A dam displacement estimation method and system based on deep learning

Abstract

The application discloses a dam displacement estimation method and system based on deep learning, which comprises the following steps: obtaining a historical displacement sequence of a dam, a reservoir water level sequence, and an underground water level sequence and a pore water pressure sequence of a dam abutment mountain; calculating the reservoir water level change rate, the rising and falling direction, and the lag time; constructing a dam abutment constraint feature set according to the lag time and the pore water pressure sequence; performing time sequence alignment resampling and fusion on the reservoir water level sequence and the dam abutment constraint feature set to obtain a multi-dimensional driving feature; inputting the multi-dimensional driving feature into a physical driving sub-network and a data driving sub-network, wherein the two sub-networks both take the dam abutment constraint feature set as a dynamic boundary condition and respectively output a physical predicted displacement and a data predicted displacement; and adaptively adjusting the confidence according to the historical prediction errors of the two sub-networks and weighting and fusing the confidence into a final displacement estimation value. The application quantifies the dynamic constraint state of the dam abutment mountain in the process of the reservoir water level rising and falling and integrates the dynamic constraint state into a displacement prediction model, so that the precision and reliability of the dam displacement estimation are improved.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering monitoring technology, and in particular to a method and system for estimating dam displacement based on deep learning. Background Technology

[0002] Dam displacement prediction is a key technical means to ensure the safe operation of water conservancy projects. The dam body exhibits displacement response under the influence of multiple factors such as reservoir water level fluctuations, temperature changes, and temporal evolution. Accurate prediction of dam displacement is crucial for dam safety monitoring and risk early warning. However, the dam body is not an isolated structure; its left and right ends are embedded in the abutment hills, which constitute lateral constraint boundaries on the dam body. During reservoir water level fluctuations, reservoir water seeps into or out of the abutment hills, altering the seepage field distribution and pore water pressure state within the hills, thus affecting the lateral constraint capacity of the abutment hills on the dam body. This dynamic constraint mechanism makes the relationship between dam displacement and environmental factors highly nonlinear and time-varying, posing significant technical challenges to displacement prediction.

[0003] Existing methods for predicting dam displacement mainly fall into two categories: statistical regression models and deep learning models. Statistical regression models predict dam displacement by establishing empirical regression relationships between environmental variables such as reservoir water level, temperature, and time-related factors. While these methods have clear physical meanings, they struggle to capture the complex nonlinear relationships between environmental variables and displacement. Deep learning models, such as Long Short-Term Memory (LSTM) networks and gated recurrent units (GRUs), have been introduced into dam displacement prediction to capture the nonlinear temporal relationships between environmental variables and dam displacement. Some schemes are beginning to focus on the time-lag effect of environmental variables on dam displacement, calculating the lag time between water level changes and dam displacement using methods such as the maximum information coefficient. Regarding physical constraints, existing research combines mechanical mechanism constraints with deep neural networks to construct physical data fusion prediction models. By embedding physical equations as penalty terms in the form of loss functions during network training, the prediction results are made to satisfy physical laws to a certain extent.

[0004] However, existing methods have significant shortcomings in dynamically representing dam abutment constraints. Current models typically simplify the dam abutment hillside to fixed boundary conditions, failing to incorporate changes in the seepage state of the abutment hillside as input features for displacement prediction. This results in the model's inability to perceive the dynamic changes in the dam abutment constraint state during reservoir water level fluctuations. Furthermore, while some methods have addressed the time-lag effect of water level changes on dam displacement, the lag time is usually treated as a fixed parameter, neglecting the dynamic variation of the lag time with the reservoir water level change rate. Therefore, quantifying the dynamic constraint state of the dam abutment hillside during reservoir water level fluctuations and integrating it into the displacement prediction model, enabling the displacement prediction network to possess the dual advantages of physical constraints and data-driven modeling, remains a focus for those skilled in the art. Consequently, a deep learning-based dam displacement estimation method and system are needed to address these issues. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a dam displacement estimation method and system based on deep learning. It aims to overcome the shortcomings of existing dam displacement prediction methods, which simplify the dam abutment hillside as a fixed boundary condition, cannot quantify the impact of dynamic changes in the seepage state of the dam abutment on the dam displacement, and cannot dynamically adjust the lag time as a fixed parameter with the rate of change of reservoir water level. It has the characteristics of using the lag time feature and pore water pressure feature as dynamic boundary condition inputs of the physical driving sub-network and the data driving sub-network through the dam abutment constraint feature set, and adaptively weighting the confidence of the historical prediction error of the two sub-networks to output the result.

[0006] To achieve the above technical solution, the technical solution adopted by the present invention is as follows: A deep learning-based method for estimating dam displacement includes the following steps: Acquire monitoring data, including the historical displacement sequence of the dam body, the reservoir water level sequence, and the groundwater level sequence and pore water pressure sequence of the dam abutment hills; Calculate the rate of change and direction of rise and fall of reservoir water level based on reservoir water level sequence, and calculate the lag time based on groundwater level sequence; A dam shoulder constraint feature set is constructed based on the lag time and pore water pressure sequences. The dam shoulder constraint feature set includes lag time features and pore water pressure features. Using the historical displacement sequence as a unified time axis, the reservoir water level sequence and the dam abutment constraint feature set are time-aligned and resampled, and then fused into multi-dimensional driving features. Multidimensional driving features are input into the displacement prediction network, which includes a physical driving subnetwork and a data driving subnetwork. Both subnetworks use the dam abutment constraint feature set as dynamic boundary condition input. The physical driving subnetwork outputs the physical predicted displacement based on the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics. The data driving subnetwork outputs the data predicted displacement. The data driving subnetwork contains a gated recurrent unit layer and a temporal attention layer. The gated recurrent unit layer performs temporal encoding on the multidimensional driving features. The temporal attention layer calculates the attention weight based on the reservoir water level change amplitude at each historical moment, and then performs weighted fusion on the hidden states at each moment output by the gated recurrent unit layer to output the data predicted displacement. The confidence level is adaptively adjusted based on the historical prediction errors of the physical and data-based displacement predictions, and then weighted and fused to obtain the final displacement estimate.

[0007] Furthermore, the calculation process of the gated loop unit layer includes: Let the current time be The hidden state at the previous moment was The input feature vector at the current time is , update door The formula for calculation is: ; Reset door The formula for calculation is: ; Candidate hidden state The formula for calculation is: ; Current hidden state The formula for calculation is: ; in, , , This is the weight matrix. , , For bias terms, It is the sigmoid activation function. This represents element-wise multiplication. It is the hyperbolic tangent activation function.

[0008] Furthermore, each historical moment in the temporal attention layer The formula for calculating the range of reservoir water level change is: ; Attention weight The formula for calculation is: ; fused context vector The formula for calculation is: ; The data predicts displacement The formula for calculation is: ; in, for The reservoir water level at that moment. The time length of the input sequence. For the gated loop unit layer in The hidden state that is constantly being output. This is the weight matrix of the fully connected output layer. This is the bias term for the fully connected output layer.

[0009] Furthermore, the calculation method for the reservoir water level change rate is as follows: taking the current moment as the reference, the reservoir water level sequence within a preset time window length is subjected to least squares linear fitting, and the slope of the fitted line is used as the reservoir water level change rate at the current moment. When the slope of the fitted line is positive, it is determined that the reservoir water level is in the rising stage, and when it is negative, it is determined that the reservoir water level is in the falling stage.

[0010] Preferably, the physical driving sub-network uses a mechanical relationship model to calculate the physical predicted displacement. This mechanical relationship model is in the form of a second-order differential equation, where the second derivative of the dam displacement with respect to time is related to the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics. The mechanical relationship model is expressed as follows: ; in, To predict displacement in physics, For time, The rate of change of reservoir water level The characteristic value of pore water pressure. The characteristic value of the lag time. The coefficient representing the influence of the rate of change of reservoir water level. The pore water pressure influence coefficient is... The lag time influence coefficient is... For constant terms; The physics-driven subnetwork discretizes the mechanical relationship model and then recursively calculates the physical predicted displacement. The discretization method is as follows: Let the time step be The current time is the [number]th [time]. Step, the previous moment was the first step. Step 1: The first derivative of displacement with respect to time is approximated using a first-order backward difference, and the second derivative of displacement with respect to time is approximated using a difference of the first-order backward difference. The recursive formula for the physically predicted displacement is: ; in, For the first The physical prediction displacement of the step. For the first The rate of change of reservoir water level in the step, For the first The characteristic value of pore water pressure in the step, For the first The lag time characteristic value of the step; the initial value required for recursive calculation. and The measured displacement values ​​at the initial moment and the first moment in the historical displacement sequence are taken respectively.

[0011] Furthermore, the mechanical relationship model adopts a simplified first-order difference form, and the physical prediction of displacement change is directly proportional to the current values ​​of the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics: ; in, , for the first Step and the first The difference between the physical predicted displacement and the actual displacement.

[0012] Furthermore, The value range is from 0.5 to 5.0. The value range is from 0.01 to 0.50. The value range is from 0.001 to 0.100. The range of values ​​is Up to 0.10 ^2; For concrete gravity dams, The value ranges from 0.5 to 2.0; for earth-rock dams, The value range is from 2.0 to 5.0; and Based on the permeability characteristics of the dam abutment hill, it is determined that: when the permeability coefficient of the dam abutment hill is less than... At m / s The value range is from 0.01 to 0.10. The value ranges from 0.001 to 0.010; when the permeability coefficient of the dam abutment hillside is greater than... At m / s The value range is from 0.20 to 0.50. The value ranges from 0.050 to 0.100; when the permeability coefficient of the dam abutment is between m / s to When the speed is between m / s The value range is from 0.10 to 0.20. The value range is from 0.010 to 0.050.

[0013] Furthermore, each input variable undergoes scale alignment before participating in the calculation of the mechanical relationship model: each input variable is divided by its own reference value to convert it into a dimensionless relative value; the reference value for the reservoir water level change rate is the maximum historical reservoir water level change rate at the dam site, the reference value for the pore water pressure is the maximum historical pore water pressure at the dam abutment, and the reference value for the lag time is the maximum historical lag time at the dam abutment.

[0014] Preferably, the adaptive adjustment of confidence level is performed as follows: The absolute value of the difference between the predicted displacement and the measured displacement of the physical driving subnetwork and the data driving subnetwork at each moment within the historical time window is calculated as the prediction error. The confidence levels of the two subnetworks are determined based on the prediction errors of the physical-driven subnetwork and the data-driven subnetwork within the historical time window. The final displacement estimate is obtained by weighting the physical predicted displacement and the data predicted displacement according to the confidence level. When the fluctuation amplitude of the pore water pressure characteristic in the dam abutment constraint feature set exceeds a preset fluctuation threshold, the confidence level of the physical driving sub-network is reduced and the confidence level of the data driving sub-network is increased; the calculation method for the fluctuation amplitude of the pore water pressure characteristic is as follows: Take the pore water pressure sequence within a preset fluctuation statistical window length before the current time, and calculate the standard deviation of the sequence. ;when Exceeding the preset fluctuation threshold When this occurs, a confidence level switch is triggered, lowering the confidence level of the physical-driven sub-network to 0.2 and raising the confidence level of the data-driven sub-network to 0.8; when Falling back to After the preset recovery time is reached, the confidence level switching mechanism will be deactivated, and the confidence level will be restored to be adaptively determined by the average prediction error.

[0015] Preferably, the confidence level is determined as follows: the average prediction error of the physical driving sub-network within the historical time window is: ; The average prediction error of the data-driven subnetwork within the historical time window is: ; The confidence level of the physical-driven subnetwork is: ; The confidence level of the data-driven subnetwork is: ; The final displacement estimate is: ; in, The length of the historical time window. For the current moment, For indexing moments within a historical time window, For physical driving subnetworks in The physical predicted displacement at time t. For data-driven subnetworks in Predict displacement based on data at any given time. for Measured value of dam displacement at time t. and These are the average prediction errors of the physics-driven subnetwork and the data-driven subnetwork, respectively. and The confidence levels of the physical-driven subnetwork and the data-driven subnetwork are respectively. , for The final displacement estimate at time [time].

