A Risk Prediction Method for High Arch Dams Based on Digital Twin Model

Abstract

This invention discloses a risk prediction method for high arch dams based on a digital twin model, relating to the field of high arch dam risk prediction. The method includes: constructing a monitoring model of measuring points; constructing a mapping model of measuring points and training the model using a dynamic monitoring loss function; constructing a digital twin model; constructing a deformation-stress prediction model; and constructing a failure risk analysis model. This invention solves the problems of existing technologies, such as large errors in global state prediction, insufficient spatial consistency constraints in safety assessment, and insufficient failure assessment functions.

Description

Technical Field

[0001] This invention relates to the field of high arch dam risk prediction, and specifically to a method for high arch dam risk prediction based on a digital twin model. Background Technology

[0002] High arch dams are widely used in global water resource engineering due to their superior overload capacity, seismic performance, and lightweight, flexible dam structure. In recent years, China has constructed a number of ultra-high arch dams, such as Jinping I, Xiaowan, Baihetan, and Wudongde, whose safety is crucial for watershed water resource management. However, the complex structure, time-varying characteristics, and spatial variability of high arch dams make structural health monitoring and risk assessment extremely challenging. Currently, statistical models based on HST theory and shallow machine learning methods are the mainstream data-driven approaches, but they struggle to capture complex nonlinear relationships and have limited generalization capabilities. Deep learning (such as LSTM, GRU, and GCN) improves deformation prediction capabilities, but it relies on limited measurement point data and cannot globally analyze the dam's state. Mechanism-driven methods, with finite element methods at their core, can assess dam safety from a physical mechanism perspective, but suffer from computational complexity and poor real-time performance. Furthermore, the impact of engineering uncertainties has not been fully quantified. Digital twin technology in the Industry 4.0 era holds promise for integrating the advantages of data-driven and mechanism-driven approaches to achieve intelligent and precise monitoring of the structural health of high arch dams. However, the technology is still in the exploratory stage and faces challenges such as modeling accuracy, data fusion, and real-time interaction. Summary of the Invention

[0003] To address the aforementioned shortcomings in existing technologies, this invention provides a risk prediction method for high arch dams based on a digital twin model, which solves problems such as large errors in global state prediction, insufficient spatial consistency constraints in safety assessment, and inadequate damage assessment functions in existing technologies.

[0004] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for predicting the risk of high arch dams based on a digital twin model, comprising the following steps:

[0005] S1: Construct a monitoring model for measuring points, and use the environmental impact factors extracted based on the HST model as input to the monitoring model to obtain the predicted values ​​of the deformation process of the high arch dam.

[0006] S2: Construct a measurement point mapping model and train the model using a dynamic monitoring loss function;

[0007] S3: Construct a digital twin model, input the predicted values ​​of the deformation process of the high arch dam into the trained measurement point mapping model through the measurement point monitoring model, and obtain the three-dimensional global deformation situation;

[0008] S4: Construct a deformation-stress prediction model, use principal component analysis to reduce the dimensionality of the three-dimensional global deformation, and obtain the predicted values ​​of the major principal stress, intermediate principal stress and minor principal stress of the three-dimensional global deformation through the deformation-stress prediction model;

[0009] S5: Construct a failure risk analysis model, use the Monte Carlo Dropout method to simulate the predicted distribution of the model under different weight configurations, and perform random Dropout sampling to obtain the predicted value of failure risk of high arch dams, thus completing the risk prediction of high arch dams based on the digital twin model.

[0010] Furthermore, the monitoring model for the measuring points in S1 is as follows:

[0011]

[0012] Among them, H 0 The environmental features are input to the graph convolution feature extractor. and Let W be the set of water load influencing factors, temperature influencing factors, and time-effect influencing factors at time t. 0 H is a trainable weight matrix. (·) Let be the feature matrix of the influence factors of the nodes, l be the number of node layers, and σ be the nonlinear activation function. Let A be the normalized degree matrix, A ~ be the adjacency matrix with self-connected influence factors, and W be the adjacency matrix. (·) is a trainable weight matrix.

