Earth and rockfill dam seepage pressure prediction method based on mechanism and data dual drive

Abstract

The invention provides an earth and rockfill dam seepage pressure prediction method based on mechanism and data dual drive, and the method comprises the following steps: generating multi-working-condition seepage pressure data, optimizing a radial basis function neural network through an aurora optimization algorithm, and constructing an efficient proxy model; on the basis of a proxy model prediction result, fitting is carried out by combining measured data, and a seepage pressure prediction mechanism model based on a hysteresis effect function is established; with the predicted value of the mechanism model as a label, supervising the training model to learn the physical law of the seepage field; and freezing the physical feature extraction layer in the trained model, and finely adjusting the time sequence modeling layer by using actually measured data to realize approximation of an actually measured value. According to the method, mechanism driving and data driving are organically combined, the interpretability of a physical rule is reserved, complex factors which are not considered by a mechanism model are complemented by utilizing actually measured data, and the generalization ability of a deep learning model under extreme working conditions is remarkably improved while the prediction precision of the model is ensured.

Description

Technical Field

[0001] The present invention relates to the field of dam safety monitoring, and specifically to a seepage pressure prediction method for earth-rock dams based on the dual drive of mechanism and data. Background Art

[0002] Dam safety monitoring is an important technical means to ensure the long-term stable operation of water conservancy projects. As one of the core indicators of earth-rock dam safety monitoring, seepage pressure directly reflects the operation status of the dam body's anti-seepage system and is a key parameter for evaluating the safety of earth-rock dams. Therefore, using the seepage pressure monitoring data of earth-rock dams to establish a reasonable monitoring model helps to analyze and monitor the current seepage behavior of the dam and provides decision-making support for the safe operation and scientific management of earth-rock dams.

[0003] The engineering practice of earth-rock dam seepage pressure monitoring mainly relies on piezometers and osmometers arranged inside the dam body. These sensors can continuously collect pore water pressure data at different locations to form high-precision time-series monitoring records. The monitoring data shows that the change of seepage pressure is mainly affected by the comprehensive influence of three types of environmental factors: (1) The change of reservoir water level, as the dominant driving factor, directly affects the seepage pressure distribution; (2) Rainfall infiltration has a secondary influence by changing the water content of the dam body; (3) The aging factor causes the seepage pressure to show a long-term evolution trend. The coupled action of these three factors jointly determines the dynamic change characteristics of seepage pressure.

[0004] In terms of monitoring technology, automated monitoring systems have generally been established for newly built earth-rock dams in recent years, which can realize the real-time collection and remote transmission of seepage pressure. These monitoring data provide a basis for the establishment of seepage pressure prediction models. In terms of modeling methods, currently, it is mainly divided into two major systems: mechanism models and data-driven models. Mechanism models are based on the basic principles of seepage mechanics and describe the seepage process by establishing control equations and constitutive relations, with clear physical meanings. Data-driven models directly start from the monitoring data and use statistical or machine learning methods to mine the correlation relationships between variables, with stronger adaptive capabilities.

[0005] Although the mechanism model has a clear physical meaning, it still has the following common problems: 1) Low computational efficiency: Numerical methods such as the finite element method (FEM) require fine mesh division and complex mathematical formula solving, resulting in high computational costs. 2) Parameter sensitivity and uncertainty: The model prediction results are extremely sensitive to input parameters such as permeability coefficients and boundary conditions, and these parameters are often difficult to accurately obtain. At the same time, there are uncertainties in material properties, initial conditions, etc., and parameter time-variation (such as material aging, crack development) is not fully considered. 3) Difficulty in multi-field coupling: There are complex coupling effects between the actual seepage field and the stress field and temperature field, but existing models often adopt simplified assumptions (such as one-way coupling), resulting in deviations in simulations under complex working conditions such as sudden water level changes and earthquakes. 4) Poor static adaptability: Most models are based on quasi-static assumptions and cannot effectively adapt to environmental dynamic changes or real-time data updates, making it difficult to meet the monitoring requirements of dynamic systems. 5) Insufficient characterization of hysteresis effects: Traditional deterministic equations approximate unsteady seepage by superimposing the previous average water level component. Although this method is simple and intuitive, it has obvious defects: The selection of the time window depends on experience and is relatively rough (usually fixed at 7 - 30 days), and it cannot accurately reflect the dynamic hysteresis characteristics of seepage pressure, resulting in insufficient prediction accuracy.

[0006] Although data-driven models perform excellently in processing monitoring data, they have the following key problems: 1) Lack of physical consistency: The model completely relies on data training and may output results that violate physical laws such as mass conservation and Darcy's law, posing risks in engineering applications. 2) Strong dependence on data quality: The model performance strongly depends on data quality, and there are many problems in monitoring data in actual engineering: sparse measurement points, data heterogeneity, noise interference, etc., which require complex preprocessing. 3) Poor overfitting and extrapolation ability: The model performs well on training data, but the prediction deviation for extreme working conditions beyond the training distribution is significant, and there is a problem of catastrophic forgetting. 4) Poor interpretability: The internal mechanisms of "black box" models such as deep learning are complex, and it is difficult to associate prediction results with specific engineering factors, which is not conducive to engineering personnel's understanding and decision-making. 5) Difficulty in online updating: Existing models are mostly in an offline training mode and are difficult to adapt to the degradation of material properties and environmental changes during the long-term service of dams.

