Water conservancy risk prediction method based on evolution characteristics

Abstract

The invention discloses a water conservancy risk prediction method based on evolution characteristics, and belongs to the technical field of water conservancy projects. The method comprises the following steps: dividing a to-be-detected area into a plurality of grids, collecting and normalizing surface and internal water conservancy risk factors of each grid, constructing a surface and internal evolution incidence matrix, extracting collaborative strength and total evolution strength, and splicing to form an evolution characteristic matrix; determining an evolution characteristic transfer coefficient based on the distance between the to-be-predicted grid and the neighborhood grid, and obtaining a neighborhood evolution characteristic transfer matrix; and inputting the multi-moment evolution characteristic three-dimensional matrix and the neighborhood evolution characteristic transmission three-dimensional matrix into a water conservancy risk prediction network, and outputting a grid water conservancy risk score. According to the method, high-precision prediction of the water conservancy risk is realized by capturing a risk factor space-time evolution rule and a neighborhood transfer effect.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering technology, and specifically to a water conservancy risk prediction method based on evolutionary characteristics. Background Technology

[0002] Water conservancy risk prediction is a crucial technical foundation for river basin flood control scheduling, water conservancy project operation and management, and regional disaster prevention and mitigation decision-making. With the expansion of water conservancy projects and the increasing complexity of the hydrological environment, water conservancy risk prediction is gradually evolving from experience-based judgment to data-driven and model-based analysis. Among existing technologies, a commonly used method is a comprehensive evaluation approach based on water conservancy risk indicators.

[0003] Existing technologies of this type typically use regions or fixed spatial units as prediction targets. For each region, multiple water conservancy risk factors, such as rainfall, runoff, and soil moisture content, are collected and normalized. Subsequently, according to pre-set weights or evaluation rules, the water conservancy risk factors are weighted and aggregated to obtain the corresponding region's water conservancy risk value or risk level, which is used to characterize the region's water conservancy risk status at the current moment or in the short term.

[0004] In practical applications, water conservancy risk prediction methods based on index weighting usually assume that the relationship between various risk factors is relatively fixed and fail to fully consider the relationship between different risk factors in the evolution process. Therefore, it is difficult to accurately reflect the inherent correlation characteristics in the development process of water conservancy risks, which can easily lead to prediction bias in complex hydrological situations and result in low accuracy of water conservancy risk prediction. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides a water conservancy risk prediction method based on evolutionary characteristics, which solves the problem of low accuracy in water conservancy risk prediction in existing technologies.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a water conservancy risk prediction method based on evolutionary characteristics, comprising the following steps:

[0007] The area to be tested is divided into multiple grids, and surface and internal water risk factors are collected and normalized for each grid.

[0008] Based on the surface and internal hydraulic risk factors of each grid, construct the surface evolution correlation matrix and the internal evolution correlation matrix for each grid;

[0009] The co-existence strength and total evolution strength are obtained from the surface evolution correlation matrix and the internal evolution correlation matrix respectively, and then concatenated to obtain the evolution feature matrix;

[0010] Based on the distance between the grid to be predicted and its neighboring grids, the evolution feature transfer coefficients are obtained, and the evolution feature transfer matrix is ​​obtained.

[0011] Multiply the evolution feature transfer matrix by the evolution feature matrix of the corresponding neighborhood grid to obtain the neighborhood evolution feature transfer matrix.

[0012] The evolution feature matrices of each time step of the grid to be predicted are spliced ​​together to form a three-dimensional evolution feature matrix, and the evolution feature transfer matrices of each neighborhood are spliced ​​together to form a three-dimensional neighborhood evolution feature transfer matrix.

[0013] The three-dimensional matrix of evolutionary features and the three-dimensional matrix of neighborhood evolutionary features are input into the water conservancy risk prediction network to obtain the grid water conservancy risk score.

[0014] Furthermore, the surface water risk factors for each grid include: rainfall intensity, cumulative rainfall, and surface runoff;

[0015] The hydraulic risk factors within each grid include: seepage flow, pore water pressure, and sediment content in seepage water.

[0016] Furthermore, the process of constructing the surface evolution correlation matrix and the internal evolution correlation matrix includes:

[0017] In each grid, the water conservancy risk factor at time t is subtracted from the water conservancy risk factor at time t-1 to obtain the evolution variable at time t, where t is the number of the time dimension;

[0018] In each grid, the evolution variables of each surface hydraulic risk factor are used to construct a surface evolution correlation matrix. The elements in the surface evolution correlation matrix are the product of the evolution variables of any two surface risk factors.

