Small and medium-sized reservoir operation safety evaluation system based on sky-ground water conservancy project

Abstract

The invention provides a small and medium-sized reservoir operation safety evaluation system based on sky-ground work, and belongs to the technical field of hydrological monitoring. The device comprises a space-time alignment layer, a feature weighting layer and a dynamic coupling layer which are connected in sequence, the space-time alignment layer is used for performing time alignment and space alignment on the input data of the sky-land water conservancy project; the feature weighting layer is used for quantifying contribution degrees of different monitoring features to safety assessment according to the data after time-space alignment; and the dynamic coupling layer adopts a dynamic weight distribution strategy for different modules through a multi-module coordination engine, fuses feature weighting to obtain a fusion result, and outputs a small and medium-sized reservoir operation safety index according to the fusion result. The evaluation result is more accurate, and the timeliness is better.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrological monitoring, and in particular to a safety assessment system for the operation of medium and small-sized reservoirs based on sky-earth-hydraulic engineering. Background Art

[0002] Medium and small-sized reservoirs play an important role in China's water conservancy project system. Especially those with earth-rock dams are numerous and widely distributed, undertaking multiple functions such as flood control, irrigation, water supply, and power generation, and playing an indispensable role in regional economic development and people's livelihood. However, limited by factors such as the construction era, technical level, and capital investment, some medium and small-sized reservoirs face many potential safety hazards during operation. Traditional monitoring means often have problems such as untimely information acquisition, insufficient data accuracy, and limited monitoring scope, making it difficult to meet the real-time and comprehensive monitoring requirements for the safe operation of reservoirs. Against the background of frequent extreme weather events, the safety risks of medium and small-sized reservoirs are further highlighted. Once accidents such as dam breaches occur, they will cause huge losses to the lives and property of people downstream, seriously affecting social stability and sustainable economic development.

[0003] Currently, the evaluation methods for the operation safety of medium and small-sized reservoirs mainly rely on image recognition, and a dam safety assessment model system constructed based on the static basic data of the reservoir dam, the real-time monitoring data of the project entity, and the runoff conditions. The source data is relatively single, only able to reflect the working conditions of local parts of the project, and has not been coordinated and integrated in multiple aspects such as time and space, scope, accuracy, and frequency, and has not realized the comprehensive evaluation of all elements, the whole process, and the entire time domain of the reservoir project.

[0004] Traditional dam safety assessment models have two major defects:

[0005] Single data dimension: Only relying on limited factors such as the working conditions of the dam body (displacement, settlement) and water conditions (reservoir water level, rainfall), ignoring the systematic correlation between the basin - reservoir area - dam body;

[0006] Poor monitoring timeliness: Relying on manual inspections and fixed sensors, it is impossible to achieve high-frequency, high-precision, and full-space coverage monitoring, resulting in a lag in the discovery of potential hazards. Summary of the Invention

[0007] To solve the technical problems existing in the prior art, the present invention provides a safety assessment system for the operation of medium and small-sized reservoirs based on sky-earth-hydraulic engineering, which realizes.

[0008] The present application provides a safety assessment system for the operation of medium and small-sized reservoirs based on sky-earth-hydraulic engineering, including a time-space alignment layer, a feature weighting layer, and a dynamic coupling layer connected in sequence;

[0009] The time-space alignment layer performs time alignment and space alignment on the input sky-earth-hydraulic engineering data;

[0010] A feature weighting layer that quantifies the contribution of different monitoring features to safety assessment based on the spatio-temporally aligned data;

[0011] A dynamic coupling layer that, through a multi-module coordination engine, adopts a dynamic weight allocation strategy for different modules, fuses the feature weighting to obtain a fusion result, and outputs the operation safety index of small and medium-sized reservoirs according to the fusion result, where the multi-modules include a short-term risk prediction module, a medium- and long-term stability prediction module, and an extreme condition deduction module; the short-term risk prediction module is used for the safety assessment of the dam in real time or within a preset first time period, the medium- and long-term stability prediction module assesses the stability of the dam within a preset number of years, and the extreme condition deduction module evaluates the structural safety under extreme conditions by simulating the impact of overloading events on the dam body.

[0012] Preferably, the time alignment of the input data is specifically achieved by dynamically bending the time axis for data with different sampling frequencies to realize time series alignment, and cubic spline interpolation is used for low-frequency data below the preset frequency threshold to generate preset high-frequency time point data.

[0013] Preferably, dynamic time warping is adopted for data with different sampling frequencies, and the specific formula is as follows:

[0014]

[0015] Wherein,

[0016] α m is the weight coefficient of the m-th time point of the data to be aligned;

[0017] γ is the time decay coefficient;

[0018] t m is the m-th time point of the reference time series;

[0019] t k is the time point of the data to be aligned;

[0020] t m′ is the time point of the reference time series;

[0021] M is the total number of time points of the reference sequence.

[0022] Preferably, the following formula is used to generate preset high-frequency time point data for low-frequency data below the preset frequency threshold:

[0023]

[0024] Wherein,

[0025] a, b, c, d are the coefficients of the cubic polynomial, determined by cubic spline interpolation to ensure that the interpolation function is in the interval [ti ,t i+1

[0026] be continuous within and have continuous first and second derivatives;

[0027] t is the target time point to be interpolated, located within the interval [t i , t i+1 ;

[0028] t i is the previous time point in the original low-frequency data;

[0029] t i+1 is the next time point adjacent to t i in the original low-frequency data;

[0030] is the estimated value obtained by cubic spline interpolation at time t.

[0031] Preferably, a, b, c, and d are determined by the cubic spline interpolation conditions to achieve smooth high-frequency data interpolation. The calculation methods of a, b, c, and d are as follows:

[0032] When t = t k , substituting into the formula gives d = X(t k ), that is, d is the value of the original low-frequency data at t k . Then, a three-moment equation is constructed to solve for M k . Cubic spline interpolation requires continuous second derivatives. The second derivative values M k (M k is the second derivative value at t k , and the interval length h k = t k+1 - t k ) are obtained through the three-moment equation;

[0033] Then, a, b, and c are calculated respectively using the following formulas:

[0034]

[0035] The quadratic term coefficient is directly determined by the second derivative value at t k :

[0036]

[0037] Combining the interval function value difference and the second derivative value, the linear term coefficient is determined.

[0038] Preferably, the following method is used for spatial alignment of data: All data are transformed into the same coordinate system, the Gauss-Krüger projection is used to eliminate deformation, and the spatial interpolation kernel function is used for discrete monitoring point data to generate a continuous spatial distribution.​

[0039] Preferably, the spatial interpolation kernel function is specifically as follows:

[0040]

[0041] wherein,

[0042] r is the normalized spatial distance;

[0043] K(r) is the spatial interpolation weight.

[0044] Preferably, the feature weighting layer adopts feature selection based on the attention mechanism, and dynamically adjusts the feature weights by combining context-aware attention with the environmental conditions.

