Dam operation state intelligent prediction method based on feature fusion and model migration

Abstract

The application discloses a dam operation state intelligent prediction method based on feature fusion and model migration, belongs to the technical field of dam operation safety monitoring and management, and comprises the following steps: S1, acquiring environment quantity monitoring data and multi-measuring point effect quantity monitoring data, and screening typical measuring points to construct effect quantity causal analysis data set; S2, extracting local time features by using a multi-scale time sequence convolution network, extracting global frequency features by using a fast Fourier transform, and fusing the two features through a dynamic fusion strategy; and S3, constructing a dam effect quantity intelligent prediction model by using a Transformer with a LoRA adaptive adapter, and then fine-tuning model parameters through migration learning, so that the model is adapted to intelligent prediction of effect quantities of other measuring points. The application adopts the above method, so that the model trained by data measuring points can efficiently adapt to data-deficient measuring points, high-precision prediction of multiple measuring points is realized, and a feasible method is provided for dam operation state analysis.

Description

Technical Field

[0001] This invention relates to the field of dam operation safety monitoring and management technology, and in particular to an intelligent prediction method for dam operation performance based on feature fusion and model transfer. Background Technology

[0002] As dams age, they are prone to material aging and overall performance degradation. Therefore, timely identification and elimination of safety hazards, strengthening monitoring and early warning capabilities, and ensuring the long-term safe operation of reservoirs and dams are core priorities in water conservancy project safety management. Dam safety monitoring is a crucial component of the dam safety management system and a key technical means for dam risk prevention and control and safe operation and maintenance. Dam operational performance monitoring and analysis, as a core means of assessing the healthy operating status of dams, requires comprehensive understanding of the dynamic changes in dam performance during operation through routine monitoring and scientific forecasting.

[0003] Currently, common models for monitoring the operational safety of dams include statistical models, time series models, machine learning models, and deep learning models. These models all use environmental variables such as upstream water level and rainfall as independent variables and monitored quantities as dependent variables, establishing the interaction relationships between these variables. Based on the deviation between the model's calculated values ​​and the measured values, they determine whether the earth-rock dam's operational status is safe. However, these models predict operational performance based on monitoring points with abundant data, and cannot effectively address the problem of high-precision prediction when monitoring data contains significant errors or is missing.

[0004] However, dam monitoring data exhibits significant multi-timescale characteristics, with a superposition of short-term fluctuations (such as the instantaneous response of seepage pressure caused by rainfall), medium-term cycles (such as the influence of seasonal temperature changes), and long-term trends (such as material aging). Existing single models struggle to simultaneously capture these multi-scale characteristics. Furthermore, the physical response patterns differ across different monitoring points, and conventional transfer learning methods, by adapting the source domain model as a whole to the target domain, carry the risk of parameter redundancy and overfitting, especially when monitoring data in the target domain is scarce. Therefore, how to construct a prediction method that can simultaneously capture multi-scale temporal features and frequency domain periodic characteristics, while also being lightweight enough to adapt to data-scarce monitoring points, is a pressing technical problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent prediction method for dam operation performance based on feature fusion and model transfer, which solves the problem that it is difficult to predict the operation performance of atypical dam monitoring points with high accuracy due to the lack of monitoring data. The method achieves intelligent and accurate prediction under the condition of scarce monitoring data through feature fusion and model transfer.

[0006] To achieve the above objectives, this invention provides an intelligent prediction method for dam operational performance based on feature fusion and model transfer, comprising the following steps: S1. Data Preparation and Dataset Construction: Acquire dam safety monitoring data, including environmental quantity monitoring data and multi-point effect quantity monitoring data; based on the dam safety monitoring principle, select monitoring points from the multi-point data with monitoring data accuracy higher than a preset threshold and data volume greater than a preset number as typical monitoring points; construct an effect quantity causal analysis dataset for typical monitoring points; the input variables of the effect quantity causal analysis dataset should include at least water pressure factor, temperature factor, and time factor, and the output variable should be the time series of effect quantity monitoring data, and the dataset should be divided into training set and test set according to a preset ratio; S2. Feature Extraction and Fusion: Multi-scale Temporal Convolutional Network (MS-TCN) is used to mine the local temporal features and dynamic fluctuation patterns of input variables at different temporal scales in the effect size causal analysis dataset. At the same time, Fast Fourier Transform (FFT) is used to extract the global frequency features of input variables, capturing the hidden periodic change patterns and long-term evolution trends in the input variable sequence. Then, a dynamic fusion strategy is used to fuse the local temporal features and global frequency features, integrating temporal dynamics and global periodic information. S3. Model Building and Transfer Prediction: Using a Transformer with LoRA adaptation adapter, the nonlinear functional relationship between typical measurement point effect quantity monitoring data and the fusion features of input variables is captured to build an intelligent prediction model for dam effect quantities. Then, transfer learning technology is used to fine-tune the parameters of the built intelligent prediction model so that the fine-tuned model can be adapted to the intelligent prediction of other measurement point effect quantities of the dam.

