A method for predicting seepage pressure in concrete dams based on scaled preprocessing and dual-model fusion
By using EEMD decomposition and dual-model fusion, high, medium and low frequency signals are screened out. Ridge regression and Transformer-BiLSTM models are trained in parallel to achieve high-precision and strong generalization prediction of seepage pressure in concrete dams. It has engineering physical interpretation capabilities and solves the problems of multi-scale signal coupling interference and anomaly tracing in seepage pressure prediction in existing technologies.
Patent Information
- Application Number
- CN202611257453.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-19
- Publication Date
- 2026-11-10
AI Technical Summary
Existing concrete dam seepage pressure prediction technologies struggle to simultaneously achieve high accuracy, strong generalization ability, and engineering interpretability. They are unable to effectively capture nonlinear short-term seepage pressure fluctuations, suffer from severe multi-scale signal coupling interference, and lack the ability to trace the source of seepage anomalies.
Adaptive noise-assisted decomposition is performed using ensemble empirical mode decomposition (EEMD) to screen out high, medium and low frequency signals. Ridge regression and Transformer-BiLSTM dual-model parallel training are combined, and the optimal model is selected by normalized comprehensive selection index (COI). Seepage pressure prediction is reconstructed using a multi-layer fully connected network to realize the nonlinear coupling relationship mining of multi-scale signals.
It achieves high-precision and strong generalization prediction of seepage pressure in concrete dams, has engineering physics interpretation capabilities, can accurately capture the causes of seepage anomalies, adapts to the characteristics of different seepage areas, reduces information loss in the reconstruction process, and improves the overall prediction accuracy.
Smart Images

Figure CN123069887A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of concrete dam safety monitoring and seepage pressure prediction technology, specifically involving a concrete dam seepage pressure prediction method based on multi-scale preprocessing and dual-model fusion. Background Technology
[0002] Concrete dams are core water conservancy projects for flood control, water supply, power generation, and ecological regulation. The seepage status of the dam body and foundation directly determines the structural safety during service. Seepage pressure monitoring data is a key indicator for judging defects such as curtain aging, rock erosion, contact scouring, and seepage instability. Accurate prediction of seepage pressure sequence can detect abnormal seepage trends in advance and avoid major dam failures and leakage safety accidents.
[0003] Existing concrete dam seepage pressure prediction technologies are mainly divided into three categories:
[0004] The first type is the traditional hydraulic statistical model, which uses the linear superposition of water pressure, rainfall, temperature and time to construct the fitting formula. It can only be adapted to steady-state simple seepage conditions and cannot capture the nonlinear short-term seepage pressure fluctuations caused by sudden increases in rainfall and sudden drops in water level. It has large prediction errors and poor generalization ability.
[0005] The second category consists of single deep learning time series models, including networks such as LSTM, BiLSTM, and Transformer, which directly input the original seepage pressure time series for global fitting. However, the original monitoring sequences are simultaneously mixed with instrument noise, short-term rainfall disturbances, seasonal water level cycles, and multi-scale signals from decades of long-term curtain evolution. These multi-scale features are deeply coupled, causing mutual interference during network fitting, which easily leads to overfitting or underfitting. At the same time, the model is a "black box" output, only providing predicted values and unable to distinguish whether the anomaly originates from short-term rainfall, seasonal water levels, or long-term curtain degradation, resulting in severely insufficient engineering interpretability.
[0006] The third type is the decomposition ensemble prediction model, which uses modal decomposition algorithms such as EEMD and VMD to split the original time series. Each IMF component after decomposition is then predicted using the same machine learning / deep learning model, and finally, the predicted values of each component are linearly added together to obtain the global seepage pressure. This approach has the following drawbacks:
[0007] 1. The difference in evolution patterns between high and low frequency components is not distinguished. The advantages of linear models in adapting to long-term linear trends and deep learning in adapting to nonlinear fluctuations cannot be brought into play. All components use a single model, resulting in distorted fitting of low-frequency trends and loss of details in high-frequency fluctuations.
[0008] 2. The lack of a standardized component screening mechanism leads to the direct participation of the decomposition of meaningless noise IMFs in the modeling process, introducing additional prediction errors.
[0009] 3. The reconstruction method only uses simple linear superposition, ignoring the nonlinear coupling between seepage signals of different frequency bands and losing multi-scale interactive information;
[0010] 4. Relying on fixed frequency band matching models based on human experience, without a quantitative selection mechanism, the same set of models cannot be adapted to different seepage areas such as curtain zones, drainage corridors, and dam toes, thus limiting generalization performance.
[0011] Among the existing improvement schemes, a few studies adopt the approach of "manually dividing high and low frequencies and pairing them with linear / neural networks respectively," but there are still obvious shortcomings: the model's merits are judged solely by a single error index, without comprehensively evaluating performance from multiple dimensions such as goodness of fit, absolute error, and relative error; there is no normalized weighted comprehensive selection system, making it impossible to quantitatively and automatically match the optimal model for each component; at the same time, all models are reconstructed using linear superposition, without utilizing neural networks to explore the nonlinear coupling relationships between components; the driving factors only use daily environmental monitoring values, without considering the multi-period hysteresis response of seepage pressure to water level and rainfall, resulting in low matching degree with the actual seepage mechanics mechanism.