[0016] Preferably, the abutment hill includes a left abutment hill and a right abutment hill; the groundwater level sequence and pore water pressure sequence of the left abutment hill and the groundwater level sequence and pore water pressure sequence of the right abutment hill are obtained respectively; constraint feature sets for the left abutment and constraint feature sets for the right abutment are constructed respectively; the constraint feature sets for the left abutment and the constraint feature sets for the right abutment are time-series aligned and fused as independent dynamic boundary condition features; The lag time calculation steps are performed independently for the left and right abutments to obtain the dynamically adjusted lag time for the left and right abutments respectively; the proportional coefficients for the left and right abutments are independently determined based on the geological permeability characteristics of their respective abutment hills.

[0017] Preferably, during the displacement prediction process, the lag time features and pore water pressure features in the dam abutment constraint feature set are updated based on the deviation between the real-time monitored dam displacement sequence and the final displacement estimate; the updated features are then re-input into the physical driving sub-network to generate the updated physical predicted displacement. The update trigger condition is as follows: calculate the moving average of the displacement deviation sequence within the length of the deviation statistical window before the current time. When the absolute value of the moving average exceeds a preset deviation threshold, online correction is triggered. The correction amount is determined by taking the monitoring data within the length of the correction window before the trigger time, and constructing a least squares problem with the objective of minimizing the sum of squares of the deviations between the measured dam displacement and the physical predicted displacement at each time point within the correction window. ; The lag time characteristic correction quantity is obtained by solving. and pore water pressure characteristic correction amount The corrected hysteresis time characteristic values ​​and pore water pressure characteristic values ​​are as follows: ; ; in, and These are the original hysteresis time characteristic value and pore water pressure characteristic value at the trigger correction time, respectively. To correct the time intervals within the window, The physical predicted displacement is calculated using the corrected hysteresis time characteristic value and pore water pressure characteristic value as inputs; online correction only updates the characteristic input values ​​and does not update the parameters of the mechanical relationship model.

[0018] Preferably, the lag time is calculated as follows: Calculate the cross-correlation function between the groundwater level sequence at the dam abutment and the reservoir water level sequence under multiple different time lags, and take the time lag corresponding to the maximum cross-correlation function value as the basic lag time; Using the aforementioned basic lag time as a benchmark, multiply it by an exponential decay function with the absolute value of the reservoir water level change rate as the exponent to obtain the dynamically adjusted lag time. Cross-correlation function The formula for calculation is: ; Basic lag time The value corresponding to the maximum value of the cross-correlation function : ; Dynamically adjusted lag time The formula for calculation is: ; in, This is the time lag. For sequence length, For reservoir water level sequence, The groundwater level sequence at the dam abutment. and These are the mean values ​​of the reservoir water level sequence segment and the dam abutment groundwater level sequence segment, respectively, used in the calculation. The rate of change of reservoir water level This is the proportionality coefficient. The value ranges from 0.5h / m to 5.0h / m and is determined based on the geological permeability characteristics of the dam shoulder mountain.

[0019] Preferably, a displacement change rate sequence is calculated based on the final displacement estimate; Within a set time window, the number of times the displacement rate of change changes from positive to negative or from negative to positive in the rate of change sequence is used as the direction alternation frequency. When the frequency of the alternation of directions exceeds a preset frequency threshold or the amplitude of the rate of change sequence exceeds a preset amplitude threshold, it is determined that the dam displacement trend has an abnormal reversal. rate of change of displacement The formula for calculation is: ; Alternating frequency of directions The formula for calculation is: ; Displacement change rate amplitude The formula for calculation is: ; in, for The final displacement estimate at time [time]. for The final displacement estimate at time [time]. To determine the time window The total number of times the value changes from positive to negative and from negative to positive; when Exceeding the preset frequency threshold or Exceeding the preset amplitude threshold The dam displacement trend was determined to have shown an abnormal reversal. Take 0.02 times / hour. Take 1.5mm.

[0020] Preferably, the multidimensional driving features further include temperature sequences and rainfall sequences; The temperature sequence and the rainfall sequence are interpolated and resampled along the unified time axis, and then time-series aligned and fused. The interpolation and resampling uses a linear interpolation method, and the linear interpolation formula is as follows: ; in, For the target time point, and These are the closest known data time points before and after the target time point. and They are respectively and Temperature or rainfall value at any given time.

[0021] Preferably, a deep learning-based dam displacement estimation system is used to execute the deep learning-based dam displacement estimation method, comprising: The data acquisition module is used to acquire monitoring data, which includes the historical displacement sequence of the dam body, the reservoir water level sequence, and the groundwater level sequence and pore water pressure sequence of the dam abutment hill. The water level rise and fall and lag analysis module is used to calculate the rate of change and direction of rise and fall of the reservoir water level based on the reservoir water level sequence, and to calculate the lag time based on the groundwater level sequence; A dam shoulder constraint feature construction module is used to construct a dam shoulder constraint feature set based on the lag time and the pore water pressure sequence. The dam shoulder constraint feature set includes lag time features and pore water pressure features. The temporal alignment and fusion module is used to perform temporal alignment and resampling on the reservoir water level sequence and the dam abutment constraint feature set, using the time reference of the historical displacement sequence as a unified time axis, and fuse them into multi-dimensional driving features. A dual-drive displacement prediction module is used to input the multi-dimensional driving features into a displacement prediction network. The displacement prediction network includes a physical drive sub-network and a data drive sub-network. Both sub-networks are input with the dam abutment constraint feature set as dynamic boundary conditions. The physical drive sub-network outputs the physical predicted displacement based on the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics. The data drive sub-network outputs the data predicted displacement. The confidence fusion output module is used to adaptively adjust the confidence level based on the historical prediction errors of the physical predicted displacement and the data predicted displacement, and then weight and fuse them into a final displacement estimate.

[0022] The beneficial effects of this invention are as follows: 1. This invention establishes a dam abutment constraint feature set that includes lag time and pore water pressure characteristics. This feature set is then used as the common dynamic boundary condition input for both the physical-driven and data-driven sub-networks, enabling the displacement prediction network to perceive the dynamic changes in the dam abutment constraint state during reservoir water level fluctuations. The lag time of the groundwater level in the dam abutment hill to the reservoir water level reflects the hill's sensitivity to reservoir water level changes, while pore water pressure directly characterizes the hydraulic state within the hill. Together, they constitute a quantitative representation of the dam abutment constraint state. The physical-driven and data-driven sub-networks perform dual-path prediction based on the same dam abutment constraint state information, allowing the model to output displacement estimation results reflecting differences in the dam abutment constraint state under the same reservoir water level conditions. This overcomes the shortcomings of existing methods that simplify the dam abutment to fixed boundary conditions and fail to perceive dynamic changes in the constraint state.

[0023] 2. This invention employs a confidence-adaptive adjustment mechanism. It dynamically calculates and weights the confidence scores of the physical-driven and data-driven sub-networks based on their prediction errors within a historical time window, allowing the sub-network with higher prediction accuracy to have a greater weight in the fusion process. Simultaneously, when the pore water pressure fluctuation at the dam abutment exceeds a preset threshold, the system determines that the dam abutment is in a non-steady-state condition. It automatically reduces the confidence score of the physical-driven sub-network based on a steady-state mechanical relationship model and increases the confidence score of the data-driven sub-network. This ensures that the prediction results remain reliable even under conditions of drastic changes in the seepage state at the dam abutment, solving the problem of decreased prediction accuracy of a single model under non-steady-state conditions and effectively improving the engineering reliability of dam displacement estimation.

[0024] 3. This invention determines the basic lag time of reservoir water level changes transmitted to the dam abutment through cross-correlation function analysis, and dynamically adjusts the basic lag time using an exponential decay function with the absolute value of the reservoir water level change rate as the exponent. This ensures that the lag time reflects the physical process of accelerated head transmission during rapid reservoir water level changes. When the reservoir water level change rate is large, the lag time automatically shortens, making the model more sensitive to the dam abutment constraint state; when the reservoir water level change rate is small, the lag time approaches the statistical average, and the model maintains stable predictive performance. The lag time for the left and right dam abutments is calculated independently, and their respective constraint feature sets are constructed. This accurately reflects the different constraint capabilities of the left and right dam abutments due to differences in geological conditions, solving the problem that fixed lag time parameters cannot adapt to dynamic changes in the reservoir water level change rate. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a schematic diagram of the overall system flow of the present invention. Detailed Implementation

[0026] Example 1: A deep learning-based method for estimating dam displacement includes the following steps: Acquire monitoring data, including the historical displacement sequence of the dam body, the reservoir water level sequence, and the groundwater level sequence and pore water pressure sequence of the dam abutment hills; Calculate the rate of change and direction of rise and fall of reservoir water level based on reservoir water level sequence, and calculate the lag time based on groundwater level sequence; A dam shoulder constraint feature set is constructed based on the lag time and pore water pressure sequences. The dam shoulder constraint feature set includes lag time features and pore water pressure features. Using the historical displacement sequence as a unified time axis, the reservoir water level sequence and the dam abutment constraint feature set are time-aligned and resampled, and then fused into multi-dimensional driving features. Multidimensional driving features are input into the displacement prediction network, which includes a physical driving subnetwork and a data driving subnetwork. Both subnetworks use the dam abutment constraint feature set as dynamic boundary condition input. The physical driving subnetwork outputs the physical predicted displacement based on the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics. The data driving subnetwork outputs the data predicted displacement. The data driving subnetwork contains a gated recurrent unit layer and a temporal attention layer. The gated recurrent unit layer performs temporal encoding on the multidimensional driving features. The temporal attention layer calculates the attention weight based on the reservoir water level change amplitude at each historical moment, and then performs weighted fusion on the hidden states at each moment output by the gated recurrent unit layer to output the data predicted displacement. The confidence level is adaptively adjusted based on the historical prediction errors of the physical and data-based displacement predictions, and then weighted and fused to obtain the final displacement estimate.

[0027] Furthermore, the calculation process of the gated loop unit layer includes: Let the current time be The hidden state at the previous moment was The input feature vector at the current time is , update door The formula for calculation is: ; Reset door The formula for calculation is: ; Candidate hidden state The formula for calculation is: ; Current hidden state The formula for calculation is: ; in, , , This is the weight matrix. , , For bias terms, It is the sigmoid activation function. This represents element-wise multiplication. It is the hyperbolic tangent activation function.

[0028] Furthermore, each historical moment in the temporal attention layer The formula for calculating the range of reservoir water level change is: ; Attention weight The formula for calculation is: ; fused context vector The formula for calculation is: ; The data predicts displacement The formula for calculation is: ; in, for The reservoir water level at that moment. The time length of the input sequence. For the gated loop unit layer in The hidden state that is constantly being output. This is the weight matrix of the fully connected output layer. This is the bias term for the fully connected output layer.

[0029] Furthermore, the calculation method for the reservoir water level change rate is as follows: taking the current moment as the reference, the reservoir water level sequence within a preset time window length is subjected to least squares linear fitting, and the slope of the fitted line is used as the reservoir water level change rate at the current moment. When the slope of the fitted line is positive, it is determined that the reservoir water level is in the rising stage, and when it is negative, it is determined that the reservoir water level is in the falling stage.

[0030] Preferably, the physical driving sub-network uses a mechanical relationship model to calculate the physical predicted displacement. This mechanical relationship model is in the form of a second-order differential equation, where the second derivative of the dam displacement with respect to time is related to the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics. The mechanical relationship model is expressed as follows: ; in, To predict displacement in physics, For time, The rate of change of reservoir water level The characteristic value of pore water pressure. The characteristic value of the lag time. The coefficient representing the influence of the rate of change of reservoir water level. The pore water pressure influence coefficient is... The lag time influence coefficient is... For constant terms; The physics-driven subnetwork discretizes the mechanical relationship model and then recursively calculates the physical predicted displacement. The discretization method is as follows: Let the time step be The current time is the [number]th [time]. Step, the previous moment was the first step. Step 1: The first derivative of displacement with respect to time is approximated using a first-order backward difference, and the second derivative of displacement with respect to time is approximated using a difference of the first-order backward difference. The recursive formula for the physically predicted displacement is: ; in, For the first The physical prediction displacement of the step. For the first The rate of change of reservoir water level in the step, For the first The characteristic value of pore water pressure in the step, For the first The lag time characteristic value of the step; the initial value required for recursive calculation. and The measured displacement values ​​at the initial moment and the first moment in the historical displacement sequence are taken respectively.

