Reservoir dam-break risk dynamic deduction method and system
Patent Information
- Application Number
- CN202611080665.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-21
- Publication Date
- 2026-10-30
Smart Images
Figure CN123012962A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of water conservancy disaster prevention and mitigation technology, and in particular to a method and system for dynamic simulation of reservoir dam failure risk. Background Technology
[0002] As core water conservancy hubs for flood control, urban water supply, and ecological water replenishment, some reservoirs also play a crucial strategic role in reducing and staggering flood peaks in the Haihe River Basin. These reservoirs generally have simple dam structures, soft soil foundations, and dense surrounding populations and urban facilities. Coupled with the influence of multiple factors such as extreme rainfall and tidal backwater, the risk of dam failure is significantly more difficult to prevent and control than that of reservoirs in mountainous areas.
[0003] Currently, the simulation and assessment of dam failure risks in coastal plain reservoirs mainly rely on traditional two-dimensional hydrodynamic numerical simulation technology. This type of technology requires solving complex physical equations to simulate flood evolution, which has core shortcomings such as long computation time, high software and hardware costs, and high professional application threshold. A single complete simulation of a working condition takes several hours to several days, which cannot meet the real-world needs of emergency decision-making under extreme weather conditions for rapid risk prediction at the minute level, making it difficult for grassroots reservoir management units to implement.
[0004] At the same time, a large number of cases of dam failures and major emergencies in plain reservoirs at home and abroad have long been stored in a fragmented state. Key experience and knowledge such as engineering entities, disaster-causing factors, damage modes, evolution processes, and disposal measures have not been structured, integrated, and digitally reused. The risk simulation process cannot rely on the patterns of historical cases to make accurate predictions, resulting in insufficient adaptability of simulation results to actual working conditions.
[0005] Furthermore, existing simulation techniques only use single monitoring data or topographic data as input, failing to achieve deep integration of real-time operational data, geospatial information, and historical case knowledge. The lack of input feature dimensions further reduces the accuracy of the simulation. Moreover, the simulation results can only output visualized inundation maps, lacking a dual verification mechanism for the authenticity and physical consistency of the results. This can easily lead to problems that violate hydrodynamic laws, such as water volume imbalance and discrepancies between water flow direction and topographic slope. The reliability of risk prediction results cannot be guaranteed. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings and defects of the prior art and provide a method and system for dynamic simulation of reservoir dam failure risk. It aims to solve the problems of low efficiency of traditional numerical simulation, inability to reuse historical knowledge, and lack of physical consistency verification of simulation results in the prior art.
[0007] According to one aspect of this application, a method for dynamically predicting the risk of reservoir dam failure is provided, comprising the following steps:
[0008] Real-time monitoring data of reservoir operation is collected, and digital elevation model data, land use type data and building distribution data of the downstream area of the reservoir are obtained. The real-time monitoring data and the digital elevation model data, land use type data and building distribution data are processed respectively to form an input feature tensor.
[0009] Using the current working condition as the query condition, similar historical cases are retrieved from the pre-constructed dam failure risk knowledge graph. The evolution path and final failure mode of the similar historical cases are extracted as embedding vectors and used as knowledge graph enhancement vectors.
[0010] The knowledge graph enhancement vector and the input feature tensor are fused and then input into the generator network. The generator network predicts the breach morphology parameters and the downstream flood inundation dynamic grid sequence within a preset time period based on the fused features, and outputs the prediction results.
[0011] Based on the prediction results, the discriminator network performs authenticity and physical consistency judgments; when the judgment results meet preset conditions, it outputs a valid risk prediction result.
[0012] Preferably, the dam failure risk knowledge graph is constructed by extracting engineering entities, disaster-causing factors, damage modes, evolution processes, disposal measures, and consequences from domestic and international cases of dam failures and emergencies in plain reservoirs as nodes of the knowledge graph, defining causal relationships, temporal relationships, and similarity relationships between nodes as edges of the knowledge graph, and constructing the knowledge graph based on the nodes and edges.
[0013] The dam failure risk knowledge graph performs in-depth analysis on each case and extracts the initial location of the hazard, its development and expansion process, and the timeline of the final collapse as entities in the evolution process. It also extracts early warning failure nodes, rescue failure nodes, and successful handling nodes as key decision point entities. It establishes temporal relationship edges between the key decision point entities and the evolution process entities, so that the dam failure risk knowledge graph can fully present the detailed evolution path from the initial cause to the final collapse, as well as the key turning points in the path. The graph embedding algorithm is used to transform nodes and edges into low-dimensional vector storage, and to map entities and relationships to the low-dimensional vector space.
[0014] Preferably, the step of retrieving similar historical cases from a pre-constructed dam failure risk knowledge graph using the current operating conditions as query criteria includes:
[0015] Calculate the cosine similarity between the real-time operating data of the current operating condition and the engineering entity feature vectors of each historical case in the dam failure risk knowledge graph, and select the cases with the highest similarity as similar cases.
[0016] The evolution path sequence and final destruction mode label corresponding to the similar cases are extracted. The evolution path sequence is converted into a fixed-length path feature vector through a sequence encoder. The final destruction mode label is converted into a mode feature vector through one-hot encoding. The path feature vector and the mode feature vector are concatenated to form the knowledge graph enhancement vector.