[0013] Furthermore, the measurement point mapping model in S2 is as follows:

[0014]

[0015] θ = (θ1, θ2)

[0016] in, Let W be the predicted value of the measurement point mapping model, W be the learnable weights, and b be the learnable bias. This is the output state of the LSTM. f is the input point for the sensor measurement value. LSTM (·) represents the LSTM network, θ1 represents the parameters of the LSTM network, θ2 represents the parameters of W and b, and θ represents all parameters of the measurement point mapping model.

[0017] Furthermore, the dynamic monitoring loss function L in S2 Digital twin Including the accuracy loss function L accuracy and spatial variability loss function L space The formula is:

[0018] L Digitaltwin =λ1L accuracy +λ2L space

[0019]

[0020] Where λ1 and λ2 are weight parameters, N is the total number of global computation points, i is the i-th global computation point, and l is the loss function calculation method. and The spatial gradient between the predicted values ​​from the measurement point mapping model and the numerical simulation results is represented by α, where α is the loss balance parameter. The loss of the finite element node corresponding to the sensor measurement point. For the loss of sensor measuring points, This represents the loss of all finite element nodes.

[0021] Furthermore, the digital twin model in S3 is as follows:

[0022]

[0023] in, f is the digital twin prediction value for time t. Synapse (·) represents the measurement point mapping model, f Cell body (·) represents the monitoring model for measurement points, θ Digital twin These are the pre-trained parameters for the measurement point mapping model. and It is a set of factors influencing water load, temperature, and time.

[0024] Furthermore, the deformation-stress prediction model in S4 is as follows:

[0025]

[0026]

[0027] in, f is the predicted stress value. σ (·) is the deformation-stress prediction function, ω σ L represents the parameters of the deformation-stress prediction model. σ For the deformation-stress prediction model loss, σ1, σ2, and σ3 are the calculated values ​​of the major principal stress, intermediate principal stress, and minor principal stress, respectively, and δ is the stress value. Digital twin ω represents the predicted value from the digital twin, and ω is a parameter of the deformation-stress prediction model. The transformed data is after dimensionality reduction; PCA(·) is the dimensionality reduction operation; h t To be in a hidden state, o tFor the memory state, r is a state related to h. t For Bernoulli random variables of the same dimension, the probability of their value being 1 is p. LSTM(·) is the LSTM framework, h t-1 To hide the state for the next moment, o t-1 To remember the state at the next moment, ⊙ represents element-wise multiplication, and Bernoulli(·) is Bernoulli randomness. and These are the predicted values ​​for the major principal stress, intermediate principal stress, and minor principal stress.

[0028] Furthermore, the damage risk analysis model in S5 is as follows:

[0029]

[0030] in, To disrupt the risk analysis model's i-th sampling result at time t, and Let be the major principal stress, intermediate principal stress, and minor principal stress at time t. Random Dropout sampling is performed on the deformation-stress prediction function, where T is the number of random Dropout samples. Let P(·) be the predicted failure risk value for high arch dams, and let P(·) be the indicator function. is the global Drucker-Prager yield function value, and n is the number of connections between plastic elements.

[0031] The beneficial effects of this invention are:

[0032] (1) This invention constructs a digital twin model of a high arch dam based on the fusion of neural network and finite element method, realizing the mapping from measurement point monitoring to global deformation, and further combining deep learning and finite element simulation to predict the stress distribution of the dam body. This method improves the prediction accuracy while avoiding the dependence of traditional finite element method on a large number of physical parameters and reducing the computational resource consumption caused by complex iterative calculations, thereby significantly improving the real-time performance and computational efficiency of the model.

[0033] (2) Compared with existing digital twin methods, this invention introduces spatial variability constraints and a skip connection strategy, and verifies its effectiveness through ablation experiments. This optimization strategy can more accurately capture the spatial distribution characteristics of dam deformation, optimize data transmission paths, and avoid information attenuation in deep networks, thereby improving the stability and generalization ability of predictions. In addition, the deformation-stress prediction model constructed based on dimensionality reduction technology realizes accurate extrapolation of the stress state of the entire dam body. At the same time, this method supports dynamic model updates and can extrapolate predictions for unknown extreme working conditions, making up for the insufficient generalization ability of traditional deep learning methods when dealing with working conditions that exceed the distribution of the training set, thereby enhancing the adaptability of structural safety monitoring.