[0031] Furthermore, the mechanical relationship model adopts a simplified first-order difference form, and the physical prediction of displacement change is directly proportional to the current values ​​of the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics: ; in, , for the first Step and the first The difference between the physical predicted displacement and the actual displacement.

[0032] Furthermore, The value range is from 0.5 to 5.0. The value range is from 0.01 to 0.50. The value range is from 0.001 to 0.100. The range of values ​​is Up to 0.10 ^2; For concrete gravity dams, The value ranges from 0.5 to 2.0; for earth-rock dams, The value range is from 2.0 to 5.0; and Based on the permeability characteristics of the dam abutment hill, it is determined that: when the permeability coefficient of the dam abutment hill is less than... At m / s The value range is from 0.01 to 0.10. The value ranges from 0.001 to 0.010; when the permeability coefficient of the dam abutment hillside is greater than... At m / s The value range is from 0.20 to 0.50. The value ranges from 0.050 to 0.100; when the permeability coefficient of the dam abutment is between m / s to When the speed is between m / s The value range is from 0.10 to 0.20. The value range is from 0.010 to 0.050.

[0033] Furthermore, each input variable undergoes scale alignment before participating in the calculation of the mechanical relationship model: each input variable is divided by its own reference value to convert it into a dimensionless relative value; the reference value for the reservoir water level change rate is the maximum historical reservoir water level change rate at the dam site, the reference value for the pore water pressure is the maximum historical pore water pressure at the dam abutment, and the reference value for the lag time is the maximum historical lag time at the dam abutment.

[0034] Preferably, the adaptive adjustment of confidence level is performed as follows: The absolute value of the difference between the predicted displacement and the measured displacement of the physical driving subnetwork and the data driving subnetwork at each moment within the historical time window is calculated as the prediction error. The confidence levels of the two subnetworks are determined based on the prediction errors of the physical-driven subnetwork and the data-driven subnetwork within the historical time window. The final displacement estimate is obtained by weighting the physical predicted displacement and the data predicted displacement according to the confidence level. When the fluctuation amplitude of the pore water pressure characteristic in the dam abutment constraint feature set exceeds a preset fluctuation threshold, the confidence level of the physical driving sub-network is reduced and the confidence level of the data driving sub-network is increased; the calculation method for the fluctuation amplitude of the pore water pressure characteristic is as follows: Take the pore water pressure sequence within a preset fluctuation statistical window length before the current time, and calculate the standard deviation of the sequence. ;when Exceeding the preset fluctuation threshold When this occurs, a confidence level switch is triggered, lowering the confidence level of the physical-driven sub-network to 0.2 and raising the confidence level of the data-driven sub-network to 0.8; when Falling back to After the preset recovery time is reached, the confidence level switching mechanism will be deactivated, and the confidence level will be restored to be adaptively determined by the average prediction error.

[0035] Preferably, the confidence level is determined as follows: the average prediction error of the physical driving sub-network within the historical time window is: ; The average prediction error of the data-driven subnetwork within the historical time window is: ; The confidence level of the physical-driven subnetwork is: ; The confidence level of the data-driven subnetwork is: ; The final displacement estimate is: ; in, The length of the historical time window. For the current moment, For indexing moments within a historical time window, For physical driving subnetworks in The physical predicted displacement at time t. For data-driven subnetworks in Predict displacement based on data at any given time. for Measured value of dam displacement at time t. and These are the average prediction errors of the physics-driven subnetwork and the data-driven subnetwork, respectively. and The confidence levels of the physical-driven subnetwork and the data-driven subnetwork are respectively. , for The final displacement estimate at time [time].

[0036] Preferably, the abutment hill includes a left abutment hill and a right abutment hill; the groundwater level sequence and pore water pressure sequence of the left abutment hill and the groundwater level sequence and pore water pressure sequence of the right abutment hill are obtained respectively; constraint feature sets for the left abutment and constraint feature sets for the right abutment are constructed respectively; the constraint feature sets for the left abutment and the constraint feature sets for the right abutment are time-series aligned and fused as independent dynamic boundary condition features; The lag time calculation steps are performed independently for the left and right abutments to obtain the dynamically adjusted lag time for the left and right abutments respectively; the proportional coefficients for the left and right abutments are independently determined based on the geological permeability characteristics of their respective abutment hills.

[0037] Preferably, during the displacement prediction process, the lag time features and pore water pressure features in the dam abutment constraint feature set are updated based on the deviation between the real-time monitored dam displacement sequence and the final displacement estimate; the updated features are then re-input into the physical driving sub-network to generate the updated physical predicted displacement. The update trigger condition is as follows: calculate the moving average of the displacement deviation sequence within the length of the deviation statistical window before the current time. When the absolute value of the moving average exceeds a preset deviation threshold, online correction is triggered. The correction amount is determined by taking the monitoring data within the length of the correction window before the trigger time, and constructing a least squares problem with the objective of minimizing the sum of squares of the deviations between the measured dam displacement and the physical predicted displacement at each time point within the correction window. ; The lag time characteristic correction quantity is obtained by solving. and pore water pressure characteristic correction amount The corrected hysteresis time characteristic values ​​and pore water pressure characteristic values ​​are as follows: ; ; in, and These are the original hysteresis time characteristic value and pore water pressure characteristic value at the trigger correction time, respectively. To correct the time intervals within the window, The physical predicted displacement is calculated using the corrected hysteresis time characteristic value and pore water pressure characteristic value as inputs; online correction only updates the characteristic input values ​​and does not update the parameters of the mechanical relationship model.

[0038] Preferably, the lag time is calculated as follows: Calculate the cross-correlation function between the groundwater level sequence at the dam abutment and the reservoir water level sequence under multiple different time lags, and take the time lag corresponding to the maximum cross-correlation function value as the basic lag time; Using the aforementioned basic lag time as a benchmark, multiply it by an exponential decay function with the absolute value of the reservoir water level change rate as the exponent to obtain the dynamically adjusted lag time. Cross-correlation function The formula for calculation is: ; Basic lag time The value corresponding to the maximum value of the cross-correlation function : ; Dynamically adjusted lag time The formula for calculation is: ; in, This is the time lag. For sequence length, For reservoir water level sequence, The groundwater level sequence at the dam abutment. and These are the mean values ​​of the reservoir water level sequence segment and the dam abutment groundwater level sequence segment, respectively, used in the calculation. The rate of change of reservoir water level This is the proportionality coefficient. The value ranges from 0.5h / m to 5.0h / m and is determined based on the geological permeability characteristics of the dam shoulder mountain.

[0039] Preferably, a displacement change rate sequence is calculated based on the final displacement estimate; Within a set time window, the number of times the displacement rate of change changes from positive to negative or from negative to positive in the rate of change sequence is used as the direction alternation frequency. When the frequency of the alternation of directions exceeds a preset frequency threshold or the amplitude of the rate of change sequence exceeds a preset amplitude threshold, it is determined that the dam displacement trend has an abnormal reversal. rate of change of displacement The formula for calculation is: ; Alternating frequency of directions The formula for calculation is: ; Displacement change rate amplitude The formula for calculation is: ; in, for The final displacement estimate at time [time]. for The final displacement estimate at time [time]. To determine the time window The total number of times the value changes from positive to negative and from negative to positive; when Exceeding the preset frequency threshold or Exceeding the preset amplitude threshold The dam displacement trend was determined to have shown an abnormal reversal. Take 0.02 times / hour. Take 1.5mm.

[0040] Preferably, the multidimensional driving features further include temperature sequences and rainfall sequences; The temperature sequence and the rainfall sequence are interpolated and resampled along the unified time axis, and then time-series aligned and fused. The interpolation and resampling uses a linear interpolation method, and the linear interpolation formula is as follows: ; in, For the target time point, and These are the closest known data time points before and after the target time point. and They are respectively and Temperature or rainfall value at any given time.

[0041] Preferably, a deep learning-based dam displacement estimation system is used to execute the deep learning-based dam displacement estimation method, comprising: The data acquisition module is used to acquire monitoring data, which includes the historical displacement sequence of the dam body, the reservoir water level sequence, and the groundwater level sequence and pore water pressure sequence of the dam abutment hill. The water level rise and fall and lag analysis module is used to calculate the rate of change and direction of rise and fall of the reservoir water level based on the reservoir water level sequence, and to calculate the lag time based on the groundwater level sequence; A dam shoulder constraint feature construction module is used to construct a dam shoulder constraint feature set based on the lag time and the pore water pressure sequence. The dam shoulder constraint feature set includes lag time features and pore water pressure features. The temporal alignment and fusion module is used to perform temporal alignment and resampling on the reservoir water level sequence and the dam abutment constraint feature set, using the time reference of the historical displacement sequence as a unified time axis, and fuse them into multi-dimensional driving features. A dual-drive displacement prediction module is used to input the multi-dimensional driving features into a displacement prediction network. The displacement prediction network includes a physical drive sub-network and a data drive sub-network. Both sub-networks are input with the dam abutment constraint feature set as dynamic boundary conditions. The physical drive sub-network outputs the physical predicted displacement based on the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics. The data drive sub-network outputs the data predicted displacement. The confidence fusion output module is used to adaptively adjust the confidence level based on the historical prediction errors of the physical predicted displacement and the data predicted displacement, and then weight and fuse them into a final displacement estimate.

[0042] Example 2: like Figures 1 to 2 As shown, this invention provides a deep learning-based method for estimating dam displacement, comprising the following steps: Acquire monitoring data, which includes the historical displacement sequence of the dam body, the reservoir water level sequence, and the groundwater level sequence and pore water pressure sequence of the dam abutment hill; Calculate the rate of change and direction of rise and fall of reservoir water level based on the reservoir water level sequence, and calculate the lag time based on the groundwater level sequence; A dam shoulder constraint feature set is constructed based on the lag time and the pore water pressure sequence. The dam shoulder constraint feature set includes lag time features and pore water pressure features. Using the time reference of the historical displacement sequence as a unified time axis, the reservoir water level sequence and the dam abutment constraint feature set are time-aligned and resampled, and fused into multi-dimensional driving features. The multidimensional driving features are input into the displacement prediction network, which includes a physical driving sub-network and a data driving sub-network. Both sub-networks use the dam abutment constraint feature set as dynamic boundary conditions as input. The physical driving sub-network outputs the physical predicted displacement based on the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics. The data driving sub-network outputs the data predicted displacement. The confidence level is adaptively adjusted based on the historical prediction errors of the physical predicted displacement and the data predicted displacement, and then weighted and fused to obtain the final displacement estimate.

[0043] Specifically, in the data acquisition step, displacement monitoring equipment, including any one or a combination of vibrating wire displacement gauges, differential resistance displacement gauges, or global navigation satellite system receivers, is deployed on key monitoring sections of the dam body. Displacement data at each measuring point of the dam body are collected according to a preset sampling frequency to form a historical displacement sequence. A water level gauge is installed in front of the dam, and reservoir water level data is collected at the same sampling frequency as the displacement monitoring equipment to form a reservoir water level sequence. Groundwater level monitoring wells and pore water pressure gauges were installed on key sections of the dam abutment. Water level sensors were installed in the groundwater level monitoring wells, and vibrating wire pore water pressure gauges were used. Groundwater level data and pore water pressure data from the dam abutment were collected at preset sampling frequencies to form a groundwater level sequence. and pore water pressure sequence All the above monitoring data are aggregated through the data acquisition terminal and then transmitted to the data processing center.

[0044] In the calculation of reservoir water level change characteristics, after receiving the reservoir water level sequence, the data processing center uses the current time as a reference and performs least-squares linear fitting on the reservoir water level sequence within a preset time window. The slope of the fitted line is the rate of change of the reservoir water level at the current time. .like If the value is positive, the reservoir water level is determined to be in a rising phase; if... If the value is negative, it indicates that the reservoir water level is in a downward phase.

[0045] In the lag time calculation step, the data processing center will use the groundwater level sequence of the dam abutment hillside. With reservoir water level sequence Comparative analysis was conducted to calculate the time required for changes in reservoir water level to propagate to the interior of the dam abutment hillside, which was used as the lag time. .