[0017] Preferably, the real-time monitoring data, the digital elevation model data, the land use type data, and the building distribution data are processed separately to form an input feature tensor, including:
[0018] The real-time monitoring data is normalized to form a real-time operating condition tensor. The digital elevation model data, land use type data, and building distribution data are encoded into a multi-dimensional matrix to form a geographic environment tensor. The real-time operating condition tensor and the geographic environment tensor are concatenated to form the input feature tensor.
[0019] Preferably, the generator network adopts an improved Attention U-Net architecture, including an encoder, a decoder, and a multi-head attention mechanism. The encoder performs multi-level downsampling on the input fused features to extract deep semantic features, and the decoder upsamples the deep semantic features to restore spatial resolution. A multi-head attention mechanism is introduced at the skip connections corresponding to the encoder and the decoder. The multi-head attention mechanism calculates the attention weight between the encoder feature map and the location encoding of historical high-risk areas retrieved from the dam breach risk knowledge graph. The encoder feature map is weighted and enhanced according to the attention weight, so that the decoder gives higher attention to the historical high-risk areas when generating the inundation range. Finally, the predicted results of the breach morphology parameters and the downstream flood inundation dynamic grid sequence are output.
[0020] Preferably, the discriminator network is a dual-output discriminator that performs both authenticity discrimination and physical consistency discrimination. The authenticity discrimination is used to determine whether the generated inundation map is visually and statistically similar to the sample generated by the high-fidelity hydrodynamic model. The physical consistency discrimination calculates the water volume conservation error and the consistency score between the water flow direction and the terrain slope in the generated result by introducing a physical constraint loss function.
[0021] Preferably, the physical constraint loss function includes a water volume conservation loss term and a flow direction consistency loss term; wherein, the water volume conservation loss term includes the error between the change in water volume between different time steps calculated based on the inundation depth raster sequence output by the generator network and the actual breach outflow; the flow direction consistency loss term includes the proportion of pixels with inconsistent flow directions calculated based on the flow direction of each pixel calculated based on the digital elevation model data and compared with the inundation evolution direction output by the generator network.
[0022] According to another aspect of this application, a dynamic simulation system for reservoir dam failure risk is provided, comprising:
[0023] The data acquisition module is used to collect real-time monitoring data of reservoir operation and obtain digital elevation model data, land use type data and building distribution data of the downstream area of the reservoir.
[0024] The feature tensor generation module is used to process the real-time monitoring data, the digital elevation model data, the land use type data, and the building distribution data respectively to form an input feature tensor;
[0025] The knowledge graph retrieval module is used to retrieve similar historical cases from a pre-built dam failure risk knowledge graph using the current working conditions as the query criteria.
[0026] An enhanced vector generation module is used to extract the embedding vectors of the evolution path and final destruction mode of the similar historical cases, as enhanced vectors for the knowledge graph;
[0027] The feature integration module is used to fuse the knowledge graph enhancement vector with the input feature tensor to form a fused feature.
[0028] The prediction module is used to use a generator network to predict the breach morphology parameters and the downstream flood inundation dynamic grid sequence within a preset time period based on fused features, and output the prediction results.
[0029] The discrimination module is used to perform authenticity discrimination and physical consistency discrimination based on the prediction results using a discriminator network; when the discrimination results meet preset conditions, it outputs a valid risk prediction result.
[0030] Preferably, it also includes an emergency simulation module connected to the discrimination module. When the effective risk prediction result output by the discrimination module exceeds the preset risk threshold, it calls the knowledge graph module to retrieve historical successful disposal measures cases under similar working conditions and similar damage modes, combines real-time traffic data and population heat map data to dynamically adapt and optimize historical evacuation plans, generate graded early warning instructions and personnel evacuation route plans, and push them to decision-makers through a visualization platform.
[0031] Preferably, the knowledge graph module includes a case deep analysis unit, used to perform deep analysis on each case and extract the initial location of the hazard, the process of its development and expansion, and the timeline of the final collapse as evolutionary process entities; extract early warning failure nodes, rescue failure nodes, and successful handling nodes as key decision point entities; and establish temporal relationship edges between the key decision point entities and the evolutionary process entities, so that the dam failure risk knowledge graph can fully present the detailed evolutionary path from the initial cause to the final collapse, as well as the key turning points in the path; and use a graph embedding algorithm to convert nodes and edges into low-dimensional vector storage, mapping entities and relationships to a low-dimensional vector space.
[0032] The present invention relates to a method and system for dynamic simulation of reservoir dam break risk. By constructing a dam break risk knowledge graph, it achieves the structured integration and digital empowerment of historical case knowledge. It integrates real-time monitoring data and digital elevation model data to form multi-dimensional input features. Relying on a generator and discriminator collaborative network, it completes the dynamic simulation of dam break risk. It fundamentally solves the core problems of low efficiency, inability to reuse historical knowledge, and lack of physical consistency verification of simulation results in traditional numerical simulation. It can complete the prediction of breach morphology parameters and the dynamic simulation of downstream flood inundation within seconds, fully meeting the rapid response requirements in emergency scenarios, while significantly reducing the technical application threshold and software and hardware costs.