[0019] In each grid, an internal evolution correlation matrix is ​​constructed using the evolution variables of each internal water conservancy risk factor. The elements in the internal evolution correlation matrix are the product of the evolution variables of any two internal risk factors.

[0020] Furthermore, the process of obtaining the evolutionary feature matrix includes:

[0021] The surface evolution correlation matrix is ​​solved to obtain multiple eigenvalues, and the largest eigenvalue is selected as the surface co-intensity.

[0022] The total intensity of surface evolution is obtained by adding the elements in the surface evolution correlation matrix. The surface co-intensity and the total intensity of surface evolution are used as elements to form the surface evolution feature vector.

[0023] The internal evolution correlation matrix is ​​solved to obtain multiple eigenvalues, and the largest eigenvalue is selected as the internal synergy strength.

[0024] The total internal evolution intensity is obtained by summing the elements in the internal evolution correlation matrix. The internal synergy intensity and the total internal evolution intensity are used as elements to form the internal evolution feature vector.

[0025] The surface evolution feature vector and the internal evolution feature vector are concatenated to form the evolution feature matrix.

[0026] Furthermore, the process of obtaining the evolutionary feature transfer matrix includes:

[0027] Extract the neighborhood grid of the grid to be predicted;

[0028] Calculate the distance between the neighboring grid and the grid to be predicted;

[0029] Based on the distance between the neighboring grid and the grid to be predicted, calculate the evolution feature transfer coefficient corresponding to each element in the evolution feature matrix of the neighboring grid;

[0030] The evolution feature transfer coefficients belonging to each neighboring grid are used to form the evolution feature transfer matrix of the corresponding neighboring grid.

[0031] Furthermore, the formula for calculating the evolutionary characteristic transmission coefficient is as follows:

[0032] ,

[0033] in, For the first The first neighborhood grid Evolutionary feature transmission coefficients For the first The distance between each neighboring grid and the grid to be predicted. For the first In the evolution feature matrix of the nth neighborhood grid One element, The number of neighboring grids, and It is a positive integer.

[0034] Furthermore, the formula for obtaining the neighborhood evolution feature transfer matrix is:

[0035] ,

[0036] in, The neighborhood evolution feature transfer matrix, For the first The evolutionary feature transfer matrix of a neighborhood grid. For the first The evolutionary feature matrix of a neighborhood grid, The number of neighboring grids, It is a positive integer. This is element-wise multiplication.

[0037] Furthermore, the water conservancy risk prediction network includes: a first time-series average pooling layer, a second time-series average pooling layer, a first time-series maximum pooling layer, a second time-series maximum pooling layer, a first evolution-transfer feature fusion unit, a second evolution-transfer feature fusion unit, a mean feature aggregation layer, a maximum feature aggregation layer, and a BP neural network;

[0038] Furthermore, the processing steps of the water conservancy risk prediction network include:

[0039] The evolution feature three-dimensional matrix is ​​subjected to time-dimensional average pooling processing using the first time-series average pooling layer to obtain the time-series mean feature.

[0040] A second temporal average pooling layer is used to perform time-dimensional average pooling on the three-dimensional matrix of neighborhood evolution feature transfer to obtain the mean feature of neighborhood temporal transfer.

[0041] The first temporal max pooling layer is used to perform time-dimensional max pooling on the three-dimensional matrix of evolution features to obtain the temporal maximum feature.

[0042] A second temporal max pooling layer is used to perform temporal max pooling on the three-dimensional matrix of neighborhood evolution feature propagation to obtain the neighborhood temporal propagation maximum value feature.

[0043] The first evolution-transmission feature fusion unit is used to fuse the temporal mean feature and the neighborhood temporal transmission mean feature to obtain the evolution-transmission fused mean feature;

[0044] A second evolution-transfer feature fusion unit is used to fuse the temporal maximum feature with the neighborhood temporal transfer maximum feature to obtain the evolution-transfer fused maximum feature.

[0045] A mean feature aggregation layer is used to aggregate the evolution-transmission fusion mean features to obtain mean aggregated features.

[0046] A maximum value feature aggregation layer is used to aggregate the evolution-transmission fusion maximum value features to obtain the maximum value aggregated features;

[0047] The mean aggregation feature and the maximum aggregation feature are input into the BP neural network to obtain the grid-based water conservancy risk score.

[0048] Furthermore, the expression for the first evolution-transfer feature fusion unit is:

[0049] ,

[0050] in, For evolution-transmission fusion mean characteristics, It is a time series mean feature. For Sigmoid layer, The mean feature is propagated over time in the neighborhood. For element-wise multiplication;

[0051] The expression for the second evolution-transfer feature fusion unit is:

[0052] ,

[0053] in, For the evolution-transfer fusion maximum feature, This is a time-series maximum value feature. The maximum value feature is propagated in the neighborhood time series.