[0007] There are still obvious limitations in the current research on the integration of mechanism and data-driven methods: Most methods only adopt simple weighted average or serial connection strategies, resulting in loose physical constraints, insufficient dynamic adaptation ability, and limited coupling depth, making it difficult to meet the prediction requirements of the seepage pressure of earth-rock dams under extreme working conditions. Summary of the Invention

[0008] Aiming at the key problems existing in the existing seepage pressure monitoring technology of earth-rock dams, the present invention proposes a seepage pressure prediction method for earth-rock dams based on double drive of mechanism and data. By means of agent modeling guided by finite element simulation and optimization of the hysteresis effect function based on normal distribution, an efficient and dynamic improved mechanism model is constructed to overcome the defects of high calculation cost, large parameter sensitivity and poor static adaptability of the traditional mechanism model. At the same time, a deep learning migration framework with physical constraints is designed, and a two-stage training strategy is adopted, so that the model can not only follow the physical laws of the seepage field, but also adapt to the local non-linear characteristics of the measured data, thus solving the problems of lack of physical consistency of the pure data-driven model and weak extrapolation ability under extreme working conditions. Finally, the hyperparameters of the agent model and deep learning are optimized by the PLO algorithm, and the dynamic hysteresis component and aging factor are fused to realize the complementary advantages of the interpretability of the mechanism model and the self-adaptability of the data-driven model, and construct a seepage pressure prediction model with high precision, strong robustness and engineering applicability, providing reliable technical support for dam safety monitoring and risk warning. The technical solutions provided by the present invention are as follows:

[0009] A seepage pressure prediction method for earth-rock dams based on double drive of mechanism and data, comprising the following steps:

[0010] Step 1, rejecting outliers and filling missing values in the monitoring data, and performing preprocessing by wavelet denoising;

[0011] Step 2, generating a water level-seepage pressure data set based on finite element simulation, and optimizing the radial basis function neural network RBFNN by the aurora optimization algorithm PLO to construct an agent model; based on the prediction results of the agent model, fitting with the measured data, and establishing a seepage pressure prediction mechanism model based on the hysteresis effect function;

[0012] Step 3, physical constraint-based deep learning migration, including two stages: the first stage of training, using the predicted values of the mechanism model in Step 2 as labels, and supervising the training of the PLO-Transformer-BiLSTM network to learn the physical laws of the seepage field; the second stage of fine-tuning, freezing the physical feature extraction layer in the PLO-Transformer-BiLSTM network completed in the first stage of training, fine-tuning the time series modeling layer BiLSTM with the measured data, and optimizing the time series residual module with the measured data;

[0013] Step 4, calculating regression evaluation indexes through the data set, comparing the training effects of the theoretical labels and the measured labels; calculating the model evaluation indexes on the virtual working conditions generated by finite element, and comparing and verifying the extrapolation ability of the model; saving the optimal parameters and normalized parameters, and embedding them into the safety monitoring system to implement seepage pressure prediction and warning.

[0014] Preferably, Step 2 specifically includes the following steps:

[0015] Step 21: Based on finite element simulation, simulate the seepage field of the earth-rock dam to generate seepage pressure datasets under different working conditions; cover possible boundary conditions of the actual project through parametric modeling to ensure data diversity;

[0016] Step 22: Define input variables and output variables, and construct an initial RBFNN surrogate model; use the sum of the mean square errors of the dataset as the fitness function to optimize the hyperparameters of the RBFNN; evaluate the accuracy of the surrogate model through error metrics to enable it to efficiently replace the finite element model.

[0017] Step 23: Introduce the lag effect function of the normal distribution, optimize the lag parameters through PLO to generate the lag component; combine the prediction results of the surrogate model, the lag component and the aging factor to establish a seepage pressure prediction mechanism model based on the lag effect function.

[0018] Preferably, the RBFNN surrogate model is a three-layer feedforward network, and its network structure includes an input layer, a hidden layer and an output layer. The hidden layer uses a Gaussian kernel function as the activation function, and its input expression is:

[0019]

[0020] Among them, is the input, is the center of the j-th hidden layer neuron, is the expansion speed, which controls the width of the basis function; the network output is the weighted sum of the hidden layer outputs:

[0021]

[0022] Among them, is the network output, is the weight, is the number of hidden layer neurons.

[0023] Preferably, the seepage pressure prediction mechanism model based on the lag effect function is specifically:

[0024]

[0025] Among them, is the constant term, t is the time calculated from the starting point of the calculation, and to are the regression coefficients of each component; is the prediction result of the surrogate model at time t, is the lag effect function of the normal distribution, and its form is:

[0026]

[0027] Among them, For adjusting parameters, is the peak value of lag days, is the number of influencing days, and is obtained by the PLO optimization algorithm.

[0028] Preferably, when the seepage pressure prediction mechanism model based on the lag effect function is calculated, the continuous integral is converted into a discrete integral, and the window size is taken as 2 to 3 times of to cover the main lag period.

[0029] Preferably, step 3 specifically includes the following steps:

[0030] Step 31, the supervised training of the first-stage mechanism model is specifically as follows: The input features are the upstream water level, rainfall, and aging factor, and the label data is the theoretical seepage pressure value predicted by the mechanism model. According to the ratio of 70%, 10%, and 20%, the data set is divided into a training set, a validation set, and a test set; Build a Transformer-BiLSTM model, including a Transformer module, a BiLSTM module, and a fully connected layer; Through the theoretical label pre-training model, guide the Transformer-BiLSTM model to learn the physical laws of the seepage field, establish the mapping relationship between the input features and the theoretical seepage pressure value, and realize the migration of physical laws;

[0031] Step 32, the fine-tuning of the second-stage measured data is specifically as follows: Replace the theoretical label with the measured seepage pressure value, and keep the input features unchanged; Fix the position embedding layer and self-attention layer weights of the Transformer module, and only fine-tune the parameters of the BiLSTM module and the subsequent fully connected layer. On the basis of retaining the global features of the mechanism model, adapt to the local deviation of the measured data, improve the generalization ability of the model, and realize the integration of physical constraints and data-driven.

[0032] Preferably, the Transformer-BiLSTM model connects the Transformer module and the BiLSTM module in series, and finally outputs the prediction value through the fully connected layer. The Transformer module includes a position embedding layer and a self-attention layer. The position embedding layer introduces position encoding to retain the timing information, and the encoding formula is:

[0033]

[0034] where pos is the sequence position, i is the dimension index, and d model is the feature dimension;

[0035] The self-attention layer captures the global dependence through multi-head attention, and the calculation formula is:

[0036]

[0037] Where Q, K, and V are the query, key, and value matrices, is the dimension of the key, and the multi-head outputs are concatenated and fused through a linear transformation;

[0038] BiLSTM captures the context dependencies of sequence data through bidirectional temporal modeling, consisting of a forward LSTM and a backward LSTM in parallel, which process forward and backward sequence information respectively, and finally output comprehensive features through a merging strategy.

[0039] Preferably, the PLO algorithm is used to optimize four hyperparameters in the Transformer-BiLSTM model, namely the number of self-attention mechanism heads, learning rate, regularization coefficient, and the number of BiLSTM neurons, with the mean squared error on the validation set as the optimization objective.