[0007] Preferably, in step S1, the environmental monitoring data includes water level monitoring data, temperature monitoring data, and rainfall monitoring data; the multi-point effect monitoring data includes dam deformation monitoring data and seepage pressure monitoring data.

[0008] Preferably, in step S1, the water pressure factor and temperature factor are functional transformations of the water level monitoring data and temperature monitoring data, respectively, and the time factor is a function of time.

[0009] Preferably, in step S1, the selection criteria for typical measurement points are: measurement points whose monitoring data accuracy is higher than a preset threshold and whose data volume is greater than a preset number.

[0010] Preferably, step S2 specifically includes: S21. Temporal Feature Extraction: Multi-scale temporal convolutional networks are used to mine the local temporal features and dynamic fluctuation patterns of input variables in the effect size causal analysis dataset. Specifically, multiple convolutional branches are constructed by using one-dimensional convolutions with different kernel sizes and different dilation rates in parallel. The outputs of all branches are concatenated in the channel dimension and mapped back to the original dimension by 1×1 convolution. Then, they are added to the original input to obtain the temporal feature tensor. S22. Frequency Domain Feature Extraction: The global frequency features of the input variables are extracted using the Fast Fourier Transform. Specifically, the time-domain signal is converted to the frequency domain and the spectral amplitude is calculated. The average value is taken in the frequency dimension to obtain the global feature vector. After enhancement by a Multilayer Perceptron (MLP) and nonlinear transformation, it is expanded into a format with the same length as the sequence to obtain the frequency domain feature tensor. S23. Dynamic feature fusion: For the time-domain feature tensor and the frequency-domain feature tensor, calculate the mean respectively, generate gating weights through splicing operation, and then perform weighted summation on the two to obtain the fused feature tensor.

[0011] Preferably, in step S21, the time-domain feature tensor The calculation formula is: ; ; ; Among them, the set of convolution kernel sizes , Expansion rate set , , This represents a convolution with a kernel size of 3 and an inflation rate of 1. Indicates the expansion rate One-dimensional convolution is used to expand the receptive field. For activation function, Indicates input features, This represents one of the convolution branches. This represents the spliced ​​temporal feature tensor. This indicates a splicing operation. This represents a convolutional projection of size 1×1. This indicates normalization.

[0012] Preferably, in step S22, the frequency domain feature tensor The calculation formula is: ; ; ; ; ; in, This represents the complex number result after the Fourier transform. Represents the Fast Fourier Transform of real numbers. Represents absolute value. The index variable represents the first f One frequency, Indicates the number of frequency components. , For sequence length, This means rounding L / 2 down. Indicates the spectral amplitude. Represents the frequency domain modulation vector. Represents the global feature vector. This refers to a multilayer perceptron, used for nonlinear feature enhancement of input features. This indicates that the global feature vector will be copied and expanded into a sequence format.

[0013] Preferably, the calculation formula for dynamic fusion in step S23 is: ; ; in, For gating weights, Represents the learnable weight matrix. Indicates global average pooling. is the feature tensor after fusion, and [;] indicates splicing.

[0014] Preferably, step S3 specifically includes: S31. Feature Alignment and Position Encoding: Perform dimension mapping on the fused features obtained in step S2 to make the fused features in step S2 consistent with the high-dimensional space dimension inside the Transformer, and use rotation position encoding to capture the relative positional relationships in the time series. S32, Transformer Encoding and Prediction: The features obtained after dimensional mapping with location information are input into the Transformer encoder module. Each encoder module is processed through layer normalization, multi-head attention mechanism, and residual connections. Specifically, after layer normalization, the input features are processed through three LoRA linear layers in the multi-head attention mechanism to generate a query matrix, key matrix, and value matrix. The LoRA linear layer is composed of the sum of the product of the original pre-trained weight matrix and the low-rank decomposition matrix. Parameter fine-tuning is achieved by training only the low-rank decomposition matrix. After passing through the output projection layer, the model is calculated by a feedforward network, and the features from the last time step are input into the prediction head to obtain the effect size prediction value. The Transformer encoder and prediction head are trained using training data from typical measurement points to obtain a pre-trained model. S33. Transfer Learning Fine-tuning: Using the pre-trained model trained in step S32 as the base model, freeze all parameters in the pre-trained model, retaining only the LoRA parameters and prediction head parameters; use monitoring data from other measurement points different from typical measurement points to fine-tune the unfrozen LoRA parameters and prediction head parameters, so that the fine-tuned model can be adapted to the intelligent prediction of the effect quantities of other measurement points of the dam.

