Water engineering system deduction data construction method and system based on feature analysis
By using a feature-based analysis method, graph neural networks and generative adversarial networks are employed to generate water engineering data samples that conform to data distribution and physical laws, thus solving the problem of data shortage in water engineering systems and improving the reliability and intelligence level of scheduling decisions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BUREAU OF HYDROLOGY CHANGJIANG WATER RESOURCES COMMISSION
- Filing Date
- 2025-12-31
- Publication Date
- 2026-04-17
AI Technical Summary
Historical operational data for water engineering systems are limited to small samples, lacking samples of key scheduling scenarios such as major floods and severe droughts. This leads to blind spots in the AI model's understanding, and real-world scheduling does not allow for risky operations, resulting in a lack of samples of high-risk, high-efficiency scheduling strategies.
We employ a feature parsing-based approach, utilizing graph neural networks for spatiotemporal feature extraction and an improved attention mechanism to calculate data contribution. This is combined with generative adversarial networks to generate data samples that conform to data distribution characteristics and physical laws.
High-quality data samples were generated, which can provide reliable data support for the simulation of water engineering systems, solve the problem of data shortage, and improve the level of intelligence in scheduling decisions.
Smart Images
Figure CN121880930A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of hydrological and water resources forecasting technology, and in particular to a method and system for constructing data for water engineering system deduction based on feature analysis. Background Technology
[0002] The precise and scientific scheduling decisions of water conservancy projects are directly related to the optimal allocation of water resources in the basin. Therefore, in-depth study of the operational laws of water conservancy project scheduling is of great significance for realizing intelligent water conservancy project scheduling. However, the historical operational data of water conservancy project systems is essentially a small sample. On the one hand, key scheduling scenarios such as major floods and severe droughts occur infrequently but have extremely high decision-making value, and AI models have cognitive blind spots due to the lack of such samples. On the other hand, existing hydrological sequences cannot cover the new combinations and extreme boundaries that may occur under future climate change, and real-world scheduling cannot allow for risky operations for the sake of "data collection," resulting in a severe lack of samples of high-risk, high-efficiency scheduling strategies. Therefore, the construction of extrapolation data for water conservancy project systems has crucial practical significance and strategic value. Summary of the Invention
[0003] The purpose of this invention is to address the shortcomings of the prior art by providing a method and system for constructing water engineering system simulation data based on feature analysis. This method can generate high-quality data that conforms to both data distribution characteristics and physical laws, and can also provide reliable data samples for water engineering system simulation.
[0004] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for constructing simulation data for water engineering systems based on feature analysis, including: S1. Determine the research object: Determine the scope of the water project, the specific engineering object, and the watershed system where the project is located; S2. Basic Data Preparation: Collect and organize basic data for the research object, including geographical information, meteorological data, hydrological data, engineering characteristic parameters and scheduling operation data related to water projects; S3. Data Resource Attribute Analysis: Attribute analysis is conducted on the basic data resources involved in the research object, including data volume, data type, and data format. S4. Data Spatiotemporal Feature Analysis: Data feature analysis is performed using a spatiotemporal feature extraction method based on graph neural networks.
[0005] S5. Data Contribution Analysis: An improved attention mechanism calculation framework is used to quantitatively calculate the contribution of data from different sources to production goals. S6. Data Sample Generation: Generate data samples using generative adversarial networks; S7, Output dataset.
[0006] Furthermore, S4 specifically includes: S401. Abstract the water engineering system into a graph, which contains graphical elements such as nodes and edges; the nodes represent actual elements such as reservoirs, pumping stations, and monitoring points; the edges represent the connection relationships of actual elements such as rivers and pipelines, and each node contains multimodal features. S402. Perform water engineering system diagram construction and spatiotemporal graph convolution, specifically: The graph construction is as follows: ;in, A set of nodes; Let it be the set of edges; It is an adjacency matrix, defined based on geographical distance or hydraulic connection; If a node in a water engineering system is a reservoir or a flood storage and detention area controlled by gates, then the adjacency matrix is as follows: It means that, among them, Indicates the output water volume or flow rate at the cross-section; Indicates the length of the river channel; Indicates the average propagation time; The parameter variables that are affected by the river channel gradient and roughness; If the nodes in the water engineering system are other water engineering projects in flood storage and detention areas without gate control, then the adjacency matrix is as follows: It means that, among them, Indicates the cross-sectional water carrying capacity or flow rate; This indicates the distance of the river channel from the water project access point to the previous node; Indicates the efficiency of flood diversion or flood intake; The parameter variable representing the influence of the flood diversion area; Spatiotemporal graph convolution: Spatiotemporal graph convolutional networks are used to capture spatial relationships and temporal dynamics simultaneously. However, considering the parallel and mixed scenarios of different tributaries flowing into the same main stream in water engineering systems, as well as the characteristics of the dynamic changes of water flow processes over time, the spatial convolution formula is improved, and the improved graph convolution and gated recurrent units are coupled to perform spatial and temporal convolution. The spatial convolution: ,in, , is an adjacency matrix with added self-connections; yes The degree matrix; Calculate functions for different levels of association relationships; , It is the first Layer node characteristics; It is the learning weight; It is a non-linear activation function; The temporal convolution: inputs the temporal series features of each node into the GRU unit; , Indicates the same Inflow or outflow rate that is constantly related to the operating water level; Reservoir The operating water level at any given moment represents any point within the calculation time. It is the calculation function for the gated loop unit.
