A dam displacement interval prediction method and system based on an adaptive temporal network

Abstract

The application discloses a dam displacement interval prediction method and system based on an adaptive time sequence network, belongs to the technical field of dam safety monitoring, and comprises the following steps: embedding a differentiable feature selection layer at an input end, learning sparse weights for each influence factor by adopting Gumbel-Softmax continuous relaxation technology, and realizing dynamic optimization of features; a time sequence convolution network and a multi-head self-attention path are constructed in parallel, and local time sequence modes and global dependency relationships are extracted respectively; an heteroscedastic regression branch is designed to estimate observation noise variance, the posterior distribution of model parameters is approximated in combination with Monte Carlo Dropout, Markov distance anomaly detection is introduced, and the joint quantification of accidental uncertainty, cognitive uncertainty and abnormal working condition uncertainty is realized; and the application can improve the prediction accuracy and interval coverage reliability of dam displacement points, and provides support for engineering risk decision-making.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy project safety monitoring and artificial intelligence technology, specifically involving a method and system for predicting dam displacement intervals based on adaptive feature optimization, temporal feature encoding and multi-source uncertainty joint quantification. Background Technology

[0002] As a core infrastructure for flood control, water resource regulation, and energy development, the long-term operational safety of dams is of paramount importance. Displacement, as the most direct macroscopic indicator of dam service performance, plays an irreplaceable role in risk warning and safety decision-making through accurate prediction and uncertainty quantification.

[0003] Existing dam displacement prediction models are mainly divided into three paradigms: deterministic models (such as the finite element method), statistical models (such as the HTT / HST model), and data-driven models (such as machine learning and deep learning models). Among these, while deep learning models demonstrate strong nonlinear fitting capabilities, they still suffer from the following common technical bottlenecks: First, the influencing factors are statically fixed, lacking a dynamic adaptive optimization mechanism. Current models generally use water level polynomials, temperature measurements, and time-related terms as fixed input features. However, the contribution of each factor varies significantly over time under different operating conditions—the water level term dominates during periods of high water levels, the weight of annual temperature measurement points surges during extreme temperatures, and the signal-to-noise ratio of corresponding features decreases when sensors age or drift. This static feature combination leads to a decline in the model's generalization ability when operating conditions change, and it cannot actively suppress noise pollution introduced by low-quality sensors.

[0004] Second, the quantification of uncertainty is incomplete, neglecting the joint modeling of cognitive uncertainty and heteroscedastic noise. Although quantile regression has been successfully applied to generate prediction intervals and effectively quantifies random uncertainty, existing methods generally ignore cognitive uncertainty—that is, uncertainty caused by limited training samples and the randomness of model parameters. In addition, monitoring data generally contain heteroscedastic noise, and traditional quantile regression cannot explicitly model the time-varying characteristics of noise.

[0005] Therefore, there is an urgent need to develop an intelligent displacement interval prediction method that can adaptively select optimal features and jointly quantify multi-source uncertainties. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for predicting dam displacement intervals based on adaptive time series networks. This method achieves dynamic optimization of influencing factors and combines quantitative data uncertainty, model parameter uncertainty, and abnormal operating condition uncertainty to improve the accuracy of dam displacement point prediction and the reliability of interval coverage.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a method for predicting dam displacement intervals based on adaptive time-series networks, comprising the following steps: Step 1: Obtain the historical impact factor sequence of the dam and the corresponding displacement monitoring data, and construct training and testing sets. Normalize all data to the [0,1] interval to eliminate the influence of dimensions; Step 2: Construct an adaptive feature selection layer. Input the historical influence factor sequence into this layer, and use a differentiable gating network to dynamically generate sparse weights for each influence factor at each time step to obtain a weighted feature sequence. Step 3: Construct a dual-path temporal feature encoding module, input the weighted feature sequence into the temporal convolutional network path and the multi-head self-attention path respectively, extract local temporal patterns and global dependencies, and fuse the features output by the two paths to obtain a temporal depth representation; Step 4: Construct an uncertainty joint quantification module. Based on the time series deep representation, estimate the observation noise variance through heteroscedastic regression branch, quantify the cognitive uncertainty by approximating the posterior distribution of the model parameters through Monte Carlo Dropout, and quantify the uncertainty of abnormal working conditions through Mahalanobis distance anomaly detection mechanism. Step 5: Based on the time-series depth representation, output the displacement prediction value corresponding to the preset quantile through the quantile regression branch, and fuse the observation noise variance, cognitive uncertainty and abnormal working condition uncertainty to improve the accuracy of the prediction interval width and generate the final dam displacement interval prediction result.