[0015] Preferably, in step S32, the query matrix, key matrix, and value matrix in the multi-head attention mechanism are generated as follows: after the input features are normalized by the layers, they are obtained by linear transformation through three independent LoRA linear layers; the output projection layer is also implemented using LoRA linear layers; in step S33, the loss function for fine-tuning training includes a domain adaptation loss term, which includes maximum mean difference loss and correlation alignment loss.

[0016] Therefore, the present invention employs the above-mentioned intelligent prediction method for dam operation performance based on feature fusion and model transfer, which has the following beneficial effects: (1) Obtain dam safety monitoring data, such as environmental monitoring data like water level and temperature, and multi-point effect monitoring data like dam deformation and seepage pressure. Combined with the dam safety monitoring principle, construct an effect causal analysis dataset for typical monitoring points with monitoring data, and divide the modeling dataset into training set and test set according to a certain proportion. (2) Multi-scale temporal convolutional network (MS-TCN) is used to mine the local temporal features and dynamic fluctuation patterns of input variables such as environmental load at different temporal scales. At the same time, fast Fourier transform (FFT) is used to efficiently extract the global frequency features of input variables, accurately capture the hidden periodic change patterns and long-term evolution trends in the input variable sequence, and then the local temporal features and global frequency features are fused through a dynamic fusion strategy to integrate temporal dynamics and global periodic information.

[0017] (3) By using a Transformer with a LoRA adapter to capture the nonlinear functional relationship between the monitoring data of typical measurement point effects and the fusion characteristics of input variables, an intelligent prediction model for dam effects is constructed. Then, transfer learning technology is used to fine-tune the parameters of the constructed intelligent prediction model to achieve efficient and intelligent prediction of other measurement point effects. This invention provides a feasible approach for analyzing the impact of changes in dam operation performance and accurately predicting the future trends of other measurement point effects.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the method construction of an embodiment of the intelligent prediction method for dam operation performance based on feature fusion and model transfer of the present invention. Figure 2 This invention relates to different dates for monitoring water levels and rainfall at the dam source area measuring point (K14A) and upstream and downstream monitoring points. Figure 3This invention relates to the monitoring points (K05A, K05B, K06A) of the dam target area on different dates, and the upstream and downstream water levels and rainfall. Figure 4 This is a comparison chart of the fitted value and the actual value of the osmotic pressure at the source domain measuring point K14A in this embodiment of the invention; Figure 5 This is a curve showing the fitted value and actual value of the osmotic pressure at the target domain measuring point K05A in this embodiment of the invention. Figure 6 This is a curve showing the fitted value and the true value of the osmotic pressure at the target domain measuring point K05B in this embodiment of the invention. Figure 7 This is a curve showing the fitted value and the true value of the osmotic pressure at the target domain measuring point K06A in an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0022] Example like Figure 1 As shown, the intelligent prediction method for dam operation performance based on feature fusion and model transfer includes the following steps: S1. Data Preparation and Dataset Construction: Acquire dam safety monitoring data, including environmental quantity monitoring data and multi-point effect quantity monitoring data; based on the dam safety monitoring principle, select monitoring points from the multi-point data with monitoring data accuracy higher than a preset threshold and data volume greater than a preset number as typical monitoring points; construct an effect quantity causal analysis dataset for typical monitoring points; the input variables of the effect quantity causal analysis dataset should include at least water pressure factor, temperature factor, and time factor, and the output variable should be the time series of effect quantity monitoring data, and the dataset should be divided into training set and test set according to a preset ratio; Environmental monitoring data includes water level monitoring data, temperature monitoring data, and rainfall monitoring data; multi-point effect monitoring data includes dam deformation monitoring data and seepage pressure monitoring data.

[0023] The water pressure factor and temperature factor are functional transformations of the water level monitoring data and temperature monitoring data, respectively, and the time factor is a function of time.

[0024] The selection criteria for typical measurement points are: measurement points whose monitoring data accuracy is higher than the preset threshold and whose data volume is greater than the preset number.