[0012] In summary, existing technologies cannot simultaneously meet the three major requirements of high-precision prediction, strong generalization ability, and engineering interpretability. This invention proposes a completely new prediction framework to address all of the above-mentioned shortcomings. Summary of the Invention
[0013] The purpose of this invention is to overcome the shortcomings of existing technologies by utilizing the adaptive noise-assisted decomposition capability of ensemble empirical mode decomposition (EEMD) to decouple the original time series data at multiple scales. A multi-scale preprocessing system is constructed to separate high-frequency random noise, mid-frequency environmental load response, and low-frequency long-term trend components. The normalized comprehensive selection index (COI) is used to achieve dynamic optimization of the ridge regression and Transformer-BiLSTM dual models. The low-frequency trend signal is fitted and predicted by leveraging the L2 regularized linear training advantage of ridge regression. At the same time, the high-frequency nonlinear fluctuation signal is trained by utilizing the long-range correlation capture and bidirectional time series mining capabilities of the Transformer-BiLSTM network. Finally, the optimal results of each component are nonlinearly integrated and reconstructed through a multi-element fusion reconstruction layer. This establishes a concrete dam seepage pressure prediction model based on multi-scale preprocessing and dual-model fusion, which overcomes mode aliasing and multi-scale coupling interference in the monitoring sequence, achieves high-precision and strong generalization prediction of dam seepage pressure, and endows the deep learning model with engineering physics interpretation capabilities.
[0014] To achieve the above objectives, the present invention is implemented through the following solution:
[0015] A method for predicting seepage pressure in concrete dams based on scaled preprocessing and dual-model fusion includes the following steps:
[0016] Step S1: Perform scale-adaptive decoupling on the original seepage pressure monitoring time series of the concrete dam, and after filtering out invalid fluctuation signals, obtain multiple effective scale components.
[0017] Step S2: For each type of effective scale component, two differentiated prediction models are trained and calculated in parallel, and the optimal prediction output of a single component is automatically selected based on the normalized comprehensive selection index.
[0018] Step S3: Input the optimal prediction results of all components into the nonlinear fusion network, adaptively fit the coupling relationship of multi-scale components, and reconstruct the complete global seepage pressure prediction time series.
[0019] Step S4: Use multi-dimensional quantitative indicators to verify the model prediction effect, and trace the cause of seepage anomalies in layers according to the prediction deviation of each scale component.
[0020] Further optimization, step S1 specifically includes:
[0021] Step S101: Preprocessing of the original time series and construction of seepage driving factors: Outliers in the time series are removed using the 3σ criterion, and missing monitoring data are supplemented by linear interpolation; based on the statistical model of seepage pressure in concrete dams, four types of lagged environmental driving factors—water pressure, rainfall, temperature, and time aging—are constructed. The expression of the seepage pressure statistical model is as follows:
[0022]
[0023] In the formula: H is the seepage pressure measured on the monitoring day; z is a constant term in the model; These are water pressure component, rainfall component, temperature component, and time-dependent component, respectively; t is the cumulative number of days from the start date of monitoring. This represents the cumulative number of days since the initial monitoring date; u represents the lag component numbers for water pressure, rainfall, and temperature. The water pressure component uses a multi-period average upstream and downstream head difference factor to characterize the lag response of seepage pressure to water level changes; the rainfall component uses a multi-period average rainfall factor to characterize the lag effect of rainfall infiltration; the temperature component uses a double harmonic function and a long-term climate warming drift term to characterize the periodic influence of temperature; and the time-dependent component uses a combination of linear and logarithmic terms to characterize the irreversible evolution of the long-term seepage field.
[0024] Step S102, Ensemble Empirical Mode Decomposition: To the original osmotic pressure sequence Zero-mean white noise was added multiple times. m independent noisy time series were obtained. Its expression is:
[0025]
[0026] For each noisy time series Perform empirical mode decomposition on each component to obtain the f-th IMF component. and the residuals of the sth sequence decomposition , cumulative The results of the group decomposition are obtained, and the ensemble mean is calculated for the IMF components and residual terms of the same order in the m groups to obtain the final population mean. The overall average of the residuals To suppress mode aliasing, its expression is:
[0027]
[0028]
[0029] In the formula: s is the noise superposition group number; f is the IMF order number obtained by decomposing a single noisy sequence; m is the total number of EEMD noise superpositions.
[0030] Step S103, Joint Screening of Component Validity: Component validity is jointly determined by four thresholds: effective length, energy, variance, and Pearson correlation coefficient with the original sequence. Only components that simultaneously meet all four thresholds are considered valid components. The determination equations are as follows:
[0031]
[0032] In the formula: The effective data length of the l-th order IMF component. This is the preset minimum effective length threshold; The energy of the l-th order IMF component; This represents the total number of IMF component data. The variance of the l-th order IMF component; The l-th order IMF component and the original osmotic pressure sequence The Pearson correlation coefficient; l is the global order specific number of the IMF component; k is the internal sampling point index of a single IMF sequence.
[0033] Step S104, Component frequency band division:
[0034] High-frequency components: with a period of less than 7 days, corresponding to short-term random monitoring noise and instantaneous rainfall disturbances;
[0035] Mid-frequency component: period of 7~30 days, corresponding to the periodic response of environmental loads such as reservoir water level fluctuation and temperature change;
[0036] Low-frequency component: period greater than 30 days, corresponding to the long-term evolution trend of dam foundation curtain, rock mass permeability and seepage channels.
[0037] Further optimization, step S2 specifically includes:
[0038] Step S201, Parallel training of dual models: Simultaneously feed the same effective components into Ridge Regression and Transformer-BiLSTM for prediction;
[0039] 1) The ridge regression model adds an L2 regularization penalty term to the loss function of classic linear regression. This parameter shrinkage mechanism suppresses overfitting caused by feature multicollinearity, and is used to fit the linear long-term evolution trend of low-frequency components. Its expression is:
[0040]
[0041] In the formula: , used to balance the residual fitting accuracy and coefficient constraint strength, is the L2 regularization penalty coefficient; β is the vector of all feature regression coefficients; y is the regression coefficient of the g-th feature; n is the total number of training samples; τ X is the measured osmotic pressure label for the τth sample; τ Let τ be the input driving feature vector of the τth group; p is the total dimension of the input features.