[0033] The dam failure risk knowledge graph transforms entities and relationships into low-dimensional vectors through graph embedding algorithms, fully reconstructing the evolution path and key decision nodes of dam failure from initial causes to final collapse. This transforms implicit engineering experience into searchable and reusable explicit knowledge. Combined with similarity matching, it accurately filters similar historical cases, and the resulting knowledge-enhanced vectors can effectively improve the inference model's ability to learn and adapt to historical risk patterns.
[0034] The integration of real-time monitoring data normalization processing and geospatial information coding allows the input features to fully cover real-time operating conditions, topography, land use and building distribution elements, which is more in line with the engineering characteristics of flat terrain and complex underlying surfaces of coastal plain reservoirs.
[0035] The generator network introduces a multi-head attention mechanism, which can focus on historically high-risk areas to enhance feature extraction and significantly improve the accuracy of breach parameter prediction and flooding grid sequence generation. The dual-output discriminator simultaneously performs authenticity and physical consistency discrimination. By using the constraint loss function of water volume conservation and matching water flow direction with terrain slope, it completely avoids the problem of the inference results violating hydrodynamic laws and ensures that the prediction results are scientific and reliable.
[0036] The simulation system built on this method can also link historical case studies and real-time traffic and population data to automatically optimize tiered early warning instructions and personnel evacuation routes, achieving a closed-loop support for the entire chain from risk simulation to emergency response, and comprehensively improving the intelligence and practicality of dam failure risk prevention and control in coastal plain reservoirs. Attached Figure Description
[0037] Figure 1 This is a flowchart illustrating the dynamic simulation method for reservoir dam failure risk according to the present invention.
[0038] Figure 2 This is a schematic diagram of the architecture of the reservoir dam failure risk dynamic simulation system of the present invention. Detailed Implementation
[0039] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0040] See Figure 1 As shown in the exemplary embodiment of this application, the method for dynamically extrapolating the risk of dam failure includes the following steps:
[0041] S1. Collect real-time monitoring data of reservoir operation and obtain digital elevation model data, land use type data and building distribution data of the downstream area of the reservoir. Process the real-time monitoring data and the digital elevation model data, land use type data and building distribution data respectively to form an input feature tensor.
[0042] S2. Using the current working condition as the query condition, retrieve similar historical cases in the pre-constructed dam failure risk knowledge graph, and extract the embedding vectors of the evolution path and final failure mode of the similar historical cases as knowledge graph enhancement vectors;
[0043] S3. The knowledge graph enhancement vector and the input feature tensor are fused and then input into the generator network. The generator network predicts the breach morphology parameters and the downstream flood inundation dynamic grid sequence within a preset time period based on the fused features, and outputs the prediction results.
[0044] S4. The discriminator network performs authenticity and physical consistency judgments based on the prediction results; when the judgment results meet the preset conditions, it outputs a valid risk prediction result.
[0045] In the embodiments of this application, the dam failure risk knowledge graph is constructed by extracting engineering entities, disaster-causing factors, damage modes, evolution processes, disposal measures and consequences from domestic and foreign plain reservoir dam failure and danger cases as nodes of the knowledge graph, defining the causal relationship, temporal relationship and similarity relationship between each node as the edges of the knowledge graph, and constructing it based on the nodes and edges.
[0046] The dam failure risk knowledge graph can perform in-depth analysis of each case and extract the initial location of the hazard, the process of its development and expansion, and the timeline of the final collapse as entities in the evolution process. It also extracts early warning failure nodes, rescue failure nodes, and successful handling nodes as key decision point entities, and establishes temporal relationship edges between the key decision point entities and the evolution process entities. This allows the dam failure risk knowledge graph to fully present the detailed evolution path from the initial cause to the final collapse, as well as the key turning points in the path. The graph embedding algorithm transforms nodes and edges into low-dimensional vector storage, mapping entities and relationships to a low-dimensional vector space.
[0047] When constructing the dam failure risk knowledge graph of this application, the following steps or methods can be adopted:
[0048] Based on dam failures and major emergencies at coastal plain reservoirs both domestically and internationally, the data collection scope covers academic literature, engineering safety assessment reports, emergency response archives, industry conference materials, and publicly available emergency reports. It considers relevant cases of overtopping dam failures, local instability, and cascading risks of expansion at typical coastal plain reservoirs, while also including examples of dam failures and emergency response at similar plain reservoirs both domestically and internationally. During the knowledge graph construction process, two core tasks were completed: entity extraction and relationship definition. Entity extraction covers six categories: engineering entities, disaster-causing factors, damage modes, evolution processes, response measures, and consequences.
[0049] The engineering entity includes core attributes such as dam type, dam height, total reservoir capacity, dam length, dam foundation geological type, construction year, and monitoring facility layout; disaster-causing factors include key inducing factors such as extreme rainfall intensity, duration of high water level operation during the flood season, tidal backwater value, dam seepage pressure, and dam displacement and deformation; failure modes include typical types such as overtopping dam failure, local instability dam failure, progressive dam failure, and multi-factor coupled dam failure; evolution process includes full-process information on the initial stage of danger, the stage of damage development, and the final state of failure; response measures include practical means such as flood discharge regulation, emergency reinforcement of the dam body, personnel evacuation and relocation, and regional flood control; consequences and impacts include quantitative and qualitative information such as inundation range, water depth distribution, arrival time, property loss, personnel safety risks, and ecological and environmental impacts.