[0025] S2. Feature Extraction and Fusion: Multi-scale Temporal Convolutional Network (MS-TCN) is used to mine the local temporal features and dynamic fluctuation patterns of input variables at different temporal scales in the effect size causal analysis dataset. At the same time, Fast Fourier Transform (FFT) is used to extract the global frequency features of input variables, capturing the hidden periodic change patterns and long-term evolution trends in the input variable sequence. Then, a dynamic fusion strategy is used to fuse the local temporal features and global frequency features, integrating temporal dynamics and global periodic information. The aforementioned technical solution utilizes the parallel multi-scale convolutional structure of MS-TCN to simultaneously extract local dynamic patterns across different time spans, overcoming the limitations of fixed receptive fields in single-scale convolution. Simultaneously, it leverages FFT to map time-domain signals to the frequency domain, extracting dam response patterns caused by periodic changes in environmental loads (such as annual temperature variations and seasonal rainfall). Furthermore, through a gated fusion mechanism, the model adaptively learns the fusion weights of time-domain details and frequency-domain trends, rather than simply concatenating them, thereby achieving an organic integration of multi-scale dynamics and global periodic features, providing a more complete feature representation for downstream predictions.

[0026] S21. Temporal Feature Extraction: Utilizing a multi-scale temporal convolutional network, this study mines the local temporal features and dynamic fluctuation patterns of input variables in the effect size causal analysis dataset. Specifically, multiple convolutional branches are constructed in parallel using one-dimensional convolutions with different kernel sizes and dilation rates. The outputs of all branches are concatenated along the channel dimension and mapped back to the original dimension via a 1×1 convolution. This mapping is then added to the original input to obtain the temporal feature tensor. The calculation formula is: ; ; ; Among them, the set of convolution kernel sizes , Expansion rate set , , This represents a convolution with a kernel size of 3 and an inflation rate of 1. Indicates the expansion rate One-dimensional convolution is used to expand the receptive field. For activation function, Indicates input features, This represents one of the convolution branches. This represents the spliced ​​temporal feature tensor. This indicates a splicing operation. This represents a convolutional projection of size 1×1. This indicates normalization.

[0027] S22. Frequency Domain Feature Extraction: Global frequency features of the input variables are extracted using Fast Fourier Transform (FFT). Specifically, the time-domain signal is converted to the frequency domain, and the spectral amplitude is calculated. The average value along the frequency dimension is taken to obtain the global feature vector. After enhancement by a Multilayer Perceptron (MLP) and nonlinear transformation, it is expanded to a format with the same length as the sequence, resulting in a frequency domain feature tensor. The calculation formula is: ; ; ; ; ; in, This represents the complex number result after the Fourier transform. Represents the Fast Fourier Transform of real numbers. Represents absolute value. The index variable represents the first f One frequency, Indicates the number of frequency components. , For sequence length, This means rounding L / 2 down. Indicates the spectral amplitude. Represents the frequency domain modulation vector. Represents the global feature vector. This refers to a multilayer perceptron, used for nonlinear feature enhancement of input features. This indicates that the global feature vector will be copied and expanded into a sequence format.

[0028] S23. Dynamic Feature Fusion: For the time-domain and frequency-domain feature tensors, the means are calculated separately, and then gating weights are generated through a concatenation operation. Finally, the two are weighted and summed to obtain the fused feature tensor. The formula for dynamic fusion is: ; ; in, For gating weights, Represents the learnable weight matrix. Indicates global average pooling. is the feature tensor after fusion, and [;] indicates splicing.

[0029] S3. Model Building and Transfer Prediction: Using a Transformer with LoRA adaptation adapter, the nonlinear functional relationship between typical measurement point effect quantity monitoring data and the fusion features of input variables is captured to build an intelligent prediction model for dam effect quantities. Then, transfer learning technology is used to fine-tune the parameters of the built intelligent prediction model so that the fine-tuned model can be adapted to the intelligent prediction of other measurement point effect quantities of the dam.

[0030] The above technical solution utilizes the Transformer to capture the complex nonlinear mapping relationship between effect sizes and multi-source features, and employs a LoRA low-rank adaptation strategy to achieve model transfer. Unlike conventional fine-tuning methods that require updating all or a large number of parameters, this method only adjusts the low-rank decomposition matrix to achieve adaptive model adjustment for different measurement points. This design has two key effects: first, it significantly reduces the number of trainable parameters required for fine-tuning in the target domain, mitigating the risk of overfitting under data scarcity conditions; second, it retains the general environmental load-effect size mapping knowledge learned in the pre-training stage, adapting only to the individual differences of each measurement point, achieving a balance between universality and specificity.