[0008] In a preferred embodiment of the present invention, in step 1: According to the theory of safety monitoring of concrete dam deformation, the total deformation Decomposed into water pressure components Temperature components and time-sensitive components Construct the following factor set: Water pressure factor: based on upstream water depth A first-order, second-order, and third-order polynomial is denoted as . , , ; Temperature factor: The periodic temperature change is described using a Fourier series form, and a... to , respectively represent , , , Incremental form; Time factor: Describes concrete creep and bedrock creep using a combination of logarithmic and linear forms, structural... to , respectively represent , , , ,in Time is dimensionless; Normalize all data to the interval [0,1].

[0009] As a preferred embodiment of the present invention, the specific steps of the adaptive feature selection layer in step 2 include: Step 2-1: Calculate the log probability of feature importance of the original input vector through a gated network consisting of two fully connected layers; Step 2-2: Use the Gumbel-Softmax continuous relaxation technique to perform a differentiable approximation of the discrete feature selection, and generate weights for each feature; Steps 2-3: Multiply the generated weights element-wise with the original input vector to obtain the weighted feature vector; Steps 2-4: Apply L1 sparse regularization to the gating weights, forcing the network to assign weights close to 0 to most features, retaining only a few key features, thus achieving an active noise reduction mechanism at the input. The weighted feature sequence serves as the input to subsequent modules.

[0010] As a preferred embodiment of the present invention, the specific steps of the dual-path temporal feature encoding module in step 3 include: Step 3-1: Temporal convolutional network path extracts local temporal patterns by stacking causal dilated convolutional layers; Step 3-2: Multi-head self-attention paths extract global dependencies through positional encoding and scaled dot product attention calculation; Step 3-3: Concatenate the dual-path outputs along the feature dimension, and then fuse and reduce the dimensionality through a fully connected layer to obtain a temporal deep representation containing local and global information.

[0011] As a preferred embodiment of the present invention, the specific steps of the uncertainty joint quantization module in step 4 include: Step 4-1: Based on the time-series deep representation, a heteroscedastic regression branch is connected in parallel to directly estimate the log-variance of the observation noise, and then optimized using negative log-likelihood loss; Step 4-2: Insert a Dropout layer after the fully connected layer of the model. During the inference phase, keep the Dropout active and perform multiple random forward propagations. Calculate the sample variance of the predicted values ​​as cognitive uncertainty. Step 4-3: Extract the feature vectors of the training set samples in the temporal deep representation layer, and calculate their mean and covariance matrix; for the test samples, calculate their Mahalanobis distance and map it to uncertainty weights to obtain the uncertainty of abnormal working conditions.

[0012] As a preferred embodiment of the present invention, the model training adopts a multi-task joint optimization strategy, and the total loss function is a weighted sum of three parts: quantile loss, negative log-likelihood loss, and sparse regularization loss.

[0013] In a preferred embodiment of the present invention, the historical influence factors include at least a water pressure component factor, a temperature component factor, and a time-dependent component factor; wherein the water pressure component factor is a polynomial of the upstream water depth, the temperature component factor is in the form of a Fourier series, and the time-dependent component factor is in the form of a logarithmic and linear combination.