[0007] Furthermore, S5 specifically includes: ; in, Data representing a modality; and This indicates one or more other modes; This represents the weighting coefficients for the combined local sliding window attention and global attention. m = [ 0 , 1 ] ; For attention; Attention within a local sliding window; Global attention; global sparse attention is: ; ; in, This indicates that the center position of attention is being calculated, i.e., the query position index; Indicates and Calculate the relevance at other locations, i.e., the key position index; The sliding window radius is a preset hyperparameter that defines the size of the local attention range. Location and The original attention score; This represents an improved soft max function, which only calculates the maximum value during the solution process. The soft max function is used to calculate the values of each data point. For sparsification hyperparameters; The dimension of the key vector is represented by the attention weights. The interaction between different modal data is obtained by calculating the attention weights, and then the value of different modal data is evaluated, which is indeed the contribution of the generated target. Considering the periodicity of water project scheduling, periodic awareness attention should be increased, that is, in Add one more component.
[0008] Furthermore, S6 specifically includes: S601. The spatiotemporal feature matrix obtained in S4 is fused with the data contribution weight obtained in S5 to form a condition vector. A noise vector is randomly generated and concatenated with the condition vector to form the initial solution of the target. S602. Input the fused conditional vector into the generator G, and output the preliminary generation target through the forward propagation of the graph neural network. Substitute the preliminary generation target into the known physical equations. S603. Input the initial generated target and the actual data within the same time period into the discriminator D to judge and evaluate the similarity in data distribution and obtain a similarity score. At the same time, the discriminator evaluates the physical rationality of the initial generated target and the actual data within the same time period and obtains a consistency score. S604. Calculate the standard generative adversarial loss based on similarity scoring. And calculate physical loss based on physical residuals. Simultaneously considering the periodicity of water project scheduling, the periodic consistency loss is calculated. The generator's total loss is a weighted sum of adversarial loss, physical loss, and periodic consistency loss, i.e. And optimize the generator and discriminator through backpropagation; S605. Repeat S601 to S604 until the total loss is minimized, i.e., the model is balanced, and the data output by the generator is the generation target.
[0009] Furthermore, the standard's generative adversarial loss for: minutes G max D V ( D , G ) = E x ∼ Pdata ( x ) [ log D ( x ) ] + E z ∼ Pz ( z ) [ log ( 1 − D ( G ( z ) ) ) ] ; in, This represents the value function, i.e., the objective of calculation; Mathematical expectation of a certain distribution ; This is actual data; Let be the probability distribution function that the actual data satisfies; This is noise data; Let be the probability distribution function satisfied by the noisy data; The physical residual is used to calculate physical loss. for: ; in, To generate variables; For the physical constraint functions associated with the generated variables; Spatial coordinates related to physical constraints; These are constant terms related to physical constraints; The length of the data sequence; The periodic consistency loss for: ; in, express Constantly schedule state variables or decision variables; This indicates the total number of time periods in the scheduling cycle. ; This represents the penalty weight coefficient, used to control the intensity of periodic satisfaction.
[0010] Furthermore, a water engineering system simulation data construction system based on feature parsing includes at least one processor; and a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor to implement the feature-based water engineering system deduction data construction method.