[0031] S31. Feature Alignment and Position Encoding: Perform dimension mapping on the fused features obtained in step S2 to make the fused features in step S2 consistent with the high-dimensional space dimension inside the Transformer, and use rotation position encoding to capture the relative positional relationships in the time series. S32, Transformer Encoding and Prediction: The features obtained after dimensional mapping with location information are input into the Transformer encoder module. Each encoder module is processed through layer normalization, multi-head attention mechanism, and residual connections. Specifically, after layer normalization, the input features are used in the multi-head attention mechanism to generate a query matrix, key matrix, and value matrix through three LoRA linear layers. After passing through the output projection layer, they are calculated by a feedforward network. The features of the last time step are taken and input into the prediction head to obtain the effect size prediction value. The Transformer encoder and prediction head are trained using training data of typical measurement points to obtain a pre-trained model. The hyperparameter configurations of the model used in this embodiment are shown in Table 1. The time window length is set to 128, meaning that monitoring data from the past 128 time steps is used to predict future effect sizes; the batch size is set to 32; the number of pretraining epochs is set to 50; and the number of fine-tuning epochs is set to 20. The learning rate (lr) during the pretraining phase is set to 1×10⁻⁶. -3 The learning rate ft_lr for the fine-tuning phase is set to 5 × 10. -4 The weight decay factor, weight_decay, is set to 1×10. -4 This is to control overfitting.

[0032] For the Transformer encoder part, the internal feature dimension d_model is set to 64, the number of encoder layers n_layers is set to 4, the number of multi-head attention heads n_heads is set to 4, the internal dimension of the feedforward network d_ff is set to 256, and the random dropout rate is set to 0.1.

[0033] In the low-rank adaptation part of LoRA, the rank lora_r of the low-rank decomposition is set to 8, and the scaling factor lora_alpha is set to 16. During the fine-tuning phase, only the LoRA parameters and prediction head parameters are trained; the rest are frozen.

[0034] For the domain adaptation part, the maximum mean difference (MMD) loss weight lambda_mmd is set to 0.5, the correlation alignment (CORAL) loss weight lambda_coral is set to 0.5, and the MMD kernel function bandwidth mmd_sigma is set to 1.0.

[0035] Table 1. Hyperparameter Configuration Table for Model Training

[0036] The query matrix, key matrix, and value matrix in the multi-head attention mechanism are generated as follows: after the input features are normalized by the layers, they are obtained by linear transformation through three independent LoRA linear layers; the output projection layer is also implemented using LoRA linear layers.

[0037] LoRA linear layers are low-rank adaptive linear layers. By introducing two low-rank matrices, their product is used to represent the incremental update of the weights. This allows for efficient linear transformation operations for fine-tuning parameters while significantly reducing the number of trainable parameters. It can be represented as: ; in W 0 represents the original weight matrix that remains constant in the pre-trained model. A and B It is a low-rank matrix; only the low-rank matrix is ​​updated during training. A andB This enables efficient linear transformation of parameters.

[0038] S33. Transfer Learning Fine-tuning: Using the pre-trained model trained in step S32 as the base model, freeze all other parameters except for the LoRA parameters and the prediction head parameters; use monitoring data from other measurement points that are different from typical measurement points to fine-tune the unfrozen LoRA parameters and prediction head parameters, so that the fine-tuned model can be adapted to the intelligent prediction of the effect quantities of other measurement points of the dam.

[0039] Furthermore, during the fine-tuning training process, a domain adaptation loss term is introduced as a constraint. The domain adaptation loss term is a regularization constraint term introduced in addition to the prediction loss during fine-tuning training, used to measure and minimize the difference between the feature distributions of the source and target domains. The domain adaptation loss term includes the maximum mean difference (MMD) loss and the correlation alignment (CORAL) loss, each weighted with a weight of 0.5.

[0040] The maximum mean difference loss maps the features of the source and target domains to the reproducing kernel Hilbert space, measuring the distance between the mean embeddings of the features in the two domains, thus narrowing their marginal distributions. The correlation alignment loss aligns the second-order statistics, i.e., the covariance matrix, of the features in the source and target domains, reducing the conditional distribution difference between the feature distributions of the two domains. The two are used together to comprehensively align the feature distributions of the source and target domains from both first-order and second-order statistical perspectives.

[0041] The effect of introducing a domain adaptation loss term is that, due to the relative scarcity of monitoring data for target domain measurement points, if only the prediction loss is used for fine-tuning, the model is prone to overfitting to a limited number of samples. The domain adaptation loss term, acting as a regularization constraint, guides the model to retain the general feature representations learned during the source domain pre-training phase while learning the characteristics of the target domain data. This alleviates the overfitting problem under conditions of scarce data and improves the model's prediction accuracy for the effect size of the target measurement points.