[0040] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0041] By replacing the traditional finite element calculation with a surrogate model (PLO-RBFNN), the model can complete the calculation and output the predicted value in an extremely short time, meeting the real-time monitoring requirements while maintaining physical accuracy.

[0042] The two-stage training strategy forces the model to learn the physical laws of the seepage field, avoiding problems such as pure data-driven models violating physical laws. Under virtual extreme working conditions, the MAE, MAPE, MSE, and RMSE of the fusion model are significantly lower than those of pure data-driven models, and the multiple correlation coefficient R can still maintain a high value, and the prediction trend is highly consistent with the theoretical seepage process.

[0043] By innovatively introducing the dynamic hysteresis effect function into the mechanism model framework, the problem of inaccurate characterization of the hysteresis effect caused by the traditional deterministic model using a fixed time window method is solved, achieving a double improvement in accuracy and interpretability.

[0044] The PLO algorithm realizes the collaborative optimization of the surrogate model parameters and deep learning hyperparameters, reducing the dependence on manual parameter tuning; the RMSE fluctuation range of the optimized model in complex scenarios such as the test set and virtual extreme working conditions is significantly reduced.

[0045] The Transformer module learns the basic laws of the seepage field through physical constraints, and the BiLSTM module captures local features through data-driven methods, with significantly improved prediction accuracy compared to a single architecture, and improved prediction stability under extreme working conditions, realizing the deep coupling of physical laws and data features. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification, and are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0047] Figure 1 It is the main flow chart of the method of the present invention;

[0048] Figure 2 is a cross-sectional view of a dam according to an embodiment of the present invention;

[0049] Figure 3 are four measuring point process lines of the implementation scheme of the present invention;

[0050] Figure 4 is a comparison chart of theoretical values and predicted values of the proxy model on the test set of the implementation scheme of the present invention;

[0051] Figure 5 The figure shows a comparison of the prediction results of the four models under extreme virtual working conditions. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] In order to make the above-mentioned objects, features and effects of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] Embodiment 1: A method for predicting seepage pressure of earth-rock dam based on dual-driven mechanism and data, mainly including agent modeling guided by finite element model and deep learning migration of physical constraints, specifically the following steps:

[0055] Step 1: Eliminate gross errors in monitoring data and fill in missing values, and use wavelet denoising for preprocessing.

[0056] Step 2, finite element model-guided proxy modeling, that is, generating multi-condition seepage pressure data based on finite element simulation, optimizing the radial basis function neural network RBFNN through the Aurora optimization algorithm PLO to build an efficient proxy model; based on the prediction results of the proxy model, combined with the measured data for fitting, establish a seepage pressure prediction mechanism model based on the hysteresis effect function. The specific principles are as follows:

[0057] PLO is a meta-heuristic algorithm based on swarm intelligence that simulates the motion of high-energy particles in the aurora phenomenon. Specifically, the PLO algorithm achieves optimization through the following three main strategies:

[0058] Gyration motion: Simulate the gyration motion of particles in a magnetic field, mainly used for local development to help the algorithm find the optimal solution within a local range. Aurora ellipse walk: Simulate the random walk of particles within the aurora ellipse region, mainly used for global exploration to help the algorithm jump out of the local optimum and explore a broader search space. Particle collision: Simulate the collision process between particles to help the algorithm quickly jump out when it falls into the local optimum and enhance the global search ability of the algorithm.

[0059] The core steps of the PLO algorithm are as follows:

[0060] Initialization: Randomly generate the initial population X ∈ R N×dim , where N is the population size and dim is the dimension of the optimization variable.

[0061] Position update strategy: The individual position update is jointly driven by local search LS and global search GS:

[0062]

[0063] Among them: Local search term: , where it is the current iteration number.

[0064] Global search term: , X mean is the population mean, the Levy flight step size is generated by the Levy distribution, and R is a value between [0, 1].

[0065] Weight coefficient: , , where T is the maximum number of iterations, and w1 and w2 balance local development and global exploration respectively.

[0066] The Levy flight step size is generated by the following formula:

[0067]

[0068] Among them , and σ is calculated by the gamma function.

[0069] Boundary constraint and selection: Correct the out-of-bounds individuals and retain the better solutions through greedy selection.

[0070] The RBFNN surrogate model is a simplified computational model designed to obtain results close to those of a high-precision model at a relatively low computational cost. It establishes a non-linear mapping from parameters to responses by selecting sampling points in the parameter space and determining the system responses at these points. This model can be used to predict the system behavior under any parameters.

[0071] The RBFNN is a three - layer feed - forward network, and its core is to achieve high - dimensional space mapping through the linear combination of non - linear basis functions. The network structure includes an input layer, a hidden layer, and an output layer. The Gaussian kernel function is used as the activation function in the hidden layer, and its expression is:

[0072]

[0073] Among them, is the center of the j - th hidden - layer neuron, is the spread speed, which controls the width of the basis function. The network output is the weighted sum of the hidden - layer outputs:

[0074]

[0075] Among them, is the network output, is the weight, is the number of hidden - layer neurons.

[0076] To sum up, based on the finite - element calculation results, a water - level - seepage pressure data set is constructed. Combining with the PLO algorithm, with the sum of the mean square errors (MSE) of the test set as the fitness function and the optimization goal being to minimize the total MSE, the optimal hyperparameter combination of the RBFNN is determined, and a high - precision finite - element surrogate model is constructed accordingly.

[0077] Seepage pressure prediction mechanism model based on the hysteresis effect function: Since the surrogate model is constructed based on the calculation results of the finite - element model, and the calculation accuracy of the finite - element model is greatly affected by boundary conditions, model parameters, and working conditions. Even after parameter inversion and correction, its calculation results may still deviate from the measured values. In addition, the surrogate model only considers the influence of water level, while the dynamic changes of the actual seepage field of the earth - rock dam are also affected by aging factors. Therefore, in practical engineering applications, the output results of the surrogate model are usually combined with the aging component, and a deterministic equation is constructed by fitting with the measured values as the mechanism model for seepage prediction of the earth - rock dam. Its traditional equation form is generally:

[0078]

[0079] Among them, is the constant term, is the prediction result of the surrogate model, t is the time calculated from the starting point of the calculation, , to are the regression coefficients of each component.