[0034] (3) The global plastic failure risk assessment method proposed in this invention combines uncertainty quantification analysis, effectively addressing the multi-source uncertainty problem in the monitoring process of high arch dam structures, and realizing real-time assessment of the safety of the dam structure. 800,000 random sampling tests based on this method show that, under different dropout rates, the probability of plastic failure of the dam body remains within the safety threshold range, verifying the robustness of this invention under complex working conditions.

[0035] (4) This invention breaks through the limitations of the traditional "point monitoring" model, realizing real-time perception, prediction, and early warning of deformation across the entire area of ​​high arch dams, and can combine mechanistic methods such as plastic failure analysis for real-time structural safety assessment. Compared with existing numerical simulation and statistical monitoring methods, this method not only improves prediction accuracy, computational efficiency, and generalization ability, but also optimizes data security and real-time monitoring capabilities, providing an economical and efficient solution for the intelligent and refined safety management of high arch dams, and has broad engineering application prospects. Attached Figure Description

[0036] Figure 1 This is a flowchart of a risk prediction method for high arch dams based on a digital twin model.

[0037] Figure 2 Flowchart for constructing a digital twin model of a high arch dam.

[0038] Figure 3 This is a diagram showing the layout of monitoring points for deformation monitoring of the DG Dam.

[0039] Figure 4 A process line graph for monitoring the measured values ​​at the measurement points.

[0040] Figure 5 This is a comparison chart of mesh generation and partitioning in a finite element model.

[0041] Figure 6 The figure shows a comparison of the accuracy of the proposed model with that of the finite element model and existing models.

[0042] Figure 7 This is a comparison chart of the accuracy of the proposed model and the ablation model.

[0043] Figure 8 A comparison diagram showing the spatial variation of the proposed model, the finite element model, and existing models in 3D model slices.

[0044] Figure 9 A comparison of the spatial variation patterns of the proposed model and the ablation model.

[0045] Figure 10 This is a comparison chart of the prediction accuracy between the monitoring model and the baseline model.

[0046] Figure 11 This is a graph showing the analysis results of the fault risk analysis model. Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0048] like Figure 1 As shown, a method for predicting the risk of high arch dams based on a digital twin model includes the following steps:

[0049] S1: Construct a monitoring model for measuring points, and use the environmental impact factors extracted based on the HST model as input to the monitoring model to obtain the predicted values ​​of the deformation process of the high arch dam.

[0050] S2: Construct a measurement point mapping model and train the model using a dynamic monitoring loss function;

[0051] S3: Construct a digital twin model, input the predicted values ​​of the deformation process of the high arch dam into the trained measurement point mapping model through the measurement point monitoring model, and obtain the three-dimensional global deformation situation;

[0052] S4: Construct a deformation-stress prediction model, use principal component analysis to reduce the dimensionality of the three-dimensional global deformation, and obtain the predicted values ​​of the major principal stress, intermediate principal stress and minor principal stress of the three-dimensional global deformation through the deformation-stress prediction model;

[0053] S5: Construct a failure risk analysis model, use the Monte Carlo Dropout method to simulate the predicted distribution of the model under different weight configurations, and perform random Dropout sampling to obtain the predicted value of failure risk of high arch dams, thus completing the risk prediction of high arch dams based on the digital twin model.

[0054] The monitoring model for measurement points in S1 is as follows:

[0055]

[0056] Among them, H 0 The environmental features are input to the graph convolution feature extractor. and Let W be the set of water load influencing factors, temperature influencing factors, and time-effect influencing factors at time t. 0 H is a trainable weight matrix. (·) Let be the feature matrix of the influence factors of the nodes, l be the number of node layers, and σ be the nonlinear activation function. Let A be the normalized degree matrix, A ~ be the adjacency matrix with self-connected influence factors, and W be the adjacency matrix. (·) is a trainable weight matrix.