[0054] The beneficial effects of this invention are as follows:

[0055] 1. This invention collects surface and internal water conservancy risk factors respectively, constructs surface evolution correlation matrices and internal evolution correlation matrices, and extracts synergistic strength and total evolution strength. It breaks through the limitation of existing technologies that "fail to fully consider the relationship between different risk factors in the evolution process", accurately depicts the dynamic correlation characteristics of different water conservancy risk factors in the evolution process, and reduces prediction bias in complex hydrological situations.

[0056] 2. This invention introduces an "evolutionary feature transmission coefficient" (based on grid distance) to associate the risk characteristics of neighboring grids with the grid to be predicted, which fully reflects the actual law of "near-field transmission and spatial diffusion" of water conservancy risks. Compared with the traditional "regional single evaluation", it can better reflect the risk transmission effect of local grids.

[0057] 3. This invention calculates the evolution feature transfer coefficient based on the distance between the grid to be predicted and its neighboring grids, and constructs an evolution feature transfer matrix, fully considering the spatial transmission effect of water conservancy risks. Compared with the shortcomings of existing technologies that ignore spatial correlation, this invention combines the evolution features of the grid to be predicted with the transfer features of its neighboring grids, forming a dual feature dimension of "temporal evolution + spatial transfer". The three-dimensional matrix of evolution features and the three-dimensional matrix of neighboring evolution feature transfer are input into the water conservancy risk prediction network to obtain the grid water conservancy risk score, thereby improving the accuracy of water conservancy risk prediction. Attached Figure Description

[0058] Figure 1 A flowchart of a water conservancy risk prediction method based on evolutionary characteristics;

[0059] Figure 2 This is a schematic diagram of the structure of a water conservancy risk prediction network. Detailed Implementation

[0060] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0061] like Figure 1 As shown, a water conservancy risk prediction method based on evolutionary characteristics includes the following steps:

[0062] The area to be tested is divided into multiple grids, and surface and internal water risk factors are collected and normalized for each grid.

[0063] Based on the surface and internal hydraulic risk factors of each grid, construct the surface evolution correlation matrix and the internal evolution correlation matrix for each grid;

[0064] The co-existence strength and total evolution strength are obtained from the surface evolution correlation matrix and the internal evolution correlation matrix respectively, and then concatenated to obtain the evolution feature matrix;

[0065] Based on the distance between the grid to be predicted and its neighboring grids, the evolution feature transfer coefficients are obtained, and the evolution feature transfer matrix is ​​obtained.

[0066] Multiply the evolution feature transfer matrix by the evolution feature matrix of the corresponding neighborhood grid to obtain the neighborhood evolution feature transfer matrix.

[0067] The evolution feature matrices of the grid to be predicted at each time step are spliced ​​together to form a three-dimensional evolution feature matrix, and the neighborhood evolution feature transfer matrices at each time step are spliced ​​together to form a three-dimensional neighborhood evolution feature transfer matrix.

[0068] The three-dimensional matrix of evolutionary features and the three-dimensional matrix of neighborhood evolutionary features are input into the water conservancy risk prediction network to obtain the grid water conservancy risk score.

[0069] In this embodiment, the surface water risk factors for each grid include: rainfall intensity, cumulative rainfall, and surface runoff;

[0070] The hydraulic risk factors within each grid include: seepage flow, pore water pressure, and sediment content in seepage water.

[0071] After collecting surface and internal water risk factors, each water risk factor needs to be normalized.

[0072] In this embodiment, the process of constructing the surface evolution correlation matrix and the internal evolution correlation matrix includes:

[0073] In each grid, the water conservancy risk factor at time t is subtracted from the water conservancy risk factor at time t-1 to obtain the evolution variable at time t, where t is the number of the time dimension;

[0074] In each grid, the evolution variables of each surface hydraulic risk factor are used to construct a surface evolution correlation matrix. The elements in the surface evolution correlation matrix are the product of the evolution variables of any two surface risk factors.

[0075] In each grid, an internal evolution correlation matrix is ​​constructed using the evolution variables of each internal water conservancy risk factor. The elements in the internal evolution correlation matrix are the product of the evolution variables of any two internal risk factors.