[0014] The present invention also provides a dam displacement interval prediction system based on the above method, comprising: The data acquisition module is used to obtain the historical impact factor sequence of the dam; The adaptive feature selection module is used to dynamically weight and filter the input influence factor sequence; The temporal feature encoding module includes a parallel temporal convolutional network submodule and a multi-head self-attention submodule, which are used to extract multi-scale temporal features. The uncertainty quantification module includes a heteroscedasticity regression submodule, a Monte Carlo Dropout submodule, and a Mahalanobis distance anomaly detection submodule, which are used to jointly quantify multi-source uncertainties. The interval prediction output module is used to fuse multi-source uncertainties and generate the final displacement prediction interval.

[0015] VI. What are the advantages of this invention compared to the best prior art as described in Article IV?

[0016] Compared with the prior art, the present invention has the following advantages: (1) Adaptive feature selection to improve model generalization ability. By introducing a differentiable feature selection layer based on Gumbel-Softmax, dynamic and sparse selection of influencing factors is achieved. Ablation experiments show that after removing this module, the root mean square error of prediction increases by an average of 15.9%, the mean absolute error increases by 13.8%, and the coefficient of determination decreases by 5.7%, which verifies the significant effectiveness of the dynamic feature selection mechanism in suppressing low-quality sensor noise and improving the model's generalization ability.

[0017] (2) Dual-path temporal coding enhances multi-scale feature extraction. By integrating a temporal convolutional network and a multi-head attention mechanism in parallel, the local receptive field of convolution and the global interaction capability of attention are effectively combined. This architecture achieves a point prediction accuracy of 0.465 mm average root mean square error and 0.956 average coefficient of determination on the test set.

[0018] (3) Joint quantification of multi-source uncertainties improves the reliability of interval prediction. By organically integrating heteroscedastic regression, Monte Carlo Dropout, and Mahalanobis distance anomaly detection, the joint quantification of data uncertainty, model parameter uncertainty, and abnormal operating condition uncertainty is achieved. At a 95% confidence level, the predicted interval coverage probability reaches 96.2%, the normalized average width is only 0.084, and the continuous rank probability score is 0.045, achieving the best balance between interval coverage reliability and sharpness. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the overall architecture of the interval prediction model in this embodiment of the invention.

[0020] Figure 2 This is a schematic diagram of the adaptive feature selection layer in this embodiment.

[0021] Figure 3 This is a schematic diagram of the dual-path temporal feature encoding module in this embodiment.

[0022] Figure 4 This is a flowchart of the multi-source uncertainty joint quantization module in this embodiment.

[0023] Figure 5 This is a three-dimensional schematic diagram of a concrete gravity dam in this embodiment.

[0024] Figure 6 for Figure 5 A schematic diagram of the layout of measuring points in a medium-density concrete gravity dam.

[0025] Figure 7 This is a schematic diagram of point prediction comparison analysis in this embodiment.

[0026] Figure 8 This is a schematic diagram comparing the point prediction analysis indicators in this embodiment.

[0027] Figure 9 This is a schematic diagram of the test set interval prediction comparison in this embodiment.

[0028] Figure 10 This is a schematic diagram comparing the interval prediction analysis indicators in this embodiment.

[0029] Figure 11 This is a schematic diagram showing the comparison of ablation experiment verification results in this embodiment. Detailed Implementation

[0030] The present invention will be further described in detail below with reference to specific implementation examples and accompanying drawings.

[0031] This embodiment is applied to displacement monitoring data of a concrete gravity dam. The dam has 10 sections from the left bank to the right bank, and the displacement of the dam along the river is monitored by using inverted plumb lines and tension lines. Figure 5 and 6 This is a schematic diagram of the measuring point layout of a concrete gravity dam in an embodiment of the present invention, wherein... Figure 5 This is a view of the downstream section of the dam. Figure 6 This is a schematic diagram of the distribution of measuring points.

[0032] Displacement and influencing factor data from typical measuring point EX1-2 from 2018 to 2024 were selected as the training set, and data from 2025 were selected as the test set.

[0033] Step 1: Data acquisition and impact factor construction.