[0045] Preferably, the features in the sky layer are compressed in weight using the following formula to reduce the dimension of the original features:

[0046]

[0047] wherein,

[0048] is the PCA projection matrix, obtained by performing eigenvalue decomposition on the covariance matrix of the original data;

[0049] is the transpose of the l-th original feature vector after alignment, n l is the total number of original features;

[0050] is the l-th compressed feature in the sky layer, and k represents the dimension after dimension reduction.

[0051] Preferably, in the cross-layer of the context vectors of sky + ground + environment, the features are compressed in weight using the following formula:

[0052]

[0053] wherein,

[0054] β j is the attention weight of the j-th layer;

[0055] C env is the environmental context vector;

[0056] Z j is the compressed feature of the j-th layer;

[0057] p is the summation index, with a value range from 1 to 4, corresponding to different layers;

[0058] W q is the learnable weight vector, is Wq Transpose of;

[0059] || is the concatenation operator that concatenates Z j and C env along a dimension to form a comprehensive vector containing hierarchical features and environmental information.

[0060] Preferably, the feature weighting layer adopts a static and dynamic hybrid weighting guided by expert knowledge. The static weight presets the basic weight according to expert experience, and the dynamic weight is adjusted according to the real-time working conditions.

[0061] Preferably, the short-term prediction module uses the LSTM-Transformer model to mine local and global features in the reservoir operation data. The LSTM-Transformer model includes an LSTM layer, a Transformer self-attention layer, and a fully connected layer. The LSTM layer captures the temporal relationship of data features, the Transformer self-attention layer identifies the spatial attention weights between multi-sensor data features, and the fully connected layer outputs the dam-break probability using the Sigmoid activation function.

[0062] Preferably, the medium- and long-term stability prediction module constructs a physics-guided neural network architecture, inputs low-frequency cumulative data and environmental parameters, takes the embedded constitutive equation of the earth-rock dam as a physical constraint, and takes the non-linear relationship between the historical monitoring data learned by the multi-layer perceptron and the geological parameters as data-driven.

[0063] Preferably, the extreme working condition deduction module constructs a three-dimensional model of the dam body-foundation-reservoir area based on BIM, integrates the geometric model, material properties, and safety boundary conditions, establishes different parameterized scenario libraries, then performs coupled solutions for multiple scenarios to obtain the pore water pressure distribution, updates the effective stress field, and based on the elastoplastic constitutive equation, iteratively solves for displacement and strain, and extracts the safety factor FS.

[0064] Preferably, the geometric model includes the geometric models of the terrain, dam structure, and flood discharge facilities. The material properties include the constitutive parameters of earth-rock materials and the time-varying characteristics of the permeability coefficient. The safety boundary conditions include the reservoir water pressure, ground motion input, and seepage boundary.

[0065] Preferably, the dynamic weight allocation strategy defines a rule base through a fuzzy logic controller to achieve the weight adaption of different models, and updates the feature weighting parameters in real time based on the evaluation error.

[0066] Preferably, if the reservoir water level > warning water level, then the weight of the short-term prediction module = 0.8; if the seismic intensity ≥ magnitude 6, then the weight of the extreme working condition deduction module = 0.7; if the reservoir water level is extremely high and the rainfall intensity is heavy rain, then the weight of the short-term prediction module = 0.8, the weight of the medium- and long-term stability prediction module = 0.1, and the weight of the extreme working condition deduction module = 0.1.

[0067] Preferably, after the output result, compare the predicted result with the actual monitoring data, trigger the retraining of the models in each module, inject the new monitoring data into the training set, update the feature weighting parameters, and the adaptive correction equation based on the adjustment of the model parameters is as follows:

[0068]

[0069] Wherein,

[0070] s is an index used to identify the s-th data point or the s-th monitoring index;

[0071] is the s-th original monitoring data;

[0072] is the s-th predicted value based on Kalman filtering;

[0073] σ is the standard deviation of historical data;

[0074] is the s-th corrected sensor data value;

[0075] ω is the weighting coefficient. When the data deviation exceeds 3σ, balance the weight of the original data and the predicted value .

[0076] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0077] 1. The present invention establishes a three-level calculation system of "spatiotemporal alignment - feature weighting - dynamic coupling", breaking through the traditional single-point static evaluation mode, and the evaluation result is more accurate and has better timeliness.

[0078] 2. The present invention performs full-dimensional alignment of data: it supports data fusion at the scale from seconds to years in time, and the spatial resolution reaches sub-meter level, and can eliminate more than 75% of the data conflicts caused by spatiotemporal differences. For example, it realizes the real-time fusion of satellite rainfall data (with a 0.5-hour delay) and ground radar data to generate a minute-level rainfall distribution map.

[0079] 3. The present invention intelligently adjusts the feature weights: the response time of the dynamic adjustment of feature weights is < 100 ms, the recognition accuracy of key features is > 92%, and it also supports the hybrid decision-making of expert rules and data-driven, such as realizing the automatic increase of the weights of seepage flow and phreatic line pressure to 0.45.

[0080] 4. The present invention uses multiple modules for coupled evaluation, with high efficiency: the error of the collaborative evaluation of multiple modules is reduced by 40% - 60% compared with a single model, and the deduction speed of extreme working conditions is increased by 5 times. Description of the Drawings

[0081] To more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on the structures shown in these drawings.

[0082] Figure 1 It is a block diagram of a small and medium-sized reservoir operation safety assessment system based on sky-earth-hydraulic engineering according to an embodiment of the present invention;

[0083] Figure 2 It is a schematic diagram of the processing flow of a short-term risk prediction module according to an embodiment of the present invention;

[0084] Figure 3 It is a schematic diagram of the processing flow of a medium- and long-term stability prediction module according to an embodiment of the present invention;

[0085] Figure 4 It is a schematic diagram of the processing flow of an extreme condition deduction module according to an embodiment of the present invention. Detailed Embodiments

[0086] The following will make a detailed description of the specific embodiments of the present invention.

[0087] The present application provides a small and medium-sized reservoir operation safety assessment system based on sky-earth-hydraulic engineering, including a time-space alignment layer, a feature weighting layer, and a dynamic coupling layer connected in sequence;

[0088] The time-space alignment layer performs time alignment and space alignment on the input sky-earth-hydraulic engineering data;

[0089] The feature weighting layer quantifies the contribution degrees of different monitoring features to the safety assessment according to the data after time-space alignment;

[0090] The dynamic coupling layer, through a multi-module coordination engine, adopts a dynamic weight allocation strategy for different modules, fuses the feature weighting, obtains a fusion result, and outputs a small and medium-sized reservoir operation safety index according to the fusion result, where the multi-module includes a short-term risk prediction module, a medium- and long-term stability prediction module, and an extreme condition deduction module; the short-term risk prediction module is used for the safety assessment of the dam in real time or within a preset first time, the medium- and long-term stability prediction module evaluates the stability of the dam within a preset number of years, and the extreme condition deduction module evaluates the structural safety under extreme conditions by simulating the impact of overloading events on the dam body.