[0050] In the definition of relationships, the logic of the connections between entities is clearly defined. Causal relationships are used to characterize the induced association between disaster-causing factors and damage modes and evolution processes. Temporal relationships are used to characterize the sequence between the various stages of the development of the hazard, the implementation of the response measures and the failure result. Similarity relationships are used to characterize the feature matching association between cases with the same dam type, the same induced cause and the same damage mode. Through the above definition of nodes and edges, a knowledge graph structure that fully covers all elements of dam failure risk is formed.
[0051] After the knowledge graph nodes and edges are constructed, graph embedding algorithms are used to transform the entities and relationships in the graph into low-dimensional vectors for digital storage. The TransE algorithm or GraphSAGE algorithm can be used for mapping. The TransE algorithm is suitable for vector transformation of simple relationships and can quickly achieve low-dimensional spatial mapping of entities and relationships. The GraphSAGE algorithm can adapt to the topological structure features of the knowledge graph and achieve the aggregation and extraction of neighborhood entity features. Both algorithms can transform unstructured case knowledge into computable vector data. This enables in-depth analysis of each case entered into the database, accurately extracting the initial location of the hazard, the process of damage development and expansion, and the timeline of the final collapse as evolutionary process entities. Simultaneously, early warning failure nodes, rescue failure nodes, and successful handling nodes are extracted as key decision point entities. Temporal relationship edges are established between key decision point entities and evolutionary process entities according to the chronological order of hazard development. This allows the knowledge graph to fully present the detailed evolutionary path from the initial disaster trigger to the final collapse, clearly identifying the key inflection points in the path that can be intervened and controlled, providing a foundation for subsequent similar case retrieval and knowledge empowerment.
[0052] In the embodiments of this application, the step of retrieving similar historical cases from a pre-constructed dam failure risk knowledge graph using the current operating condition as the query condition includes:
[0053] Calculate the cosine similarity between the real-time operating data of the current operating condition and the engineering entity feature vectors of each historical case in the dam failure risk knowledge graph, and select the cases with the highest similarity as similar cases.
[0054] The evolution path sequence and final destruction mode label corresponding to the similar cases are extracted. The evolution path sequence is converted into a fixed-length path feature vector through a sequence encoder. The final destruction mode label is converted into a mode feature vector through one-hot encoding. The path feature vector and the mode feature vector are concatenated to form the knowledge graph enhancement vector.
[0055] In the embodiments of this application, the real-time monitoring data, the digital elevation model data, the land use type data, and the building distribution data are processed respectively to form an input feature tensor, including:
[0056] The real-time monitoring data is normalized to form a real-time operating condition tensor. The digital elevation model data, land use type data, and building distribution data are encoded into a multi-dimensional matrix to form a geographic environment tensor. The real-time operating condition tensor and the geographic environment tensor are concatenated to form the input feature tensor.
[0057] In the embodiments of this application, the generator network adopts an improved Attention U-Net architecture, including an encoder, a decoder, and a multi-head attention mechanism. The encoder performs multi-level downsampling on the input fusion features to extract deep semantic features, and the decoder upsamples the deep semantic features to restore spatial resolution. A multi-head attention mechanism is introduced at the skip connections corresponding to the encoder and the decoder. The multi-head attention mechanism calculates the attention weight between the encoder feature map and the location encoding of historical high-risk areas retrieved from the dam break risk knowledge graph. The encoder feature map is weighted and enhanced according to the attention weight, so that the decoder gives higher attention to the historical high-risk areas when generating the inundation range. Finally, the predicted results of the breach morphology parameters and the downstream flood inundation dynamic grid sequence are output.
[0058] In the embodiments of this application, the discriminator network is a dual-output discriminator that performs both authenticity discrimination and physical consistency discrimination. The authenticity discrimination is used to determine whether the generated flood map is visually and statistically similar to the sample generated by the high-fidelity hydrodynamic model. The physical consistency discrimination calculates the water volume conservation error and the consistency score between the water flow direction and the terrain slope in the generated result by introducing a physical constraint loss function.
[0059] In the embodiments of this application, the physical constraint loss function includes a water volume conservation loss term and a flow direction consistency loss term; wherein, the water volume conservation loss term includes the error between the change in water volume between different time steps calculated based on the inundation depth raster sequence output by the generator network and the actual breach outflow; the flow direction consistency loss term includes the proportion of pixels with inconsistent flow directions calculated based on the flow direction of each pixel calculated based on the digital elevation model data and compared with the inundation evolution direction output by the generator network.
[0060] The reservoir dam failure risk dynamic simulation method of this invention can rely on the engineering characteristics of reservoirs in coastal plains, such as flat terrain, widespread distribution of soft soil in dam foundations, dense surrounding urban facilities, and the dual impact of tidal backwater and extreme rainfall. It combines the knowledge accumulated from historical dam failure cases, multi-source real-time monitoring data and geospatial data, and achieves rapid and accurate simulation of dam failure risk through dam failure risk knowledge graph empowerment and generative adversarial network collaborative calculation.