[0076] The formula for the elements of the surface evolution correlation matrix is:

[0077] ,

[0078] in, For the first The first time in the surface evolution correlation matrix Line number Column surface evolution correlation values, For the first in the grid The first moment Evolutionary variables of surface water conservancy risk factors For the first in the grid The first moment The evolution variables of surface water conservancy risk factors The value range is 1, 2, 3. The value range is 1, 2, 3. When the surface water risk factor equals 1, it corresponds to the rainfall intensity. When the value equals 2, the surface water conservancy risk factor corresponds to the cumulative rainfall. When the value equals 3, the surface water conservancy risk factor corresponds to surface runoff. When the value equals 1, the internal water conservancy risk factor corresponds to the seepage flow. When the value equals 2, the internal hydraulic risk factor corresponds to pore water pressure. When the value is equal to 3, the internal water conservancy risk factor corresponds to the sediment content of seepage water. In the matrix, i is the row number and j is the column number.

[0079] The formula for the elements of the internal evolutionary correlation matrix is:

[0080] ,

[0081] in, For the first The first time in the internal evolution correlation matrix Line number The internal evolution of the column is related to the value. For the first in the grid The first moment The evolution variables of internal water conservancy risk factors For the first in the grid The first moment Evolutionary variables of internal water conservancy risk factors.

[0082] This invention uses the difference between water conservancy risk factors at adjacent time points as evolutionary variables to reflect the magnitude and direction of the changes in each water conservancy risk factor over time, enabling the constructed evolutionary correlation matrix to characterize the dynamic evolution process of water conservancy risks. This invention constructs the elements of the evolutionary correlation matrix by multiplying the evolutionary variables of any two water conservancy risk factors, so that the values ​​in the matrix simultaneously contain the magnitude of changes in both risk factors and their synergistic change characteristics. This allows it to describe the strength of the evolutionary correlation between different water conservancy risk factors at the same time point, enhancing the ability to characterize the intrinsic relationships of complex water conservancy processes.

[0083] The surface evolution correlation matrix quantifies the evolutionary coupling relationship between three types of surface factors—rainfall intensity, cumulative rainfall, and surface runoff—by multiplying the evolutionary variables at adjacent time points. The internal evolution correlation matrix characterizes the permeability-structure response relationship within the soil and rock mass by multiplying the evolutionary variables for three types of internal factors—seepage flow, pore water pressure, and sediment content in seepage water.

[0084] In this embodiment, the process of obtaining the evolution feature matrix includes:

[0085] The surface evolution correlation matrix is ​​solved to obtain multiple eigenvalues, and the largest eigenvalue is selected as the surface co-intensity.

[0086] The total intensity of surface evolution is obtained by adding the elements in the surface evolution correlation matrix. The surface co-intensity and the total intensity of surface evolution are used as elements to form the surface evolution feature vector.

[0087] The internal evolution correlation matrix is ​​solved to obtain multiple eigenvalues, and the largest eigenvalue is selected as the internal synergy strength.

[0088] The total internal evolution intensity is obtained by summing the elements in the internal evolution correlation matrix. The internal synergy intensity and the total internal evolution intensity are used as elements to form the internal evolution feature vector.

[0089] The surface evolution feature vector and the internal evolution feature vector are concatenated to form the evolution feature matrix.

[0090] In this embodiment, the surface evolution feature vector is used as the first row of the matrix, and the internal evolution feature vector is used as the second row of the matrix, forming a 4×4 evolution feature matrix.

[0091] The surface evolution correlation matrix and the internal evolution correlation matrix are symmetric matrices.

[0092] This invention solves for the eigenvalues ​​of both the surface evolution correlation matrix and the internal evolution correlation matrix, selecting the largest eigenvalue as the synergistic strength. This ensures that the obtained synergistic strength reflects the dominant synergistic change pattern during the evolution of multiple water conservancy risk factors, thus avoiding the excessive influence of a single risk factor change on the overall evolution characteristics. Furthermore, by summing the elements in the evolution correlation matrix, the total evolution strength is obtained, enabling this feature to reflect the overall evolutionary magnitude of multiple water conservancy risk factors at the same time, characterizing the overall strength of water conservancy risk changes from a macroscopic perspective.

[0093] In this embodiment, the process of obtaining the evolution feature transfer matrix includes:

[0094] Extract the neighborhood grid of the grid to be predicted. The neighborhood grid is the grid that is in contact with the grid to be predicted.

[0095] Calculate the distance between the neighboring grid and the grid to be predicted;

[0096] Based on the distance between the neighboring grid and the grid to be predicted, calculate the evolution feature transfer coefficient corresponding to each element in the evolution feature matrix of the neighboring grid;

[0097] The evolution feature transfer coefficients belonging to each neighboring grid are used to form the evolution feature transfer matrix of the corresponding neighboring grid. At the same time, a neighboring grid corresponds to 4 evolution feature transfer coefficients.