[0011] The beneficial effects of this invention are as follows: Relevant data is collected and organized for the research object, and data resource attribute analysis is performed; a spatiotemporal feature extraction method based on an improved graph neural network is used for data spatiotemporal feature analysis; a fusion module based on an improved attention mechanism is used for data contribution analysis; and generative adversarial networks are used to obtain and output the data samples to be generated. This invention's method can generate high-quality data that conforms to both data distribution characteristics and physical laws, providing reliable data samples for water engineering system simulation. Attached Figure Description
[0012] Figure 1 This is a flowchart of a method for constructing simulation data for water engineering systems based on feature analysis. Detailed Implementation
[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0014] Please see Figure 1 A method for constructing simulation data for water engineering systems based on feature analysis, comprising: S1. Determine the research object: Determine the scope of the water project, the specific engineering object, and the watershed system where the project is located; S2. Basic Data Preparation: Collect and organize basic data for the research object, including geographical information, meteorological data, hydrological data, engineering characteristic parameters and scheduling operation data related to water projects; S3. Data Resource Attribute Analysis: Attribute analysis is conducted on the basic data resources involved in the research object, including data volume, data type, and data format. S4. Data Spatiotemporal Feature Analysis: Data feature analysis is performed using a spatiotemporal feature extraction method based on graph neural networks.
[0015] S5. Data Contribution Analysis: An improved attention mechanism calculation framework is used to quantitatively calculate the contribution of data from different sources to production goals. S6. Data Sample Generation: Generate data samples using generative adversarial networks; S7, Output dataset.
[0016] Specifically, S4 is: S401. Abstract the water engineering system into a graph, which contains graphical elements such as nodes and edges; the nodes represent actual elements such as reservoirs, pumping stations, and monitoring points; the edges represent the connection relationships of actual elements such as rivers and pipelines, and each node contains multimodal features. For example, reservoir nodes include water level, inflow, outflow, and gate opening / closing status; water intake nodes include flow rate, water quality, intake elevation, and intake status.
[0017] S402. Perform water engineering system diagram construction and spatiotemporal graph convolution, specifically: The graph construction is as follows: ;in, A set of nodes; Let it be the set of edges; It is an adjacency matrix, defined based on geographical distance or hydraulic connection; If a node in a water engineering system is a reservoir or a flood storage and detention area controlled by gates, then the adjacency matrix is as follows: It means that, among them, Indicates the output water volume or flow rate at the cross-section; Indicates the length of the river channel; Indicates the average propagation time; The parameter variables that are affected by the river channel gradient and roughness; If the nodes in the water engineering system are other water engineering projects in flood storage and detention areas without gate control, then the adjacency matrix is as follows: It means that, among them, Indicates the cross-sectional water carrying capacity or flow rate; This indicates the distance of the river channel from the water project access point to the previous node; Indicates the efficiency of flood diversion or flood intake; The parameter variable representing the influence of the flood diversion area; Spatiotemporal graph convolution: Spatiotemporal graph convolutional networks are used to capture spatial relationships and temporal dynamics simultaneously. However, considering the parallel and mixed scenarios of different tributaries flowing into the same main stream in water engineering systems, as well as the characteristics of the dynamic changes of water flow processes over time, the spatial convolution formula is improved, and the improved graph convolution and gated recurrent units are coupled to perform spatial and temporal convolution. The spatial convolution: ,in, , is an adjacency matrix with added self-connections; yes The degree matrix; Calculate functions for different levels of association relationships; , It is the first Layer node characteristics; It is the learning weight; It is a non-linear activation function; The temporal convolution: inputs the temporal series features of each node into the GRU unit; , Indicates the same Inflow or outflow rate that is constantly related to the operating water level; Reservoir The operating water level at any given moment represents any point within the calculation time. It is the calculation function of the gated cycle unit; the relationship between the inflow or outflow and the operating water level can be obtained from this formula.
[0018] Specifically, S5 is: ; in, Data representing a modality; and This indicates one or more other modes; This represents the weighting coefficients for the combined local sliding window attention and global attention. m = [ 0 , 1 ] ; For attention; Attention within a local sliding window; Global attention; global sparse attention is: ; ; in, This indicates that the center position of attention is being calculated, i.e., the query position index; Indicates and Calculate the relevance at other locations, i.e., the key position index; The sliding window radius is a preset hyperparameter that defines the size of the local attention range. Location and The original attention score; This represents an improved soft max function, which only calculates the maximum value during the solution process. The soft max function is used to calculate the values of each data point. For sparsification hyperparameters; The dimension of the key vector is represented by the attention weights. The interaction between different modal data is obtained by calculating the attention weights, and then the value of different modal data is evaluated, which is indeed the contribution of the generated target. Considering the periodicity of water project scheduling, periodic awareness attention should be increased, that is, in Add one more component.