[0042] 2) The structure of the Transformer-BiLSTM model is as follows: the front end is a Transformer encoder, which completes the temporal position information through sine and cosine position encoding, and uses a multi-head self-attention mechanism to extract long-range global correlation features of the sequence. Its expression is:
[0043]
[0044]
[0045] In the formula: For query, key, value matrix; For the number of attention heads; The learnable parameter matrix; The dimension is the key vector. To compensate for the lack of temporal information, sine and cosine functions are introduced to generate positional codes:
[0046]
[0047]
[0048] In the formula: The position of the input parameter factor in the sequence; The index of the input factor in the sequence; The dimension of the input factor is encoded using a sine function at even positions and a cosine function at odd positions.
[0049] The backend is a BiLSTM network, consisting of bidirectionally stacked forward and backward LSTM units, which mines the bidirectional temporal dependency features of the sequence. Finally, the nonlinear prediction results of the output components are passed through a fully connected layer to fit the complex nonlinear fluctuations of the mid-to-high frequency components. The state update formula is:
[0050]
[0051]
[0052]
[0053] In the formula: The states are hidden in both forward and backward directions; The concatenated bidirectional features fully preserve sequence context information. Compared to LSTM, BiLSTM utilizes both historical and future information to improve its ability to fit complex fluctuations.
[0054] Step S202, Index Calculation: Simultaneously calculate R 2 The four evaluation indicators are RMSE, MAE, and MAPE.
[0055] Step S203, Construction of the Comprehensive Selection Index (COI), specifically includes:
[0056] Step S2031, Index Normalization: Perform inverse normalization on the three types of error indices: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The closer the value is to 1, the better the performance. The expression is as follows:
[0057]
[0058] In the formula, The measured values of the indicators of the model to be evaluated are: The maximum and minimum values of the model participating in component selection are among all measured values under this error index;
[0059] For the coefficient of determination R 2 Perform forward normalization; the closer the value is to 1, the better the fit. The expression is:
[0060]
[0061] In the formula, R of the model to be evaluated2 Measured value For models participating in component selection in R 2 The maximum and minimum values among all measured values under the indicator.
[0062] Step S2032, Weighted Comprehensive Calculation: Combining the engineering characteristics and index representation of dam seepage, differentiated weights are assigned to the four normalized indicators, with the coefficients summed to 1. The weighted summation yields the COI comprehensive value, expressed as:
[0063]
[0064] In the formula, A higher COI value indicates better overall model performance; To determine the weight of the coefficients; The root mean square error weight; Weighted by the average absolute error; The weight is the average absolute percentage error weight; the weight can be customized and adjusted according to the engineering characteristics of different dam seepage areas.
[0065] Step S204: Optimal selection: Compare the COI values of the lowerridge regression and Transformer-BiLSTM for the same component, and select the model output with the higher COI value as the optimal prediction result for that component. If the COI values of the two models are equal, select the model with higher computational efficiency.
[0066] Further optimization, step S3 includes the following steps:
[0067] Step S301: Align and splice the optimal prediction sequences of each component according to a unified time step to generate a multi-dimensional component feature matrix; Time step: the monitoring time interval corresponding to a single sampling.
[0068] Step S302: Input the multidimensional feature matrix into the multilayer fully connected fusion network, and use the original measured seepage pressure time series as the supervised training label;
[0069] Step S303: Adaptively learn the nonlinear coupling correlation between high-frequency, mid-frequency, and low-frequency components and the contribution weight of each component based on the network backpropagation mechanism; the contribution weight represents the proportion of the corresponding component's influence on the global seepage pressure prediction result.
[0070] Step S304: Output a global osmotic pressure prediction time series that perfectly matches the time steps and total length of the original monitoring sequence.
[0071] Further optimization, step S4 includes the following steps:
[0072] Step S401: Use the coefficient of determination R 2The overall prediction bias of the model is quantified by four indicators: root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), etc.
[0073] Step S402: Plot a comparison curve between the measured seepage pressure time series and the model prediction time series to intuitively complete the visualization verification of the model's generalization performance; the generalization performance represents the stability of the model's prediction on new monitoring data that has not participated in the training.
[0074] Step S403: Based on the prediction deviation of different frequency band components, locate the causes of anomalies in layers: high frequency component deviation corresponds to instantaneous rainfall disturbance or monitoring instrument noise, medium frequency component deviation corresponds to seasonal periodic fluctuations in reservoir water level and temperature, and low frequency component deviation corresponds to dam foundation curtain and long-term seepage deterioration of rock mass.
[0075] Further optimization is needed; the calculation formulas for the four quantitative evaluation indicators in step S401 are as follows:
[0076]
[0077]
[0078]
[0079]
[0080] Where: n is the total number of samples; For the first The true values of the measured seepage pressure were obtained. For the first Group model seepage pressure prediction values; This represents the average value of all measured osmotic pressure samples.
[0081] Further optimizations include the following dataset partitioning parameters for the model training phase: The first 75% of the time series is used as the training set to iteratively update the model's internal parameters; the remaining 25% is used as the test set to verify the model's generalization ability; and 0.5-day single-step rolling prediction: Each time a historical time series is input, only the seepage pressure value for the next sampling day is predicted.
[0082] A concrete dam seepage safety early warning device includes a memory and a processor; the memory stores a computer-executable program, and when the processor runs the computer-executable program, it executes the above-mentioned concrete dam seepage pressure prediction method based on multi-scale preprocessing and dual-model fusion; the memory stores monitoring time series, model fixed parameters, and executable program; the processor executes the complete seepage pressure prediction calculation process.