[0034] According to the theory of safety monitoring of concrete dam deformation, the total deformation Decomposed into water pressure components Temperature components and time-sensitive components Construct the following factor set: Water pressure factor: based on upstream water depth A first-order, second-order, and third-order polynomial is denoted as . , , .

[0035] Temperature factor: The periodic temperature change is described using a Fourier series form, and a... to , respectively represent , , , Incremental form.

[0036] Time factor: Describes concrete creep and bedrock creep using a combination of logarithmic and linear forms, structural... to , respectively represent , , , ,in Time is dimensionless.

[0037] After the above factors are constructed, the Min-Max normalization method is used to map all data (including the shift of the dependent variable) to the [0,1] interval to eliminate the influence of dimensions and accelerate model convergence.

[0038] Step 2: Build and train the interval prediction model.

[0039] The calculation process of the model is as follows Figure 1 As shown. Figure 1 This is a diagram illustrating the overall architecture of the interval prediction model in an embodiment of the present invention. (Refer to...) Figure 1The model consists of four main modules: an adaptive feature selection module (input layer), a dual-path temporal feature encoding module (feature extraction layer), an uncertainty joint quantization module (uncertainty analysis layer), and an interval prediction output module (output layer). Dam monitoring data, after preprocessing, forms a historical impact factor sequence. After being processed by the above four modules in sequence, the final displacement prediction interval result is generated. The following section combines... Figures 2 to 4 The specific implementation methods of each module are described in detail.

[0040] (a) Adaptive Feature Selection Layer

[0041] Figure 2 This is a schematic diagram of the adaptive feature selection layer in an embodiment of the present invention. (Refer to...) Figure 2 This module is embedded in the model input end, and its workflow is as follows: set up The original input vector at time t is ,in This represents the total number of impact factors. First, it is determined through a lightweight gating network. Calculate the log probability of feature importance:

[0042] It consists of two fully connected layers, with the number of hidden units being... The activation function is ReLU. To achieve discretized feature selection and ensure gradient propagation during training, the Gumbel-Softmax continuous relaxation technique is used to perform a differentiable approximation of the discrete selection. Specifically, the... Weights of each feature Generated by sampling using the following formula:

[0043] in Temperature coefficient. Initial training settings. To encourage exploration, an exponential annealing strategy was then gradually applied to reduce [the temperature / temperature]. This makes the weight distribution approximate the discretized one-hot form. The weighted eigenvectors are:

[0044] in This represents element-wise multiplication. To prevent the model from activating too many redundant features, gating weights are applied. Sparse regularization:

[0045] is the regularization coefficient (0.01 in this paper). This regularization term forces the gating network to assign weights close to 0 to most features, retaining only a few key features, thus achieving adaptive feature selection. The weighted feature sequence As input to the subsequent timing coding module.

[0046] (ii) Dual-path temporal feature coding module

[0047] Figure 3 This is a schematic diagram of the dual-path temporal feature encoding module in an embodiment of the present invention. (Refer to...) Figure 3 This module contains temporal convolutional network paths ( Figure 3 (left side) and multi-head self-attention path ( Figure 3 (On the right side), the two processes the weighted feature sequence in parallel, and the output is fused to obtain a temporal depth representation.

[0048] Temporal convolutional networks are constructed using stacked causal dilated convolutions to ensure that future information is not leaked into the past. Let the input sequence be... First through a Convolution maps feature dimensions to hidden layer dimensions. :

[0049] Then stacked Dilated convolutional layer, dilation factor of the l-th layer kernel size Each layer of operation can be represented as:

[0050] in This article takes Experience the wild cover Each time step. The output of the last layer is denoted as... .

[0051] Multi-head self-attention path As input, timing sequence information is injected through position encoding:

[0052] in This is a sinusoidal position encoding matrix. Linear projection to query ,key ,value :

[0053] , This article sets the number of attention heads. , The scaled dot product attention output is:

[0054] Multi-head splicing and output projection:

[0055] Output the TCN path With attention path output The components are concatenated along the feature dimension, fused through a fully connected layer, and then reduced to the hidden layer dimension. :

[0056] Take the hidden state of the last time step As a temporal depth representation of the input sequence, this vector contains key feature information, local temporal patterns, and global dependencies after adaptive weighting, providing a complete feature foundation for subsequent displacement prediction and uncertainty modeling.