[0080] It can be seen from the formula that in the traditional deterministic equation, the approximation of the actual unsteady seepage is achieved by superimposing the results under the average water level in the previous period, and through to characterize long-term effects such as soil creep and material aging. Although the previous average water level component method can simply and intuitively characterize the influence of unsteady seepage, this method has problems with insufficient accuracy. In addition, there are limitations such as rough selection of the time window and inability to reasonably reflect the cumulative hysteresis effect of seepage. Therefore, the present invention improves the traditional deterministic equation. Based on the assumption that the influence of reservoir water level on seepage change follows a normal distribution, the theoretical seepage calculation value (i.e., the result of the surrogate model) should also follow a normal distribution as the actual seepage result. On this basis, a hysteresis effect function is introduced to replace the original previous average component, and the dynamic hysteresis component and the aging factor are integrated. Finally, a multivariable regression equation is constructed, and the equation is as follows:

[0081]

[0082] where is the regression coefficient, is the prediction result of the surrogate model at time t, is the hysteresis effect function of the normal distribution, and its form is:

[0083]

[0084] where is the adjustment parameter, is the peak of the lag days, is the distribution width (influence days). It can be seen that and need to be obtained through optimization calculation. In the present invention, the PLO algorithm is used for optimization. For the monitoring data of the dam environmental variables, the reservoir water level generally has one measured value per day. Therefore, when calculating, it is necessary to convert the continuous integral into a discrete integral. According to the statistical principle, the integral interval takes 2 to 3 times of which can meet the calculation requirements.

[0085] In summary, the above method accurately captures the dynamic hysteresis effect of the seepage pressure of the earth-rock dam by coupling PLO optimization and the hysteresis effect function of the normal distribution, and combines the long-term influence of the aging factor, which can significantly improve the prediction accuracy and physical interpretability of the mechanism model.

[0086] Step 2 of this embodiment specifically includes the following steps:

[0087] Step 21, simulate the seepage field of the earth-rock dam based on finite element simulation to generate a seepage pressure data set under different working conditions; cover the possible boundary conditions of the actual project through parametric modeling to ensure data diversity.

[0088] Step 22: Define input variables and output variables, and construct an initial RBFNN model; use the sum of MSE of the test set as the fitness function to optimize the hyperparameters of the RBFNN; evaluate the accuracy of the surrogate model through error metrics to ensure that it can efficiently replace the finite element model.

[0089] Step 23: Introduce a lag effect function of normal distribution, optimize the lag parameters (lag peak x1, distribution width x2) through PLO to generate a lag component; combine the prediction results of the surrogate model, the lag component and the aging factor to establish a mechanism model; convert the integral of the continuous lag effect into discrete summation, and take the window size as 2-3 times of x2 to cover the main lag period.

[0090] Step 3: Deep learning migration of physical constraints, including two stages: The first stage of training, using the predicted values of the above mechanism model as labels, supervised training of the PLO-Transformer-BiLSTM network to learn the physical laws of the seepage field; The second stage of fine-tuning, freeze the physical feature extraction layer (positional encoding, self-attention layer) in the PLO-Transformer-BiLSTM network completed in the first stage of training, use the measured data to fine-tune the time series modeling layer (BiLSTM layer), and use the measured data to optimize the time series residual module to achieve approximation of the measured values. The specific principle is as follows:

[0091] The Transformer model is a deep learning architecture based on the self-attention mechanism. Its core advantage lies in its ability to effectively handle long-distance dependence relationships and significantly improve the model convergence speed through parallel training. To solve the complex time-dependent sequence prediction problem, the present invention proposes an improved architecture - the Transformer-BiLSTM model. By embedding the BiLSTM module after the self-attention layer in the Transformer framework, it realizes the collaborative modeling of global features and time series dynamics. The model aims to integrate the global parallel processing ability of the Transformer and the two-way time series modeling advantages of the BiLSTM to handle complex prediction tasks that require simultaneous grasp of global features and dynamic time series information. Specifically, its hierarchical feature extraction structure includes:

[0092] Transformer module: It includes a positional embedding layer and a self-attention layer. The positional embedding layer introduces positional encoding to retain time series information, and the encoding formula is:

[0093]

[0094] where pos is the sequence position, i is the dimension index, and d model is the feature dimension.

[0095] The self-attention layer captures global dependencies through multi-head attention, and the calculation formula is:

[0096]

[0097] Among them, Q, K, and V are the query, key, and value matrices, is the dimension of the key, and the multi-head outputs are fused through linear transformation after concatenation.

[0098] BiLSTM captures the context dependencies of sequence data through bidirectional temporal modeling. Its core consists of a forward LSTM and a backward LSTM in parallel, which process the forward and backward sequence information respectively, and finally output comprehensive features through a merging strategy (such as concatenation, summation, etc.). The specific formula is as follows:

[0099] Forward LSTM: The forward propagation process is the same as that of the standard LSTM, and long-term dependencies are modeled through the gating mechanism. The calculation formula is:

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106] Backward LSTM: The backward propagation starts from the end of the sequence, and the calculation formula is:

[0107]

[0108]

[0109]

[0110]

[0111]

[0112]

[0113] Bidirectional feature fusion: Concatenate the forward and backward hidden states as the final output:

[0114]

[0115] Among them, [ ; ] represents the vector concatenation operation, and the output dimension is the sum of the forward and backward hidden layer dimensions.

[0116] The complete Transformer-BiLSTM model concatenates the Transformer module and the BiLSTM module, and finally outputs the predicted value through the fully connected layer.

[0117] Two-stage transfer learning strategy: The first stage is the mechanism model fitting (theoretical label training), specifically:

[0118] Input features: Based on the optimization results of the lag coefficient of the mechanism model, select the water level, rainfall, and time effect factors within an appropriate time range.

[0119] Label data: The theoretical seepage pressure value predicted by the mechanism model.

[0120] Objective: Through pre-training the model with theoretical labels, guide the deep learning model to learn the physical laws of the seepage field, establish the mapping relationship between input features and theoretical seepage pressure values, and achieve the transfer of physical laws.

[0121] The second stage: Measured value fitting (real label fine-tuning), specifically:

[0122] Input features: The same as the first stage, but the label is replaced with the measured seepage pressure value.