[0046] In the process of constructing the dam abutment constraint feature set, the data processing center will calculate the lag time features. and pore water pressure characteristics collected from the field The combination forms the abutment constraint feature set. A longer lag time indicates a slower response of the abutment hill to reservoir water level changes, with a more gradual change in the internal seepage field. A shorter lag time indicates a more sensitive response of the abutment hill to reservoir water level changes, with a more drastic change in the internal seepage field. Higher pore water pressure indicates higher internal water pressure and reduced effective stress within the abutment hill; lower pore water pressure indicates lower internal water pressure and higher effective stress.

[0047] In the temporal alignment and feature fusion steps, the data processing center uses the historical displacement sequence as a unified time axis to perform temporal alignment and interpolation resampling on the reservoir water level sequence and the dam abutment constraint feature set, respectively. If the reservoir water level data or dam abutment constraint feature data at a certain time point is inconsistent with the time point of the historical displacement sequence, a linear interpolation method is used to interpolate using adjacent known data points to obtain the data value corresponding to that time point, ensuring that the temporal resolution of each sequence is completely consistent. After completing the interpolation resampling, the homogenized reservoir water level sequence, lag time feature sequence, and pore water pressure feature sequence are mapped one-to-one with the historical displacement sequence on the time axis and fused into a multi-dimensional driving feature. Each set of data in this multi-dimensional driving feature contains the historical displacement value of the dam body, reservoir water level value, lag time value, and pore water pressure value at the same time.

[0048] In the displacement prediction step, the fused multidimensional driving features are input into the displacement prediction network. This network includes a physical driving sub-network and a data driving sub-network. The physical driving sub-network uses the current reservoir water level change rate as the driving factor. pore water pressure and lag time As input, the physical predicted displacement is calculated according to the preset mechanical relationship model. The data-driven subnetwork takes multi-dimensional driving features as input and uses a deep learning model to calculate data and predict displacement. Both the physics-driven subnetwork and the data-driven subnetwork use the dam abutment constraint feature set as the dynamic boundary condition input, meaning that the two subnetworks use the same lag time feature and pore water pressure feature.

[0049] In the confidence fusion step, the physical predicted displacement output by the physical driving sub-network is obtained. and the data-driven subnetwork output predicted displacement Then, based on the prediction errors of the two sub-networks within the historical time window, the confidence levels of the two networks are adaptively adjusted, and a weighted fusion is performed to obtain the final displacement estimate. The confidence level is dynamically updated based on the prediction error. When a subnetwork performs better recently, its weight in the fusion process is automatically increased. Simultaneously, when the fluctuation amplitude of the pore water pressure characteristics in the dam abutment constraint feature set exceeds a preset fluctuation threshold, it indicates that the seepage field inside the dam abutment hill is in a severely unsteady state. The reliability of the physical-driven subnetwork based on the steady-state mechanical relationship model decreases. In this case, the confidence level of the physical-driven subnetwork is automatically reduced, while the confidence level of the data-driven subnetwork is increased.

[0050] After the above steps are completed, the data processing center will output the estimated displacement values ​​of the dam body for multiple future time steps.

[0051] The mechanical relationship model of the physical driving sub-network is as follows: the second derivative of the dam displacement with respect to time is related to the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics. The physical driving sub-network calculates the physical predicted displacement based on the current reservoir water level change rate, pore water pressure characteristics, and lag time characteristics.

[0052] Specifically, the physics-driven subnetwork is a differentiable computation module built on a dam displacement mechanics model. Its core function is to calculate the predicted physical displacement based on the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics at the current moment, using the mechanics model. The physics-driven subnetwork does not contain trainable network layer parameters, but instead contains a set of mechanics model parameters, which are calibrated using historical monitoring data.

[0053] The input to the physics-driven subnetwork includes three variables: the current rate of change of reservoir water level. The characteristic value of pore water pressure at the dam abutment at the current moment. and the lag time characteristic value at the current moment .in The unit is meters per hour. The unit is kilopascal (kPa). The unit is hours. The output of the physics-driven subnetwork is the predicted physical displacement at the current moment. The unit is millimeters.

[0054] The mechanical relationship model of the physical driving subnetwork is presented in the form of differential equations. The second derivative of the dam displacement with respect to time... The rate of change of reservoir water level at the current moment pore water pressure characteristic value and lag time characteristic value There is a correlation. Specifically, the mechanical relationship model is expressed as follows: ; in , , and These are the parameters of the mechanical relationship model. The reservoir water level change rate influence coefficient reflects the degree of influence of the reservoir water level rise and fall rate on the acceleration or deceleration of dam displacement, and is dimensionless. The pore water pressure influence coefficient reflects the inhibitory effect of pore water pressure on dam displacement at the dam abutment, and is dimensionless. The lag time influence coefficient reflects the regulating effect of the dam abutment response delay on dam displacement, and is dimensionless. This is a constant term that reflects the inherent trend of dam displacement, and its unit is mm / h².

[0055] In the initial state, The value range is from 0.5 to 5.0. The value range is from 0.01 to 0.50. The value range is from 0.001 to 0.100. The value range is -0.10 to 0.10. This range is preset based on the dam type and the geological conditions of the dam abutments. In one embodiment of the invention, for a concrete gravity dam, The value ranges from 0.5 to 2.0; for earth-rock dams, The value ranges from 2.0 to 5.0. The value is determined based on the lithology and permeability characteristics of the dam abutment: when the dam abutment is a dense rock mass and the permeability coefficient is less than 1×10 -7 At m / s The value ranges from 0.01 to 0.10. The value ranges from 0.001 to 0.010; when the abutment hill is a fractured rock mass or a rock mass with well-developed fissures and a permeability coefficient greater than 1×10 -5 At m / s The value ranges from 0.20 to 0.50. The value ranges from 0.050 to 0.100; when the permeability coefficient of the dam abutment hill is between 1×10⁻⁶. -7 m / s to 1×10 -5 When the speed is between m / s The value ranges from 0.10 to 0.20. The value ranges from 0.010 to 0.050.

[0056] The specific method by which the physics-driven sub-network calculates the predicted physical displacement based on the aforementioned mechanical relationship model is as follows: The second-order differential equation is discretized. Let the time step be... If the current time is the nth step and the previous time was the (n-1)th step, then the first derivative of the displacement with respect to time is... A first-order backward difference approximation is used: ; Second derivative of displacement with respect to time The difference approximation is adopted using the first-order backward difference: ; Substituting the above discretized expression into the mechanical relationship model, we obtain the physical predicted displacement. The recursive calculation formula: ; in Let n be the rate of change of the reservoir water level at the current time and the nth step. This represents the characteristic value of the pore water pressure at the current time and step n. This represents the lag time characteristic value at the current time and the nth step. It also represents the initial value required for the recursive calculation. and The measured displacement values ​​at the initial moment and the first moment in the historical displacement sequence are taken respectively.

[0057] Mechanical relationship model parameters , , and Calibration is performed using historical monitoring data. The calibration method involves acquiring historical sequences of dam displacement, reservoir water level, pore water pressure, and lag time. Using a least-squares fitting method, the parameters of the four mechanical relationship models are optimized and determined with the goal of minimizing the mean square error between the predicted and measured displacements. After calibration, this set of parameters is fixed for subsequent calculations of the predicted physical displacement. In another embodiment of the invention, the aforementioned mechanical relationship model parameters are periodically recalibrated with newly acquired monitoring data to maintain the continuous and accurate characterization of the dam's mechanical properties by the physical driving sub-network. The periodic recalibration interval is 30 days or 60 days.

[0058] In another embodiment of the invention, the mechanical relationship model uses a simplified proportional relationship form instead of the aforementioned differential equation form, which is suitable for scenarios with low data sampling frequencies. Specifically, the physical prediction of the change in displacement... It is directly proportional to the current values ​​of the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics: ; in This simplified form reduces the second-order differential equation to a first-order difference equation, requiring less computation and making it suitable for real-time calculations at the monitoring equipment.

[0059] In another embodiment of the present invention, the input variables in the mechanical relationship model undergo scale alignment before being used in the calculation. Specifically, the scale alignment is performed by dividing each input variable by its own reference value to convert it into a dimensionless relative value. The reference value for the reservoir water level change rate is the maximum historical reservoir water level change rate at the dam site; the reference value for the pore water pressure is the maximum historical pore water pressure at the dam abutment; and the reference value for the lag time is the maximum historical lag time at the dam abutment. After scale alignment, the numerical range of each input variable is between 0 and 1, avoiding calibration deviations in the mechanical relationship model parameters due to different dimensions.

[0060] After the physics-driven subnetwork outputs the predicted physical displacement, this predicted displacement, as one of the two outputs of the displacement prediction network, participates in the subsequent confidence fusion step together with the predicted data displacement output by the data-driven subnetwork. The predicted physical displacement itself has a clear physical meaning, and its trend strictly follows the response law of dam displacement to reservoir water level change rate and dam abutment constraint state described by the mechanical relationship model. Therefore, it has strong reliability and interpretability under the condition of stable seepage at the dam abutment.

[0061] The adaptive adjustment of confidence level is as follows: The absolute value of the difference between the predicted displacement and the measured displacement of the physical driving subnetwork and the data driving subnetwork at each moment within the historical time window is calculated as the prediction error. The confidence levels of the two subnetworks are determined based on the prediction errors of the physical-driven subnetwork and the data-driven subnetwork within the historical time window. The final displacement estimate is obtained by weighting the physical predicted displacement and the data predicted displacement according to the confidence level. When the fluctuation amplitude of the pore water pressure characteristic in the dam abutment constraint feature set exceeds a preset fluctuation threshold, the confidence of the physical driving sub-network is reduced and the confidence of the data driving sub-network is increased.

[0062] Specifically, in the confidence adaptive adjustment step, after acquiring the physical predicted displacement output by the physical driving subnetwork and the data predicted displacement output by the data driving subnetwork at each time step, the data processing center simultaneously acquires the measured value of the dam displacement at that time. The measured value is collected by the displacement monitoring equipment at the same time.

[0063] Let the current time be Take the length of the historical time window In one embodiment of the present invention, Take 72 hours. For each moment within the historical time window. ( ), respectively obtain the physical predicted displacement values ​​output by the physical driving subnetwork. The data-driven subnetwork outputs the predicted displacement values. and displacement values ​​measured by displacement monitoring equipment .

[0064] For the physics-driven subnetwork, the absolute value of the difference between the predicted displacement and the measured displacement at each time point within the historical time window is calculated as the prediction error at each time point, denoted as . For the data-driven subnetwork, the same calculation is performed. Then, the average prediction error of each subnetwork within the historical time window is calculated. The average prediction error of the physics-driven subnetwork is: ; The average prediction error of the data-driven subnetwork is: ; Subnetworks with smaller average prediction errors indicate higher prediction accuracy within recent historical time windows. Subnetworks with smaller average prediction errors are assigned higher confidence levels.

[0065] The confidence level is calculated as follows: the initial confidence level for both the physical-driven and data-driven sub-networks is set to 0.5. The confidence level is then adaptively adjusted based on the average prediction error. The adjustment method is as follows: ; ; in The confidence level of the physical-driven subnetwork. The confidence level of the data-driven subnetwork, and When the historical average prediction error of the physical driving subnetwork is small, that is... hour, The physics-driven subnetwork achieves higher confidence. Conversely, when the historical average prediction error of the data-driven subnetwork is small, The data-driven subnetwork achieves a high level of confidence.

[0066] Obtain the confidence level of the physical driving subnetwork. and the confidence of data-driven subnetworks Then, the physical predicted displacement is calculated using the following formula. and data prediction displacement By performing weighted fusion, the final displacement estimate for the current moment is obtained: ; in and The normalized confidence level satisfies Final displacement estimate This is the estimated dam displacement output by the system at the current moment.

[0067] The confidence score is dynamically updated at each time step based on the prediction error within the most recent historical time window. In one embodiment of the invention, as the historical time window slides forward, new prediction data is added to the window, old data is removed from the window, and the confidence score is recalculated accordingly, thereby ensuring that the confidence score always reflects the difference in prediction performance between the two sub-networks in the current time period.