[0057] (III) Multi-source uncertainty joint quantification module

[0058] Figure 4 This is a flowchart of the multi-source uncertainty joint quantization module in an embodiment of the present invention. (Refer to...) Figure 4 This module jointly quantifies the uncertainty of prediction from three dimensions.

[0059] Data uncertainty is mainly manifested in the time-varying noise level of displacement monitoring values, i.e., heteroscedasticity. While traditional quantile regression can characterize the conditional distribution through outputs at different quantiles, it cannot explicitly model the variation of noise variance with the input. This paper addresses this issue in time-series deep representation... Based on this, a parallel heteroscedasticity regression branch is added to directly estimate the log-variance of the observation noise:

[0060] Meanwhile, the quantile regression branch output Preset quantiles (This article is taken from) =19, The corresponding predicted displacement value:

[0061] Assume that the displacement observations follow a Gaussian distribution. ,in Take the median quantile ( The predicted value is then used. The noise branch is optimized using negative log-likelihood loss.

[0062] The quantile regression branch uses the quantile loss function:

[0063] The data uncertainty is ultimately estimated by the observation noise variance from the heteroscedasticity branch. It is characterized in conjunction with the quantile span inherent in quantile regression.

[0064] Model parameter uncertainty stems from the insufficient constraint of limited training data on the massive parameters of deep networks. This paper employs Monte Carlo Dropout (MC Dropout) to approximate Bayesian inference, efficiently capturing cognitive uncertainty without altering the model structure or adding additional parameters. Specifically, a Dropout layer is inserted after all fully connected layers in the model (including gated networks, fully connected layers after TCN, attention projection layers, output layers, etc.), with a uniform dropout rate set to a certain value. Dropout is enabled normally during the training phase; it remains active during the inference phase. Sub-random forward propagation (in this paper, we take...) ),get Component number prediction and noise variance The uncertainty of the model parameters is measured by the sample variance of the predicted values ​​for each quantile:

[0065] In real-world engineering projects, dams may encounter extreme floods, earthquakes, sensor malfunctions, and other abnormal operating conditions, causing the current input features to deviate significantly from the training data distribution. In such cases, even if the model has good point prediction accuracy, the confidence level of its prediction results should be automatically lowered. To quantify the uncertainty caused by out-of-distribution samples, this paper introduces an anomaly detection mechanism based on Mahalanobis distance in the feature space.

[0066] First, extract all samples from the training set into the temporal deep representation layer. For the eigenvectors of the test samples, calculate their mean vector and covariance matrix (using a shrinkage estimate for numerical stability). Calculate its Mahalanobis distance:

[0067] A larger Mahalanobis distance indicates that the sample is more likely to be an outlier. Uncertainty weights are applied to map the Mahalanobis distance to the [0,1] interval:

[0068] in The training set uses the 95th percentile of the Mahalanobis distance. Ultimately, outlier parameter uncertainty acts on the prediction interval as an interval amplification factor:

[0069] For the standard deviation of the training set displacement, This is the amplitude coefficient (2.0 in this paper). This mechanism ensures that the model can actively widen the prediction interval when facing unknown operating conditions, avoiding overconfidence.

[0070] The aforementioned modules are organically integrated to form a complete adaptive feature optimization and uncertainty joint quantification displacement interval prediction model. The model uses historical influence factor sequences... As input, a weighted sequence is obtained through an adaptive feature selection layer. The time-series feature encoding module will Mapped to depth representation ;based on The quantile regression branch outputs the predicted values ​​for each quantile. Heteroscedastic branch output observation noise variance During the inference phase, Monte Carlo Dropout is enabled to obtain M sets of predicted samples.