[0123] Parameter freezing: Freeze the weights of the position embedding layer and the self-attention layer of the Transformer, and only fine-tune the parameters of the BiLSTM and subsequent fully connected layers.

[0124] Objective: On the basis of retaining the global features of the mechanism model, adapt to the local deviation of the measured data, improve the generalization ability of the model, and achieve the integration of physical constraints and data-driven.

[0125] The Transformer-BiLSTM model requires manual setting of parameters, which is usually determined based on engineering experience, with a large amount of work and being rather cumbersome. If the model has too many hyperparameters, it is difficult for humans to find the optimal parameter combination. Therefore, the PLO algorithm can be used to optimize four hyperparameters in the Transformer-BiLSTM model, namely the number of self-attention mechanism heads, learning rate, regularization coefficient, and the number of BiLSTM neurons, with the MSE result on the validation set as the optimization objective.

[0126] Step 3 of this embodiment specifically includes the following steps:

[0127] Step 31, Supervised training of the mechanism model in the first stage:

[0128] Dataset construction: The input features are the upstream water level, rainfall, and time effect factors, and the label data is the theoretical seepage pressure value predicted by the mechanism model. The dataset is divided into a training set, a validation set, and a test set according to the ratio of 70%, 10%, and 20%.

[0129] Transformer-BiLSTM network pre-training: build the model architecture, mainly Transformer module (position embedding layer encodes temporal information, self-attention layer captures global dependencies), BiLSTM module (bidirectional LSTM layer models temporal dynamics, outputs concatenated forward and backward hidden states) and fully connected layer (regression outputs osmosis prediction value);

[0130] PLO hyperparameter optimization: optimize the number of self-attention heads, learning rate, regularization coefficient, number of BiLSTM hidden layer units, and minimize the validation set MSE.

[0131] Step 32, second stage measured data fine-tuning:

[0132] Dataset adjustment: theoretical labels were replaced with measured osmotic pressure values, keeping input features unchanged;

[0133] Parameter freezing and fine-tuning: fix the position embedding layer and self-attention layer weights of the Transformer, retain the physical feature extraction capability, and only fine-tune the parameters of the BiLSTM layer and subsequent fully connected layers to adapt to the local deviation of the measured data;

[0134] Early stopping mechanism: monitor the validation set loss and terminate training if there is no improvement for 5 consecutive iterations to prevent overfitting.

[0135] Step 4: Model verification and deployment:

[0136] Performance evaluation: Calculate RMSE, MAE, R on the test set 2 , MAPE and other indicators to compare the training effects of theoretical labels and measured labels; by calculating model evaluation indicators on virtual working conditions generated by finite elements, the model extrapolation ability is compared and verified;

[0137] Visual analysis: draw the comparison curve between the predicted value and the measured value, the residual distribution diagram and the convergence curve to verify the generalization ability of the model;

[0138] Model solidification: Save the optimal parameters and normalized parameters, support online prediction, regular updates and engineering applications.

[0139] Example 2: The following is an example of the prediction of a total of 4 seepage pressure measurement points in two typical monitoring sections of the main and auxiliary dams of a reservoir to illustrate the implementation scheme of the present invention. The main and auxiliary dams of the reservoir are both homogeneous earth dams with a total length of 2514.00m. The main dam is 874.00m long and the maximum dam height is 22.50m; the auxiliary dam is 1640.00m long and the maximum dam height is 17.50m. The normal water storage level is 902.40m, the design flood level is 903.61m, and the verification flood level is 908.45m. Figure 2 As shown, the main dam P1 and P2 of the reservoir are selected. Figure 2 Upper) and auxiliary dams P3 and P4 (Figure 2 The monitoring data from 2019 to 2024 of four measuring points (below) are used as modeling data, including upstream and downstream water levels, rainfall, seepage pressure monitoring data and other information. After eliminating gross errors and filling missing values in the monitoring data, the data is further subjected to wavelet noise reduction processing, and the final process line is as follows: Figure 3 shown.

[0140] Multi-condition data generation: Based on the CAD drawings of the typical sections of the main and auxiliary dams of the reservoir and the inversion results of the material parameters of each partition, the finite element calculation model of the P1~P4 sections was established. According to the design parameters of the normal water level of the reservoir of 902.40m and the verified flood level of 908.45m, the calculation conditions were set at intervals of 0.2m in the elevation range of 900m to 908.6m, for a total of 45 conditions. The calculated seepage pressure values of the measuring points under each condition were obtained through finite element model calculation, and finally a water level-seepage pressure relationship data set was constructed.

[0141] Proxy model construction and optimization: Based on the constructed water level-osmotic pressure relationship data set, 30 training sets and 15 test sets were obtained by random division. RBFNN was initialized, and the initial expansion speed range was set to [1,10] and the number of neurons was set to [1,30]. The initial PLO population size was 20, the maximum number of iterations was 50, and the sum of the MSE of the test set was used as the fitness function. The total MSE was minimized to determine the optimal hyperparameter combination of RBFNN. After optimization iteration, the optimal expansion speed was 1.67 and the optimal number of neurons was 12. The comparison between the finite element theoretical value and the proxy model prediction value is shown in Figure 4 ,The accuracy index results are shown in Table 1. It can be seen that the PLO-RBF proxy model has extremely high prediction accuracy and can predict all working conditions between 900m and 908.6m.

[0142] Table 1. Accuracy index of proxy model

[0143] Establishment of the mechanism model considering the hysteresis effect: According to the prediction results of the PLO-RBF proxy model, the hysteresis effect function of the normal distribution is introduced to construct the mechanism model equation considering the hysteresis effect function. The optimal complex correlation coefficient R of the equation is used as the objective function. The hysteresis parameters x1 and x2 are optimized by PLO to generate the hysteresis component H. d (t), and the regression coefficient of the equation is solved by partial least squares method. The x1 and x2 results of the P1~P4 measurement points and the accuracy index of the regression equation are shown in Table 2. From the optimization results, it can be seen that P1 has significant hysteresis, and P2~P4 has no hysteresis, which is consistent with the law shown on the process line diagram. The mechanism model finally constructed has good prediction accuracy, and according to the hysteresis coefficient obtained by optimization, it can provide guidance and mechanism support for the data set construction in the next step of deep learning model training.