[0019] Specifically, S6 is: S601. The spatiotemporal feature matrix obtained in S4 is fused with the data contribution weight obtained in S5 to form a condition vector. A noise vector is randomly generated and concatenated with the condition vector to form the initial solution of the target. S602. Input the fused conditional vector into the generator G, and output the preliminary generation target through the forward propagation of the graph neural network. Substitute the preliminary generation target into the known physical equations, such as the Saint-Venant equations and the water balance equations, and calculate their physical residuals.
[0020] S603. Input the initial generated target and the actual data within the same time period into the discriminator D to judge and evaluate the similarity in data distribution and obtain a similarity score. At the same time, the discriminator evaluates the physical rationality of the initial generated target and the actual data within the same time period and obtains a consistency score. S604. Calculate the standard generative adversarial loss based on similarity scoring. And calculate physical loss based on physical residuals. Simultaneously considering the periodicity of water project scheduling, the periodic consistency loss is calculated. The generator's total loss is a weighted sum of adversarial loss, physical loss, and periodic consistency loss, i.e. And optimize the generator and discriminator through backpropagation; S605. Repeat S601 to S604 until the total loss is minimized, i.e., the model is balanced, and the data output by the generator is the generation target.
[0021] The standard generative adversarial loss for: minutes G max D V ( D , G ) = E x ∼ Pdata ( x ) [ log D ( x ) ] + E z ∼ Pz ( z ) [ log ( 1 − D ( G ( z ) ) ) ] ; in, This represents the value function, i.e., the objective of calculation; Mathematical expectation of a certain distribution ; This is actual data; Let be the probability distribution function that the actual data satisfies; This is noise data; Let be the probability distribution function satisfied by the noisy data; The physical residual is used to calculate physical loss. for: ; in, To generate variables; For the physical constraint functions associated with the generated variables; Spatial coordinates related to physical constraints; These are constant terms related to physical constraints; The length of the data sequence; Taking a reservoir as an example, the outflow from the reservoir needs to satisfy the one-dimensional flow continuity equation. Therefore, h represents the water level, and H represents the average cross-sectional flow rate. This refers to the source and sink terms per unit length, representing the amount of water that increases (e.g., tributary inflows and outflows, rainfall) or decreases (e.g., leakage, water intake) per unit length along the route.
[0022] The periodic consistency loss for: ; in, express Constantly schedule state variables or decision variables; This indicates the total number of time periods in the scheduling cycle. ; This represents the penalty weight coefficient, used to control the intensity of periodic satisfaction.
[0023] also, and The weighting coefficients for physical loss and periodic consistency loss are dynamically adjusted according to the stage of calculation. In the early stage of training, physical loss is given a higher weight to quickly fit the pattern; in the middle and later stages of training, the weight of adversarial loss is appropriately increased to optimize the generated data; the weight of periodic loss can be relatively stable.
[0024] A system for constructing data for water engineering system simulation based on feature parsing, comprising at least one processor; and a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor to implement a method for constructing water engineering system deduction data based on feature parsing.
[0025] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be defined by the appended claims.
Claims
1. A method for constructing simulation data for water engineering systems based on feature analysis, characterized in that: include: S1. Determine the research object: Determine the scope of the water project, the specific engineering object, and the watershed system where the project is located; S2. Basic Data Preparation: Collect and organize basic data for the research object, including geographical information, meteorological data, hydrological data, engineering characteristic parameters and scheduling operation data related to water projects; S3. Data Resource Attribute Analysis: Attribute analysis is conducted on the basic data resources involved in the research object, including data volume, data type, and data format. S4. Spatiotemporal Feature Analysis of Data: A spatiotemporal feature extraction method based on an improved graph neural network is used for data feature analysis. S5. Data Contribution Analysis: An improved attention mechanism calculation framework is used to quantitatively calculate the contribution of data from different sources to production goals. S6. Data Sample Generation: Generate data samples using generative adversarial networks; S7, Output dataset.