[0091] Further, time alignment of the input data specifically means that for data with different sampling frequencies, time series alignment is achieved by dynamically bending the time axis, and cubic spline interpolation is used for low-frequency data below the preset frequency threshold to generate data at preset high-frequency time points.

[0092] Further, for data with different sampling frequencies, dynamic time warping is adopted, and the specific formula is as follows:

[0093]

[0094] Among them,

[0095] α m is the weight coefficient at the m-th time point of the data to be aligned;

[0096] γ is the time decay coefficient;

[0097] t m is the m-th time point of the reference time series;

[0098] t k is the time point of the data to be aligned;

[0099] t m′ is the time point of the reference time series;

[0100] M is the total number of time points in the reference series.

[0101] Further, the following formula is used to generate data at preset high-frequency time points for low-frequency data below the preset frequency threshold:

[0102]

[0103] Among them,

[0104] a, b, c, d are the coefficients of the cubic polynomial, determined by cubic spline interpolation to ensure that the interpolation function is continuous in the interval [t i , t i+1

[0105] and its first derivative and second derivative are continuous;

[0106] t is the target time point to be interpolated, located in the interval [t i , t i+1 ;

[0107] t i is the previous time point in the original low-frequency data;

[0108] t i+1 is the next time point adjacent to t i in the original low-frequency data;

[0109] ​ The estimated value obtained by cubic spline interpolation at time t.

[0110] Furthermore, a, b, c, and d are determined by the cubic spline interpolation conditions, thereby realizing smooth high-frequency data interpolation. The calculation methods of a, b, c, and d are as follows:

[0111] When t = t k , substituting into the formula gives d = X(t k ), that is, d is the value of the original low-frequency data at t k . Then, a three-moment equation is constructed to solve for M k . Cubic spline interpolation requires the second derivative to be continuous. The second derivative values M k (M k is the second derivative value at t k , and the interval length h k = t k+1 - t k ) of each node are obtained through the three-moment equation;

[0112] Then, a, b, and c are calculated respectively using the following formulas:

[0113]

[0114] The quadratic term coefficient is directly determined by the second derivative value at t k :

[0115]

[0116] Combining the interval function value difference and the second derivative value, the linear term coefficient is determined.

[0117] Furthermore, the following method is used for spatial alignment of data: All data are converted to the same coordinate system, the Gauss-Krüger projection is used to eliminate deformation, and a spatial interpolation kernel function is used to generate a continuous spatial distribution for the discrete monitoring point data.

[0118] Furthermore, the spatial interpolation kernel function is specifically as follows:

[0119]

[0120] Among them,

[0121] r is the normalized spatial distance;

[0122] K(r) is the spatial interpolation weight.

[0123] Furthermore, the feature weighting layer uses feature selection based on the attention mechanism to dynamically adjust the feature weights by combining context-aware attention with the environmental conditions.

[0124] Furthermore, the features in the sky layer are compressed in weight using the following formula to reduce the dimension of the original features:

[0125]

[0126] where

[0127] is the PCA projection matrix, obtained by performing eigen decomposition on the covariance matrix of the original data;

[0128] is the transpose of the l-th original feature vector after alignment, and n l is the total number of original features;

[0129] is the l-th compressed feature in the sky layer, and k represents the dimension after dimension reduction.

[0130] Weight compression reduces the dimension of high-dimensional features through mathematical transformation, retaining key information while reducing the feature dimension for more efficient subsequent analysis (such as weight assignment).

[0131] Furthermore, is composed of several principal components (eigenvectors) with the largest variances in the data, used to map high-dimensional original features to a low-dimensional space for dimension reduction.

[0132] Furthermore, is calculated as follows:

[0133] Perform standardization processing on the original feature data;

[0134] Calculate the covariance matrix of the standardized data;

[0135] Perform eigen decomposition on the covariance matrix to obtain eigenvalues and eigenvectors;

[0136] Sort the eigenvalues from largest to smallest, and select the eigenvectors corresponding to the top preset number of largest eigenvalues to form

[0137] Furthermore, the original data includes features such as surface deformation, thermal infrared anomaly, and UAV crack width in the sky monitoring layer.

[0138] The original features in the sky layer (such as satellite surface deformation values, UAV crack widths, etc.) have a high dimension. Through the projection of W l PCA these features can be compressed into low-dimensional principal components (such as ), reducing the computational complexity while retaining the main information.

[0139] Furthermore, in the cross-level of the context vectors of sky + ground + environment, the features are weighted and compressed using the following formula:

[0140]

[0141] where

[0142] β j is the attention weight of the j-th level; used for weight allocation between the sky layer (such as satellites / drones) and the ground layer (such as sensors), etc.;

[0143] C env is the environmental context vector;

[0144] Z j is the compressed feature of the j-th level; for example, the main components of surface deformation and the time series features of seepage flow rate after dimensionality reduction, etc.;

[0145] p is the summation index, with a value range from 1 to 4, corresponding to different levels; the denominator sums the exponential terms of these 4 levels to achieve weight normalization;

[0146] W q is a learnable weight vector, used to calculate the scores in the attention mechanism, and determines the weights by multiplying with the concatenated vector and then taking the exponent is the transpose of W q ;

[0147] || is the concatenation operator, which concatenates Z j and C env dimensionally to form a comprehensive vector containing hierarchical features and environmental information.

[0148] Furthermore, the environmental context vector includes environmental information such as the real-time reservoir water level (unit: m) and rainfall intensity (unit: mm / h).

[0149] Furthermore, the feature weighting layer adopts a static and dynamic hybrid weighting guided by expert knowledge. The static weight presets the basic weight according to expert experience, and the dynamic weight is adjusted according to the real-time working conditions.

[0150] Furthermore, the short-term prediction module uses the LSTM-Transformer model to mine local and global features in the reservoir operation data. The LSTM-Transformer model includes an LSTM layer, a Transformer self-attention layer, and a fully connected layer. The LSTM layer captures the temporal relationship of data features, the Transformer self-attention layer identifies the spatial attention weights between multi-sensor data features, and the fully connected layer outputs the dam break probability using the Sigmoid activation function.

[0151] Furthermore, the medium- and long-term stability prediction module constructs a physics-guided neural network architecture, inputs low-frequency cumulative data and environmental parameters, takes the embedded constitutive equation of the earth-rock dam as a physical constraint, and takes the non-linear relationship between the historical monitoring data learned by the multi-layer perceptron and the geological parameters as data-driven.

[0152] Furthermore, the extreme condition deduction module constructs a three-dimensional model of the dam body-foundation-reservoir area based on BIM, integrates the geometric model, material properties, and safety boundary conditions, establishes different parametric scenario libraries, then performs coupled solutions for multiple scenarios to obtain the pore water pressure distribution, updates the effective stress field, and based on the elastoplastic constitutive equation, iteratively solves for displacement and strain to extract the safety factor FS.