[0083] Compared with the prior art, the present invention has the following beneficial effects:
[0084] 1. This invention uses the EEMD algorithm with auxiliary noise to decompose the original seepage pressure sequence, and uses four thresholds to screen out noise components without physical meaning. Based on the 7-day and 30-day engineering cycle thresholds, the signals are divided into high, medium and low frequency categories. Each category of components can be associated with the seepage driving mechanism, separating noise, short-term rainfall, seasonal load and long-term curtain evolution signals, eliminating the mutual interference of multi-scale signals, and reducing the difficulty of model fitting from the preprocessing stage.
[0085] 2. Existing technologies rely on manually fixed frequency band matching models, resulting in poor generalization. This invention simultaneously trains two models—Ridge Regression and Transformer-BiLSTM—for each order component, constructing a weighted COI comprehensive scoring system that includes goodness of fit, absolute error, and relative error, and quantitatively and automatically selects the best model. Ridge Regression excels at fitting low-frequency linear time-dependent trends, while Transformer-BiLSTM accurately captures mid-to-high-frequency nonlinear fluctuations, fully leveraging the advantages of both models and balancing prediction accuracy and computational efficiency.
[0086] 3. All decomposition and ensemble models in the industry use direct linear summation of component prediction values for reconstruction, ignoring the nonlinear coupling effect of seepage signals in different frequency bands; this invention uses a multi-layer fully connected network to adaptively learn the interaction weights of each component, fully retaining multi-scale interaction information, significantly reducing information loss in the reconstruction process, and significantly improving the global prediction accuracy.
[0087] 4. Unlike existing interpretable solutions that can only perform ex-post SHAP analysis, this invention completes the physical layering of signals during the data preprocessing and modeling stages. It can directly predict deviations and locate the causes of seepage anomalies through the components of each frequency band: high-frequency anomalies represent causes such as short-term rainfall / instrument failure, mid-frequency anomalies represent causes such as periodic changes in water level and air temperature, and low-frequency anomalies represent causes such as long-term aging and erosion of the curtain. This achieves the integration of "prediction and anomaly tracing" and is fully adapted to the safety management needs of water conservancy projects.
[0088] 5. Based on multi-period lag water pressure, rainfall, refined temperature, and time-dependent driving factors, it fully conforms to the seepage lag mechanics mechanism of concrete dams; R is maintained at four typical measuring points in three types of regions: curtain high head zone, drainage low fluctuation zone, and dam toe steady-state zone. 2 With a value >0.99, RMSE and MAE are reduced by 7% to 77% compared to single Transformer, BiLSTM, and fixed decomposition ensemble models. It has outstanding generalization ability and can be directly integrated into existing dam safety monitoring platforms. Attached Figure Description
[0089] Figure 1 This is the implementation process of the model of the present invention;
[0090] Figure 2 This is a diagram showing the layout of seepage pressure measuring points for a concrete dam in the case study.
[0091] Figure 3 The actual measured curves of water level and rainfall in the case study;
[0092] Figure 4 Flowchart of the Empirical Mode Decomposition (EEMD) method;
[0093] Figure 5 A typical EEMD decomposition component curve diagram for measuring points;
[0094] Figure 6 A schematic diagram of the ridge regression principle;
[0095] Figure 7 This is a schematic diagram of the basic structure of Transformer;
[0096] Figure 8 This is a schematic diagram of the basic structure of BiLSTM;
[0097] Figure 9 A comparison chart of prediction results from different models at various measurement points;
[0098] Figure 10 A scatter plot comparing the correlations of prediction accuracies for different models. Detailed Implementation
[0099] To make the objectives, technical solutions, and effects of this invention clearer, the technical solutions of this invention are described in detail below with reference to specific examples. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this invention.
[0100] To better understand the present invention, the present invention will be further described below in conjunction with embodiments and accompanying drawings.
[0101] Example 1:
[0102] The test object in this embodiment is a concrete double-curvature arch dam in the Yalong River basin of Sichuan Province, such as... Figure 2 As shown, 18 piezometers were deployed. Four representative monitoring points were selected: PZ01 (behind the right bank curtain), PZ09 (middle of the riverbed curtain), PZ10 (dam foundation drainage gallery), and PZ15 (downstream dam toe on the left bank). Data collection lasted from January 2010 to December 2019, with a sampling interval of 12 hours, resulting in 7302 valid samples. The time series was divided into training and test sets according to 75% / 25% time sequence, and a 0.5-day single-step rolling prediction was adopted. This dam project has a large scale and a complex service environment. It has a long-term safety monitoring system covering displacement, reservoir water level, temperature, and other aspects. The data is continuous and highly representative, and can reflect the characteristics of seepage pressure changes in concrete dams under complex geological and load coupling conditions. The monitoring point layout of this concrete dam is as follows. Figure 2 As shown, the actual measured curves of water level and rainfall are as follows: Figure 3 As shown, this case study is used to examine the application value and effectiveness of the prediction model established in this paper in dam safety monitoring.
[0103] like Figure 1 As shown, the method for predicting seepage pressure in concrete dams based on scaled preprocessing and dual-model fusion includes the following steps:
[0104] Step S1: Data Preprocessing and Dataset Division: In this embodiment, seepage pressure-related monitoring data from typical dam sections are selected as the research object, with the data spanning multiple years of continuous observation sequences. First, the raw data is preprocessed, using the 3σ criterion to remove abnormal monitoring values and linear interpolation to complete a small number of missing data, resulting in a continuous and stable data time series. The dataset is then divided into a training set and a test set according to time sequence. The data from the first 75% of the time periods is used as the training set for model training, and the data from the last 25% of the time periods is used as the test set for model performance validation.
[0105] Step S2, EEMD Multiscale Adaptive Decomposition: The pretreated osmotic pressure sequence is decomposed into multiple scales using EEMD. The decomposition process is as follows: Figure 4 As shown. The model hyperparameters are set as follows: white noise standard deviation 0.001, ensemble iterations m=200, sliding window 90°. Specifically:
[0106] Step S21: Add 200 sets of zero-mean white noise to the original osmotic pressure sequence to generate 200 sets of noisy sequences;
[0107] Step S22: Perform EMD decomposition on each group of noisy sequences to obtain the IMF components of each group;
[0108] Step S23: Take the ensemble average of the 200 groups of IMF components of the same order to obtain the final 8th order stable IMF components.