[0071] Finally, the first Quantile corrected predicted values ​​and confidence levels The prediction interval is generated by the following formula:

[0072]

[0073] The key parameters of the model are set as follows:

[0074] Adaptive Feature Selection Layer: The detailed computational process of the model is as follows... Figure 2 As shown, the number of hidden units in the gated network is 32, the initial temperature of Gumbel-Softmax is 5.0, the final temperature is 0.5, and the sparsity regularization coefficient is 0.01.

[0075] Dual-path encoding module: The temporal convolutional network contains 4 layers of dilated convolutions with a kernel size of 3 and a hidden layer dimension of 64; the multi-head attention has 8 heads, a model dimension of 128, and a dropout rate of 0.1.

[0076] Uncertainty quantification module: Monte Carlo Dropout dropout rate is 0.1, sampling count is 50; heteroscedasticity branch hidden unit count is 32; anomaly detection amplitude coefficient is 2.0.

[0077] Training parameters: The optimizer is Adam, the initial learning rate is 1e-3, decays by 0.5 every 20 epochs, the batch size is 64, the maximum training epochs is 200, and the early stopping patience value is 15. In the total loss function, the quantile loss weight is 1.0, the heteroscedasticity loss weight is 0.2, and the sparsity regularization weight is 0.01.

[0078] according to Figure 1 The architecture shown utilizes the training set data to perform multi-task joint training of the model until the validation set loss converges.

[0079] Step 3: Model performance evaluation.

[0080] The trained model was used to make predictions on the 2025 test set data to evaluate the performance of point prediction and interval prediction.

[0081] Point prediction results: such as Figure 7 , Figure 8 As shown, the curves of measured values ​​and predicted values ​​at each measuring point highly coincide. The root mean square error of each measuring point is 0.472 mm, and the coefficient of determination is 0.948 for all points.

[0082] Interval prediction results: such as Figure 9 , Figure 10 As shown, the 95% prediction interval automatically widens when displacement fluctuations are large, and the measured values ​​are basically within the interval. The interval coverage rate of the measuring points is 0.965; the normalized average width is 0.138. The comprehensive coverage width criterion index is the lowest at all measuring points.

[0083] Ablation experiment verification: such as Figure 11 As shown, by removing the adaptive feature selection module, heteroscedastic regression branch, Monte Carlo Dropout, and anomaly detection module respectively, the complete model achieves optimal performance across all measurement points. This demonstrates that the modules proposed in this invention are effective and synergistic.

[0084] In summary, the dam displacement interval prediction method based on adaptive temporal network proposed in this invention can effectively solve the problems of incomplete feature selection staticization and uncertainty quantification in the existing technology. It has the advantages of high accuracy, high reliability and strong generalization ability, and is suitable for promotion and application in the field of dam safety monitoring.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting dam displacement intervals based on adaptive temporal networks, characterized in that, Includes the following steps: Step 1: Obtain the historical impact factor sequence of the dam and the corresponding displacement monitoring data, and construct the training set and test set; Step 2: Construct an adaptive feature selection layer. Input the historical influence factor sequence into this layer, and use a differentiable gating network to dynamically generate sparse weights for each influence factor at each time step to obtain a weighted feature sequence. The specific steps include: Step 2-1: Calculate the log probability of feature importance of the original input vector through a gated network consisting of two fully connected layers; Step 2-2: Use the Gumbel-Softmax continuous relaxation technique to perform a differentiable approximation of the discrete feature selection, and generate weights for each feature; Steps 2-3: Multiply the generated weights element-wise with the original input vector to obtain the weighted feature vector; Steps 2-4: Apply L1 sparse regularization to the gating weights to make the weight distribution sparser; Step 3: Construct a dual-path temporal feature encoding module. Input the weighted feature sequence into the temporal convolutional network path and the multi-head self-attention path respectively to extract local temporal patterns and global dependencies. Then, fuse the features output from the two paths to obtain a temporal depth representation. Specific steps include: Step 3-1: Temporal convolutional network path extracts local temporal patterns by stacking causal dilated convolutional layers; Step 3-2: Multi-head self-attention paths extract global dependencies through positional encoding and scaled dot product attention calculation; Step 3-3: Concatenate the dual-path outputs along the feature dimension, fuse and reduce the dimensionality through a fully connected layer to obtain a temporal deep representation containing local and global information; Step 4: Construct an uncertainty joint quantification module. Based on the time series deep representation, estimate the observation noise variance through heteroscedastic regression branch, quantify the cognitive uncertainty by approximating the posterior distribution of the model parameters through Monte Carlo Dropout, and quantify the uncertainty of abnormal working conditions through Mahalanobis distance anomaly detection mechanism. Step 5: Based on the time-series depth representation, output the displacement prediction value corresponding to the preset quantile through the quantile regression branch, and fuse the observation noise variance, cognitive uncertainty and abnormal working condition uncertainty to generate the final dam displacement interval prediction result.