[0144] Table 2 Mechanistic model lag coefficient and accuracy index

[0145] Mechanistic model supervised training: Based on the established mechanistic model above, obtain the prediction results of measuring points P1 - P4 from 2019 to 2024. Based on this result, combined with water level factors, rainfall factors, and aging factors, construct the training dataset for the Transformer - BiLSTM deep learning model.

[0146] Analysis of the mechanistic model shows that among the four measuring points, except for P1, the other three measuring points (P2 - P4) have no significant lag. Specifically, the lag peak of measuring point P1 is 11 days, and the distribution width is 15 days. Therefore, when constructing the model input of measuring point P1, the rolling time window method is adopted, and the time window is set to 30 days to fully cover the influence of the lag effect. For measuring points P2 - P4, due to their insignificant lag, to simplify the model, only the factors of one week are selected as the input, and the time window is correspondingly set to 7 days. Based on the above method, complete the dataset construction of measuring points P1 - P4.

[0147] According to the ratio of 70%, 10%, and 20%, divide the dataset into a training set, a validation set, and a test set. The training set is used for model fitting and parameter iteration optimization; the validation set is used to evaluate model accuracy, adjust hyperparameters, and monitor overfitting during the training process; the test set is completely invisible during the model training and tuning stages and is only used to finally evaluate the prediction accuracy, extrapolation ability, and generalization performance of the model.

[0148] Based on the established Transformer - BiLSTM model framework, use the PLO algorithm to optimize the key hyperparameters of the model - the number of self - attention heads, learning rate, regularization coefficient, and the number of BiLSTM hidden units. The range of the number of self - attention heads is [2, 8], the range of the learning rate is [1e - 5, 1e - 3], the range of the regularization coefficient is [1e - 5, 1e - 3], and the range of the number of BiLSTM hidden units is [16, 128]. Through mechanistic model supervised training and learning, the aim is to achieve an accurate fit of the theoretical seepage pressure values of the four measuring points, thereby completing the transfer learning of the physical laws of the seepage field. The hyperparameter optimization results are shown in Table 3, and the performance comparison between the original model and the optimized model on the test set is shown in Table 4.

[0149] Table 3 Hyperparameter optimization results

[0150] Table 4 Performance comparison between the original model and the optimized model on the test set

[0151] The training results show that the Transformer-BiLSTM model after PLO parameter optimization exhibits excellent prediction performance in the first stage. The model not only achieves a high-precision fitting of the theoretical seepage pressure value, but more importantly, successfully transfers the physical laws of the seepage field characterized by the mechanism model into the deep learning framework.

[0152] Fine-tuning with measured data: Based on the model obtained from the first-stage training, the second-stage training is carried out. Although the current model can already accurately represent the laws of the theoretical seepage field, there are still differences between its prediction results and the measured seepage pressure values. Therefore, the core goal of the second-stage training is: while maintaining the global characteristics of the mechanism model, adjust the model parameters to better adapt to the local deviations of the measured data, so as to improve the generalization performance of the model and finally achieve the organic integration of physical mechanism constraints and data-driven methods.

[0153] Based on the constructed dataset, replace the labeled data from the theoretical seepage pressure value with the measured seepage pressure value, while keeping all input features unchanged. In terms of the model architecture, freeze the parameters of the Transformer part in the Transformer-BiLSTM model (including the positional embedding layer and the multi-head self-attention layer), and only fine-tune the BiLSTM layer and the subsequent fully connected layers. This design not only retains the ability of the model to extract physical features but also better adapts to the characteristics of the measured data.

[0154] An early stopping mechanism is set during the training process: terminate the training when the validation set loss does not decrease for 5 consecutive iterations to avoid overfitting. Adopt the hyperparameter configuration determined by the first-stage optimization, and start the second-stage training after loading the pre-trained model weights. The performance evaluation results of the final model on the test set are shown in Table 5.

[0155] Table 5 Performance evaluation indicators of the final model on the test set

[0156] Test set verification: To verify the superiority of the fusion model proposed in this paper, five comparison models are selected for performance comparison: PLO-Transformer-BiLSTM (pure data-driven), Transformer-BiLSTM (pure data-driven), BiLSTM (pure data-driven), statistical model, and mechanism model. The performance evaluation indicators (MAE, MAPE, MSE, RMSE, and multiple correlation coefficient R) of the six models (including the fusion model in this paper) are shown in detail in Table 6.

[0157] Table 6 Performance evaluation indicators of six models on the test set

[0158] From the evaluation results of the test set, the mechanism-data fusion model proposed in this paper shows significant comprehensive advantages in the seepage pressure prediction task, especially exceeding the pure data-driven models and traditional methods in terms of physical consistency and generalization ability.

[0159] From the perspectives of accuracy and stability, compared with the pure data-driven model, for the measuring points P1, P3, and P4, the MAE, RMSE, and MSE of the fusion model are significantly lower than those of other models, indicating that its prediction error is smaller. Although at the measuring point P2, the MAE of PLO-Transformer-BiLSTM is slightly better than that of the fusion model, its multiple correlation coefficient R is still close to that of PLO-Transformer-BiLSTM, indicating that while maintaining high accuracy, the fusion model does not sacrifice physical rationality. The pure data-driven model shows large fluctuations at some measuring points, while the fusion model performs stably at all measuring points, verifying its anti-overfitting ability. Compared with the statistical model and the mechanism model, the MAE of the fusion model is reduced by 52.1% and 54.3% on average, and the RMSE is reduced by 46.8% and 48.7% on average, with a significant improvement in accuracy. The multiple correlation coefficient R is higher than that of the traditional model at all measuring points, further confirming its stronger interpretability.

[0160] From the perspective of physical interpretability, the fusion model proposed in this invention undergoes two-stage training. First, it learns the theoretical laws of the mechanism model and then fine-tunes with the measured data. This design enables it to retain the physical mechanism of the seepage field while correcting the complex factors not considered in the theoretical model. Although PLO-Transformer-BiLSTM has slightly higher accuracy at some measuring points, it completely relies on data fitting and lacks physical constraints, which may lead to prediction deviations under extreme conditions. The fusion model significantly improves the extrapolation reliability through the transfer of physical laws.