[0153] Furthermore, the solution of the pore water pressure distribution is based on the finite element method (FEM). The three-dimensional model of the dam body-foundation-reservoir area is discretized into elements, and the seepage control equation is established. For steady seepage, it satisfies the Laplace equation; for unsteady seepage, it satisfies the diffusion equation.

[0154] Furthermore, the solution is carried out through the following steps:

[0155] Discretization: Divide the elements and derive the permeability matrix for each element.

[0156] Boundary conditions: Set the upstream and downstream water levels (known water heads), impermeable boundaries (flow rate is zero), etc.;

[0157] Assembly of equations: Integrate the element permeability matrices into an overall matrix to form an algebraic equation system [K h {h} = {Q} ([K h is the total permeability matrix, and {Q} is the flow rate vector);

[0158] Solution: Solve the equation system through numerical methods (such as Gaussian elimination method) to obtain the water head h of each node, and then calculate the pore water pressure u from u = γ w h (γ w is the unit weight of water).

[0159] Furthermore, the geometric model includes the geometric models of the terrain, dam body structure, and flood discharge facilities. The material properties include the constitutive parameters of the earth-rock materials and the time-varying characteristics of the permeability coefficient. The safety boundary conditions include the reservoir water pressure, seismic motion input, and seepage boundary.

[0160] Furthermore, the dynamic weight allocation strategy defines the rule library through a fuzzy logic controller to achieve the adaptive weight of different models and updates the feature weighting parameters in real time based on the evaluation error.

[0161] Further, if the reservoir water level > warning water level, then the weight of the short-term prediction module = 0.8; if the earthquake intensity ≥ magnitude 6, then the weight of the extreme condition deduction module = 0.7; if the reservoir water level is extremely high and the rainfall intensity is heavy rain, then the weight of the short-term prediction module = 0.8, the weight of the medium- and long-term stability prediction module = 0.1, and the weight of the extreme condition deduction module = 0.1.

[0162] Further, after the output result, compare the prediction result with the actual monitoring data, trigger the retraining of the models in each module, inject the new monitoring data into the training set, update the feature weighting parameters, and the adaptive correction equation based on the model parameter adjustment is as follows:

[0163]

[0164] Wherein,

[0165] s is an index used to identify the s-th data point or the s-th monitoring index;

[0166] is the s-th original monitoring data;

[0167] is the s-th predicted value based on the Kalman filter;

[0168] σ is the standard deviation of historical data;

[0169] is the s-th corrected sensor data value;

[0170] ω is a weighting coefficient. When the data deviation exceeds 3σ, balance the weights of the original data and the predicted value .

[0171] Embodiment 1

[0172] The present application provides a safety assessment system for the operation of medium and small reservoirs based on sky-earth-hydraulic engineering, including a time-space alignment layer, a feature weighting layer, and a dynamic coupling layer connected in sequence;

[0173] The time-space alignment layer performs time alignment and space alignment on the input sky-earth-hydraulic engineering data;

[0174] The feature weighting layer quantifies the contribution degree of different monitoring features to the safety assessment according to the data after time-space alignment;

[0175] The dynamic coupling layer, through a multi-module coordination engine, adopts a dynamic weight allocation strategy for different modules, fuses the feature weights, obtains a fusion result, and outputs the operation safety index of small and medium-sized reservoirs according to the fusion result. The multi-modules include a short-term risk prediction module, a medium- and long-term stability prediction module, and an extreme condition deduction module. The short-term risk prediction module is used for the safety assessment of the dam in real time or within a preset first time. The medium- and long-term stability prediction module evaluates the stability of the dam within a preset number of years. The extreme condition deduction module evaluates the structural safety under extreme conditions by simulating the impact of overloading events on the dam body.

[0176] Embodiment 2

[0177] This application provides a small and medium-sized reservoir operation safety assessment system based on sky-earth-hydraulic engineering, including a time-space alignment layer, a feature weighting layer, and a dynamic coupling layer connected in sequence;

[0178] The time-space alignment layer performs time alignment and space alignment on the input sky-earth-hydraulic engineering data;

[0179] The feature weighting layer quantifies the contribution degree of different monitoring features to the safety assessment according to the time-space aligned data;

[0180] The dynamic coupling layer, through a multi-module coordination engine, adopts a dynamic weight allocation strategy for different modules, fuses the feature weights, obtains a fusion result, and outputs the operation safety index of small and medium-sized reservoirs according to the fusion result. The multi-modules include a short-term risk prediction module, a medium- and long-term stability prediction module, and an extreme condition deduction module. The short-term risk prediction module is used for the safety assessment of the dam in real time or within a preset first time. The medium- and long-term stability prediction module evaluates the stability of the dam within a preset number of years. The extreme condition deduction module evaluates the structural safety under extreme conditions by simulating the impact of overloading events on the dam body.

[0181] Further, the time alignment of the input data is specifically to achieve time series alignment for data with different sampling frequencies by dynamically bending the time axis, and use cubic spline interpolation for low-frequency data below the preset frequency threshold to generate preset high-frequency time point data.

[0182] Further, for data with different sampling frequencies, dynamic time warping is adopted, and the specific formula is as follows:

[0183]

[0184] Among them,

[0185] α m is the weight coefficient of the m-th time point of the data to be aligned;

[0186] γ is the time decay coefficient;

[0187] t m is the m-th time point of the reference time series;

[0188] t k is the time point of the data to be aligned;

[0189] t m′ is the time point of the reference time series;

[0190] M is the total number of time points in the reference sequence.

[0191] Furthermore, for low-frequency data below the preset frequency threshold, the following formula is used to generate preset high-frequency time point data:

[0192]

[0193] Wherein,

[0194] a, b, c, d are the coefficients of the cubic polynomial, determined by cubic spline interpolation to ensure that the interpolation function is continuous in the interval [t i , t i+1

[0195] and the first derivative and second derivative are continuous;

[0196] t is the target time point to be interpolated, located in the interval [t i , t i+1 ;

[0197] t i is the previous time point in the original low-frequency data;

[0198] t i+1 is the next time point adjacent to t i in the original low-frequency data;

[0199] is the estimated value obtained by cubic spline interpolation at time t.

[0200] Furthermore, a, b, c, d are determined by the cubic spline interpolation conditions to achieve smooth high-frequency data interpolation. The calculation methods of a, b, c, d are as follows:

[0201] When t = t k , substituting into the formula gives d = X(t k ), that is, d is the value of the original low-frequency data at t k . Then, a three-moment equation is constructed to solve M k . Cubic spline interpolation requires the second derivative to be continuous. The second derivative values M k (M k is at t k ​The second derivative value at, interval length h k = t k+1 -t k );

[0202] Then, use the following formulas to calculate a, b, and c respectively:

[0203]

[0204] From the second derivative value at t k directly determine the quadratic term coefficient:

[0205]

[0206] Combine the interval function value difference and the second derivative value to determine the linear term coefficient.