2. The method for constructing water engineering system simulation data based on feature analysis according to claim 1, characterized in that, Specifically, S4 is: S401. Abstract the water engineering system into a graph, which contains graphical elements such as nodes and edges; the nodes represent actual elements such as reservoirs, pumping stations, and monitoring points; the edges represent the connection relationships of actual elements such as rivers and pipelines, and each node contains multimodal features. S402. Perform water engineering system diagram construction and spatiotemporal graph convolution, specifically: The graph construction is as follows: ;in, A set of nodes; Let it be the set of edges; It is an adjacency matrix, defined based on geographical distance or hydraulic connection; If a node in a water engineering system is a reservoir or a flood storage and detention area controlled by gates, then the adjacency matrix is as follows: It means that, among them, Indicates the output water volume or flow rate at the cross-section; Indicates the length of the river channel; Indicates the average propagation time; The parameter variables that are affected by the river channel gradient and roughness; If the nodes in the water engineering system are other water engineering projects in flood storage and detention areas without gate control, then the adjacency matrix is as follows: It means that, among them, Indicates the cross-sectional water carrying capacity or flow rate; This indicates the distance of the river channel from the water project access point to the previous node; Indicates the efficiency of flood diversion or flood intake; The parameter variable representing the influence of the flood diversion area; Spatiotemporal graph convolution: Spatiotemporal graph convolutional networks are used to capture spatial relationships and temporal dynamics simultaneously. However, considering the parallel and mixed scenarios of different tributaries flowing into the same main stream in water engineering systems, as well as the characteristics of the dynamic changes of water flow processes over time, the spatial convolution formula is improved, and the improved graph convolution and gated recurrent units are coupled to perform spatial and temporal convolution. The spatial convolution: ,in, , is an adjacency matrix with added self-connections; yes The degree matrix; Calculate functions for different levels of association relationships; , It is the first Layer node characteristics; It is the learning weight; It is a non-linear activation function; The temporal convolution: inputs the temporal series features of each node into the GRU unit; , Indicates the same Inflow or outflow rate that is constantly related to the operating water level; Reservoir The operating water level at any given moment represents any point within the calculation time. It is the calculation function for the gated loop unit.
3. The method for constructing water engineering system simulation data based on feature analysis according to claim 2, characterized in that, Specifically, S5 is: ； in, Data representing a modality; and This indicates one or more other modes; This represents the weighting coefficients for the combined local sliding window attention and global attention. ; For attention; Attention within a local sliding window; Global attention; global sparse attention is: ； ； in, This indicates that the center position of attention is being calculated, i.e., the query position index; Indicates and Calculate the relevance at other locations, i.e., the key position index; The sliding window radius is a preset hyperparameter that defines the size of the local attention range. Location and The original attention score; This represents an improved soft max function, which only calculates the maximum value during the solution process. The soft max function is used to calculate the values of each data point. For sparsification hyperparameters; The dimension of the key vector is represented by the attention weights. The interaction between different modal data is obtained by calculating the attention weights, and then the value of different modal data is evaluated, which is indeed the contribution of the generated target. Considering the periodicity of water project scheduling, periodic awareness attention should be increased, that is, in Add one more component.
4. The method for constructing water engineering system simulation data based on feature analysis according to claim 3, characterized in that, Specifically, S6 is: S601. The spatiotemporal feature matrix obtained in S4 is fused with the data contribution weight obtained in S5 to form a condition vector. A noise vector is randomly generated and concatenated with the condition vector to form the initial solution of the target. S602. Input the fused conditional vector into the generator G, and output the preliminary generation target through the forward propagation of the graph neural network. Substitute the preliminary generation target into the known physical equations. S603. Input the initial generated target and the actual data within the same time period into the discriminator D to judge and evaluate the similarity in data distribution and obtain a similarity score. At the same time, the discriminator evaluates the physical rationality of the initial generated target and the actual data within the same time period and obtains a consistency score. S604. Calculate the standard generative adversarial loss based on similarity scoring. And calculate physical loss based on physical residuals. Simultaneously considering the periodicity of water project scheduling, the periodic consistency loss is calculated. The generator's total loss is a weighted sum of adversarial loss, physical loss, and periodic consistency loss, i.e. And optimize the generator and discriminator through backpropagation; S605. Repeat S601 to S604 until the total loss is minimized, i.e., the model is balanced, and the data output by the generator is the generation target.
5. The method for constructing water engineering system simulation data based on feature analysis according to claim 4, characterized in that, The standard generative adversarial loss for: ； in, This represents the value function, i.e., the objective of calculation; Mathematical expectation of a certain distribution ; This is actual data; Let be the probability distribution function that the actual data satisfies; This is noise data; Let be the probability distribution function satisfied by the noisy data; The physical residual is used to calculate physical loss. for: ； in, To generate variables; For the physical constraint functions associated with the generated variables; Spatial coordinates related to physical constraints; These are constant terms related to physical constraints; The length of the data sequence; The periodic consistency loss for: ； in, express Constantly schedule state variables or decision variables; This indicates the total number of time periods in the scheduling cycle. ; This represents the penalty weight coefficient, used to control the intensity of periodic satisfaction.
6. A system for constructing data for water engineering system deduction based on feature analysis, characterized in that: At least one processor; and a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor to implement the feature-analysis-based data construction method for water engineering systems as described in any one of claims 1 to 5.