[0068] In addition to the adaptive confidence adjustment mechanism based on average prediction error, the adaptive confidence adjustment step also includes a confidence switching mechanism based on the pore water pressure fluctuation state at the dam abutment. When the fluctuation amplitude of the pore water pressure characteristics in the dam abutment constraint feature set exceeds a preset fluctuation threshold, the system determines that the seepage field inside the dam abutment hill is in a severely unsteady state. Under this condition, the reliability of the physical driving sub-network based on the steady-state mechanical relationship model decreases, and the credibility of its physical predicted displacement decreases. At this time, the system automatically reduces the confidence of the physical driving sub-network. And correspondingly increase the confidence level of the data-driven subnetwork. .

[0069] The specific calculation method for the characteristic fluctuation amplitude of pore water pressure is as follows: take the preset fluctuation statistical window length before the current time. Calculate the standard deviation of the pore water pressure sequence within the sequence. .when Exceeding the preset fluctuation threshold At that time, a confidence level switch is triggered. In one embodiment of the present invention, Take 24 hours Take 10% of the historical average pore water pressure. After triggering the switch, the confidence level of the physical drive subnetwork. The confidence level of the data-driven subnetwork was lowered to 0.2. The confidence level is adjusted to 0.8. Once the pore water pressure fluctuation falls below the preset fluctuation threshold and remains below it for a preset recovery time (e.g., 12 hours), the confidence level switching mechanism is deactivated, and the confidence level is restored to be adaptively determined by the average prediction error. In one embodiment of the invention, the confidence level switching employs a smooth transition rather than a jump approach. That is, when the fluctuation amplitude exceeds the threshold, the confidence level is gradually reduced and increased within a preset transition time window, such as 6 hours, to maintain the continuity of the fusion results.

[0070] In one embodiment of the present invention, a concrete dam is used as an example. Under normal operating conditions, the average prediction error of the physical-driven subnetwork over three months is 0.8 mm, and the average prediction error of the data-driven subnetwork over three months is 1.2 mm. Therefore, the confidence level of the physical-driven subnetwork is 1.2 / (0.8+1.2) = 0.60, and the confidence level of the data-driven subnetwork is 0.40. The physical-driven subnetwork receives a higher fusion weight due to its higher historical accuracy. When the dam experiences continuous heavy rainfall, causing the pore water pressure at the dam shoulder to rise from 50 kPa to 120 kPa within 6 hours, and the fluctuation exceeds a preset threshold, the system determines that the dam shoulder is in a non-steady state. The confidence level of the physical-driven subnetwork automatically decreases from 0.60 to 0.20, while the confidence level of the data-driven subnetwork increases from 0.40 to 0.80. At this point, the data-predicted displacement dominates the final estimated value.

[0071] The data-driven subnetwork includes a gated recurrent unit layer and a temporal attention layer; The gated loop unit layer performs temporal encoding on the multidimensional driving features; The temporal attention layer calculates attention weights based on the reservoir water level change amplitude at each historical moment, and outputs the data prediction displacement after weighted fusion of the hidden states at each moment output by the gated loop unit layer.

[0072] Specifically, the data-driven subnetwork comprises a gated recurrent unit layer and a temporal attention layer. The input to the data-driven subnetwork is a multidimensional driving feature sequence, and the output is a data prediction shift.

[0073] The multidimensional driven feature sequence uses the historical displacement sequence as a unified time axis and is composed of input feature vectors from each time step on the same time axis. Each time step Input feature vector This includes the reservoir water level at that moment, lag time characteristic values, pore water pressure characteristic values, and historical displacement values. When the multidimensional driving features also include temperature and rainfall sequences, each moment... The input feature vector also includes the temperature and rainfall values ​​at that moment. The dimension is denoted as , which is equal to the number of input feature variables.

[0074] The gated recurrent unit layer of the data-driven subnetwork receives a multidimensional driving feature sequence as input. Let the time length of the input sequence be... That is, taking from the current moment backwards The input feature vector at each historical moment. In one embodiment of the present invention, The value is 72, meaning the input is the multidimensional driving feature data from the most recent 72 hours. The gated recurrent unit layer contains multiple gated recurrent units. Each gated recurrent unit receives the input feature vector at the current time step and the hidden state from the previous time step at each time step, and outputs the hidden state at the current time step.

[0075] The internal structure of the gated loop unit is as follows. Let the current time be... The hidden state at the previous moment was , dimension The input feature vector at the current time is , dimension The gated loop unit first calculates and updates the gate. Update Gate The calculation formula is: ; in To update the gate weight matrix, To update the gate's bias term, Indicates to hide state and input feature vector The concatenated vector This is the sigmoid activation function. Update gate. The output range is between 0 and 1, controlling how much of the hidden state from the previous time step is retained in the current time step.

[0076] The gate control loop unit then calculates and resets the gate. Reset the door The calculation formula is: ; in To reset the weight matrix of the gate, This resets the door's bias. Reset the door. The output range is between 0 and 1, controlling how much of the hidden state from the previous moment is ignored.

[0077] The gated recurrent unit then calculates the candidate hidden state. Candidate hidden state The calculation formula is: ; in Let be the weight matrix of the candidate hidden states. The bias term for the candidate hidden state. This represents element-wise multiplication. Reset gate. The hidden state of the previous moment After scaling, the input feature vector at the current time step The candidate hidden states are obtained by concatenating the data and applying the tanh activation function.

[0078] The gated loop unit finally calculates the hidden state at the current moment. Current hidden state The calculation formula is: ; Update Gate Control the hidden state of the previous moment and candidate hidden state The proportional relationship between them. When updating the gate. When the value is close to 1, the current hidden state mainly adopts the candidate hidden state, that is, the current input information dominates. When updating the gate... When the value is close to 0, the current hidden state mainly uses the hidden state from the previous moment, that is, historical information is preserved.

[0079] In one embodiment of the present invention, the hidden state dimension Set the dimension to 64. (Input feature vector dimension) Take 5 (reservoir water level, lag time, pore water pressure, historical displacement, temperature). The hidden state output sequence of the gated cyclic unit layer is: This corresponds to the time-series coding results for each historical moment.

[0080] The hidden states output from the gated recurrent unit layer at each time step are input into the temporal attention layer. The temporal attention layer calculates the attention weights based on the reservoir water level change amplitude at each historical time step. Specifically, the reservoir water level change amplitude at each historical time step is first calculated. ,in . The definition of is: ; That is, the first Time and the The absolute value of the difference in reservoir water level at each time step. Then, the magnitude of the reservoir water level change at each time step is converted into attention weights using a softmax function. Attention weights The calculation formula is: ; in It is an exponential function. The greater the historical fluctuation in reservoir water level, the higher the corresponding attention weight. The larger the value, the more significant the impact of the drastic fluctuation in the reservoir water level at that moment on the current dam displacement, and it should be given greater attention in the time-series integration.

[0081] Obtain attention weights at each time step Subsequently, the temporal attention layer processes the hidden states at each time step output by the gated recurrent unit layer. Perform weighted fusion to obtain the fused context vector. : ; fused context vector The dimension is the same as the hidden state dimension, which is 1. This context vector integrates the temporal encoding information of all historical moments from the gated recurrent unit layer, and through an attention weighting mechanism, it gives greater weight to moments of dramatic reservoir water level changes in the fusion.

[0082] The temporal attention layer will fuse the context vectors. Input a fully connected layer and output data to predict displacement. The calculation method for the fully connected output layer is as follows: ; in This is the weight matrix of the fully connected output layer. This is the bias term for the fully connected output layer. The fully connected output layer has only one output node, meaning it outputs a single data prediction displacement value.

[0083] In one embodiment of the present invention, the data-driven subnetwork is trained using historical monitoring data. During training, mean squared error is used as the loss function, with the objective of minimizing the deviation between the data-predicted displacement output by the data-driven subnetwork and the measured displacement. The backpropagation algorithm and the Adam optimizer are used to train all trainable parameters of the gated recurrent unit layer and the temporal attention layer, including... , , , , , , , Optimization is then performed. After training, the data-driven subnetwork is permanently used for subsequent data-predicted displacement calculations. In another embodiment of the invention, the data-driven subnetwork is periodically incrementally trained using newly accumulated monitoring data to maintain continuous tracking of dam displacement trends.

[0084] Taking a concrete dam as an example, under normal operating conditions, the distribution of reservoir water level changes over the past 72 hours is as follows: a sudden drop of 0.5 meters at hour 40, a sudden drop of 0.3 meters at hour 55, and water level changes within 0.05 meters at other times. After calculating the attention weights based on the water level changes at each time point, the attention weights for hour 40 and hour 55 reach 0.35 and 0.25 respectively, while the total attention weight for the remaining 70 hours is only 0.40. The weighted fused context vector is mainly composed of the hidden states at hours 40 and 55, while the hidden states output by the gated recurrent unit layer contribute little to the context vector. This ensures that the data-driven subnetwork's predicted displacement primarily reflects the impact of the sudden drop in reservoir water level on the dam's displacement, rather than being overwhelmed by data from a large number of stable periods with no significant water level changes.

[0085] The dam shoulder mountain includes the left dam shoulder mountain and the right dam shoulder mountain; The groundwater level and pore water pressure sequences of the left abutment hill and the right abutment hill were obtained respectively. Construct constraint feature sets for the left abutment and the right abutment respectively; The left abutment constraint feature set and the right abutment constraint feature set are treated as independent dynamic boundary condition features and then time-series aligned and fused.

[0086] Specifically, the dam abutment consists of the left and right abutments. In actual engineering, the geological conditions of the left and right abutments often differ. For example, the left abutment may be located in a relatively fractured fault zone, while the right abutment may be located in a relatively intact rock mass. Therefore, the hydrological response characteristics of the left and right abutments to reservoir water level fluctuations may differ, and their respective lag times and pore water pressure variation patterns may also differ. To accurately characterize the lateral constraint state of the left and right abutments on the dam body, independent data acquisition and constraint feature construction are required for each abutment.

[0087] Groundwater level and pore water pressure sequences were obtained for the left abutment hillside. Groundwater level monitoring wells were deployed on key sections of the left abutment hillside. The depth of the wells was determined based on the geological conditions of the hillside, ranging from 1 / 3 to 2 / 3 of the hillside thickness. Water level sensors were installed within the monitoring wells on the left abutment hillside, and groundwater level data were collected at a preset sampling frequency to form the left abutment groundwater level sequence. Pore ​​water pressure gauges were installed on key sections of the left abutment hillside. The burial depth of the pore water pressure gauges was consistent with the depth of the groundwater level monitoring boreholes. Pore water pressure data of the left abutment hillside were collected at the same preset sampling frequency to form a pore water pressure sequence for the left abutment. .

[0088] Groundwater level and pore water pressure sequences were obtained for the right abutment hillside. Groundwater level monitoring wells were deployed on key sections of the right abutment hillside, and water level sensors were installed in these wells. Groundwater level data were collected at a preset sampling frequency to form the right abutment groundwater level sequence. Pore ​​water pressure gauges were installed on key sections of the right abutment hillside, and pore water pressure data of the right abutment hillside were collected at the same preset sampling frequency to form a pore water pressure sequence for the right abutment. .

[0089] Construct the constraint feature sets for the left abutment and the right abutment respectively.

[0090] For the left abutment, according to the groundwater level sequence of the left abutment With reservoir water level sequence Calculate the lag time of the left abutment The calculation method is as follows: calculate the groundwater level sequence on the left abutment of the dam. With reservoir water level sequence The cross-correlation function under multiple different time lags is used. The time lag corresponding to the maximum cross-correlation function value is taken as the lag time of the left abutment foundation. Then, this foundation lag time is used as a benchmark and multiplied by an exponential decay function with the absolute value of the reservoir water level change rate as the exponent to obtain the dynamically adjusted lag time of the left abutment. Delay time of the left abutment and pore water pressure on the left abutment The combination forms the left abutment constraint feature set.

[0091] For the right abutment, according to the groundwater level sequence of the right abutment With reservoir water level sequence Calculate the right abutment lag time The calculation method is the same as for the left abutment. The right abutment lag time is then factored in. and pore water pressure on the right abutment The combination forms the right abutment constraint feature set.