[0061] The following describes the complete dynamic simulation process of reservoir dam failure risk after the dam risk knowledge graph is constructed.
[0062] First, multi-source data acquisition and input feature tensor generation were carried out. Data acquisition was divided into real-time monitoring data acquisition and geospatial data acquisition. Real-time monitoring data could come from the online monitoring system of the coastal plain reservoir, including dynamic data such as reservoir water level, dam seepage pressure, dam displacement, real-time rainfall, nearshore tide level, and dam temperature. Geospatial data was obtained from the digital elevation model data of the downstream area of the reservoir. At the same time, land use type data and building distribution data of the downstream area were supplemented. Land use type was distinguished into different categories such as urban construction land, farmland, green space, and water area. Building distribution was marked with spatial attributes such as building height, density, and usage function. The collected real-time monitoring data is normalized to eliminate the differences in dimensions and numerical magnitudes of different monitoring indicators, forming a standardized real-time operating condition tensor. The digital elevation model data, land use type data, and building distribution data are spatially encoded to transform discrete geographic attributes into continuous multidimensional matrices, forming a geographic environment tensor. Subsequently, the real-time operating condition tensor and the geographic environment tensor are dimensionally concatenated to generate an input feature tensor containing both real-time operating condition and geographic environment information. This input feature tensor can comprehensively reflect the current operating status of the reservoir and the spatial characteristics of downstream topography and urban construction, adapting to the simulation needs of complex underlying surfaces of coastal plain reservoirs.
[0063] Next, knowledge graph-enhanced vector extraction and feature fusion operations are performed. Using the current real-time operating conditions of the reservoir as query conditions, similar historical cases are retrieved from the constructed dam failure risk knowledge graph. The real-time monitoring data of the current operating conditions is transformed into vector data of the same dimension as the engineering entity feature vectors in the knowledge graph. By calculating cosine similarity, the feature matching degree between the current operating conditions and each historical case in the knowledge graph is matched. A preset number of cases with the highest similarity values are selected as similar cases. The preset number can be flexibly adjusted according to the inference accuracy requirements. For the selected similar cases, their corresponding evolution path sequences and final failure mode labels are extracted. The evolution path sequences are input into a sequence encoder and transformed into fixed-length path feature vectors through sequence encoding processing. The final failure mode labels are processed using one-hot encoding to transform them into standardized mode feature vectors. The path feature vectors and mode feature vectors are concatenated to form a knowledge graph-enhanced vector that integrates the evolution patterns of historical cases and failure mode features.
[0064] Subsequently, the knowledge graph augmentation vector is fused with the previously generated input feature tensor. The fusion method can be channel-weighted concatenation, which not only preserves the real-time data and spatial information of the input feature tensor, but also incorporates the knowledge and experience of historical cases to form a fused feature for model input, achieving deep coupling of real-time data, spatial information and historical knowledge.
[0065] The fused features are input into the generator network for inference calculation. The generator network adopts an improved Attention U-Net architecture, which is adapted to the spatial feature inference requirements of flood evolution in coastal plain reservoirs. Its encoder performs multi-level downsampling processing on the input fused features. Through the alternating operation of convolutional layers and pooling layers, it gradually extracts deep semantic features from the fused features and explores the correlation between real-time working conditions, geographical environment and historical case knowledge. The decoder performs upsampling processing on the deep semantic features extracted by the encoder and gradually restores the spatial resolution of the feature map through deconvolution operation to restore the spatial distribution features of flood inundation. A multi-head attention mechanism is introduced at the skip connection positions between the encoder and decoder. This mechanism calculates the attention weight between the encoder's output feature map and the historical high-risk area location codes retrieved from the knowledge graph. Historical high-risk areas include key locations such as breach locations, severely flooded areas, and areas prone to water convergence in historical cases. The encoder feature map is weighted and enhanced based on the calculated attention weights, allowing the decoder to assign higher attention weights to historical high-risk areas when generating flood range and breach parameters. This aligns with the characteristics of wide flood spread and concentrated risk areas in coastal plain reservoirs. The generator network ultimately outputs two sets of projection results: one set of breach morphology parameters, including core parameters such as breach location, width, depth, and development rate; and the other set of downstream flood inundation dynamic raster sequences, containing spatial data such as flood inundation range, water depth distribution, and flood arrival time at different time steps, enabling dynamic and visual projection of dam breach risk.
[0066] After the generator network outputs the inference results, the results are input into the discriminator network for dual discrimination and verification. The discriminator network adopts a dual-output structure, simultaneously completing the authenticity judgment and physical consistency judgment. Among them, the authenticity judgment is used to compare the differences between the inundation map output by the generator and the standard samples generated by the high-fidelity hydrodynamic model. The high-fidelity hydrodynamic model includes commonly used industry models such as MIKE21 and HEC RAS. The discrimination process is carried out from two dimensions: visual features and statistical features, to verify the degree of matching between the contour, distribution, and values of the generated inundation map and the standard samples, and to determine the authenticity and validity of the inference results. Physical consistency judgment is achieved by introducing a physical constraint loss function, which includes a water volume conservation loss term and a flow direction consistency loss term. The water volume conservation loss term calculates the change in water volume in the downstream area at different time steps based on the inundation depth raster sequence output by the generator, compares it with the theoretical value of the breach outflow, and calculates the error between the two to ensure that the extrapolation results conform to the basic hydrodynamic law of water volume conservation. The flow direction consistency loss term calculates the topographic slope of each pixel based on the digital elevation model data of the downstream area to determine the flow direction under natural conditions, compares it with the flood evolution direction output by the generator pixel by pixel, and counts the proportion of pixels with inconsistent directions to verify the matching between the flood evolution direction and the topographic conditions.