[0207] Furthermore, the following method is used for spatial alignment of data: Convert all data to the same coordinate system, use the Gauss-Krüger projection to eliminate deformation, and use a spatial interpolation kernel function for discrete monitoring point data to generate a continuous spatial distribution.

[0208] Furthermore, the spatial interpolation kernel function is specifically as follows:

[0209]

[0210] Among them,

[0211] r is the normalized spatial distance;

[0212] K(r) is the spatial interpolation weight.

[0213] Furthermore, the feature weighting layer uses feature selection based on the attention mechanism, and dynamically adjusts the feature weights through context-aware attention combined with environmental conditions.

[0214] Furthermore, the features in the sky layer are weighted and compressed using the following formula to reduce the dimension of the original features:

[0215]

[0216] Among them,

[0217] is the PCA projection matrix, obtained by performing eigenvalue decomposition on the covariance matrix of the original data;

[0218] is the transpose of the l-th original feature vector after alignment, n l is the total number of original features;

[0219] is the l-th compressed feature in the sky layer, and k represents the dimension after dimensionality reduction.

[0220] Weight compression reduces the dimension of high-dimensional features through mathematical transformation, retaining key information while reducing the feature dimension for more efficient subsequent analysis (such as weight assignment).

[0221] Furthermore, It is composed of several principal components (eigenvectors) with the largest variances in the data and is used to map high-dimensional original features to a low-dimensional space to achieve dimensionality reduction.

[0222] Furthermore, The calculation method of is as follows:

[0223] Perform standardization processing on the original feature data;

[0224] Calculate the covariance matrix of the standardized data;

[0225] Perform eigen-decomposition on the covariance matrix to obtain eigenvalues and eigenvectors;

[0226] Sort the eigenvalues from largest to smallest, and select the eigenvectors corresponding to the top preset number of largest eigenvalues to form

[0227] Furthermore, the original data includes features such as surface deformation, thermal infrared anomalies, and UAV crack widths in the sky monitoring layer.

[0228] The original features (such as satellite surface deformation values, UAV crack widths, etc.) at the sky level have a high dimension. Through the W l PCA projection, these features can be compressed into low-dimensional principal components (such as ), reducing the computational complexity while retaining the main information.

[0229] Furthermore, in the cross-level of the context vectors of sky + ground + environment, the features are weighted compressed using the following formula:

[0230]

[0231] where

[0232] β j is the attention weight of the j-th level; it is used for weight assignment between the sky layer (such as satellite / UAV) and the ground layer (such as sensors), etc.;

[0233] C env is the environmental context vector;

[0234] Z j is the compressed feature of the j-th level; for example, the principal component of surface deformation after dimensionality reduction, the time series feature of seepage flow, etc.;

[0235] Let \(p\) be the summation index, ranging from 1 to 4, corresponding to different levels; the denominator sums the exponential terms of these 4 levels to achieve the normalization of weights.

[0236] W q is a learnable weight vector used to calculate the scores in the attention mechanism, and determines the weights by multiplying with the concatenated vector and then taking the exponent. is the transpose of W q ;

[0237] || is the concatenation operator, which concatenates \(Z\) j and \(C\) env along the dimension to form a comprehensive vector containing hierarchical features and environmental information.

[0238] Furthermore, the environmental context vector contains environmental information such as the real-time reservoir water level (unit: m) and rainfall intensity (unit: mm / h).

[0239] Furthermore, the feature weighting layer adopts a static and dynamic hybrid weighting guided by expert knowledge. The static weights preset the basic weights according to expert experience, and the dynamic weights are adjusted according to real-time working conditions.

[0240] Furthermore, the short-term prediction module uses the LSTM-Transformer model to mine local and global features in the reservoir operation data. The LSTM-Transformer model includes an LSTM layer, a Transformer self-attention layer, and a fully connected layer. The LSTM layer captures the temporal relationship of data features, the Transformer self-attention layer identifies the spatial attention weights between multi-sensor data features, and the fully connected layer outputs the dam-break probability using the Sigmoid activation function.

[0241] Furthermore, the medium- and long-term stability prediction module constructs a physics-guided neural network architecture, inputs low-frequency cumulative data and environmental parameters, takes the embedded constitutive equation of the earth-rock dam as a physical constraint, and takes the non-linear relationship between historical monitoring data and geological parameters learned by the multi-layer perceptron as data-driven.

[0242] Furthermore, the extreme condition deduction module constructs a three-dimensional model of the dam body-foundation-reservoir area based on BIM, integrates geometric models, material properties, and safety boundary conditions, and establishes different parameterized scenario libraries, and then performs coupled solutions for multiple scenarios to obtain the pore water pressure distribution, update the effective stress field, and based on the elastoplastic constitutive equation, iteratively solve for displacements and strains, and extract the safety factor FS.

[0243] Furthermore, the pore water pressure distribution is solved based on the finite element method (FEM). The three-dimensional model of the dam - foundation - reservoir area is discretized into elements, and the seepage control equation is established. For steady seepage, it satisfies the Laplace equation; for unsteady seepage, it satisfies the diffusion equation.

[0244] Furthermore, it is solved through the following steps:

[0245] Discretization: Divide the elements and derive the permeability matrix for each element.

[0246] Boundary conditions: Set the upstream and downstream water levels (known water heads), impermeable boundaries (flow rate is zero), etc.;

[0247] Assembly of equations: Integrate the element permeability matrices into an overall matrix to form an algebraic equation system [K h {h} = {Q} ([K h is the total permeability matrix, {Q} is the flow vector);

[0248] Solution: Solve the equation system through numerical methods (such as the Gauss elimination method) to obtain the water head h at each node, and then calculate the pore water pressure u from u = γ w h (γ w is the unit weight of water).

[0249] Furthermore, the geometric model includes the geometric models of the terrain, dam structure, and flood discharge facilities. The material properties include the constitutive parameters of earth and rock materials and the time - varying characteristics of the permeability coefficient. The safety boundary conditions include the reservoir water pressure, seismic ground motion input, and seepage boundary.

[0250] Furthermore, the dynamic weight allocation strategy defines the rule base through a fuzzy logic controller to achieve the weight self - adaptation of different models and update the feature weighting parameters in real - time based on the evaluation error.

[0251] Furthermore, if the reservoir water level > warning water level, then the weight of the short - term prediction module = 0.8; if the seismic intensity ≥ 6 degrees, then the weight of the extreme condition deduction module = 0.7; if the reservoir water level is extremely high and the rainfall intensity is heavy rain, then the weight of the short - term prediction module = 0.8, the weight of the medium - and long - term stability prediction module = 0.1, and the weight of the extreme condition deduction module = 0.1.

[0252] Furthermore, after the output results, compare the prediction results with the actual monitoring data, trigger the re - training of the models in each module, inject the new monitoring data into the training set, and update the feature weighting parameters. The adaptive correction equation based on the model parameter adjustment is as follows:

[0253]

[0254] Among them,

[0255] s is an index used to identify the s-th data point or the s-th monitoring index;

[0256] is the s-th original monitoring data;

[0257] is the s-th predicted value based on Kalman filtering;

[0258] σ is the standard deviation of historical data;

[0259] is the s-th corrected sensor data value;

[0260] ω is a weighting coefficient. When the data deviation exceeds 3σ, it balances the weights of the original data and the predicted value .