[0042] In the source and target domain prediction of this invention, in order to quantitatively evaluate the fitting accuracy and prediction accuracy of the prediction model, the coefficient of determination R is comprehensively used. 2 The mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE) were used as evaluation metrics to calculate the statistical metrics of the prediction model for each measurement point in the fitting segment and the prediction segment, respectively.

[0043] This example uses the seepage pressure monitoring of an earth-rock dam as an example.

[0044] I. Project Overview.

[0045] A certain earth-rock dam is located in an earthquake-prone area of ​​my country. The dam has a maximum height of 29 meters, a crest length of 260 meters, a width of 8 meters, a crest elevation of 1875.82 meters, and a total reservoir capacity of 65 million cubic meters. 3 The normal water level is 1874.40m. It is a medium-sized reservoir mainly used for irrigation, industrial and urban domestic water supply, combined with flood control and power generation. The climate in this area is continental and arid, with an annual rainfall of 107.2-259.3mm and 33-57 rainy days per year, with uneven distribution of precipitation.

[0046] A 6.9-magnitude earthquake and a 5.2-magnitude aftershock occurred in the area where a certain earth-rock dam is located, with a focal depth of approximately 10 km. The reservoir where the earth-rock dam is located is about 90 km from the epicenter and is within the earthquake's impact range. To avoid a decrease in the accuracy of monitoring data after the earthquake, the data period of this invention is selected from the period before the earthquake.

[0047] II. Selection of monitoring sections.

[0048] To monitor the seepage status of the dam body and foundation, a seepage monitoring system was installed on the dam. The system uses a shared monitoring section for seepage pressure measuring points in the dam body and foundation, and each measuring point automatically collects data through directly buried piezometers.

[0049] In this embodiment, the relative value of osmotic pressure refers to a quantitative indicator characterizing the dynamic change of osmotic pressure at a monitoring point in subsequent monitoring days, with the first day of the monitoring period as the baseline value. Specifically, it is the difference between the measured osmotic pressure head (or piezometric head) at the monitoring point and the baseline head (i.e., the monitoring value on the first day of the monitoring sequence), expressed in meters (m, equivalent water column height difference), used to visually reflect the magnitude and direction of the change in osmotic pressure at that point relative to the baseline state.

[0050] In this embodiment, the 0+275 section, which exhibits the best data continuity and integrity, is selected as the source domain. The temporal variation process of the monitoring data from this section is as follows: Figure 2 As shown. Figure 2 This data shows the changes in upstream and downstream water levels, relative seepage pressure at a typical monitoring point K14A, and concurrent rainfall data for the cross-section over a complete monitoring period. The head curve at monitoring point K14A reflects the dynamic changes in seepage pressure at this cross-section and serves as the foundational data for building the pre-trained model.

[0051] Because the earthquake caused the loss of monitoring data for some monitoring points at section 0+200, this invention selects monitoring points K05A, K05B, and K06A at this section as the target domain. The temporal variation process of their monitoring data is as follows: Figure 3 As shown. Figure 3The data also show the upstream and downstream water levels of the cross-section, the relative values ​​of seepage pressure at the three target measuring points, and the concurrent rainfall data. Due to the limited amount of data at the target domain measuring points, this invention uses transfer learning to adapt the source domain pre-trained model to the aforementioned target measuring points, thereby enabling head prediction under conditions of data scarcity.

[0052] III. Data Preprocessing.

[0053] Initial observation data was obtained based on actual engineering monitoring, including upstream and downstream water levels, rainfall, and dam operational status observations at each monitoring section. Gross errors were processed from the initial observation data, and the lagged effects of environmental factors were determined according to the principles of dam safety monitoring, with a lag of 15 days for upstream water level and 7 days for rainfall. Based on this, a causal analysis dataset for the dam seepage pressure effect was constructed and divided into training and testing sets.

[0054] IV. Model Building and Transfer Prediction.

[0055] The following steps are performed sequentially according to the method described in the embodiment: (1) A multi-scale temporal convolutional network is used to extract temporal features, a fast Fourier transform is used to extract frequency domain features, and a dynamic fusion strategy is used to fuse features. The specific implementation method is described in step S2 of the embodiment. (2) Construct an intelligent prediction model for dam effect using a Transformer with LoRA adaptive adapter. For specific implementation, refer to the relevant description of step S3 in the embodiment. (3) The model is pre-trained using the 0+275 section as the source domain. Then, all parameters except the LoRA adaptation parameters and the prediction head parameters are frozen. The unfrozen parameters are fine-tuned using the target domain data of the 0+200 section.