[0057] The aforementioned monitoring model first extracts environmental impact factors as input variables based on the HST model, and then adopts a feature extractor-decoder architecture based on GCN.

[0058] The essence of this formula is to normalize and aggregate the information of the neighboring nodes of the impact factor, so that the characteristics of each layer of impact factor nodes depend not only on themselves, but also on the information of their neighboring impact factor nodes.

[0059] The measurement point mapping model in S2 is as follows:

[0060]

[0061] θ = (θ1, θ2)

[0062] in, Let W be the predicted value of the measurement point mapping model, W be the learnable weights, and b be the learnable bias. This is the output state of the LSTM. f is the input point for the sensor measurement value. LSTM (·) represents the LSTM network, θ1 represents the parameters of the LSTM network, θ2 represents the parameters of W and b, and θ represents all parameters of the measurement point mapping model.

[0063] The dynamic monitoring loss function L in S2 Digital twin Including the accuracy loss function L accuracy and spatial variability loss function L space The formula is:

[0064] L Digital twin =λ1L accuracy +λ2L space

[0065]

[0066] Where λ1 and λ2 are weight parameters, N is the total number of global computation points, i is the i-th global computation point, and l is the loss function calculation method. and The spatial gradient between the predicted values ​​from the measurement point mapping model and the numerical simulation results is represented by α, where α is the loss balance parameter. The loss of the finite element node corresponding to the sensor measurement point. For the loss of sensor measuring points, This represents the loss of all finite element nodes.

[0067] The digital twin model in S3 is as follows:

[0068]

[0069] in, f is the digital twin prediction value for time t. Synapse (·) represents the measurement point mapping model, f Cell body (·) represents the monitoring model for measurement points, θ Digital twin These are the pre-trained parameters for the measurement point mapping model. and It is a set of factors influencing water load, temperature, and time.

[0070] The deformation-stress prediction model in S4 is as follows:

[0071]

[0072] in, f is the predicted stress value. σ (·) is the deformation-stress prediction function, ω σ L represents the parameters of the deformation-stress prediction model. σ For the deformation-stress prediction model loss, σ1, σ2, and σ3 are the calculated values ​​of the major principal stress, intermediate principal stress, and minor principal stress, respectively, and δ is the stress value. Digital twin ω represents the predicted value from the digital twin, and ω is a parameter of the deformation-stress prediction model. The transformed data is after dimensionality reduction; PCA(·) is the dimensionality reduction operation; h t To be in a hidden state, o t For the memory state, r is a state related to h. t For Bernoulli random variables of the same dimension, the probability of their value being 1 is p. LSTM(·) is the LSTM framework, h t-1 To hide the state for the next moment, o t-1 To remember the state at the next moment, ⊙ represents element-wise multiplication, and Bernoulli(·) is Bernoulli randomness. and These are the predicted values ​​for the major principal stress, intermediate principal stress, and minor principal stress.

[0073] The deformation-stress prediction model uses principal component analysis to reduce the dimensionality of the dam deformation twin data, constructs an LSTM-based prediction framework, and captures the spatiotemporal evolution characteristics of dam stress through time series modeling.

[0074] The damage risk analysis model in S5 is as follows:

[0075]

[0076] in, To disrupt the risk analysis model's i-th sampling result at time t, and Let be the major principal stress, intermediate principal stress, and minor principal stress at time t. Random Dropout sampling is performed on the deformation-stress prediction function, where T is the number of random Dropout samples. Let P(·) be the predicted failure risk value for high arch dams, and let P(·) be the indicator function. is the global Drucker-Prager yield function value, and n is the number of connections between plastic elements.