[0261] Example 3

[0262] This application provides a safety assessment system for the operation of medium and small reservoirs based on sky-earth-hydraulic engineering:

[0263] (1) System architecture

[0264] The present invention adopts a hierarchical progressive fusion strategy, including a spatio-temporal alignment layer, a feature weighting layer, and a dynamic coupling layer connected in sequence; as Figure 1 shown;

[0265] The spatio-temporal alignment layer performs time alignment and space alignment on the input sky-earth-hydraulic engineering data;

[0266] The feature weighting layer quantifies the contribution of different monitoring features to safety assessment according to the data after spatio-temporal alignment;

[0267] The dynamic coupling layer, through a multi-module coordination engine, adopts a dynamic weight allocation strategy for different modules to fuse the feature weighting, obtains a fusion result, and outputs the safety operation index of medium and small reservoirs according to the fusion result, where the multi-module includes a short-term risk prediction module, a medium and long-term stability prediction module, and an extreme condition deduction module; the short-term risk prediction module is used for the safety assessment of the dam in real time or within a preset first time, the medium and long-term stability prediction module evaluates the stability of the dam within a preset number of years, and the extreme condition deduction module evaluates the structural safety under extreme conditions by simulating the impact of overloading events on the dam body.

[0268] (2) Fusion of multi-source data

[0269] 2.1. Adopt multi-source data input

[0270] The present invention adopts multi-source data input, as shown in Table 1 specifically.

[0271] Table 1. Input multi-source data table

[0272]

[0273] 2.2. Temporal and spatial alignment processing

[0274] The spatio-temporal alignment layer solves the inconsistencies of multi-source heterogeneous monitoring data in the temporal and spatial dimensions and constructs a unified data benchmark.

[0275] In terms of temporal alignment:

[0276] For data with different sampling frequencies, time series matching is achieved by dynamically bending the time axis. For example, for satellite data (once every 6 hours) and sensor data (once every 5 minutes), dynamic time warping (DTW) is used, and the time decay coefficient γ = 0.3 is set to generate a minute-level continuous data set.

[0277] Among them, the calculation of dynamic time warping (DTW) is as follows:

[0278]

[0279] Among them, α m is the weight coefficient of the m-th time point of the data to be aligned;

[0280] γ is the time decay coefficient; t m is the m-th time point of the reference time series; t k is the time point of the data to be aligned;

[0281] t m′ is the time point of the reference time series; M is the total number of time points of the reference series.

[0282] Table 2. Input data table for dynamic time warping

[0283] Symbol Actual data input <![CDATA[α m > Weight when satellite data timestamp is aligned with ground sensor timestamp γ Empirical value (usually taken as 0.3) <![CDATA[t m > Sampling time points of satellite data (such as 06:00, 12:00, etc.) <![CDATA[t k > Sampling time points of ground sensors (such as once every 5 minutes)

[0284] For low-frequency data (such as satellites) below the preset frequency threshold, the following formula is used to generate preset high-frequency time point data, and cubic polynomial-based time interpolation is adopted:

[0285]

[0286] Among them,

[0287] a, b, c, d are the coefficients of the cubic polynomial, which are determined by cubic spline interpolation to ensure that the interpolation function is continuous in the interval [t i , t i+1

[0288] and the first derivative and second derivative are continuous;

[0289] ​t is the target time point to be interpolated, located within the interval [t i , t i+1 ;

[0290] t i is the previous time point in the original low-frequency data;

[0291] t i+1 is the next time point adjacent to t i in the original low-frequency data;

[0292] is the estimated value obtained by cubic spline interpolation at time t.

[0293] Furthermore, a, b, c, and d are determined through the cubic spline interpolation conditions to achieve smooth high-frequency data interpolation. The calculation methods of a, b, c, and d are as follows:

[0294] When t = t k , substituting into the formula gives d = X(t k ), that is, d is the value of the original low-frequency data at t k . Then, a three-moment equation is constructed to solve for M k . Cubic spline interpolation requires the second derivative to be continuous. The second derivative values M k of each node are obtained through the three-moment equation (M k is the second derivative value at t k , and the interval length h k = t k+1 - t k );

[0295] Then, a, b, and c are calculated respectively using the following formulas:

[0296]

[0297] The quadratic term coefficient is directly determined by the second derivative value at t k :

[0298]

[0299] Combined with the interval function value difference and the second derivative value, the linear term coefficient is determined.

[0300] In terms of space:

[0301] Convert all data to the same coordinate system (such as WGS-84), use the Gauss-Krüger projection to eliminate deformation and achieve the unification of geographical coordinates, and use the spatial interpolation kernel function for discrete monitoring point data to generate a continuous spatial distribution. For example, for the dam seepage pressure data, convert the discrete piezometer data into a two-dimensional distribution map of the dam saturation line with a grid accuracy of 0.5m×0.5m. For abnormal data processing, use Mahalanobis distance-based method to identify abnormal data and mark them.

[0302] The spatial interpolation kernel function is calculated as follows:

[0303]

[0304] Table 3. Input data table for spatial interpolation kernel function

[0305]

[0306] The feature weighting layer mainly quantifies the contribution of different monitoring features to the safety assessment and realizes dynamic weight allocation. Two main methods are adopted. The first one is feature selection based on the attention mechanism, which dynamically adjusts the feature weights by combining context-aware attention with environmental conditions (such as reservoir water level, rainfall intensity).

[0307] Feature weight compression in the sky layer:

[0308] The features in the sky layer are compressed using the following formula to reduce the dimension of the original features:

[0309]

[0310] where,

[0311] is the PCA projection matrix, obtained by performing eigenvalue decomposition on the covariance matrix of the original data;

[0312] is the transpose of the l-th original feature vector after alignment, and n l is the total number of original features;

[0313] is the l-th compressed feature in the sky layer, and k represents the dimension after dimension reduction.

[0314] Table 4. Input data table for sky layer feature dimension reduction

[0315]

[0316] In the cross-layer of the sky+ground+environment context vector, the features are compressed using the following formula:

[0317]

[0318] Among them,

[0319] β j is the attention weight of the j-th layer; it is used for weight allocation between the sky layer (such as satellites / drones) and the ground layer (such as sensors), etc.

[0320] C env is the environmental context vector;

[0321] Z j is the compressed feature of the j-th layer; for example, the main components of surface deformation after dimensionality reduction, the time series features of seepage flow, etc.

[0322] p is the summation index, and its value range is from 1 to 4, corresponding to different layers; the denominator realizes the normalization of the weight by summing the exponential terms of these 4 layers.