[0092] In one embodiment of the present invention, a concrete arch dam is taken as an example. The left abutment of the arch dam is located in a fault fracture zone area revealed by geological survey, with poor rock mass integrity and a permeability coefficient of 5×10⁻⁶. -4 m / s; the right abutment hill is located in a complete granite area with a permeability coefficient of 5×10 m / s. -6m / s. Under the same reservoir water level change conditions, the lag time of the groundwater level on the left abutment to the reservoir water level is 6 hours, and the lag time of the groundwater level on the right abutment to the reservoir water level is 48 hours. The constraint feature sets of the left and right abutments reflect different constraint states: the left abutment has a short lag time and a fast pore water pressure response, indicating that the left abutment has strong permeability and its constraint state is sensitive to changes in reservoir water level; the right abutment has a long lag time and a slow pore water pressure response, indicating that the right abutment has weak permeability and its constraint state is relatively stable. Both, as independent dynamic boundary condition features, can accurately reflect the different constraint capabilities of the left and right abutments on the left and right sides of the dam body.

[0093] The left and right abutment constraint feature sets are treated as independent dynamic boundary condition features and time-series aligned and fused. Using the historical displacement sequence as a unified time axis, the left and right abutment constraint feature sets are interpolated and resampled along this unified time axis to ensure their temporal resolution matches the historical displacement sequence. After interpolation and resampling, the lag time feature sequence of the left abutment is... Characteristic sequence of pore water pressure on the left abutment Right abutment lag time characteristic sequence , characteristic sequence of pore water pressure on the right abutment These are used as independent input feature variables, and then fused together with the reservoir water level sequence and historical displacement sequence to form a multidimensional driving feature.

[0094] The fused multidimensional driving features are input into the displacement prediction network. The physical driving sub-network within the displacement prediction network receives the left abutment lag time features, left abutment pore water pressure features, right abutment lag time features, and right abutment pore water pressure features as dynamic boundary condition inputs. In one embodiment of the invention, the mechanical relationship model of the physical driving sub-network multiplies the left abutment constraint features and right abutment constraint features by their respective weight coefficients and then adds them together to obtain a comprehensive abutment constraint feature, which participates in the calculation of the mechanical relationship model. The data driving sub-network also receives the aforementioned left and right abutment constraint features as part of its input features.

[0095] During the displacement prediction process, the lag time characteristics and pore water pressure characteristics in the dam abutment constraint feature set are updated based on the deviation between the real-time monitored dam displacement sequence and the final displacement estimate. The updated features are then re-input into the physics-driven subnetwork to generate updated physical prediction displacements.

[0096] Specifically, during long-term service, the mechanical properties of dam construction materials, the consolidation state of the dam foundation, and the seepage conditions of the abutment hillside may change slowly over time. Furthermore, extreme environmental events such as heavy rainfall and sudden rises or falls in reservoir water levels can cause short-term, drastic changes in the seepage field within the abutment hillside. These changes can cause deviations between the initially calibrated lag time characteristics and pore water pressure characteristics and the current actual constraint state of the dam body, thus affecting the prediction accuracy of the physical driving subnetwork. Therefore, it is necessary to perform online correction of the abutment constraint characteristics based on real-time monitoring data.

[0097] During the displacement prediction process, the data processing center acquires the measured values ​​of dam displacement collected in real time by the displacement monitoring equipment at every moment. Simultaneously, the final displacement estimate at the current moment is obtained from the confidence fusion step output. Calculate the deviation between the two. Take the deviation sequence within the length of the deviation statistics window up to the current time, and calculate the moving average of the deviation sequence. The deviation statistical window length is 24 hours, and the moving average is calculated using an equal-weighted average method. When the moving average... The absolute value exceeds the preset deviation threshold. At that time, online correction is triggered. Take 2mm.

[0098] After triggering online correction, the lag time and pore water pressure characteristics in the abutment constraint feature set are updated based on the deviation between the real-time monitored dam displacement sequence and the final displacement estimate. The update method is as follows: monitoring data within the correction window length prior to the triggering time is used to establish a correction relationship between the measured dam displacement values ​​and the input characteristics of the physical driving sub-network. The correction window length is 72 hours. Using the deviation sequence between the measured dam displacement values ​​and the physically predicted displacement at each time point within the correction window as the target, and the lag time and pore water pressure characteristic values ​​at each time point within the same time window as independent variables, the correction amounts for the lag time and pore water pressure characteristics are determined using the least squares fitting method.

[0099] Corrected hysteresis characteristics and corrected pore water pressure characteristic value The calculation formula is: ; ; in To trigger the correction, the original lag time characteristic value, This is the correction amount for the lag time characteristic. The original pore water pressure characteristic value at the time of triggering correction, This is the correction value for the pore water pressure characteristics. The unit is h. The unit is kPa.

[0100] Correction amount and The determination method is as follows: Construct a least squares problem with the objective of minimizing the sum of squares of the deviations between the predicted and measured displacements at each time point within the correction window. ; in To correct the time intervals within the window, The physical predicted displacement is calculated by the physics-driven subnetwork using the corrected hysteresis time characteristic value and pore water pressure characteristic value as inputs. The optimal solution is obtained by solving the aforementioned least squares problem. and During the solution process, the parameters of the mechanical relationship model remain constant.

[0101] The updated features are re-input into the physics-driven sub-network to generate updated physical prediction displacements. The corrected lag time feature values ​​are then used. and corrected pore water pressure characteristic value Together with the current rate of change of reservoir water level The mechanical relationship model of the physics-driven subnetwork is input together with the model. The physics-driven subnetwork then recalculates the predicted physical displacement at the current moment according to the mechanical relationship model. : ; in and These are the corrected pore water pressure characteristic value and hysteresis time characteristic value at the current moment, respectively.

[0102] Online calibration only updates the feature input values ​​of the physics-driven subnetwork, without updating the mechanical relationship model parameters of the physics-driven subnetwork. The data-driven subnetwork does not participate in online calibration. The mechanical relationship model parameters of the physics-driven subnetwork are calibrated based on monitoring data over a long historical period, reflecting the long-term average mechanical response characteristics of the dam abutment. Short-term fluctuations in the seepage field only affect the current abutment constraint state and should not change the intrinsic parameters of the mechanical relationship model. Updating only the feature input values ​​without updating the model parameters allows for lightweight calibration of the physics-driven subnetwork while maintaining the long-term stability of the mechanical relationship model.

[0103] Generate updated physical prediction displacement Then, the updated physical predicted displacement replaces the original physical predicted displacement. After participating in subsequent confidence fusion steps, the final displacement estimate at the current moment is updated as follows: ; Confidence and It remains unchanged.

[0104] In one embodiment of the present invention, the altered seepage conditions at the dam abutment after continuous heavy rainfall are taken as an example. During the initial calibration phase, the response relationship between the dam abutment pore water pressure characteristics and the reservoir water level was as follows: a 1m rise in reservoir water level resulted in a 10kPa increase in pore water pressure at the dam abutment. After continuous heavy rainfall, the saturation of the dam abutment increased, and the same 1m rise in reservoir water level led to a 15kPa increase in pore water pressure at the dam abutment. The original pore water pressure characteristic value was 5kPa lower than the actual value. This deviation caused the physical predicted displacement of the physical driving subnetwork to be systematically 1.2mm smaller than the measured displacement. The average deviation exceeded the 2mm threshold, triggering online correction. The correction window length was 72 hours, and least squares fitting was used to determine the correction amount for the pore water pressure characteristics. With a value of +5 kPa, the corrected pore water pressure characteristic value increases by 5 kPa. After re-inputting into the physical driving sub-network, the physical predicted displacement increases by 1.1 mm, and the deviation from the measured displacement is controlled within 0.2 mm.

[0105] The lag time is calculated as follows: Calculate the cross-correlation function between the groundwater level sequence at the dam abutment and the reservoir water level sequence under multiple different time lags, and take the time lag corresponding to the maximum cross-correlation function value as the basic lag time; Using the aforementioned basic lag time as a benchmark, multiply it by an exponential decay function with the absolute value of the reservoir water level change rate as the exponent to obtain the dynamically adjusted lag time.

[0106] Specifically, the lag time is calculated by the data processing center at each sampling moment. The data processing center obtains the historical reservoir water level sequence and the historical groundwater level sequence at the dam abutment prior to the current moment, both with the same sampling frequency and the same time axis. Let the current moment be... Take the length of time before the current moment as Reservoir water level sequence and groundwater level sequence at the dam abutment ,in , The sequence length is given. Take 720 hours, that is, the monitoring data of the most recent 30 days.

[0107] Calculate the cross-correlation function between the groundwater level sequence at the dam abutment and the reservoir water level sequence under multiple different time lags. Let the time lag be... , The value range is from 0h to , The maximum search latency is 10 days, calculated as 240 hours. The step size for the value is consistent with the sampling interval of the monitoring data. When the sampling interval is 1 hour, Take the values ​​0h, 1h, 2h, ..., 240h. For each... Values, calculate the groundwater level sequence and hysteresis at the dam abutment. Cross-correlation function values ​​between reservoir water level sequences at time points : ; in This represents the mean of the reservoir water level sequence segments used in the calculation. This represents the average of the groundwater level sequence segments at the dam abutment used in the calculation. The value of is in the range of -1 to 1. The closer the absolute value is to 1, the stronger the correlation between the two sequences under this lag.

[0108] Take the cross-correlation function value The time lag corresponding to the maximum is used as the base lag time. : ; The unit is h. The basic lag time reflects the time delay during which, in a statistical sense, changes in reservoir water level have the most significant impact on changes in groundwater level at the dam abutment.

[0109] In obtaining the basic lag time Then, the basic lag time is dynamically adjusted based on the current rate of change of the reservoir water level. Let the current rate of change of the reservoir water level be... The unit is m / h. The dynamically adjusted lag time. The calculation formula is: ; in This is a preset proportionality coefficient, with units of h / m and a value range of 0.5h / m to 5.0h / m. The specific value is determined based on the geological permeability characteristics of the dam abutment hill. When the reservoir water level changes at a certain rate... When it is large, the exponent term Smaller The shortening indicates that the head transfer is more rapid when the reservoir water level changes quickly, and the groundwater level response at the dam abutment is faster; when the reservoir water level change rate... When it is small, the exponent term Approaching 1, near This indicates that the groundwater level at the dam abutment responds with a statistically averaged lag time when the reservoir water level changes slowly.

[0110] In one embodiment of the present invention, the right abutment of an earth-rock dam is taken as an example. The abutment hillside is composed of gravel and sand strata with a permeability coefficient of 1×10⁻⁶.- ³m / s. The data processing center performs cross-correlation analysis on the reservoir water level sequence and the groundwater level sequence at the dam abutment for the previous 720 hours to calculate the cross-correlation function value. exist The maximum value of 0.89 is obtained when the time is 12h, therefore the basic lag time is... The current time is 12 hours. The reservoir water level is rising at a rate of 0.08 m / h, which means... =0.08 m / h. Based on the geological seepage characteristics of the dam abutment, a preset proportionality coefficient is determined. Take 2.5 h / m. The dynamically adjusted lag time. =12·e^(-2.5×0.08)=12·e^(-0.20)=12×0.819=9.8h. This dynamically adjusted lag time of 9.8h is less than the basic lag time of 12h, reflecting the accelerated physical process of water head transmission into the dam abutment hillside during a rapid rise in reservoir water level. When the dam abutment experiences prolonged drought, and the reservoir water level change rate drops to 0.01m / h, =12·e^(-2.5×0.01)=12·e^(-0.025)=12×0.975=11.7h, which is close to the basic lag time of 12h.

[0111] In another embodiment of the present invention, when there is a significant difference in geological conditions between the left and right abutments, the aforementioned lag time calculation steps are performed independently for the left and right abutments respectively, to obtain the dynamically adjusted lag time for the left abutment. Lag time after dynamic adjustment of the right abutment Preset proportional coefficients for the left and right abutments. and Each value is independently determined based on the geological permeability characteristics of its respective dam shoulder mountain.

[0112] In another embodiment of the invention, the time length of the cross-correlation analysis and maximum search latency The adjustment is adaptive based on the permeability characteristics of the dam abutment hillside. For areas with a high permeability coefficient (greater than 1×10⁻⁶), the adjustment is made accordingly. - The dam shoulder mountain with a speed of 3m / s, Take 360h, Take 120h; for those with a small permeability coefficient (less than 1×10⁻⁶), -5 The dam shoulder mountain (m / s), Take 1440h, Use 480h. Adaptive adjustment ensures that the cross-correlation analysis window covers the complete reservoir water level fluctuation cycle, avoiding bias in the basic lag time estimation caused by a window that is too short or too long.