[0109] Step S24: Then, perform validity screening and frequency classification:
[0110] Validity screening: Invalid components with no physical meaning are eliminated by a combination of four indicators: effective length, energy, variance, and Pearson correlation coefficient with the original sequence.
[0111] Frequency classification: Based on the period threshold, the effective components are divided into three categories: IMF1~IMF3 are high-frequency components (period < 7 days, corresponding to short-term random monitoring noise and instantaneous rainfall disturbance), IMF4~IMF6 are medium-frequency components (period 7~30 days, corresponding to the environmental load periodic response of reservoir water level fluctuation and temperature change), and IMF7~IMF8 are low-frequency components (period > 30 days, corresponding to the long-term evolution trend of dam foundation curtain, rock mass permeability and seepage channels).
[0112] Taking four typical measuring points as examples, the decomposition results are as follows: Figure 5 As shown: Figure 5 (a) shows the EEMD decomposition results of the PZ01 measuring point sequence, (b) shows the EEMD decomposition results of the PZ09 measuring point sequence, (c) shows the EEMD decomposition results of the PZ10 measuring point sequence, and (d) shows the EEMD decomposition results of the PZ15 measuring point sequence. PZ01 and PZ09 are located behind the curtain, with high overall head and significant mid-frequency components. IMF4~IMF6 show obvious annual cycle oscillations, confirming the dominant driving role of the reservoir water level annual cycle on the seepage pressure behind the curtain. PZ10 is located in the drainage area, with a significantly reduced head amplitude and an increased proportion of high-frequency components, indicating the peak-shaving and smoothing effect of the drainage holes on seepage pressure. PZ15 is located at the dam toe, with the lowest head. All components are mainly characterized by high-frequency micro-fluctuations without obvious periodicity or trend, indicating that the seepage field in this area has become stable.
[0113] Step S3, Dual-Model Parallel Training and COI Dynamic Optimization: All effective IMF components decomposed by EEMD are fed into the dual-model parallel optimization stage. Each component is simultaneously input into both Ridge Regression and Transformer-BiLSTM models. Ridge Regression uses L2 regularization to overcome the multicollinearity problem in linear fitting of low-frequency components, obtaining the linear prediction sequence for that component. Transformer-BiLSTM extracts long-range global correlation features using a multi-layer multi-head self-attention encoder, then connects to a bidirectional long short-term memory network to mine bidirectional temporal dependencies, and finally outputs the nonlinear prediction sequence for that component through a fully connected layer, accurately capturing the complex nonlinear temporal dependencies of mid-to-high-frequency components. The two sets of prediction results for the same component and their corresponding RMSE, MAE, MAPE, and R are presented. 2 The four indicators are simultaneously fed into the COI comprehensive selection module. After eliminating dimensional differences, the comprehensive selection value is calculated by weighting according to the engineering adaptation weights, and the result with the higher score is dynamically selected as the optimal output for that component. Specifically:
[0114] Step S31, Ridge Regression Model: The ridge regression principle diagram is shown below. Figure 6As shown, an L2 regularization penalty term is added to the classic linear regression loss function, with the L2 regularization coefficient set to 0.001. This model suppresses feature multicollinearity through a parameter contraction mechanism, exhibiting robust fitting of low-frequency linear trends, high computational efficiency, and clear physical meaning. The fusion model established in this invention uses the IMF components of EEMD decomposition as input. The low-frequency components are mainly the long-term evolution trends of dam foundation curtain, rock mass permeability, and seepage channels, exhibiting strong collinearity among feature factors. Nonlinear models are prone to overfitting. However, ridge regression is simple in structure and computationally efficient, possessing both the efficient fitting ability of linear regression and the robustness of regularization. It can overcome the model instability caused by the collinearity of monitoring data and ensure robust training of low-frequency components, which is the key reason why ridge regression was chosen as the core model in this study.
[0115] Step S32, Transformer-BiLSTM Model: The basic structure settings of Transformer and BiLSTM are as follows. Figure 7-8 As shown, the model parameters were set as follows: Transformer encoder with 6 layers, multi-head attention with 8 heads, and self-attention feature dimension of 256; BiLSTM network with 4 layers and 256 neurons per layer; learning rate of 0.0005, Dropout coefficient of 0.3, and prediction step size of 1 day. The Transformer's multi-head self-attention mechanism dynamically allocates attention weights based on the characteristics across time steps within the sequence through matrix operations, solving the problem of insufficient long-range dependency capture in traditional time-series models; BiLSTM mines the bidirectional temporal dependencies of the sequence, and LSTM uses gating to retain and update temporal information, solving the gradient vanishing problem.
[0116] Step S33, Construction of COI Comprehensive Selection Criteria: In this invention, R is selected 2 The four evaluation indicators, RMSE, MAE, and MAPE, are used to construct a normalized comprehensive selection index, COI.
[0117] Inverse normalization was performed on the three error indices: RMSE, MAE, and MAPE. 2 Perform forward normalization, and the normalized values will all be in the range of [0,1]. The closer the value is to 1, the better the performance.
[0118] In this embodiment, the weight is determined by combining the scores from hydraulic engineering monitoring experts with the engineering characteristics: R 2 The weights are 0.40 for RMSE, 0.30 for MAE, and 0.15 for MAPE. The COI value is the weighted sum of the normalization indices, with a higher value indicating better performance. When comparing the COI values of two models for the same component, the model output with the higher COI is selected as the optimal prediction sequence for that component.