[0323] W q is a learnable weight vector, which is used to calculate the score in the attention mechanism and determines the weight by multiplying with the concatenated vector and then taking the exponent. is the q transpose of W;

[0324] || is the concatenation operator, which concatenates Z j and C env by dimension to form a comprehensive vector containing hierarchical features and environmental information.

[0325] Table 5. Input data of context vector dimensionality reduction for sky + ground + environment

[0326]

[0327] The second is the hybrid weighting guided by expert knowledge. The static weight presets the basic weight according to expert experience (such as seepage flow weight 0.3, displacement weight 0.4), and the dynamic correction is adjusted through real-time working conditions (such as when an earthquake is triggered, the vibration spectrum weight is increased).

[0328] The dynamic coupling layer is an effective integration for realizing feature weighting. Through the multi-module coordination engine, a dynamic weight allocation strategy is adopted for different modules to fuse feature weighting, obtain the fusion result, and output the operation safety index of small and medium-sized reservoirs according to the fusion result, and feedback to the optimization loop to realize the dynamic iteration of risk assessment. Among them, the multi-modules include a short-term risk prediction module, a medium- and long-term stability prediction module, and an extreme condition deduction module; the short-term risk prediction module is used for the safety assessment of the dam within real-time or a preset first time, the medium- and long-term stability prediction module evaluates the stability of the dam within the preset years, and the extreme condition deduction module evaluates the structural safety under extreme conditions by simulating the impact of overloading events on the dam body.

[0329] Table 6. List of processing ideas for the dynamic coupling layer

[0330]

[0331]

[0332] The short-term risk prediction module is used for the dynamic assessment of the safety risks of earth-rock dams in real-time or in the near future. It can quickly identify sudden hidden dangers, predict the probability of dam failure and weak parts within the next few hours. In this embodiment, by building an LSTM-Transformer model, the local and global features in the reservoir operation data are fully mined, the changing rules of indicators such as water level and flow rate over time and their complex correlations are accurately captured, and the future short-term (such as the next 1-3 days) operation status of the reservoir is predicted.

[0333] The LSTM-Transformer model consists of an LSTM layer, a Transformer self-attention layer, and an output layer. The LSTM layer captures temporal dependencies (such as continuous changes in displacement and seepage flow rate), the Transformer self-attention layer identifies the spatial correlations between multi-sensor data (such as the linkage between crack propagation and rising phreatic line), and the output layer uses the Sigmoid activation function to output the dam failure probability (0-1). Its operation process is as Figure 2 shown.

[0334] The medium- and long-term stability prediction module evaluates the stability of the dam on the time scale of several years or even decades, and focuses on progressive risks such as material aging, geological settlement, and seepage erosion. In this embodiment, by constructing a PINNs model, the future medium- and long-term (such as the next 1 month to 2 years) operation status of the reservoir is predicted. When designing the architecture of the physics-informed neural network (PINNs), the constitutive equation of the earth-rock dam is embedded as a physical constraint, and the multi-layer perceptron (MLP) learns the non-linear relationship between historical monitoring data and geological parameters as data-driven, and the data fitting error and the physical equation residual (weight λ = 1.0) are used as the loss function. The operation process is as Figure 3 shown.

[0335] The extreme condition deduction module evaluates the structural safety under extreme conditions and generates emergency plans by simulating the impacts of rare or extreme load events (such as the possible maximum flood and earthquake intensity VIII) on the dam body. In this embodiment, by building a digital twin engine for reservoir projects, a three-dimensional refined model of the dam body - foundation - reservoir area is constructed based on BIM, integrating geometric models such as topography, dam body structure, and flood discharge facilities, material properties such as constitutive parameters of earth and rock materials and time-varying characteristics of permeability coefficients, and boundary conditions such as reservoir water pressure, earthquake ground motion input, and seepage boundaries. Different parameterized scenario libraries are established as shown in the following table, and then coupled solutions for multiple scenarios are carried out to solve the pore water pressure distribution, update the effective stress field, and iteratively solve displacements and strains based on the elastoplastic constitutive equation. Finally, the safety factor FS, as well as the contour maps of the maximum principal stress and strain distributions, the three-dimensional coordinates of the critical slip surface, and the recommended emergency flood discharge plan are extracted. The operation process is as Figure 4 shown.

[0336] Table 7. Example Table of Extreme Condition Parameter Settings and Purposes

[0337]

[0338]

[0339] (3) Dynamic weight allocation strategy

[0340] For the dynamic weight allocation strategy, the model weight adaption is mainly achieved by defining a rule base through a fuzzy logic controller, and the weights are updated in real time based on the evaluation error. In this embodiment:

[0341] IF the reservoir water level > warning water level, the short-term model weight of TNHE = 0.8

[0342] IF the earthquake intensity ≥ magnitude 6, the extreme condition model weight of TNHE = 0.7

[0343] IF the reservoir water level is too high AND the rainfall intensity is heavy rain, the short-term module weight of TNHE = 0.8, medium- and long-term = 0.1, extreme = 0.1

[0344] After that, the feedback optimization loop verifies the evaluation results, compares the model prediction with the actual monitoring data, triggers the retraining of the model, and injects the new monitoring data into the training set to update the feature weighting parameters.

[0345] Adaptive correction equation based on model parameter adjustment:

[0346]

[0347] Among them,

[0348] s is the index used to identify the s-th data point or the s-th monitoring index;

[0349] is the s-th original monitoring data;

[0350] is the s-th predicted value based on Kalman filtering;

[0351] σ is the standard deviation of historical data;

[0352] is the s-th corrected sensor data value;

[0353] ω is the weighting coefficient. When the data deviation exceeds 3σ, it balances the original data and the predicted value weights.

[0354] Table 7. List of input data for the adaptive correction equation

[0355]

[0356] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A safety assessment system for the operation of medium and small-sized reservoirs based on sky-earth-water conservancy projects, characterized in that, It includes a spatio-temporal alignment layer, a feature weighting layer, and a dynamic coupling layer that are connected in sequence; The spatio-temporal alignment layer performs temporal alignment and spatial alignment on the input data of sky, ground, water conservancy projects; The feature weighting layer quantifies the contribution of different monitoring features to safety assessment according to the data after spatio-temporal alignment; The dynamic coupling layer, through a multi-module coordination engine, adopts a dynamic weight allocation strategy for different modules, fuses the feature weighting, obtains a fusion result, and outputs the operation safety index of small and medium-sized reservoirs according to the fusion result. The multi-modules include a short-term risk prediction module, a medium- and long-term stability prediction module, and an extreme condition deduction module; the short-term risk prediction module is used for the safety assessment of the dam in real time or within a preset first time, the medium- and long-term stability prediction module evaluates the stability of the dam within a preset number of years, and the extreme condition deduction module evaluates the structural safety under extreme conditions by simulating the impact of overloading events on the dam body.

2. The operation safety assessment system for small and medium-sized reservoirs based on sky-earth-hydraulic engineering according to claim 1, characterized in that The specific method for performing temporal alignment on the input data is to perform time series alignment on data with different sampling frequencies by dynamically bending the time axis, and use cubic spline interpolation for low-frequency data below the preset frequency threshold to generate data at preset high-frequency time points.