[0161] From the perspectives of generalization ability and engineering applicability, the performance fluctuation range of the fusion model at the 4 measuring points is the smallest, indicating that it has a stronger adaptability to spatial heterogeneity and is more suitable for predicting the complex and variable seepage field in actual engineering. The MAPE of the fusion model is generally lower than that of other models, indicating that it is less sensitive to small errors and the prediction results are closer to the actual observed values. The fusion model proposed in this invention performs consistently under multiple measuring points and multiple working conditions, with stronger engineering applicability.

[0162] Verification of extreme virtual working conditions: In order to verify the extrapolation ability of the model, an extreme virtual working condition test scheme is designed: taking the end point of the test set time series as the starting day, the water level rises linearly to the verified flood level (908.45m) in the first 30 days, and then drops linearly to the normal water level (901.6m) in the next 30 days. Random noise is superimposed throughout the process to simulate actual fluctuations, and the theoretical value of the seepage pressure at the corresponding measuring point is obtained through finite element calculation. Compared with the actual water level range (900.5-903.2m) used in modeling, the water level change range of this virtual working condition (901.6-908.45m) significantly exceeds the distribution of training data, which can effectively verify the extrapolation performance of the model under unknown working conditions. In order to systematically evaluate the performance of the model, three pure data-driven models, PLO-Transformer-BiLSTM, Transformer-BiLSTM, and BiLSTM, are selected for comparative analysis with the fusion model proposed in this invention. The comparison of prediction results is shown in Figure 5 ,The performance evaluation indicators are detailed in Table 7.

[0163] Table 7 Performance evaluation indicators of four models under extreme virtual working conditions

[0164] Depend on Figure 5 It can be seen from the test results in Table 7 that the mechanism-data fusion model proposed in this paper shows significant advantages under extreme virtual working conditions, and its extrapolation ability and physical consistency are all superior to the pure data-driven model. Specifically, the fusion model shows excellent performance in both extrapolation accuracy and stability, and the MAE, RMSE and MSE indicators of all measuring points are significantly better than those of the pure data-driven model. It is particularly noteworthy that the fusion model can still maintain a high complex correlation coefficient R value (all higher than 0.85) under extreme working conditions, while the R value of the pure data-driven model is generally lower than 0.6. This difference fully demonstrates that the prediction trend of the fusion model is highly consistent with the theoretical seepage physical process. In addition, the pure data-driven model shows obvious performance fluctuations under extreme working conditions, while the error growth of the fusion model is relatively small, which further verifies its strong robustness to unknown working conditions.

[0165] In terms of physical consistency, the fusion model migrates the physical laws trained in the first stage so that its prediction results strictly follow the theoretical characteristics of the seepage field. By observing the prediction curves in the stage of rapid water level changes, it can be found that the prediction results of the fusion model not only have a smooth transition, but also are highly consistent with the theoretical values; in contrast, the pure data-driven model shows systematic deviations under water level conditions beyond the distribution range because it completely relies on the statistical laws of the training data. This difference highlights the importance of physical constraints. It is by introducing physical mechanisms that the fusion model effectively avoids the failure problem of the pure data-driven model under extreme conditions. This physical consistency ensures the interpretability and reliability of the model's prediction results.

[0166] As can be seen from the above model verification and analysis, the two-stage training fusion model proposed by the present invention realizes the complementary advantages of mechanism and data. In the first stage of training, the model is forced to learn the deterministic physical laws of the seepage field, establishing a reliable framework foundation for extrapolation prediction; in the second stage of training, the local non-linear effects not covered by the theoretical model are corrected through learning from measured data, thereby improving the model's adaptability to actual noise. This synergistic effect enables the model to not only maintain high-precision prediction within the range of existing measured data, but also make reliable predictions that conform to physical laws for unknown extreme working conditions. This fusion modeling method provides an effective way to solve the extrapolation prediction problem in complex engineering problems.

[0167] Online monitoring system integration: The trained PLO-Transformer-BiLSTM model is solidified into a prediction engine module and deployed to the dam safety monitoring cloud platform, which is connected in real time with the automated acquisition system. It receives environmental data such as reservoir water level and rainfall in real time, and outputs the predicted value of seepage pressure and the confidence interval. It supports dynamic adjustment of the prediction frequency, can set the prediction period according to actual needs, and at the same time provides a visualization warning function. When the prediction residual exceeds the set threshold, it automatically triggers a hierarchical alarm mechanism to guide engineering personnel to check for abnormalities in time.

[0168] Model adaptive update strategy: Newly added monitoring data is automatically collected regularly (such as monthly), and incremental training is performed on the BiLSTM time series modeling layer to optimize the model's adaptability to long-term effects such as material aging; at the same time, an intelligent trigger mechanism is introduced. When a drastic fluctuation in reservoir water level or an extreme rainfall event is detected, the online fine-tuning process is immediately started to correct the prediction deviation in real time. The update process uses a rolling time window technique, only retaining the most recent N periods of data to control the computational load, and ensuring that the model still meets the accuracy requirements through validation set evaluation. This mechanism effectively solves the problem of performance degradation of traditional static models during long-term service.

[0169] Example 3:

[0170] The computer-readable storage medium of this embodiment stores a computer program, which when executed by a processor implements the steps in a method for predicting seepage pressure of an earth-rock dam based on dual drive of mechanism and data in Example 1.

[0171] The computer-readable storage medium of this embodiment can be the internal storage unit of the terminal, such as the hard disk or memory of the terminal; the computer-readable storage medium of this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card, a secure digital card, a flash card, etc. equipped on the terminal; further, the computer-readable storage medium can also include both the internal storage unit and the external storage device of the terminal.

[0172] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store the data that has been output or will be output.

[0173] Embodiment 4:

[0174] The computer device of this embodiment includes a processor, a memory, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in a method for predicting seepage pressure of an earth-rock dam based on dual driving of mechanism and data in Embodiment 1.

[0175] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays, or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.; the memory can include a read-only memory and a random access memory, and provide instructions and data to the processor. A part of the memory can also include a non-volatile random access memory. For example, the memory can also store information about the device type.