[0098] In this embodiment, the formula for calculating the evolutionary feature transfer coefficient is:

[0099] ,

[0100] in, For the first The first neighborhood grid Evolutionary feature transmission coefficients For the first The distance between each neighboring grid and the grid to be predicted. For the first In the evolution feature matrix of the nth neighborhood grid One element, The number of neighboring grids, and It is a positive integer.

[0101] In this embodiment, the formula for obtaining the neighborhood evolution feature transfer matrix is:

[0102] ,

[0103] in, The neighborhood evolution feature transfer matrix, For the first The evolutionary feature transfer matrix of a neighborhood grid. For the first The evolutionary feature matrix of a neighborhood grid, The number of neighboring grids, It is a positive integer. This is element-wise multiplication.

[0104] Neighborhood evolution feature transfer matrix The size is 4×4.

[0105] This invention introduces a neighboring grid of the grid to be predicted and constructs an evolution feature transfer coefficient based on the spatial distance between the neighboring grid and the grid to be predicted. This allows the evolution features of the neighboring grid to have different degrees of influence on the grid to be predicted according to their spatial proximity, thereby reflecting the spatial diffusion and transmission characteristics of water conservancy risks.

[0106] The evolution feature transmission coefficient is related to both the evolution feature matrix elements of the neighboring grid and the neighborhood distance. This results in neighboring grids that are closer and have stronger evolution features having a higher weight in the transmission process, while the influence of neighboring grids that are farther away or have weaker evolution features is weakened accordingly.

[0107] This invention first Give the neighborhood grid Each risk feature is assigned a specific transfer weight, and then the matrices obtained by multiplying each feature element-wise are summed element-wise to obtain the final result. By using the Hadamard product, each core risk feature of the neighborhood grid (surface cooperative strength, total surface strength, internal cooperative strength, and total internal strength) is given a dedicated transfer weight, avoiding cross-interference between different types of features. Then, by summing and integrating the risk contributions of all neighborhoods element by element, the constructed neighborhood evolution feature transfer matrix can simultaneously reflect the strength of neighborhood evolution features and their spatial transfer relationships, thus enhancing the integrity and stability of neighborhood evolution information expression.

[0108] like Figure 2 As shown, the water conservancy risk prediction network includes: a first time-series average pooling layer, a second time-series average pooling layer, a first time-series maximum pooling layer, a second time-series maximum pooling layer, a first evolution-transfer feature fusion unit, a second evolution-transfer feature fusion unit, a mean feature aggregation layer, a maximum feature aggregation layer, and a BP neural network.

[0109] like Figure 2 As shown, the processing steps of the water conservancy risk prediction network include:

[0110] The evolution feature three-dimensional matrix is ​​subjected to time-dimensional average pooling processing using the first time-series average pooling layer to obtain the time-series mean feature.

[0111] A second temporal average pooling layer is used to perform time-dimensional average pooling on the three-dimensional matrix of neighborhood evolution feature transfer to obtain the mean feature of neighborhood temporal transfer.

[0112] The first temporal max pooling layer is used to perform time-dimensional max pooling on the three-dimensional matrix of evolution features to obtain the temporal maximum feature.

[0113] A second temporal max pooling layer is used to perform temporal max pooling on the three-dimensional matrix of neighborhood evolution feature propagation to obtain the neighborhood temporal propagation maximum value feature.

[0114] The first evolution-transmission feature fusion unit is used to fuse the temporal mean feature and the neighborhood temporal transmission mean feature to obtain the evolution-transmission fused mean feature;

[0115] A second evolution-transfer feature fusion unit is used to fuse the temporal maximum feature with the neighborhood temporal transfer maximum feature to obtain the evolution-transfer fused maximum feature.

[0116] A mean feature aggregation layer is used to aggregate the evolution-transmission fusion mean features to obtain mean aggregated features.

[0117] A maximum value feature aggregation layer is used to aggregate the evolution-transmission fusion maximum value features to obtain the maximum value aggregated features;

[0118] Using the mean aggregation feature and the maximum aggregation feature as input features, we obtain 8 input elements, which are then input into a BP neural network to obtain a grid-based water conservancy risk score.

[0119] This invention employs time-dimensional average pooling and max pooling on both the three-dimensional matrix of evolutionary features and the three-dimensional matrix of neighborhood evolutionary feature transmission. This allows for the simultaneous preservation of the overall change level and extreme value change characteristics during the evolution of water conservancy risks. Furthermore, the local evolutionary features and neighborhood transmission features are fused in the evolution-transmission feature fusion unit, thus simultaneously characterizing the temporal evolution and spatial transmission impact of water conservancy risks at the feature level. Subsequently, mean-type features and maximum-type features are aggregated to effectively compress and enhance risk representations under different statistical meanings. Finally, a BP neural network is used to perform nonlinear mapping on the aggregated multi-source features to obtain a gridded water conservancy risk score that comprehensively reflects both temporal evolution characteristics and neighborhood influence characteristics, thereby improving the completeness of risk feature expression and the accuracy of prediction results.