3. The operation safety assessment system for small and medium-sized reservoirs based on sky-earth-water conservancy as claimed in claim 2, wherein For data with different sampling frequencies, dynamic time warping is adopted, and the specific formula is as follows: Where, α m is the weight coefficient at the m-th time point of the data to be aligned; γ is the time decay coefficient; t m is the m-th time point of the reference time series; t k is the time point of the data to be aligned; t m′ is the time point of the reference time series; M is the total number of time points in the reference sequence.

4. The operation safety assessment system for small and medium-sized reservoirs based on sky-earth-hydraulic engineering according to claim 2, characterized in that, The following formula is used to generate data at preset high-frequency time points for low-frequency data below the preset frequency threshold: Where, a, b, c, and d are the coefficients of a cubic polynomial, determined by cubic spline interpolation to ensure that the interpolation function is continuous in the interval [t i , t i+1 , and its first and second derivatives are also continuous; t is the target time point to be interpolated, located within the interval [t i , t i+1 ; t i is the previous time point in the original low-frequency data; t i+1 is the next time point adjacent to t in the original low-frequency data; i adjacent; The estimated value obtained by cubic spline interpolation at time t.

5. The operation safety assessment system for small and medium-sized reservoirs based on sky-earth-hydraulic engineering according to claim 1, characterized in that, The following method is used for spatial alignment of data: convert all data to the same coordinate system, use the Gauss-Kruger projection to eliminate deformation, and use a spatial interpolation kernel function for discrete monitoring point data to generate a continuous spatial distribution.

6. The safety assessment system for the operation of small and medium-sized reservoirs based on sky-earth-water conservancy according to claim 5, wherein, The spatial interpolation kernel function is specifically as follows: Where, r is the normalized spatial distance; K(r) is the spatial interpolation weight.

7. The operation safety assessment system for small and medium-sized reservoirs based on sky-earth-water conservancy works according to claim 1, characterized in that The feature weighting layer adopts feature selection based on the attention mechanism, and dynamically adjusts the feature weights by combining context-aware attention with environmental conditions.

8. The operation safety assessment system for small and medium-sized reservoirs based on sky-earth-hydraulic engineering according to claim 7, characterized in that, For features in the sky level, the following formula is used for weight compression to reduce the dimension of the original features: Where, is the PCA projection matrix, obtained by performing eigen decomposition on the covariance matrix of the original data; is the transpose of the l-th original feature vector after alignment, n l is the total number of original features; It is the l-th compressed feature in the sky layer, and k represents the dimension after dimensionality reduction.

9. The operation safety assessment system for small and medium-sized reservoirs based on sky-earth-hydraulic engineering according to claim 7, wherein In the cross-level of the sky+ground+environment context vector, the following formula is used for weight compression of features: Where, β j is the attention weight of the j-th level; C env is the environmental context vector; Z j is the compressed feature of the j-th level; p is the summation index, and the value range is from 1 to 4, corresponding to different levels; W q is a learnable weight vector, is the transpose of W q ; || is the concatenation operator that concatenates Z j and C env along a dimension to form a comprehensive vector containing hierarchical features and environmental information.

10. The small and medium-sized reservoir operation safety assessment system based on sky-earth-hydraulic engineering according to claim 1, characterized in that, The feature weighting layer adopts static and dynamic hybrid weighting guided by expert knowledge. The static weight presets the basic weight according to expert experience, and the dynamic weight is adjusted according to real-time working conditions.

11. The operation safety assessment system for medium and small-sized reservoirs based on sky-earth-hydraulic engineering according to claim 1, wherein The short-term prediction module uses the LSTM-Transformer model to mine local features and global features in reservoir operation data. The LSTM-Transformer model includes an LSTM layer, a Transformer self-attention layer, and a fully connected layer. The LSTM layer captures the temporal relationship of data features, the Transformer self-attention layer identifies the spatial attention weights between multi-sensor data features, and the fully connected layer uses the Sigmoid activation function to output the dam-break probability.

12. The small and medium-sized reservoir operation safety assessment system based on sky-earth-water conservancy according to claim 1, characterized in that, The medium- and long-term stability prediction module constructs a physics-guided neural network architecture, inputs low-frequency cumulative data and environmental parameters, takes the embedded constitutive equation of the earth-rock dam as a physical constraint, and takes the non-linear relationship between the historical monitoring data learned by the multi-layer perceptron and the geological parameters as data-driven.

13. The small and medium-sized reservoir operation safety assessment system based on sky-earth-water conservancy as claimed in claim 1, wherein, The extreme condition deduction module constructs a three-dimensional model of the dam body-foundation-reservoir area based on BIM, integrates geometric models, material properties, and safety boundary conditions, establishes different parameterized scenario libraries, then performs coupled solutions for multiple scenarios to obtain the pore water pressure distribution, updates the effective stress field, and based on the elastoplastic constitutive equation, iteratively solves for displacement and strain, and extracts the safety factor FS.

14. The operation safety assessment system for small and medium-sized reservoirs based on sky-earth-hydraulic engineering according to claim 13, characterized in that, The geometric models include the geometric models of the terrain, dam structure, and flood discharge facilities. The material properties include the constitutive parameters of the earth-rock materials and the time-varying characteristics of the permeability coefficient. The safety boundary conditions include the reservoir water pressure, seismic ground motion input, and seepage boundary.

15. The small and medium-sized reservoir operation safety assessment system based on sky-earth-hydraulic engineering according to claim 1, characterized in that, The dynamic weight allocation strategy defines a rule base through a fuzzy logic controller to achieve the adaptive weight of different models, and updates the feature weighting parameters in real time based on the evaluation error.

16. The safety assessment system for the operation of small and medium-sized reservoirs based on sky-earth-hydraulic engineering according to claim 15, characterized in that, If the reservoir water level > warning water level, then the weight of the short-term prediction module = 0.8; if the seismic intensity ≥ 6, then the weight of the extreme condition deduction module = 0.7; if the reservoir water level is extremely high and the rainfall intensity is heavy rain, then the weight of the short-term prediction module = 0.8, the weight of the medium- and long-term stability prediction module = 0.1, and the weight of the extreme condition deduction module = 0.

1.

17. The operation safety assessment system for small and medium-sized reservoirs based on sky-earth-hydraulic engineering according to claim 15, wherein After the output results, compare the prediction results with the actual monitoring data, trigger the re-training of the models in each module, inject the new monitoring data into the training set, update the feature weighting parameters, and the adaptive correction equation based on the adjustment of the model parameters is as follows: Where s is the index used to identify the s-th data point or the s-th monitoring index; is the s-th original monitoring data; is the s-th predicted value based on the Kalman filter; σ is the standard deviation of the historical data; is the s-th corrected sensor data value; ω is a weighting coefficient. When the data deviation exceeds 3σ, the weights of the original data and the predicted value are balanced.

Images