[0067] Specifically, when both the authenticity judgment result and the physical consistency judgment result output by the discriminator network meet the preset conditions, the inference result of the generator network is output as the effective risk prediction result. If the preset conditions are not met, the discrimination error is fed back to the generator network to iteratively optimize the generator's parameters and feature extraction logic until the output of the effective result meets the requirements.
[0068] See Figure 2 As shown in the figure, this application provides a dynamic simulation system for reservoir dam failure risk, including:
[0069] The data acquisition module is used to collect real-time monitoring data of reservoir operation and obtain digital elevation model data, land use type data and building distribution data of the downstream area of the reservoir.
[0070] The feature tensor generation module is used to process the real-time monitoring data, the digital elevation model data, the land use type data, and the building distribution data respectively to form an input feature tensor;
[0071] The knowledge graph retrieval module is used to retrieve similar historical cases from a pre-built dam failure risk knowledge graph using the current working conditions as the query criteria.
[0072] An enhanced vector generation module is used to extract the embedding vectors of the evolution path and final destruction mode of the similar historical cases, as enhanced vectors for the knowledge graph;
[0073] The feature integration module is used to fuse the knowledge graph enhancement vector with the input feature tensor to form a fused feature.
[0074] The prediction module is used to use a generator network to predict the breach morphology parameters and the downstream flood inundation dynamic grid sequence within a preset time period based on fused features, and output the prediction results.
[0075] The discrimination module is used to perform authenticity discrimination and physical consistency discrimination based on the prediction results using a discriminator network; when the discrimination results meet preset conditions, it outputs a valid risk prediction result.
[0076] The reservoir dam failure risk dynamic simulation system described in this application is built based on the dynamic simulation method. It can realize the fully automated operation of data collection, knowledge construction, feature fusion, simulation calculation, discrimination and verification and emergency response. The various components of the system work together to adapt to the actual engineering needs of reservoir dam failure risk prevention and control in coastal plains.
[0077] The data acquisition module is responsible for collecting and preprocessing multi-source data. It collects real-time dynamic monitoring data of the reservoir's operation, covering core indicators such as reservoir water level, dam seepage pressure, dam displacement and deformation, real-time rainfall intensity, and nearshore tide level. The data acquisition frequency is adapted to the real-time requirements of reservoir safety monitoring. Simultaneously, it acquires digital elevation model data, land use type data, and building distribution data for the downstream area of the reservoir. The digital elevation model data uses a high-precision resolution adapted to plain terrain, and the land use type data and building distribution data achieve accurate spatial coordinate matching. The data acquisition module performs preprocessing operations on the collected raw data, including outlier removal, data denoising, spatiotemporal alignment, and normalization, eliminating data errors and spatiotemporal deviations to ensure the accuracy and consistency of the input data, providing a high-quality data foundation for subsequent extrapolation and calculation.
[0078] The system includes a dam break risk knowledge graph module, which is responsible for maintaining the dam break risk knowledge graph. This module collects dam break and hazard cases from domestic and international coastal plain reservoirs through multiple channels, filters, cleans, and structures the case data. Knowledge graph nodes are extracted according to six categories: engineering entity, disaster-causing factor, damage mode, evolution process, response measures, and consequences. Causal relationships, temporal relationships, and similarity relationships between nodes are defined to form knowledge graph edges, constructing a complete dam break risk knowledge graph. The knowledge graph construction module includes a case in-depth analysis unit. This unit performs detailed analysis on each case, extracting the initial location of the hazard, the process of damage development and expansion, and the timeline of the final collapse as evolution process entities. It also extracts warning failure nodes, rescue failure nodes, and successful response nodes as key decision point entities. Temporal relationship edges are established between key decision point entities and evolution process entities according to the temporal logic of hazard development, fully reconstructing the evolution path and key turning points of dam break risk. The dam break risk knowledge graph module uses a graph embedding algorithm to transform entities and relationships in the dam break risk knowledge graph into low-dimensional vector storage. Depending on the application scenario, the TransE algorithm or GraphSAGE algorithm is selected to complete the vector mapping, realizing the digitization and computability of case knowledge, and providing support for similar case retrieval and knowledge empowerment.