[0120] In this embodiment, the pooling kernel size of the first time-series average pooling layer, the second time-series average pooling layer, the first time-series max pooling layer, and the second time-series max pooling layer is 1×5. The mean (local long-term trend) and the maximum (local peak fluctuation) within the window are extracted using "five consecutive moments" as a time window, with a step size of 3 for each.

[0121] The time-series mean feature, neighborhood time-series transitive mean feature, time-series maximum feature, and neighborhood time-series transitive maximum feature are all three-dimensional matrices. Pooling extracts features along the time dimension, which is equivalent to reducing the number of parameters in the time dimension.

[0122] In this embodiment, the expression for the first evolution-transfer feature fusion unit is:

[0123] ,

[0124] in, For evolution-transmission fusion mean characteristics, It is a time series mean feature. For Sigmoid layer, The mean feature is propagated over time in the neighborhood. This is element-wise multiplication.

[0125] In this embodiment, the expression for the second evolution-transfer feature fusion unit is:

[0126] ,

[0127] in, For the evolution-transfer fusion maximum feature, This is a time-series maximum value feature. The maximum value feature is propagated in the neighborhood time series.

[0128] In the first evolution-transfer feature fusion unit, this invention uses the nonlinear mapping of the Sigmoid function to dynamically generate adaptive weights for the temporal mean feature and the neighborhood temporal transfer mean feature, respectively. Then, the two types of features are weighted and fused, so that the temporal mean feature of the feature itself (reflecting the long-term evolution trend of grid risk) and the temporal transfer mean feature of the neighborhood (reflecting the long-term spatial contribution of surrounding grid risk) can automatically adjust their proportions in the fusion result according to their respective feature strengths. When the risk trend of the feature itself is significant, its weight is increased to dominate the fused feature; when the risk contribution of the neighborhood is stronger, the weight of the neighborhood feature is amplified to strengthen the spatial influence. In the end, the "risk trend of the feature itself" and the "spatial transfer contribution" are accurately and dynamically complemented, avoiding the feature imbalance problem caused by fixed weight fusion.

[0129] In the second evolution-transmission feature fusion unit, by leveraging the nonlinear mapping of the Sigmoid function, adaptive weights are dynamically generated and weighted for the temporal maximum feature and the neighborhood temporal transmission maximum feature, respectively. This allows the proportion of the self-timer maximum feature (reflecting the sudden peak fluctuation of grid risk) and the neighborhood temporal transmission maximum feature (reflecting the spatial transmission impact of sudden risks in surrounding grids) to automatically adjust according to the strength of their respective peak signals. When the self-timer sudden risk signal is significant, its weight is increased to highlight local anomalies; when the transmission impact of the neighborhood sudden risk is stronger, the weight of the neighborhood feature is amplified to strengthen the linkage effect of spatial sudden risks. Ultimately, this achieves dynamic complementarity between "self-timer sudden risk peak" and "transmission of surrounding sudden risks," effectively amplifying the abnormal signals of extreme and sudden water conservancy risks and solving the problem of insufficient capture of sudden risks by fixed-weight fusion.

[0130] In this embodiment, the expression for the mean feature aggregation layer is:

[0131] , ,

[0132] in, This is a mean aggregation feature. for The Middle Two-dimensional feature slices at time step It is a 2-norm. for Time dimension, For the first Mean weights in the time dimension, The length is the time dimension.

[0133] In this embodiment, the expression for the maximum value feature aggregation layer is:

[0134] , ,

[0135] in, For maximum aggregation features, for The Middle Two-dimensional feature slices at time step For the first The maximum weight of the time dimension.

[0136] In this embodiment, the grid-based water conservancy risk score is set to a quantitative range of 0-100 points. A score of 0-20 indicates that the grid is in a low-risk state, with gradual changes in water conservancy risk factors, weak evolution characteristics and neighborhood transmission characteristics, and relatively stable water conservancy system operation. A score of 21-50 indicates that the grid is in a low-to-medium-risk state, with some water conservancy risk factors in the area showing some evolution or neighborhood influence, but no significant risk accumulation has yet occurred. A score of 51-80 indicates that the grid is in a medium-to-high-risk state, with obvious evolution characteristics of water conservancy risk factors in the area, and enhanced neighborhood transmission effects, indicating a possibility of water conservancy risk events. A score of 81-100 indicates that the grid is in a high-risk state, with high levels of water conservancy risk evolution intensity and neighborhood transmission influence, a high probability of water conservancy risk occurrence, and requiring close monitoring.