2. The method according to claim 1, characterized in that, In step 1: According to the theory of safety monitoring of concrete dam deformation, the total deformation Decomposed into water pressure components Temperature components and time-sensitive components Construct the following factor set: Water pressure factor: based on upstream water depth A first-order, second-order, and third-order polynomial is denoted as . , , ; Temperature factor: The periodic temperature change is described using a Fourier series form, and a... to , respectively represent , , , Incremental form; Time factor: Describes concrete creep and bedrock creep using a combination of logarithmic and linear forms, structural... to , respectively represent , , , ,in Time is dimensionless; Normalize all data to the interval [0,1].

3. The method according to claim 1, characterized in that, The specific steps of the uncertainty joint quantification module in step 4 include: Step 4-1: Based on the time-series deep representation, a heteroscedastic regression branch is connected in parallel to directly estimate the log-variance of the observation noise, and then optimized using negative log-likelihood loss; Step 4-2: Insert a Dropout layer after the fully connected layer of the model. During the inference phase, keep the Dropout active and perform multiple random forward propagations. Calculate the sample variance of the predicted values ​​as cognitive uncertainty. Step 4-3: Extract the feature vectors of the training set samples in the temporal deep representation layer, and calculate their mean and covariance matrix; for the test samples, calculate their Mahalanobis distance and map it to uncertainty weights to obtain the uncertainty of abnormal working conditions.

4. The method according to claim 1, characterized in that, The specific formula for generating the interval prediction in step 5 is as follows: ； in, It is the standard normal inverse cumulative distribution function. Let M be the mean of the Dropout predictions. The sample variance of the predicted values ​​for each quantile. To observe the noise variance, The variance of the predicted values ​​for the outlier parameters is the uncertainty. Confidence level The prediction interval To predict the lower limit of the interval, To predict the upper limit of the interval, , where is the confidence level.

5. The method according to claim 1, characterized in that: The model training adopts a multi-task joint optimization strategy, and the total loss function consists of a weighted sum of three parts: quantile loss, negative log-likelihood loss, and sparse regularization loss.

6. The method according to claim 1, characterized in that: The historical influencing factors include at least a water pressure component factor, a temperature component factor, and a time-dependent component factor; wherein the water pressure component factor is a polynomial of the upstream water depth, the temperature component factor is in the form of a Fourier series, and the time-dependent component factor is in the form of a logarithmic and linear combination.

7. A dam displacement interval prediction system based on the method of any one of claims 1 to 6, characterized in that, include: The data acquisition module is used to obtain the historical impact factor sequence of the dam; The adaptive feature selection module is used to dynamically weight and filter the input influence factor sequence; The temporal feature encoding module includes a parallel temporal convolutional network submodule and a multi-head self-attention submodule, which are used to extract multi-scale temporal features. The uncertainty quantification module includes a heteroscedasticity regression submodule, a Monte Carlo Dropout submodule, and a Mahalanobis distance anomaly detection submodule, which are used to jointly quantify multi-source uncertainties. The interval prediction output module is used to fuse multi-source uncertainties and generate the final displacement prediction interval.

Images