[0079] The system incorporates a feature fusion module, which establishes data interaction connections with the data acquisition module and the dam failure risk knowledge graph module. This module receives preprocessed data from the data acquisition module, generates real-time operating condition tensors and geographic environment tensors, and concatenates them into the input feature tensor. Simultaneously, it receives knowledge graph data from the dam failure risk knowledge graph module, performs similar historical case searches using the current reservoir operating conditions as query criteria, and filters the most closely matched similar cases through cosine similarity calculation. It then extracts the evolution path and failure mode features of these similar cases, and generates an enhanced knowledge graph vector after sequence encoding and one-hot encoding. The feature fusion module then weights and fuses the enhanced knowledge graph vector with the input feature tensor to generate a fused feature that integrates real-time operating conditions, geospatial information, and historical knowledge. This fused feature is then transmitted to the subsequent inference module, achieving coupled input of multi-dimensional features.
[0080] The system includes a prediction module employing an improved AttentionU Net architecture. This module receives fused features from the feature fusion module, performs multi-level downsampling through the encoder to extract deep semantic features, and upsampling through the decoder to restore spatial resolution. A multi-head attention mechanism is introduced at the skip connections between the encoder and decoder. Attention weights are calculated based on the location encoding of historical high-risk areas in the knowledge graph, and the encoder feature map is weighted and enhanced to focus on high-risk areas for inference calculations. After the generator network module completes its calculations, it outputs breach morphology parameters and a dynamic raster sequence of downstream flood inundation. The breach morphology parameters accurately reflect the location, size, and development pattern of the breach, while the dynamic raster sequence of flood inundation visually presents the flood evolution at different time steps, providing core data for risk assessment.
[0081] The system includes a discrimination module connected to the prediction module. Implemented using a discriminator network module, this module receives the inference results output from the generator network of the prediction module. It employs a dual-output structure to simultaneously perform realism and physical consistency discrimination. The discrimination module internally includes realism discrimination units and physical consistency discrimination units. The realism discrimination unit compares the visual and statistical similarity of the generated results with standard samples of a high-fidelity hydrodynamic model to verify the realism of the results. The physical consistency discrimination unit performs calculations using physical constraint loss functions. The water volume conservation loss term calculates the error between water volume change and breach outflow, while the flow direction consistency loss term compares the matching degree between the inferred flow direction and the terrain slope direction, statistically analyzing the proportion of inconsistent pixels. The discriminator network module determines the validity of the inference results based on the dual discrimination results, outputting valid results to subsequent modules and simultaneously feeding back the discrimination error to the generator network module for iterative model optimization.
[0082] The system also includes an emergency simulation module, which connects to the judgment module. When the effective risk prediction results output by the judgment module exceed a preset risk threshold, the emergency simulation module automatically activates. The emergency simulation module accesses case data from the knowledge graph construction module, retrieves historical successful response cases under similar working conditions and damage patterns, and dynamically adapts and optimizes historical evacuation plans by combining real-time traffic data and population heatmap data. It generates tiered early warning instructions based on risk levels and plans optimal evacuation routes by considering downstream population distribution, transportation networks, and the location of refuge sites. The tiered early warning instructions and evacuation route planning results are then pushed to emergency decision-makers through a visualization platform, providing precise guidance for on-site handling and personnel transfer, achieving closed-loop management across the entire chain from risk simulation to emergency response.
[0083] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention.
[0084] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the claims be included within the invention.
[0085] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A method for dynamically predicting the risk of dam failure in reservoirs, characterized in that, Includes the following steps: Real-time monitoring data of reservoir operation is collected, and digital elevation model data, land use type data and building distribution data of the downstream area of the reservoir are obtained. The real-time monitoring data and the digital elevation model data, land use type data and building distribution data are processed respectively to form an input feature tensor. Using the current working condition as the query condition, similar historical cases are retrieved from the pre-constructed dam failure risk knowledge graph. The evolution path and final failure mode of the similar historical cases are extracted as embedding vectors and used as knowledge graph enhancement vectors. The knowledge graph enhancement vector and the input feature tensor are fused and then input into the generator network. The generator network predicts the breach morphology parameters and the downstream flood inundation dynamic grid sequence within a preset time period based on the fused features, and outputs the prediction results. Based on the prediction results, the discriminator network performs authenticity and physical consistency judgments; when the judgment results meet preset conditions, it outputs a valid risk prediction result.
2. The method for dynamically extrapolating the risk of dam failure according to claim 1, characterized in that, The dam failure risk knowledge graph is constructed by extracting engineering entities, disaster-causing factors, damage modes, evolution processes, disposal measures, and consequences from domestic and international cases of dam failures and emergencies in plain reservoirs as nodes of the knowledge graph, and defining the causal relationships, temporal relationships, and similarity relationships between nodes as edges of the knowledge graph. The dam failure risk knowledge graph performs in-depth analysis on each case and extracts the initial location of the danger, the process of its development and expansion, and the timeline of the final collapse as entities in the evolution process. It also extracts the early warning failure node, the emergency rescue failure node, and the successful handling node as key decision point entities. It establishes temporal relationship edges between the key decision point entities and the evolution process entities, so that the dam failure risk knowledge graph can fully present the detailed evolution path from the initial cause to the final collapse, as well as the key turning points in the path. The graph embedding algorithm transforms nodes and edges into low-dimensional vectors for storage, and maps entities and relationships to the low-dimensional vector space.