[0137] This invention compresses a three-dimensional matrix in the time dimension, quantifies the overall intensity of the feature slice at each time step using the second norm, and then normalizes it into weights—the stronger the risk signal (the larger the second norm of the feature slice), the higher the corresponding weight, and the greater its proportion in the aggregation result. By weighting each slice, the three-dimensional matrix is ​​compressed into a two-dimensional matrix.

[0138] This invention collects surface and internal hydraulic risk factors, constructs surface evolution correlation matrices and internal evolution correlation matrices, and extracts synergistic strength and total evolution strength. It overcomes the limitation of existing technologies that "fail to fully consider the relationship between different risk factors in the evolution process", accurately depicts the dynamic correlation characteristics of different hydraulic risk factors in the evolution process, and reduces prediction bias in complex hydrological situations.

[0139] This invention introduces an "evolutionary feature transmission coefficient" (based on grid distance) to associate the risk characteristics of neighboring grids with the grid to be predicted, fully reflecting the actual law of "near-field transmission and spatial diffusion" of water conservancy risks. Compared with the traditional "regional single evaluation", it can better reflect the risk transmission effect of local grids.

[0140] This invention calculates the evolution feature transfer coefficient based on the distance between the grid to be predicted and its neighboring grids, constructs an evolution feature transfer matrix, and fully considers the spatial transmission effect of water conservancy risks. Compared with the shortcomings of existing technologies that ignore spatial correlation, this invention combines the evolution features of the grid to be predicted with the transfer features of its neighboring grids, forming a dual feature dimension of "temporal evolution + spatial transfer". The three-dimensional matrix of evolution features and the three-dimensional matrix of neighboring evolution feature transfer are input into the water conservancy risk prediction network to obtain the grid water conservancy risk score, thereby improving the accuracy of water conservancy risk prediction.

[0141] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A water conservancy risk prediction method based on evolutionary characteristics, characterized in that, Includes the following steps: The area to be tested is divided into multiple grids, and surface and internal water risk factors are collected and normalized for each grid. Based on the surface and internal hydraulic risk factors of each grid, construct the surface evolution correlation matrix and the internal evolution correlation matrix for each grid; The co-existence strength and total evolution strength are obtained from the surface evolution correlation matrix and the internal evolution correlation matrix respectively, and then concatenated to obtain the evolution feature matrix; Based on the distance between the grid to be predicted and its neighboring grids, the evolution feature transfer coefficients are obtained, and the evolution feature transfer matrix is ​​obtained. Multiply the evolution feature transfer matrix by the evolution feature matrix of the corresponding neighborhood grid to obtain the neighborhood evolution feature transfer matrix. The evolution feature matrices of each time step of the grid to be predicted are spliced ​​together to form a three-dimensional evolution feature matrix, and the evolution feature transfer matrices of each neighborhood are spliced ​​together to form a three-dimensional neighborhood evolution feature transfer matrix. The three-dimensional matrix of evolutionary features and the three-dimensional matrix of neighborhood evolutionary features are input into the water conservancy risk prediction network to obtain the grid water conservancy risk score.

2. The water conservancy risk prediction method based on evolutionary characteristics according to claim 1, characterized in that, The surface water risk factors for each grid include: rainfall intensity, cumulative rainfall, and surface runoff; The hydraulic risk factors within each grid include: seepage flow, pore water pressure, and sediment content in seepage water.

3. The water conservancy risk prediction method based on evolutionary characteristics according to claim 1, characterized in that, The process of constructing the surface evolution correlation matrix and the internal evolution correlation matrix includes: In each grid, the water conservancy risk factor at time t is subtracted from the water conservancy risk factor at time t-1 to obtain the evolution variable at time t, where t is the number of the time dimension; In each grid, the evolution variables of each surface hydraulic risk factor are used to construct a surface evolution correlation matrix. The elements in the surface evolution correlation matrix are the product of the evolution variables of any two surface risk factors. In each grid, an internal evolution correlation matrix is ​​constructed using the evolution variables of each internal water conservancy risk factor. The elements in the internal evolution correlation matrix are the product of the evolution variables of any two internal risk factors.