[0056] V. Evaluation of Prediction Results

[0057] Using the coefficient of determination (R) 2 The mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE) were used as evaluation indicators to calculate the statistical indicators for the prediction segment at each measuring point. The results are shown in Table 2, and the fitted curves are shown in Table 2. Figures 4-7 As shown.

[0058] Table 2 Model Evaluation Indicators

[0059] As can be seen from Table 2, the R of the source domain measuring point K14A is... 2 The value reached 0.9718, and the R value at each measurement point in the target domain was [value missing]. 2All values ​​reached above 0.9058, with the lowest RMSE being only 0.0744, indicating that the method of the present invention can still achieve high-precision prediction under conditions of scarce data.

[0060] Therefore, the present invention employs the above-mentioned intelligent prediction method for dam operation performance based on feature fusion and model transfer, which has the following beneficial effects: (1) In view of the problem of superimposed features of multiple time scales in dam monitoring data, this invention combines multi-scale temporal convolutional networks with fast Fourier transform to simultaneously extract dynamic features of multiple time scales and periodic features of frequency domain, and adopts a gating fusion mechanism to achieve adaptive information integration, thus solving the technical problem that a single feature extraction method is difficult to fully characterize the complex relationship between environmental load and effect.

[0061] (2) To address the problem that data is scarce and measurement points are difficult to train high-precision prediction models independently, this invention uses LoRA low-rank adaptation technology to achieve lightweight transfer of pre-trained models, which can still achieve high-precision prediction under small sample conditions.

[0062] (3) This invention organically combines the above feature fusion method with the model transfer strategy to form a complete technical solution of "pre-training of data-sufficient measurement points - lightweight adaptation of data-scarce measurement points", providing an efficient and feasible path for multi-measurement point operation status monitoring of dams.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent prediction of dam operational performance based on feature fusion and model transfer, characterized in that, Includes the following steps: S1. Data Preparation and Dataset Construction: Obtain dam safety monitoring data, including environmental quantity monitoring data and multi-point effect quantity monitoring data; Based on the principle of dam safety monitoring, monitoring points with monitoring data accuracy higher than a preset threshold and data volume greater than a preset number are selected from multiple monitoring points as typical monitoring points; for typical monitoring points, an effect size causal analysis dataset is constructed; the input variables of the effect size causal analysis dataset include at least water pressure factor, temperature factor and time factor, and the output variable is the time series of effect size monitoring data, and the dataset is divided into training set and test set according to a preset ratio; S2. Feature Extraction and Fusion: Multi-scale temporal convolutional networks are used to mine the local temporal features and dynamic fluctuation patterns of input variables in the effect size causal analysis dataset. At the same time, fast Fourier transform is used to extract the global frequency features of input variables, capturing the hidden periodic change patterns and long-term evolution trends in the input variable sequence. Then, a dynamic fusion strategy is used to fuse the local temporal features and global frequency features, integrating temporal dynamics and global periodic information. S3. Model Building and Transfer Prediction: Using a Transformer with LoRA adaptation adapter, the nonlinear functional relationship between typical measurement point effect quantity monitoring data and the fusion features of input variables is captured to build an intelligent prediction model for dam effect quantities. Then, transfer learning technology is used to fine-tune the parameters of the built intelligent prediction model so that the fine-tuned model can be adapted to the intelligent prediction of other measurement point effect quantities of the dam.

2. The intelligent prediction method for dam operational performance based on feature fusion and model transfer as described in claim 1, characterized in that, In step S1, the environmental monitoring data includes water level monitoring data, temperature monitoring data, and rainfall monitoring data; the multi-point effect monitoring data includes dam deformation monitoring data and seepage pressure monitoring data.

3. The intelligent prediction method for dam operation performance based on feature fusion and model transfer as described in claim 2, characterized in that, In step S1, the water pressure factor and temperature factor are the functional transformation forms of the water level monitoring data and temperature monitoring data, respectively, and the time factor is a function of time.

4. The intelligent prediction method for dam operation performance based on feature fusion and model transfer as described in claim 1, characterized in that, In step S1, the selection criteria for typical measurement points are: measurement points whose monitoring data accuracy is higher than a preset threshold and whose data volume is greater than a preset number.