3. The method for dynamically extrapolating the risk of dam failure according to claim 1, characterized in that, The process of retrieving similar historical cases from a pre-constructed dam failure risk knowledge graph using the current operating conditions as query criteria includes: Calculate the cosine similarity between the real-time operating data of the current operating condition and the engineering entity feature vectors of each historical case in the dam failure risk knowledge graph, and select the cases with the highest similarity as similar cases. The evolution path sequence and final destruction mode label corresponding to the similar cases are extracted. The evolution path sequence is converted into a fixed-length path feature vector through a sequence encoder. The final destruction mode label is converted into a mode feature vector through one-hot encoding. The path feature vector and the mode feature vector are concatenated to form the knowledge graph enhancement vector.
4. The method for dynamically extrapolating the risk of dam failure according to claim 1, characterized in that, The real-time monitoring data, digital elevation model data, land use type data, and building distribution data are processed separately to form an input feature tensor, including: The real-time monitoring data is normalized to form a real-time operating condition tensor. The digital elevation model data, land use type data, and building distribution data are encoded into a multi-dimensional matrix to form a geographic environment tensor. The real-time operating condition tensor and the geographic environment tensor are concatenated to form the input feature tensor.
5. The method for dynamically extrapolating the risk of dam failure according to claim 1, characterized in that, The generator network employs an improved Attention U-Net architecture, including an encoder, a decoder, and a multi-head attention mechanism. The encoder performs multi-level downsampling on the input fused features to extract deep semantic features, and the decoder upsamples the deep semantic features to restore spatial resolution. A multi-head attention mechanism is introduced at the skip connections corresponding to the encoder and decoder. The multi-head attention mechanism calculates the attention weight between the encoder feature map and the location encoding of historical high-risk areas retrieved from the dam breach risk knowledge graph. The encoder feature map is weighted and enhanced according to the attention weight, so that the decoder gives higher attention to the historical high-risk areas when generating the inundation range. Finally, the predicted results of the breach morphology parameters and the downstream flood inundation dynamic raster sequence are output.
6. The method for dynamically extrapolating the risk of dam failure according to claim 1, characterized in that, The discriminator network is a dual-output discriminator that performs both authenticity and physical consistency discrimination. The authenticity discrimination is used to determine whether the generated inundation map is visually and statistically similar to the sample generated by the high-fidelity hydrodynamic model. The physical consistency discrimination calculates the water volume conservation error and the consistency score between the water flow direction and the terrain slope in the generated result by introducing a physical constraint loss function.
7. The method for dynamically extrapolating the risk of dam failure according to claim 6, characterized in that, The physical constraint loss function includes a water volume conservation loss term and a flow direction consistency loss term; wherein, the water volume conservation loss term includes the error between the change in water volume between different time steps calculated based on the inundation depth raster sequence output by the generator network and the actual breach outflow; the flow direction consistency loss term includes the proportion of pixels whose flow direction is inconsistent, calculated based on the flow direction of each pixel calculated based on the digital elevation model data and compared with the inundation evolution direction output by the generator network.
8. A dynamic simulation system for reservoir dam failure risk, characterized in that: include: The data acquisition module is used to collect real-time monitoring data of reservoir operation and obtain digital elevation model data, land use type data and building distribution data of the downstream area of the reservoir. The feature tensor generation module is used to process the real-time monitoring data, the digital elevation model data, the land use type data, and the building distribution data respectively to form an input feature tensor; The knowledge graph retrieval module is used to retrieve similar historical cases from a pre-built dam failure risk knowledge graph using the current working conditions as the query criteria. An enhanced vector generation module is used to extract the embedding vectors of the evolution path and final destruction mode of the similar historical cases, as enhanced vectors for the knowledge graph; The feature integration module is used to fuse the knowledge graph enhancement vector with the input feature tensor to form a fused feature. The prediction module is used to use a generator network to predict the breach morphology parameters and the downstream flood inundation dynamic grid sequence within a preset time period based on fused features, and output the prediction results. The discrimination module is used to perform authenticity discrimination and physical consistency discrimination based on the prediction results using a discriminator network; When the judgment result meets the preset conditions, the effective risk prediction result is output.
9. The reservoir dam failure risk dynamic simulation system according to claim 8, characterized in that, It also includes an emergency simulation module, connected to the discrimination module, which is used to call the knowledge graph module when the effective risk prediction result output by the discrimination module exceeds the preset risk threshold, retrieve historical successful disposal measures cases under similar working conditions and similar damage modes, dynamically adapt and optimize historical evacuation plans by combining real-time traffic data and population heat map data, generate graded early warning instructions and personnel evacuation route plans, and push them to decision-makers through a visualization platform.
10. The reservoir dam failure risk dynamic simulation system according to claim 8, characterized in that, The knowledge graph module includes a case deep analysis unit, which is used to perform deep analysis on each case and extract the initial location of the danger, the process of development and expansion, and the timeline of the final collapse as evolutionary process entities. It also extracts the early warning failure node, the emergency rescue failure node, and the successful handling node as key decision point entities, and establishes the temporal relationship edge between the key decision point entities and the evolutionary process entities, so that the dam failure risk knowledge graph can fully present the detailed evolution path from the initial cause to the final collapse, as well as the key turning points in the path. The graph embedding algorithm transforms nodes and edges into low-dimensional vectors for storage, and maps entities and relationships to the low-dimensional vector space.