[0113] Calculate the displacement change rate sequence based on the final displacement estimate; Within a set time window, the number of times the displacement rate of change changes from positive to negative or from negative to positive in the rate of change sequence is used as the direction alternation frequency. When the frequency of the alternation of directions exceeds a preset frequency threshold or the amplitude of the rate of change sequence exceeds a preset amplitude threshold, it is determined that the dam displacement trend has an abnormal reversal.

[0114] Specifically, the abnormal turning point determination step is executed after the data processing center obtains the final displacement estimate at each time step. The data processing center continuously receives the final displacement estimate at each time step output by the confidence fusion step. The values ​​are then stored in the displacement estimation buffer. The displacement estimation buffer stores the sequence of final displacement estimates for the most recent 720 hours.

[0115] Calculate the displacement rate of change sequence. Based on the final displacement estimate at the current time. And the final displacement estimate of the previous moment Calculate the rate of change of displacement at the current moment using the following formula. : ; The unit is mm. It should be noted that, since the sampling interval of the monitoring data is fixed, the rate of change of displacement with respect to time is essentially the displacement difference between adjacent moments. When When this occurs, it indicates that the dam displacement is increasing and the dam is accelerating its movement downstream; when This indicates that the dam displacement is decreasing and the dam body is rebounding upstream.

[0116] Within a set time window, the number of times the displacement rate changes from positive to negative or from negative to positive is counted as the direction alternation frequency. Let the length of the judgment time window be... , Take 168 hours, which is 7 days. Take the time before the current moment. Displacement change rate sequence within the time window Check the sign of the rate of change of displacement between adjacent time points one by one. and Then it is determined as an alternation from positive to negative; if and If a certain value is found, it is considered a single alternation from negative to positive. The total number of these two alternations within the time window is recorded as follows: Alternating frequency of directions The calculation method is as follows: ; The unit is times per hour.

[0117] Simultaneously, the amplitude characteristics of the displacement rate of change sequence are calculated. The amplitude of the displacement rate of change sequence is taken within the judgment time window. Maximum absolute value: ; The unit is mm.

[0118] When the direction alternates frequency Exceeding the preset frequency threshold or displacement change rate amplitude Exceeding the preset amplitude threshold At that time, it was determined that the dam displacement trend had an abnormal reversal. The alarm is triggered when the number of direction changes exceeds 3.36 times within 7 days, with a set value of 0.02 times / h. In actual judgment, an integer threshold is used, meaning that an abnormality is judged when the number of direction changes is ≥4 times within 7 days. Set the value to 1.5mm. An abnormal turning point is triggered if either the alternating frequency or amplitude condition is met; both conditions do not need to be met simultaneously.

[0119] In one embodiment of the present invention, a concrete gravity dam is taken as an example. Under normal operating conditions, the dam's displacement change rate alternates direction 1 to 2 times within a 7-day time window, with an amplitude typically ranging from 0.006 times / h to 0.012 times / h. On a certain day, the dam experiences a sudden drop in reservoir water level, decreasing by 2 meters within 12 hours. Monitoring data shows that from day 1 to day 4, the dam displacement continuously increases downstream, with a positive displacement change rate; from day 5 onwards, the dam displacement rebounds upstream, and the displacement change rate becomes negative; on day 6, the dam displacement again shifts downstream, and the displacement change rate becomes positive again. Within the 7-day time window, the displacement change rate changes from positive to negative once; from day 4 to day 5, it changes from negative to positive once; and from day 5 to day 6, the total number of directional alternations is [not specified]. =2, frequency of alternating directions =2 / 168 = 0.012 times / h, which does not exceed the preset frequency threshold of 0.02 times / h. However, the amplitude of the displacement change rate within this time window is... =2.3mm, exceeding the preset amplitude threshold of 1.5mm, triggering an abnormal turning point judgment, and the system outputs an abnormal turning point warning signal for the dam body displacement trend.

[0120] In another embodiment of the present invention, when an abnormal turning point determination is triggered, the data processing center generates early warning information. This early warning information includes the triggering conditions: the direction alternation frequency or amplitude exceeds a threshold, the current direction alternation frequency value, the current displacement change rate amplitude, and the start time of the abnormal turning point. The early warning information is sent to the dam safety monitoring center via wired or wireless communication for further analysis and decision-making by management personnel. After the early warning information is sent, the abnormal turning point determination process continues. When the direction alternation frequency falls below a preset frequency threshold and the displacement change rate amplitude falls below a preset amplitude threshold for 24 hours, the abnormal turning point state is automatically lifted, and the early warning is canceled.

[0121] In another embodiment of the present invention, a preset frequency threshold is used. and preset amplitude threshold The determination is based on historical monitoring data of the dam. Specifically, the method involves obtaining displacement monitoring data from the dam under normal operating conditions over the past three years, calculating the daily directional alternation frequency and displacement change rate amplitude, and using the 95th quantile of the statistical distribution as a preset threshold. For example, if the 95th quantile of the daily directional alternation frequency under normal operating conditions over the past three years is 0.018 times / hour, then... Take 0.02 times / h; if the 95th percentile of the daily displacement change rate amplitude is 1.3 mm, then A threshold of 1.5 mm is used. Determining the threshold based on historical statistics allows the judgment of abnormal turning points to adapt to the individual displacement change characteristics of different dams, avoiding false alarms or missed alarms caused by a uniform threshold.

[0122] The multidimensional driving features also include temperature sequences and rainfall sequences; The temperature sequence and the rainfall sequence are interpolated and resampled according to the unified time axis, and then time-series aligned and fused.

[0123] Specifically, the multidimensional driving characteristics also include temperature and rainfall sequences. The temperature sequence reflects the impact of ambient temperature changes on the thermal expansion and contraction of dam materials, as well as the temperature component in the relationship between reservoir water level and dam displacement. The rainfall sequence reflects the replenishment effect of rainfall infiltration on the seepage field and pore water pressure of the dam abutment.

[0124] The temperature sequence was collected as follows: an air temperature monitoring station was set up in the dam site area, and a temperature sensor was installed in the station. The temperature sensor used was a platinum resistance thermometer or a thermocouple thermometer. Ambient temperature data was collected according to a preset sampling frequency to form a temperature sequence. The location of the temperature monitoring station should avoid direct sunlight and local heat sources, and is usually set up in an open area near the dam shoulder, no more than 500m away from the dam. In one embodiment of the invention, the temperature sensor has a sampling frequency of 1 time / h, a temperature measurement accuracy of ±0.5℃, and a measurement range of -30℃ to 60℃.

[0125] The rainfall sequence was collected as follows: rain gauges, either tipping bucket or weighing type, were installed in the dam site area. Rainfall data were collected according to a preset sampling frequency to form a rainfall sequence. The resolution of a tipping bucket rain gauge is 0.1 mm, while that of a weighing rain gauge is 0.01 mm. The diameter of the rain gauge's collection port is typically 200 mm, and the installation height is 1.5 m to 2.0 m above the ground. The installation location should avoid obstruction by buildings and trees. In one embodiment of the invention, the rain gauge's sampling frequency is 1 time / h, and the rainfall unit is mm / h, meaning the rainfall at each moment is the cumulative rainfall for that hour.

[0126] Temperature and rainfall sequences are interpolated and resampled along a unified time axis. Using the historical displacement sequence as the time reference, for each time point on the unified time axis, if there is no data for either the temperature or rainfall sequence at that time point, a linear interpolation method is used to interpolate using adjacent known data points to obtain the corresponding temperature and rainfall values ​​for that time point. The linear interpolation formula is: ; in For the target time point, and These are the closest known data time points before and after the target time point. and They are respectively and Temperature or rainfall value at any given time.

[0127] After interpolation and resampling, the temperature and rainfall sequences are time-series aligned and fused. The specific time-series alignment method is as follows: the interpolated and resampled temperature sequence... Rainfall sequence The homogeneous reservoir water level sequence, homogeneous lag time characteristic sequence, homogeneous pore water pressure characteristic sequence, and historical displacement sequence are all correlated one-to-one on the time axis. Each set of data includes the historical displacement value of the dam body, reservoir water level value, lag time characteristic value, pore water pressure characteristic value, temperature value, and rainfall value at the same moment. The fused multidimensional driving feature vector is as follows: ; in for The input feature vector at time step 6 has a dimension of 6. This multidimensional driving feature vector is input into the displacement prediction network, replacing the multidimensional driving features that do not include temperature and rainfall features.

[0128] In one embodiment of the present invention, a certain earth-rock dam is taken as an example. The area where the dam is located experiences hot and rainy summers and cold and dry winters, with a seasonal temperature difference exceeding 50°C. Under the condition that the reservoir water level remains stable, the dam displacement still exhibits obvious annual periodic fluctuations: the downstream displacement of the dam increases during the high temperatures of summer, and the upstream displacement of the dam shrinks during the low temperatures of winter. The amplitude of this displacement fluctuation is approximately 2mm to 4mm, entirely caused by temperature changes. After the temperature sequence is input into the data driving subnetwork as a component of the multidimensional driving feature, the gated recurrent unit layer can learn the nonlinear mapping relationship between temperature changes and dam displacement during the time-series encoding process. This allows for automatic compensation for temperature-induced displacement increments during the high temperatures of summer, avoiding systematic biases in the prediction results due to neglecting the temperature effect.

[0129] In one embodiment of the invention, a dam experiencing continuous heavy rainfall is taken as an example. The area where the dam is located experienced daily rainfall exceeding 50 mm / d for three consecutive days, with the rainfall sequence showing a cumulative rainfall of 180 mm. Rainfall infiltration led to an increase in the saturation of the dam abutment, causing the pore water pressure at the abutment to gradually rise from an initial 30 kPa to 85 kPa. There is a temporal sequence between the rainfall sequence and the pore water pressure sequence: 6 to 12 hours after the heavy rainfall, the pore water pressure at the dam abutment began to respond and gradually increased. The gated recurrent unit layer in the data-driven subnetwork simultaneously receives the rainfall sequence and the pore water pressure sequence as input through temporal encoding, enabling it to automatically learn the temporal dependency between rainfall infiltration and the pore water pressure response, and predict the displacement trend during heavy rainfall, without relying solely on the hysteresis response characteristics of the pore water pressure sequence.

[0130] In one embodiment of the present invention, the sampling frequencies of the temperature and rainfall sequences are inconsistent with those of the displacement monitoring device. When the sampling frequency of the displacement monitoring device is 1 time / hour and the sampling frequency of the temperature sensor is 10 minutes / time, the number of raw data points for the temperature sequence is greater than that for the displacement sequence. In this case, using the time reference of the historical displacement sequence as a unified time axis, the temperature sequence is downsampled at the hourly intervals, and the average of the temperature data within 30 minutes before and after the hourly interval is taken as the temperature value at that hourly interval, thus maintaining the same sampling frequency as the displacement sequence. When the sampling frequency of the rainfall meter is 1 time / day and the sampling frequency of the displacement sequence is 1 time / hour, the number of raw data points for the rainfall sequence is less than that for the displacement sequence. In this case, using the time reference of the historical displacement sequence as a unified time axis, the rainfall sequence is upsampled at the hourly intervals, distributing the daily cumulative rainfall evenly across 24 hourly intervals, or using a zero-order hold method to assign the daily rainfall value to all hourly intervals within that day. In one embodiment of the present invention, for short-duration heavy rainfall events, the zero-order hold method can more accurately reflect the duration effect of rainfall infiltration.

[0131] Temperature and rainfall sequences, as supplementary components of the multi-dimensional driving features, together with reservoir water level sequences and dam abutment constraint feature sets, constitute the complete input features. During the training phase of the displacement prediction network, the weights of the input nodes corresponding to the temperature and rainfall sequences are automatically optimized using a backpropagation algorithm to determine the degree of influence of temperature changes and rainfall infiltration on dam displacement. During the prediction phase, real-time data from the temperature and rainfall sequences are continuously collected and input into the network, enabling the displacement prediction network to adapt to the dam displacement variation patterns under different seasons and weather conditions.