[0176] Those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general-purpose hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0177] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A seepage pressure prediction method for earth-rock dams based on the dual drive of mechanism and data, characterized in that, It includes the following steps: Step 1: Eliminate gross errors from the monitoring data and fill in the missing values, and perform preprocessing using wavelet denoising; Step 2: Generate a water level-seepage pressure data set based on finite element simulation, and optimize the radial basis function neural network (RBFNN) using the aurora optimization algorithm (PLO) to construct a surrogate model; Based on the prediction results of the surrogate model, combine with the measured data for fitting to establish a seepage pressure prediction mechanism model based on the hysteresis effect function; Step 3: Deep learning transfer with physical constraints, including two stages: The first stage of training is to use the predicted values of the mechanism model in Step 2 as labels to supervise the training of the PLO-Transformer-BiLSTM network to learn the physical laws of the seepage field; The second stage of fine-tuning is to freeze the physical feature extraction layer in the PLO-Transformer-BiLSTM network completed in the first stage of training, fine-tune the BiLSTM in the time series modeling layer using the measured data, and optimize the time series residual module using the measured data; Step 4: Calculate the regression evaluation index through the data set to compare the training effects of the theoretical label and the measured label; Calculate the model evaluation index on the virtual working conditions generated by the finite element to compare and verify the extrapolation ability of the model; Save the optimal parameters and normalization parameters, and embed them into the safety monitoring system to implement seepage pressure prediction and warning.

2. The seepage pressure prediction method for an earth-rock dam based on dual drive of mechanism and data according to claim 1, characterized in that Step 2 specifically includes the following steps: Step 21: Simulate the seepage field of the earth-rock dam based on finite element simulation to generate a seepage pressure data set under different working conditions; Cover the possible boundary conditions of the actual project through parametric modeling to ensure data diversity; Step 22: Define the input variables and output variables, and construct an initial RBFNN surrogate model; Use the sum of the mean square errors of the data set as the fitness function to optimize the hyperparameters of the RBFNN; Evaluate the accuracy of the surrogate model through error indicators to enable it to efficiently replace the finite element model; Step 23: Introduce the hysteresis effect function of the normal distribution, optimize the hysteresis parameter through PLO to generate the hysteresis component; Combine the prediction results of the surrogate model, the hysteresis component and the aging factor to establish a seepage pressure prediction mechanism model based on the hysteresis effect function.

3. A seepage pressure prediction method for earth-rock dams based on dual driving of mechanism and data according to claim 2, characterized in that, The RBFNN surrogate model is a three-layer feedforward network, and its network structure includes an input layer, a hidden layer and an output layer. The hidden layer uses a Gaussian kernel function as the activation function, and its input expression is: ； Among them, is the input, is the center of the j-th hidden layer neuron, is the expansion speed, which controls the width of the basis function; the network output is the weighted sum of the hidden layer outputs: ； Among them, is the network output, is the weight, is the number of neurons in the hidden layer.

4. A seepage pressure prediction method for earth-rock dams based on dual drive of mechanism and data according to claim 3, characterized in that The seepage pressure prediction mechanism model based on the hysteresis effect function is specifically: ； Among them, is the constant term, t is the time counted from the calculation starting point, and to are the regression coefficients of each component; is the prediction result of the surrogate model at time t, is the lag effect function of the normal distribution, and its form is: ； Among them, is the adjustment parameter, is the peak value of the lag days, is the influence days, and is obtained by the PLO optimization algorithm.

5. A seepage pressure prediction method for earth-rock dams based on dual-driving of mechanism and data according to claim 4, characterized in that, When calculating, the percolation pressure prediction mechanism model based on the hysteresis effect function converts the continuous integral into a discrete integral, and the window size is taken as 2 to 3 times that to cover the main hysteresis period.

6. The seepage pressure prediction method for earth-rock dams based on the dual drive of mechanism and data according to claim 1, characterized in that Step 3 specifically includes the following steps: Step 31: The specific supervision training of the mechanism model in the first stage is as follows: The input features are the upstream water level, rainfall and aging factor, and the label data is the theoretical seepage pressure value predicted by the mechanism model. According to the ratio of 70%, 10% and 20%, the data set is divided into a training set, a validation set and a test set; Build a Transformer-BiLSTM model, including a Transformer module, a BiLSTM module and a fully connected layer; Pretrain the model through the theoretical label to guide the Transformer-BiLSTM model to learn the physical laws of the seepage field, establish the mapping relationship between the input features and the theoretical seepage pressure value, and realize the transfer of physical laws; Step 32. The fine-tuning of the actual measured data in the second stage is specifically as follows: Replace the theoretical label with the actually measured seepage pressure value, while keeping the input features unchanged; fix the weights of the position embedding layer and the self-attention layer of the Transformer module, and only fine-tune the parameters of the BiLSTM module and the subsequent fully connected layers. On the basis of retaining the global features of the mechanism model, adapt to the local deviation of the actual measured data, improve the generalization ability of the model, and realize the integration of physical constraints and data-driven.

7. A seepage pressure prediction method for earth-rock dams based on dual drive of mechanism and data according to claim 6, characterized in that, The Transformer-BiLSTM model connects the Transformer module and the BiLSTM module in series, and finally outputs the predicted value through the fully connected layer. The Transformer module includes a position embedding layer and a self-attention layer. The position embedding layer introduces position encoding to retain the timing information, and the encoding formula is: ； where pos is the sequence position, i is the dimension index, and d model is the feature dimension; The self-attention layer captures the global dependency relationship through multi-head attention, and the calculation formula is: ； where Q, K, and V are query, key, and value matrices, is the dimension of the key, and the multi-head outputs are fused through a linear transformation after concatenation; BiLSTM captures the context dependency relationship of sequence data through bidirectional timing modeling, and is composed of a forward LSTM and a backward LSTM in parallel, which process the forward and backward sequence information respectively, and finally output the comprehensive features through the merging strategy.

8. A seepage pressure prediction method for earth-rock dams based on dual drive of mechanism and data according to claim 6, characterized in that The PLO algorithm is used to optimize four hyperparameters in the Transformer-BiLSTM model, namely the number of heads of the self-attention mechanism, the learning rate, the regularization coefficient, and the number of BiLSTM neurons, with the mean square error on the validation set as the optimization target.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in a method for predicting seepage pressure of an earth-rock dam based on mechanism and data dual-driving as described in any one of claims 1-8.

10. A computer device, comprising a processor, a memory, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in a method for predicting seepage pressure of an earth-rock dam based on mechanism and data dual-driving as described in any one of claims 1-8.

Images