[0119] Taking measurement point PZ01 as an example, the optimal selection results for each component are shown in Table 1. The COI values of the high-frequency and mid-frequency components (IMF1~IMF6) using Transformer-BiLSTM are significantly higher than those of ridge regression, with R²... 2 All values were above 0.98, with COI significantly outperforming ridge regression, indicating its accurate ability to capture nonlinear time-series dependencies; the COI values of the low-frequency components (IMF7~IMF8) of the ridge regression all exceeded 0.94, and R... 2 With a value above 0.99, the linear trend fitting is more stable and computationally efficient. The optimal results are highly consistent with the physical characteristics of the components, verifying the objectivity of the COI index and the scientific validity of the differentiated modeling strategy.
[0120] Table 1 Evaluation Indicators for Component Training at Measurement Point PZ01
[0121]
[0122] Step S4, Multi-dimensional Fusion Nonlinear Reconstruction: The optimal prediction sequences of each component are aligned by time step and concatenated to form a multi-dimensional optimal prediction feature matrix. This matrix is input into the multi-dimensional fusion reconstruction layer, using the original seepage pressure time series as training labels. A fully connected network adaptively learns the nonlinear coupling relationships and global contribution weights between different frequency components, ultimately reconstructing the output sequence of the dam's global seepage pressure prediction sequence with the same length as the original sequence and perfectly aligned time steps. This fusion method overcomes the limitations of traditional linear superposition, accurately capturing the potential correlations between multi-scale components and preserving effective information to the maximum extent.
[0123] Step S5, Model Performance Validation and Analysis, specifically includes:
[0124] Step S51: Prediction accuracy analysis; The prediction results of the complete model in this implementation are shown in Table 2. Test set R² for each measurement point. 2 All exceed 0.99. Taking PZ01 as an example, the test set R² 2 The accuracy was 0.9902, RMSE was 0.8111m, MAE was 0.7033m, and MAPE was 3.1664%. Compared with the baseline models of single Transformer, BiLSTM, and Transformer-BiLSTM, the RMSE decreased by 70.85%, 66.10%, and 56.67%, respectively, and the MAE decreased by 33.43%, 23.29%, and 7.70%, respectively. The model accurately reproduced the long-term evolution trend and accurately captured the short-term fluctuations caused by sudden changes in water level and heavy rainfall infiltration, with no significant overfitting.
[0125] Table 2. Prediction and evaluation indicators of the model at different measurement points in this study.
[0126]
[0127] Step S52: Comparison of generalization of prediction results: In this embodiment, four measurement points with different stress characteristics were selected: the right bank behind the curtain (PZ01), the middle of the riverbed behind the curtain (PZ09), the dam foundation drainage area (PZ10), and the downstream left bank dam toe (PZ15). A horizontal comparison was conducted with three comparative models: BiLSTM, Transformer, and Transformer-BiLSTM. The results are shown in Tables 3-4. Figures 9-10 As shown. Figure 9 (a) is a comparison of the prediction results of different models at measurement point PZ01, (b) is a comparison of the prediction results of different models at measurement point PZ09, (c) is a comparison of the prediction results of different models at measurement point PZ10, and (d) is a comparison of the prediction results of different models at measurement point PZ15. Figure 10 (a) is a scatter plot comparing the correlation of different models in the prediction accuracy at measurement point PZ01; (b) is a scatter plot comparing the correlation of different models in the prediction accuracy at measurement point PZ09; (c) is a scatter plot comparing the correlation of different models in the prediction accuracy at measurement point PZ10; and (d) is a scatter plot comparing the correlation of different models in the prediction accuracy at measurement point PZ15.
[0128] In regions PZ01 and PZ09 (behind the seepage barrier), the seepage pressure head amplitude is large and exhibits significant annual periodic fluctuations due to seasonal changes in reservoir water level. The proposed model accurately reproduces the long-term evolution trend and annual fluctuation patterns, outperforming the three benchmark models in terms of amplitude and phase. This method effectively mitigates the large deviations of the Transformer at water level abrupt changes and the phase lag defects of BiLSTM. For region PZ10, located in the dam foundation drainage area and regulated by drainage holes, the seepage pressure head is significantly reduced, while the proportion of high-frequency, small-amplitude fluctuations increases. Thanks to the EEMD multi-scale decoupling technique, the proposed model outperforms the benchmark models in capturing local fluctuation details and achieves higher local fitting accuracy. For region PZ15, located at the dam toe with low and stable seepage pressure, the model still maintains minimal prediction bias and reliably reflects the seepage field state. Experiments show that the proposed fusion model can simultaneously achieve accurate long-term trend prediction and refined short-term fluctuation capture, exhibiting higher prediction accuracy and robustness at monitoring points in different seepage regions. This model fully meets the actual needs of dam seepage safety management, and its prediction error is far lower than that of the benchmark model. The model of this invention achieves R0 at all measuring points. 2 All values exceeded 0.99, indicating that the error index was significantly better than the comparison model. At point PZ01 behind the impermeable curtain, the MAE was only 0.6697m, and R... 2The R² value reached 0.9956, and the RMSE decreased to 0.3116m, indicating that the model has a high-accuracy fitting capability for seepage series in areas with high water levels and strong annual periodic fluctuations. At point PZ10 in the drainage zone of the dam foundation, the RMSE decreased from 0.3108m to 0.1001m (a reduction of 67.79%), and the R² value... 2 The error rate of 0.9963 confirms the model's ability to accurately capture seepage patterns under high-frequency disturbances caused by drainage regulation. At the low-head steady-state seepage point PZ15 at the downstream dam toe, the model maintains the lowest error level, further validating its ability to adapt to low-head, low-amplitude stable seepage regions.