5. The intelligent prediction method for dam operation performance based on feature fusion and model transfer as described in claim 1, characterized in that, Step S2 specifically includes: S21. Temporal Feature Extraction: Multi-scale temporal convolutional networks are used to mine the local temporal features and dynamic fluctuation patterns of input variables in the effect size causal analysis dataset. Specifically, multiple convolutional branches are constructed by using one-dimensional convolutions with different kernel sizes and different dilation rates in parallel. The outputs of all branches are concatenated in the channel dimension and mapped back to the original dimension by 1×1 convolution. Then, they are added to the original input to obtain the temporal feature tensor. S22. Frequency Domain Feature Extraction: The global frequency features of the input variables are extracted using the Fast Fourier Transform. Specifically, the time-domain signal is converted to the frequency domain and the spectral amplitude is calculated. The average value is taken in the frequency dimension to obtain the global feature vector. After enhancement by a multilayer perceptron and nonlinear transformation, it is expanded into a format with the same length as the sequence to obtain the frequency domain feature tensor. S23. Dynamic feature fusion: For the time-domain feature tensor and the frequency-domain feature tensor, calculate the mean respectively, generate gating weights through splicing operation, and then perform weighted summation on the two to obtain the fused feature tensor.

6. The intelligent prediction method for dam operation performance based on feature fusion and model transfer as described in claim 5, characterized in that, In step S21, the time-domain feature tensor The calculation formula is: ； ； ； Among them, the set of convolution kernel sizes , Expansion rate set , , This represents a convolution with a kernel size of 3 and an inflation rate of 1. Indicates the expansion rate One-dimensional convolution is used to expand the receptive field. For activation function, Indicates input features, This represents one of the convolution branches. This represents the spliced ​​temporal feature tensor. This indicates a splicing operation. This represents a convolutional projection of size 1×1. This indicates normalization.

7. The intelligent prediction method for dam operational performance based on feature fusion and model transfer as described in claim 6, characterized in that, In step S22, the frequency domain feature tensor The calculation formula is: ； ； ； ； ； in, This represents the complex number result after the Fourier transform. Represents the Fast Fourier Transform of real numbers. Represents absolute value. As an index variable, it represents the first... f One frequency, Indicates the number of frequency components. , For sequence length, This means rounding L / 2 down. Indicates the spectral amplitude. Represents the frequency domain modulation vector. Represents the global feature vector. This refers to a multilayer perceptron, used for nonlinear feature enhancement of input features. This indicates that the global feature vector will be copied and expanded into a sequence format.

8. The intelligent prediction method for dam operation performance based on feature fusion and model transfer as described in claim 7, characterized in that, The calculation formula for dynamic fusion in step S23 is as follows: ； ； in, For gating weights, Represents the learnable weight matrix. Indicates global average pooling. is the feature tensor after fusion, and [;] indicates splicing.

9. The intelligent prediction method for dam operation performance based on feature fusion and model transfer as described in claim 1, characterized in that, Step S3 specifically includes: S31. Feature Alignment and Position Encoding: Perform dimension mapping on the fused features obtained in step S2 to make the fused features in step S2 consistent with the high-dimensional space dimension inside the Transformer, and use rotation position encoding to capture the relative positional relationships in the time series. S32, Transformer Encoding and Prediction: The features obtained after dimensional mapping with location information are input into the Transformer encoder module. Each encoder module is processed through layer normalization, multi-head attention mechanism, and residual connections. Specifically, after layer normalization, the input features are processed through three LoRA linear layers in the multi-head attention mechanism to generate a query matrix, key matrix, and value matrix. The LoRA linear layer is composed of the sum of the product of the original pre-trained weight matrix and the low-rank decomposition matrix. Parameter fine-tuning is achieved by training only the low-rank decomposition matrix. After passing through the output projection layer, the model is calculated by a feedforward network, and the features from the last time step are input into the prediction head to obtain the effect size prediction value. The Transformer encoder and prediction head are trained using training data from typical measurement points to obtain a pre-trained model. S33. Transfer Learning Fine-tuning: Using the pre-trained model trained in step S32 as the base model, freeze all parameters in the pre-trained model, retaining only the LoRA parameters and prediction head parameters; use monitoring data from other measurement points different from typical measurement points to fine-tune the unfrozen LoRA parameters and prediction head parameters, so that the fine-tuned model can be adapted to the intelligent prediction of the effect quantities of other measurement points of the dam.

10. The intelligent prediction method for dam operation performance based on feature fusion and model transfer as described in claim 9, characterized in that, In step S32, the query matrix, key matrix, and value matrix in the multi-head attention mechanism are generated as follows: after the input features are normalized by the layers, they are obtained by linear transformation through three independent LoRA linear layers; the output projection layer is also implemented using LoRA linear layers; in step S33, the loss function for fine-tuning training includes a domain adaptation loss term, which includes maximum mean difference loss and correlation alignment loss.

Images