Kalman filtering algorithm to eliminate measurement noise, and Z-Score standardization to unify the data dimensions of different measurement points.
4. The method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN as described in claim 1, characterized in that, When determining the spatial adjacency relationship between the measuring points: when two measuring points are located on the same monitoring section, they are determined to be adjacent; when two measuring points are located on adjacent monitoring sections, they are determined to be cross-section proximity; when the distance between two measuring points is less than a preset threshold, they are determined to be adjacent.
5. The method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN according to claim 1, characterized in that, The calculated edge weights refer to those calculated based on the cross-sectional spacing, spatial distance, and distance attenuation parameters between measuring points. The formula for calculating the weight of an edge is: ,when Otherwise, it is 0; in, For measuring points and Spatial distance between them For distance attenuation parameters, This is a preset distance threshold.
6. The method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN according to claim 1, characterized in that, The aforementioned multi-order spatial propagation refers to enhancing the capture of the spatial propagation patterns of rockfill dam deformation from the dam crest to downstream and from the main body to the bank slope through propagation paths of different orders. The specific multi-order spatial propagation calculation formula is as follows: ; in, For the first First-order propagation coefficient, For the maximum propagation order, This is the initial node feature matrix; The spatial feature attention refers to automatically learning the correlation strength between different measurement points to enhance the ability to capture key deformation propagation paths. The specific formula for calculating spatial feature attention is as follows: ; ; in, and They are nodes and eigenvectors, Let i be the set of neighboring nodes. This is the weight matrix. For attention vectors, Let be the transpose of vector a. Sum the exponential scores of all neighboring nodes of node i.
7. The method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN according to claim 1, characterized in that, Extracting the long-term time-series dependency features includes capturing the cumulative time effect of rockfill dam deformation and aging conditions; the attention mechanism includes focusing on the historical moment that contributes the most to the current prediction and adaptively assigning weights to different historical moments.
8. The method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN according to claim 1, characterized in that, The fusion of spatial and temporal features includes: projecting the spatial features onto a dimensional space with the same dimensions as the temporal features using a projection matrix, and then fusing them using a residual connection method; the projection matrix is... , bias is .
9. The method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN according to claim 1, characterized in that, The combined loss function is composed of a weighted average absolute error loss term and an average deviation loss term, and the loss function is specifically defined as follows: ; ; ; in, For the sample size, The true value of the i-th sample. The predicted value of the i-th sample; and The weighting coefficient has a range of values. ∈0.5,1.0, ∈0.1,0.5; This ensures that the squared error between the predicted and actual values is minimized. This enhances the robustness of anomaly prediction.
10. The method for predicting the spatiotemporal deformation of rockfill dams by integrating GCN and RNN according to claim 1, characterized in that, The accuracy verification includes: using a five-fold time series cross-validation method to evaluate the stability and generalization ability of graph convolutional networks and recurrent neural networks, calculating the mean relative error δ, root mean square error RMSE and coefficient of determination R2, and introducing cross-year extrapolation test to verify the prediction generalization ability of graph convolutional networks and recurrent neural networks beyond the training period.
Citation Information
Patent Citations
Prediction method for settlement deformation of high-core-wall rockfill dam
CN111811468A
Method for monitoring deformation of concrete faced rockfill dam based on component separation
CN117272872A