[0129] In summary, through multi-scale preprocessing and differentiated modeling, the proposed model achieves detailed seepage pressure prediction for the entire dam area, demonstrating excellent performance, high prediction accuracy, and good robustness. All monitoring points exhibit the best error index, R0. 2 The values all exceed 0.99, demonstrating excellent fitting performance. This model maintains stable and high-precision fitting performance across various scenarios, including gentle dam foundations, periodic fluctuations in the middle and upper parts of the dam body, and complex near-shore constraints, overcoming the limitations of traditional models. Its core advantages stem from the collaborative design of multi-scale feature decoupling and differentiated modeling, as well as the method of guiding model selection based on the COI comprehensive optimization index, effectively combining machine learning techniques with the physical mechanisms of dam seepage. Noise-assisted EEMD effectively suppresses mode aliasing, while IMF component classification, consistent with engineering-driven mechanisms, provides physical input for differentiated modeling. The dual-model adaptive system trains frequency components in parallel to achieve precisely matched model components. The COI index overcomes the limitations of single-index evaluation, comprehensively capturing and covering full-scale seepage pressure characteristics.
[0130] Table 3 Evaluation Indicators for Measurement Point Prediction
[0131]
[0132] Table 4 Performance Improvement Results at Measurement Points
[0133]
[0134] Example 2:
[0135] A concrete dam seepage safety early warning device includes a memory and a processor; the memory stores a computer-executable program, and when the processor runs the computer-executable program, it executes the above-mentioned concrete dam seepage pressure prediction method based on multi-scale preprocessing and dual-model fusion; the memory stores monitoring time series, model fixed parameters, and executable program; the processor executes the complete seepage pressure prediction calculation process.
[0136] The above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting seepage pressure in concrete dams based on multi-scale preprocessing and dual-model fusion, characterized in that, Includes the following steps: Step S1: Perform scale-adaptive decoupling on the original seepage pressure monitoring time series of the concrete dam, and after filtering out invalid fluctuation signals, obtain multiple effective scale components. Step S2: For each type of effective scale component, two differentiated prediction models are trained and calculated in parallel, and the optimal prediction output of a single component is automatically selected based on the normalized comprehensive selection index. Step S3: Input the optimal prediction results of all components into the nonlinear fusion network, adaptively fit the nonlinear coupling correlation between multi-scale components, and reconstruct the complete global seepage pressure prediction sequence. Step S4: Use multi-dimensional quantitative indicators to verify the model prediction effect, and trace the cause of seepage anomalies in layers based on the prediction deviation of each scale component.
2. The method for predicting seepage pressure in concrete dams based on multi-scale preprocessing and dual-model fusion as described in claim 1, characterized in that, Step S1 specifically includes: Step S101: Preprocessing of the original time series and construction of seepage driving factors: Outliers in the time series are removed using the 3σ criterion, and missing monitoring data are supplemented by linear interpolation; based on the statistical model of seepage pressure in concrete dams, four types of lagged environmental driving factors—water pressure, rainfall, temperature, and time aging—are constructed. The expression of the seepage pressure statistical model is as follows: In the formula: H is the seepage pressure measured on the monitoring day; z is a constant term in the model; These are water pressure component, rainfall component, temperature component, and time-dependent component, respectively; t is the cumulative number of days from the start date of monitoring. It represents the cumulative number of days since the initial monitoring date; u represents the lag item number for water pressure, rainfall, and temperature. Step S102, Ensemble Empirical Mode Decomposition: To the original osmotic pressure sequence Zero-mean white noise was added multiple times. m independent noisy time series were obtained. Its expression is: For each noisy time series Perform empirical mode decomposition on each component to obtain the f-th IMF component. and the residuals of the sth sequence decomposition , cumulative The results of the group decomposition are obtained, and the ensemble mean is calculated for the IMF components and residual terms of the same order in the m groups to obtain the final population mean. The overall average of the residuals To suppress mode aliasing, its expression is: In the formula: s is the noise superposition group number; f is the IMF order number obtained by decomposing a single noisy sequence; m is the total number of noise superpositions; Step S103, Joint Screening of Component Validity: Component validity is jointly determined by four thresholds: effective length, energy, variance, and Pearson correlation coefficient with the original sequence. Only components that simultaneously meet all four thresholds are considered valid components. The determination equations are as follows: In the formula: The effective data length of the l-th order IMF component. This is the preset minimum effective length threshold; The energy of the l-th order IMF component; This represents the total number of IMF component data. The variance of the l-th order IMF component; The l-th order IMF component and the original osmotic pressure sequence Pearson correlation coefficient; l is the global order specific number of the IMF component; k is the internal sampling point index of a single IMF sequence; Step S104, Component frequency band division: High-frequency components: with a period of less than 7 days, corresponding to short-term random monitoring noise and instantaneous rainfall disturbances; Mid-frequency component: period of 7~30 days, corresponding to the periodic response of environmental loads such as reservoir water level fluctuation and temperature change; Low-frequency component: period greater than 30 days, corresponding to the long-term evolution trend of dam foundation curtain, rock mass permeability and seepage channels.