4. The water conservancy risk prediction method based on evolutionary characteristics according to claim 1, characterized in that, The process of obtaining the evolutionary feature matrix includes: The surface evolution correlation matrix is ​​solved to obtain multiple eigenvalues, and the largest eigenvalue is selected as the surface co-intensity. The total intensity of surface evolution is obtained by adding the elements in the surface evolution correlation matrix. The surface co-intensity and the total intensity of surface evolution are used as elements to form the surface evolution feature vector. The internal evolution correlation matrix is ​​solved to obtain multiple eigenvalues, and the largest eigenvalue is selected as the internal synergy strength. The total internal evolution intensity is obtained by summing the elements in the internal evolution correlation matrix. The internal synergy intensity and the total internal evolution intensity are used as elements to form the internal evolution feature vector. The surface evolution feature vector and the internal evolution feature vector are concatenated to form the evolution feature matrix.

5. The water conservancy risk prediction method based on evolutionary characteristics according to claim 1, characterized in that, The process of obtaining the evolutionary feature transfer matrix includes: Extract the neighborhood grid of the grid to be predicted; Calculate the distance between the neighboring grid and the grid to be predicted; Based on the distance between the neighboring grid and the grid to be predicted, calculate the evolution feature transfer coefficient corresponding to each element in the evolution feature matrix of the neighboring grid; The evolution feature transfer coefficients belonging to each neighboring grid are used to form the evolution feature transfer matrix of the corresponding neighboring grid.

6. The water conservancy risk prediction method based on evolutionary characteristics according to claim 1 or 5, characterized in that, The formula for calculating the evolutionary characteristic transmission coefficient is: ， in, For the first The first neighborhood grid Evolutionary feature transmission coefficients For the first The distance between each neighboring grid and the grid to be predicted. For the first In the evolution feature matrix of the nth neighborhood grid One element, The number of neighboring grids, and It is a positive integer.

7. The water conservancy risk prediction method based on evolutionary characteristics according to claim 1, characterized in that, The formula for obtaining the neighborhood evolution feature transfer matrix is: ， in, The neighborhood evolution feature transfer matrix, For the first The evolutionary feature transfer matrix of a neighborhood grid. For the first The evolutionary feature matrix of a neighborhood grid, The number of neighboring grids, It is a positive integer. This is element-wise multiplication.

8. The water conservancy risk prediction method based on evolutionary characteristics according to claim 1, characterized in that, The water conservancy risk prediction network includes: a first time-series average pooling layer, a second time-series average pooling layer, a first time-series maximum pooling layer, a second time-series maximum pooling layer, a first evolution-transfer feature fusion unit, a second evolution-transfer feature fusion unit, a mean feature aggregation layer, a maximum feature aggregation layer, and a BP neural network.

9. The water conservancy risk prediction method based on evolutionary characteristics according to claim 1 or 8, characterized in that, The processing steps of the water conservancy risk prediction network include: The evolution feature three-dimensional matrix is ​​subjected to time-dimensional average pooling processing using the first time-series average pooling layer to obtain the time-series mean feature. A second temporal average pooling layer is used to perform time-dimensional average pooling on the three-dimensional matrix of neighborhood evolution feature transfer to obtain the mean feature of neighborhood temporal transfer. The first temporal max pooling layer is used to perform time-dimensional max pooling on the three-dimensional matrix of evolution features to obtain the temporal maximum feature. A second temporal max pooling layer is used to perform temporal max pooling on the three-dimensional matrix of neighborhood evolution feature propagation to obtain the neighborhood temporal propagation maximum value feature. The first evolution-transmission feature fusion unit is used to fuse the temporal mean feature and the neighborhood temporal transmission mean feature to obtain the evolution-transmission fused mean feature; A second evolution-transfer feature fusion unit is used to fuse the temporal maximum feature with the neighborhood temporal transfer maximum feature to obtain the evolution-transfer fused maximum feature. A mean feature aggregation layer is used to aggregate the evolution-transmission fusion mean features to obtain mean aggregated features. A maximum value feature aggregation layer is used to aggregate the evolution-transmission fusion maximum value features to obtain the maximum value aggregated features; The mean aggregation feature and the maximum aggregation feature are input into the BP neural network to obtain the grid-based water conservancy risk score.

10. The water conservancy risk prediction method based on evolutionary characteristics according to claim 9, characterized in that, The expression for the first evolution-transfer feature fusion unit is: ， in, For evolution-transmission fusion mean characteristics, It is a time series mean feature. For Sigmoid layer, The mean feature is propagated over time in the neighborhood. For element-wise multiplication; The expression for the second evolution-transfer feature fusion unit is: ， in, For the evolution-transfer fusion maximum feature, This is a time-series maximum value feature. The maximum value feature is propagated in the neighborhood time series.

Images