[0077] In one embodiment of the invention, the primary function of the DG hydropower station is power generation. The dam site controls an area of ​​62,727 square kilometers, accounting for 81% of the total area of ​​the Dadu River basin. The main structures of the power station include a concrete double-curvature arch dam, spillway energy dissipation structures, and water diversion and power generation structures, with a maximum dam height of 210.00 meters. The normal reservoir water level is 1130.00 meters, and the dead water level is 1120.00 meters. The reservoir is approximately 32.1 kilometers long, with a storage capacity of approximately 742 million cubic meters below the normal water level and a regulating capacity of 117 million cubic meters, capable of daily regulation. Figure 3 As shown, the horizontal displacement monitoring of the DG dam includes the vertical line method, conducted in dam sections 6, 10, and 14 (arch crown beam), dam sections 19 and 24, and the two abutment grouting galleries. A total of 7 vertical lines were set up to monitor the horizontal displacement of the dam body. Figure 4 As shown, this study uses horizontal deformation monitoring data from January 1, 2016 to September 1, 2022, and numerical simulation data from the finite element method as the dataset. Training, validation, and test data are divided proportionally. The ratio of training, validation, and test data is shown in Table 1. To evaluate the proposed method, the finite element model and the following ablation models were selected for comparison: general neural network (LSTM framework only), Synapse model, ablation model I (eliminating spatial variability loss), and ablation model II (eliminating skip connections). The parameters of all models were kept consistent, with 192 hidden layers.

[0078] Table 1 Dataset Information

[0079]

[0080] The specific steps for establishing a digital twin model of DG using this invention are as follows:

[0081] (1) Based on the structural characteristics, topography, and geological conditions of the DG project, a three-dimensional finite element model of the arch dam and foundation was established. In the direction of water flow, the simulation range extends outward from the dam heel and toe to 2.5 times the dam height. In the crossflow direction, it extends outward from the left and right dam shoulders to twice the dam height. In the vertical direction, it extends downward from the foundation surface to twice the dam height. The arch dam and foundation are divided into 22,019 eight-node three-dimensional solid elements, including 4,902 dam body elements, 16,414 bedrock elements, and 703 weak structural interface elements. Through parameter sensitivity analysis, it was found that the elastic modulus of the three dam body sections and the deformation modulus of the three bedrock sections are sensitive parameters of the arch dam displacement. Therefore, in this case, these six parameters were selected as time-varying inversion variables. The simulation structure interface and the three-dimensional finite element model mesh are as follows: Figure 5 As shown in Table 2, based on the project design report and core sample tests, the inversion results of the physical and mechanical parameters of the dam concrete and foundation rock used in this case study are as follows.

[0082] Table 2 Key Material Parameters of DG Dam

[0083]

[0084]

[0085] (2) Establishment of the monitoring model for measuring points. Based on step S1, a monitoring model for measuring points is constructed. The input is the hydraulic load factor (H, H) constructed based on the HST model. 2 H 3 H 4 Temperature factor and the time factor (θ, ln(θ), e) -θ The system outputs predicted values ​​of the longitudinal and transverse deformation along the river at 18 measuring points, including PL6-1-PL24-3. The experiment was conducted using PyCharm IDE and PyTorch 2.5.1 based on Python 3.10 within a Windows 10 environment. The server configuration was as follows: Intel Core i7-13700 CPU and Nvidia RTX 4070 Super GPU. In the experiment, the learning rate was set to 0.001, the batch size was 64, the Adam optimizer was selected to update the network parameters, and the number of iterations was set to 500. The model had 64 hidden layers, 2 graph convolutional layers, and 2 multilayer perceptron layers.

[0086] (3) Establishment of the measuring point mapping section model. Based on step S2, a measuring point mapping section model was constructed. The finite element calculation values ​​of the longitudinal and transverse deformations of 18 measuring points, including PL6-1-PL24-3, were obtained, and the output was the predicted values ​​of all finite element nodes. During training, the loss function consisted of two parts: spatial variability loss and accuracy loss (loss on the prediction of actual monitored quantities, loss on the prediction of node calculation values ​​corresponding to the arranged measuring points, and loss on the prediction of all finite element node calculation values), to dynamically adjust the comprehensive learning of the actual monitoring situation and the finite element calculation situation. The experiment was conducted using PyCharm IDE and PyTorch 2.5.1 based on Python 3.10 under a Windows 10 system environment. The server configuration was as follows: CPU: Intel Core i7-13700; GPU: Nvidia RTX 4070SSUPER. In the experiment, the learning rate was set to 0.001, the batch size was 64, the Adam optimizer was selected to update the network parameters, the number of iterations was set to 500, and the model had 192 hidden layers.