3. The method for predicting seepage pressure in concrete dams based on multi-scale preprocessing and dual-model fusion as described in claim 1, characterized in that, Step S2 specifically includes: Step S201, Parallel training of dual models: Simultaneously feed the effective components of the same order into Ridge Regression and Transformer-BiLSTM for prediction respectively; 1) The ridge regression model adds an L2 regularization penalty term to the loss function of classic linear regression. This parameter shrinkage mechanism suppresses overfitting caused by feature multicollinearity, and is used to fit the linear long-term evolution trend of low-frequency components. Its expression is: In the formula: , used to balance the residual fitting accuracy and coefficient constraint strength, is the L2 regularization penalty coefficient; β is the vector of all feature regression coefficients; y is the regression coefficient of the g-th feature; n is the total number of training samples; τ X is the measured osmotic pressure label for the τth sample; τ Let τ be the input driving feature vector of the τth group; p is the total dimension of the input features; 2) The structure of the Transformer-BiLSTM model is as follows: the front end is a Transformer encoder, which completes the temporal position information through sine and cosine position encoding, and uses a multi-head self-attention mechanism to extract long-range global correlation features of the sequence. Its expression is: In the formula: For query, key, value matrix; For the number of attention heads; The learnable parameter matrix; The dimension of the single-head attention key vector; To compensate for the lack of timing information, sine and cosine functions are introduced to generate position codes: In the formula: d represents the position of the input parameter factor in the sequence; d is the index of the input factor in the sequence. The dimension of the input factor is encoded using a sine function at even positions and a cosine function at odd positions; The backend is a BiLSTM network, consisting of bidirectionally stacked forward and backward LSTM units, which mines the bidirectional temporal dependency features of the sequence. Finally, the nonlinear prediction results of the output components are passed through a fully connected layer to fit the complex nonlinear fluctuations of the mid-to-high frequency components. The state update formula is: In the formula: The states are hidden in both forward and backward directions; The spliced bidirectional features; Step S202, Index Calculation: Simultaneously calculate R 2 The four evaluation indicators are RMSE, MAE, and MAPE. Step S203, Construction of the Comprehensive Selection Index (COI), specifically includes: Step S2031, Index Normalization: Perform inverse normalization on the three types of error indices: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The expression is as follows: In the formula: The measured values of the indicators of the model to be evaluated are: The maximum and minimum values of the model participating in component selection are the maximum and minimum values among all measured values under this error index; For the coefficient of determination R 2 To perform forward normalization, the expression is: In the formula: R of the model to be evaluated 2 Measured value For models participating in component selection in R 2 The maximum and minimum values among all measured values for the indicator; Step S2032, Weighted Comprehensive Calculation: Combining the engineering characteristics and index representation of dam seepage, differentiated weights are assigned to the four normalized indicators. All weight coefficients are summed to 1, and the weighted sum is used to obtain the COI comprehensive value, the expression of which is: In the formula: A higher COI value indicates better overall model performance; To determine the weight of the coefficients; The root mean square error weight; Weighted by the average absolute error; Weighted by the average absolute percentage error; Step S204: Optimal selection: Compare the COI values of the lowerridge regression and Transformer-BiLSTM for the same component, and select the model output with the higher COI value as the optimal prediction result for that component; if the COI values of the two models are equal, use the inference time of a single batch of samples as the basis for judging computational efficiency, and select the model output with the shorter inference time as the optimal component result.
4. The method for predicting seepage pressure in concrete dams based on multi-scale preprocessing and dual-model fusion as described in claim 1, characterized in that, Step S3 includes the following steps: Step S301: Align and splice the optimal prediction sequences of each component according to a unified time step to generate a multi-dimensional component feature matrix; the time step is defined as the monitoring time interval corresponding to a single sampling. Step S302: Input the multidimensional feature matrix into the multilayer fully connected fusion network, and use the original measured seepage pressure time series as the supervised training label; Step S303: Adaptively learn the nonlinear coupling correlation between high-frequency, mid-frequency, and low-frequency components and the contribution weight of each component based on the network backpropagation mechanism; the contribution weight represents the proportion of the corresponding component's influence on the global seepage pressure prediction result. Step S304: Output a global osmotic pressure prediction time series that perfectly matches the time steps and total length of the original monitoring sequence.
5. The method for predicting seepage pressure in concrete dams based on segmented preprocessing and dual-model fusion as described in claim 1, characterized in that, Step S4 includes the following steps: Step S401: Use the coefficient of determination R 2 The overall prediction bias of the model is quantified by four indicators: root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), etc. Step S402: Plot a comparison curve between the measured seepage pressure time series and the model prediction time series to intuitively complete the visualization verification of the model's generalization performance; the generalization performance represents the stability of the model's prediction on new monitoring data that has not participated in the training. Step S403: Based on the prediction deviation of different frequency band components, locate the abnormal causes in layers: When the high frequency component deviates significantly from the baseline, it is judged into two scenarios: continuous and irregular small deviations are instrument noise; If the peak deviation is synchronized with the rainfall sequence, it indicates an instantaneous rainfall infiltration disturbance. The mid-frequency component deviation corresponds to the seasonal periodic fluctuations in reservoir water level and temperature, while the low-frequency component deviation corresponds to the dam foundation curtain and long-term seepage deterioration of the rock mass.
6. The method for predicting seepage pressure in concrete dams based on segmented preprocessing and dual-model fusion as described in claim 5, characterized in that, The calculation formulas for the four quantitative evaluation indicators in step S401 are as follows: In the formula: n is the total number of samples; For the first The true values of the measured seepage pressure were obtained. For the first Group model seepage pressure prediction values; This represents the average value of all measured osmotic pressure samples.
7. The method for predicting seepage pressure in concrete dams based on multi-scale preprocessing and dual-model fusion as described in claim 1, characterized in that, Explanation of dataset partitioning parameters during model training: The first 75% of the time series samples from the seepage pressure monitoring are used as the training set to iteratively update the internal parameters of the model; the remaining 25% of the time series are used as the test set to test the model's generalization ability; 0.5-day single-step rolling prediction: Each time a historical time series is input, the seepage pressure value for the next sampling day is predicted only.
8. A seepage safety early warning device for concrete dams, characterized in that, It includes a memory and a processor; the memory stores a computer-executable program, and when the processor runs the computer-executable program, it executes the concrete dam seepage pressure prediction method based on subscale preprocessing and dual-model fusion as described in any one of claims 1 to 7; Memory: Stores monitoring timing data, fixed model parameters, and executable programs; Processor: Executes the complete seepage pressure prediction calculation process.