[0132] A deep learning-based dam displacement estimation system includes: The data acquisition module is used to acquire monitoring data, which includes the historical displacement sequence of the dam body, the reservoir water level sequence, and the groundwater level sequence and pore water pressure sequence of the dam abutment hill. The water level rise and fall and lag analysis module is used to calculate the rate of change and direction of rise and fall of the reservoir water level based on the reservoir water level sequence, and to calculate the lag time based on the groundwater level sequence; A dam shoulder constraint feature construction module is used to construct a dam shoulder constraint feature set based on the lag time and the pore water pressure sequence. The dam shoulder constraint feature set includes lag time features and pore water pressure features. The temporal alignment and fusion module is used to perform temporal alignment and resampling on the reservoir water level sequence and the dam abutment constraint feature set, using the time reference of the historical displacement sequence as a unified time axis, and fuse them into multi-dimensional driving features. A dual-drive displacement prediction module is used to input the multi-dimensional driving features into a displacement prediction network. The displacement prediction network includes a physical drive sub-network and a data drive sub-network. Both sub-networks are input with the dam abutment constraint feature set as dynamic boundary conditions. The physical drive sub-network outputs the physical predicted displacement based on the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics. The data drive sub-network outputs the data predicted displacement. The confidence fusion output module is used to adaptively adjust the confidence level based on the historical prediction errors of the physical predicted displacement and the data predicted displacement, and then weight and fuse them into a final displacement estimate.

[0133] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0134] In the description of this application, it should be noted that the terms "first", "second", and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0135] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0136] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the shown or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0137] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0138] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0139] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

[0140] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

Claims

1. A deep learning-based method for estimating dam displacement, characterized in that, Includes the following steps: Acquire monitoring data, including the historical displacement sequence of the dam body, the reservoir water level sequence, and the groundwater level sequence and pore water pressure sequence of the dam abutment hills; Calculate the rate of change and direction of rise and fall of reservoir water level based on reservoir water level sequence, and calculate the lag time based on groundwater level sequence; A dam shoulder constraint feature set is constructed based on the lag time and pore water pressure sequences. The dam shoulder constraint feature set includes lag time features and pore water pressure features. Using the historical displacement sequence as a unified time axis, the reservoir water level sequence and the dam abutment constraint feature set are time-aligned and resampled, and then fused into multi-dimensional driving features. Multidimensional driving features are input into the displacement prediction network, which includes a physical driving sub-network and a data driving sub-network. Both sub-networks use the dam abutment constraint feature set as the dynamic boundary condition input. The physical driving sub-network outputs the physical predicted displacement based on the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics. The data driving sub-network outputs the data predicted displacement. The data driving sub-network contains a gated recurrent unit layer and a temporal attention layer. The gated recurrent unit layer performs temporal encoding of multidimensional driving features; The temporal attention layer calculates attention weights based on the reservoir water level change amplitude at each historical moment, and then performs weighted fusion of the hidden states at each moment output by the gated recurrent unit layer to predict displacement. The confidence level is adaptively adjusted based on the historical prediction errors of the physical and data-based displacement predictions, and then weighted and fused to obtain the final displacement estimate.

2. The dam displacement estimation method based on deep learning according to claim 1, characterized in that, The physical driving sub-network uses a mechanical relationship model to calculate the predicted physical displacement. This mechanical relationship model is in the form of a second-order differential equation. The second derivative of the dam displacement with respect to time is related to the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics. The mechanical relationship model is expressed as follows: ； in, To predict displacement in physics, For time, The rate of change of reservoir water level The characteristic value of pore water pressure. The characteristic value of the lag time. The coefficient representing the influence of the rate of change of reservoir water level. The pore water pressure influence coefficient is... The lag time influence coefficient is... For constant terms; The physics-driven subnetwork discretizes the mechanical relationship model and then recursively calculates the physical predicted displacement. The discretization method is as follows: Let the time step be The current time is the [number]th [time]. Step, the previous moment was the first step. Step 1: The first derivative of displacement with respect to time is approximated using a first-order backward difference, and the second derivative of displacement with respect to time is approximated using a difference of the first-order backward difference. The recursive formula for the physically predicted displacement is: ； in, For the first The physical prediction displacement of the step. For the first The rate of change of reservoir water level in the step, For the first The characteristic value of pore water pressure in the step, For the first The lag time characteristic value of the step; the initial value required for recursive calculation. and The measured displacement values ​​at the initial moment and the first moment in the historical displacement sequence are taken respectively.

3. The dam displacement estimation method based on deep learning according to claim 1, characterized in that, The adaptive adjustment of confidence level is as follows: The absolute value of the difference between the predicted displacement and the measured displacement of the physical driving subnetwork and the data driving subnetwork at each moment within the historical time window is calculated as the prediction error. The confidence levels of the two subnetworks are determined based on the prediction errors of the physical-driven subnetwork and the data-driven subnetwork within the historical time window. The final displacement estimate is obtained by weighting the physical predicted displacement and the data predicted displacement according to the confidence level. When the fluctuation amplitude of the pore water pressure characteristic in the dam abutment constraint feature set exceeds a preset fluctuation threshold, the confidence level of the physical driving sub-network is reduced and the confidence level of the data driving sub-network is increased; the calculation method for the fluctuation amplitude of the pore water pressure characteristic is as follows: Take the pore water pressure sequence within a preset fluctuation statistical window length before the current time, and calculate the standard deviation of the sequence. ;when Exceeding the preset fluctuation threshold When, confidence level switching is triggered; when Falling back to After the preset recovery time is reached, the confidence level switching mechanism will be deactivated, and the confidence level will be restored to be adaptively determined by the average prediction error.

4. The dam displacement estimation method based on deep learning according to claim 3, characterized in that, The confidence level is determined as follows: the average prediction error of the physical driving subnetwork within the historical time window is: ； The average prediction error of the data-driven subnetwork within the historical time window is: ； The confidence level of the physical-driven subnetwork is: ； The confidence level of the data-driven subnetwork is: ； The final displacement estimate is: ； in, The length of the historical time window. For the current moment, For indexing moments within a historical time window, For physical driving subnetworks in The physical predicted displacement at time t. For data-driven subnetworks in Predict displacement based on data at any given time. for Measured value of dam displacement at time t. and These are the average prediction errors of the physics-driven subnetwork and the data-driven subnetwork, respectively. and The confidence levels of the physical-driven subnetwork and the data-driven subnetwork are respectively. , for The final displacement estimate at time [time].

5. The dam displacement estimation method based on deep learning according to claim 1, characterized in that, The dam abutment mountain includes a left dam abutment mountain and a right dam abutment mountain; the groundwater level sequence and pore water pressure sequence of the left dam abutment mountain and the groundwater level sequence and pore water pressure sequence of the right dam abutment mountain are obtained respectively; constraint feature sets for the left dam abutment and constraint feature sets for the right dam abutment are constructed respectively; the constraint feature sets for the left dam abutment and the constraint feature sets for the right dam abutment are used as independent dynamic boundary condition features for temporal alignment and fusion; The lag time calculation steps are performed independently for the left and right abutments to obtain the dynamically adjusted lag time for the left and right abutments respectively; the proportional coefficients for the left and right abutments are independently determined based on the geological permeability characteristics of their respective abutment hills.

6. The dam displacement estimation method based on deep learning according to claim 1, characterized in that, During the displacement prediction process, the lag time characteristics and pore water pressure characteristics in the dam abutment constraint feature set are updated based on the deviation between the real-time monitored dam displacement sequence and the final displacement estimate. The updated features are re-input into the physical driving sub-network to generate the updated physical prediction displacement; The update trigger condition is as follows: calculate the moving average of the displacement deviation sequence within the length of the deviation statistical window before the current time. When the absolute value of the moving average exceeds a preset deviation threshold, online correction is triggered. The correction amount is determined by taking the monitoring data within the length of the correction window before the trigger time, and constructing a least squares problem with the objective of minimizing the sum of squares of the deviations between the measured dam displacement and the physical predicted displacement at each time point within the correction window. ； The lag time characteristic correction quantity is obtained by solving. and pore water pressure characteristic correction amount The corrected hysteresis time characteristic values ​​and pore water pressure characteristic values ​​are as follows: ； ； in, and These are the original hysteresis time characteristic value and pore water pressure characteristic value at the trigger correction time, respectively. To correct the time intervals within the window, The physical predicted displacement is calculated using the corrected hysteresis time characteristic value and pore water pressure characteristic value as inputs; Online calibration only updates the feature input values ​​and does not update the parameters of the mechanical relationship model.

7. The dam displacement estimation method based on deep learning according to claim 1, characterized in that, The lag time is calculated as follows: Calculate the cross-correlation function between the groundwater level sequence at the dam abutment and the reservoir water level sequence under multiple different time lags, and take the time lag corresponding to the maximum cross-correlation function value as the basic lag time; Using the aforementioned basic lag time as a benchmark, multiply it by an exponential decay function with the absolute value of the reservoir water level change rate as the exponent to obtain the dynamically adjusted lag time. Cross-correlation function The formula for calculation is: ； Basic lag time The value corresponding to the maximum value of the cross-correlation function : ； Dynamically adjusted lag time The formula for calculation is: ； in, This is the time lag. For sequence length, For reservoir water level sequence, The groundwater level sequence at the dam abutment. and These are the mean values ​​of the reservoir water level sequence segment and the dam abutment groundwater level sequence segment, respectively, used in the calculation. The rate of change of reservoir water level This is the proportionality coefficient.

8. The dam displacement estimation method based on deep learning according to claim 1, characterized in that, Calculate the displacement change rate sequence based on the final displacement estimate; Within a set time window, the number of times the displacement rate of change changes from positive to negative or from negative to positive in the rate of change sequence is used as the direction alternation frequency. When the frequency of the alternation of directions exceeds a preset frequency threshold or the amplitude of the rate of change sequence exceeds a preset amplitude threshold, it is determined that the dam displacement trend has an abnormal reversal. rate of change of displacement The formula for calculation is: ； Alternating frequency of directions The formula for calculation is: ； Displacement change rate amplitude The formula for calculation is: ； in, for The final displacement estimate at time [time]. for The final displacement estimate at time [time]. To determine the time window The total number of times the value changes from positive to negative and from negative to positive; when Exceeding the preset frequency threshold or Exceeding the preset amplitude threshold It was determined that the dam displacement trend had an abnormal reversal.

9. The dam displacement estimation method based on deep learning according to claim 1, characterized in that, The multidimensional driving features also include temperature sequences and rainfall sequences; The temperature sequence and the rainfall sequence are interpolated and resampled along the unified time axis, and then time-series aligned and fused. The interpolation and resampling uses a linear interpolation method, and the linear interpolation formula is as follows: ； in, For the target time point, and These are the closest known data time points before and after the target time point. and They are respectively and Temperature or rainfall value at any given time.

10. A deep learning-based dam displacement estimation system, used to execute the deep learning-based dam displacement estimation method according to any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire monitoring data, which includes the historical displacement sequence of the dam body, the reservoir water level sequence, and the groundwater level sequence and pore water pressure sequence of the dam abutment hill. The water level rise and fall and lag analysis module is used to calculate the rate of change and direction of rise and fall of the reservoir water level based on the reservoir water level sequence, and to calculate the lag time based on the groundwater level sequence; A dam shoulder constraint feature construction module is used to construct a dam shoulder constraint feature set based on the lag time and the pore water pressure sequence. The dam shoulder constraint feature set includes lag time features and pore water pressure features. The temporal alignment and fusion module is used to perform temporal alignment and resampling on the reservoir water level sequence and the dam abutment constraint feature set, using the time reference of the historical displacement sequence as a unified time axis, and fuse them into multi-dimensional driving features. A dual-drive displacement prediction module is used to input the multi-dimensional driving features into a displacement prediction network. The displacement prediction network includes a physical drive sub-network and a data drive sub-network. Both sub-networks are input with the dam abutment constraint feature set as dynamic boundary conditions. The physical drive sub-network outputs the physical predicted displacement based on the reservoir water level change rate, pore water pressure characteristics, and lag time characteristics. The data drive sub-network outputs the data predicted displacement. The confidence fusion output module is used to adaptively adjust the confidence level based on the historical prediction errors of the physical predicted displacement and the data predicted displacement, and then weight and fuse them into a final displacement estimate.

Images