[0087] (4) Establishment of the digital twin model. Based on step S3, the predicted values ​​of the monitoring effect quantities of the longitudinal and transverse deformations of 18 measuring points, including PL6-1-PL24-3, are input into the pre-trained measuring point mapping section model, outputting the deformation of the high arch dam in three dimensions. Comparison with the finite element model, ablation model, and existing models is shown below. Figure 6 , Figure 7 , Figure 8 , Figure 9 As shown, Figure 8 (a) is 1135 meters, (b) is 1080 meters, (c) is 1030 meters, and (d) is 980 meters. Figure 9 (a) is May 1, 2021, (b) is August 1, 2021, (c) is May 1, 2022, (d) is August 1, 2022, (e) is 1135m, (f) is 1080m, (j) is 1030m, and (h) is 980m.

[0088] (5) Deformation-Stress Prediction Model Establishment. Following step S4, the three-dimensional deformation data calculated from the finite element model of the high arch dam is reduced to dimensionality using PCA and used as input. The output is the major principal stress, intermediate principal stress, and minor principal stress in the three-dimensional global domain. The experiment was conducted using PyCharm IDE and PyTorch 2.5.1 based on Python 3.10 in a Windows 10 system environment. The server configuration was as follows: CPU: Intel Core i7-13700; GPU: Nvidia RTX 4070 SSUPER. In the experiment, the learning rate was set to 0.001, the batch size was 64, the Adam optimizer was selected to update the network parameters, the number of iterations was set to 500, and the model had 192 hidden layers. Accuracy analysis is as follows: Figure 10 As shown.

[0089] (6) Damage Risk Analysis Model. Based on steps S4 and S5, the three-dimensional deformation data calculated from the digital twin model of the high arch dam using pre-trained parameters is used as input, and the output is the major principal stress, intermediate principal stress, and minor principal stress across the three-dimensional global domain. The plastic damage risk of the high arch dam is analyzed in real-time using Monte Carlo Dropout sampling of 800,000 times. The experiment was conducted using PyCharm IDE and PyTorch 2.5.1 based on Python 3.10 under a Windows 10 system environment. The server configuration was as follows: CPU: Intel Core i7-13700; GPU: Nvidia RTX 4070 SSUPER. In the experiment, the learning rate was set to 0.001, the batch size was 64, the Adam optimizer was selected to update the network parameters, the number of iterations was set to 500, and the model had 192 hidden layers. Results analysis is as follows: Figure 11 As shown in the figure, (a) represents the average value of node 25319 under different Dropout parameters, (b) represents the standard deviation of node 25319 under different Dropout parameters, (c) represents the MC sampling statistics of node 25319 under different Dropout parameters on May 1, 2021, and (d) represents the real-time risk process line of global plastic failure of the digital twin model under different Dropout parameters.

[0090] This embodiment applies a digital twin model of a high arch dam to a high arch dam project. Compared with traditional numerical simulation methods, this digital twin model not only improves prediction accuracy but also more accurately captures the structural response characteristics of the dam body. Its prediction accuracy is 47.09% higher than that of the numerical model based on parameter inversion. Ablation experiments verify the spatial variability loss function and skip connection strategy of the measuring point-mapped section model, which can effectively optimize the spatial distribution of deformation and improve prediction accuracy. In addition, the deformation-stress prediction model successfully realizes the mapping from the deformation state to the stress state of the entire dam body. The predicted stress distribution is highly consistent with the finite element calculation results. The mean absolute errors (MAE) of the three principal stresses are 0.00509 MPa, 0.00679 MPa, and 0.01243 MPa, respectively, indicating that it has high prediction accuracy. Based on the proposed digital twin framework, real-time assessment of the risk of plastic failure of the entire high arch dam is also realized. In 800,000 random sampling tests, the triaxial principal stress distribution tended to a normal distribution as the dropout rate increased. Under all dropout rates tested, from May 1, 2021 to September 1, 2022, the probability of plastic failure of the high arch dam was always 0, indicating that its safety margin was high, which is consistent with the actual observation.

[0091] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the invention.

Claims

1. A method for predicting the risk of high arch dams based on a digital twin model, characterized in that, Includes the following steps: S1: Construct a monitoring model for measuring points, and use the environmental impact factors extracted based on the HST model as input to the monitoring model to obtain the predicted values ​​of the deformation process of the high arch dam. S2: Construct a measurement point mapping model and train the model using a dynamic monitoring loss function; The measurement point mapping model in S2 is as follows: ； ； ； in, The predicted value of the measurement point mapping model. For learnable weights, For learnable biases, This is the output state of the LSTM. For the input point of the sensor measurement value, For LSTM networks, For the parameters of the LSTM network, for and The parameters, For all parameters of the measurement point mapping model, It is a non-linear activation function; The dynamic monitoring loss function in S2 Including accuracy loss function and spatial variability loss function The formula is: ； ； ； in, and For weight parameters, This represents the total number of calculation points across the entire region. For the first One global computation point, For loss function calculation methods, and The spatial gradient between the predicted values ​​of the measurement point mapping model and the numerical simulation results. For loss balancing parameters, The loss of the finite element node corresponding to the sensor measurement point. For the loss of sensor measuring points, For the loss of all finite element nodes; S3: Construct a digital twin model, input the predicted values ​​of the deformation process of the high arch dam into the trained measurement point mapping model through the measurement point monitoring model, and obtain the three-dimensional global deformation situation; The digital twin model in S3 is as follows: ； ； in, For time The digital twin prediction value, For the measurement point mapping model, For the monitoring model of measurement points, These are the pre-trained parameters for the measurement point mapping model. , and It is a collection of factors influencing water load, temperature, and time. S4: Construct a deformation-stress prediction model, use principal component analysis to reduce the dimensionality of the three-dimensional global deformation, and obtain the predicted values ​​of the major principal stress, intermediate principal stress and minor principal stress of the three-dimensional global deformation through the deformation-stress prediction model; The deformation-stress prediction model in S4 is as follows: ； ； ； ； ； in, This is the predicted stress value. For deformation-stress prediction function, The parameters of the deformation-stress prediction model, For the deformation-stress prediction model loss, , and These are the calculated values ​​for the major principal stress, intermediate principal stress, and minor principal stress. For digital twin predictions, The parameters of the deformation-stress prediction model, The deformed data after dimensionality reduction. For dimensionality reduction processing, In hidden state, For memory state, For a with For Bernoulli random variables of the same dimension, the probability that the value is 1 is... , For the LSTM framework, To hide the state for the next moment, To remember the state in the next moment. This indicates element-wise multiplication. For Bernoulli randomization, , and These are the predicted values ​​for the major principal stress, intermediate principal stress, and minor principal stress; S5: Construct a failure risk analysis model, use the Monte Carlo Dropout method to simulate the predicted distribution of the model under different weight configurations, and perform random Dropout sampling to obtain the predicted value of failure risk of high arch dams, thus completing the risk prediction of high arch dams based on the digital twin model.

2. The method for predicting the risk of high arch dams based on a digital twin model according to claim 1, characterized in that, The monitoring model for measurement points in S1 is as follows: ； ； in, The environmental features are input to the graph convolution feature extractor. , and For time The set of factors influencing water conservancy load, temperature, and time-dependent effects. For a trainable weight matrix, The influence factor feature matrix of the node, The number of node layers. It is a non-linear activation function. The normalized degree matrix, To add the influence factor adjacency matrix of self-connected components, is a trainable weight matrix.

3. The method for predicting the risk of high arch dams based on a digital twin model according to claim 1, characterized in that, The damage risk analysis model in S5 is as follows: ； ； in, To disrupt the risk analysis model in The first moment Second sampling results , and for The major principal stress, intermediate principal stress, and minor principal stress at any given moment. Random dropout sampling is performed on the deformation-stress prediction function. This represents the number of random Dropout samples. This represents the predicted failure risk value for high arch dams. For indicator functions, The Drucker-Prager yield function value over the entire domain. This represents the number of connections present in the plastic element.

Images