Reservoir safety risk factor quantification method based on causal knowledge graph

Abstract

The present application relates to the technical field of reservoir safety management and artificial intelligence fusion, and particularly relates to a reservoir safety risk factor quantification method based on a causal knowledge graph, which maps entities and relationships into vector representations using a graph embedding algorithm, and quantitatively infers the causal effect between the treatment variable and the response variable by combining a causal inference model to calculate the average treatment effect and the individual treatment effect. The present application further introduces a centrality index in the graph structure as a weighting factor to construct a risk score function, achieving quantitative identification and ordering of key hidden trouble factors. At the same time, the system supports dynamic updating of the graph structure and model inference results, and has a model verification mechanism and automatic risk level division function. Compared with the prior art, the present application has the advantages of clear structure expression, clear causal mechanism, strong result interpretability and good dynamic adaptability, and is suitable for risk identification and early warning management of multiple types of reservoirs.

Description

Technical Field

[0001] This invention relates to the field of reservoir safety management and artificial intelligence integration technology, and in particular to a method for quantifying reservoir safety risk factors based on causal knowledge graphs. Background Technology

[0002] With the increasing frequency of extreme weather events and the continuous increase in the operating years of reservoirs, the safety situation of reservoirs is becoming increasingly complex. On the one hand, risk factors such as floods, landslides, and structural aging are intertwined and superimposed; on the other hand, the population density and asset scale affected by reservoir inundation scenarios continue to rise, placing higher demands on safety assessment and early warning. Accurately identifying key disaster-causing factors and clarifying the transmission path between them and disaster consequences has become a core technical challenge in building a scientific and efficient reservoir risk management system.

[0003] Therefore, in recent years, a large number of studies have explored data-driven risk identification methods. For example, Chinese invention publication CN113807570A discloses a method and system for assessing the risk level of reservoir dams based on XGBoost. This method proposes using the XGBoost algorithm to assess the risk level of reservoir dams, extracting structural and operational status data through feature engineering, and optimizing model parameters through cross-validation to achieve automatic prediction of risk levels. While such methods improve the efficiency of risk assessment to some extent, their core remains based on "correlation modeling," making it difficult to explain causal relationships between variables, especially in scenarios with cross-disturbances or complex triggers, where they lack effective reasoning capabilities.

[0004] On the other hand, some technologies focus on expanding monitoring accuracy and indicator systems. For example, Chinese invention patent CN118411021A discloses an internet-based reservoir dam construction safety monitoring system. By analyzing multi-source data such as dam offset, damage level, and water level pressure, it constructs a safety assessment coefficient system, achieving automated hazard alarms. Another Chinese invention patent CN118781773A discloses a reservoir safety early warning method based on risk development monitoring and prediction. This method attempts to introduce a multi-physics development trend model and utilizes artificial intelligence to extrapolate future risk evolution paths, providing data support for real-time safety early warnings. However, a common limitation of these methods is that, although the information sources are richer, they still lack a structured and logical explanatory framework for the core question of "how risks occur."

[0005] Some existing research has begun to move towards "interpretive modeling." For example, Chinese invention patent CN119622557A discloses a method and system for constructing and operating a multimodal large-scale model for reservoir dam safety. This method attempts to build a multimodal large-scale model system for dam safety, integrating BIM geometry, historical hazards, monitoring sensor data, and engineering documentation to train a Transformer network to predict hazard events. While this method enhances the model's expressive power, its modeling still falls under the "black box mechanism" of deep learning, making it difficult to directly present the causal paths and logical explanations between risk factors.

[0006] In summary, current reservoir risk assessment technology exhibits three significant shortcomings: First, although it can access multi-source monitoring data, it lacks a unified and structured semantic expression mechanism, making it difficult to clearly present the complex relationships between disaster-causing factors, environmental conditions, and disaster consequences; second, most existing models focus on predictive performance and lack causal reasoning capabilities, making it impossible to determine whether a change in a certain factor truly triggered a risk outcome; third, risk identification results are difficult to interpret and track, making them difficult for relevant departments to directly accept or use for responsibility allocation and emergency response in actual management.

[0007] Causal knowledge graphs offer a potential breakthrough for solving the aforementioned challenges. On the one hand, they possess powerful structural representation capabilities, able to abstract multi-source heterogeneous information into semantic networks in the form of "entity-relationship-entity" triples. On the other hand, by integrating causal inference models (such as causal forests and the DoWhy framework), they can identify causal paths between variables based on the graph structure, quantify the mechanism and intensity of disaster-causing factors, and form interpretable, visualized, and quantifiable risk reasoning results.

[0008] While this "graph + causal" approach has seen initial applications in fields such as general artificial intelligence and medical analysis, it lacks a specific implementation path in the field of reservoir safety, particularly in areas such as graph construction, causal variable definition, model integration methods, and dynamic update mechanisms. There is a lack of systematic research and application examples. Therefore, there is an urgent need to construct a reservoir safety hazard identification and risk analysis method that integrates causal knowledge graphs and possesses structured expression and causal effect quantification capabilities. This would fill the gaps in the interpretability, logic, and real-time performance of existing technologies, and promote the evolution of reservoir risk identification towards "mechanism-driven, intelligent reasoning." Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a method for quantifying reservoir safety risk factors based on causal knowledge graphs. It aims to overcome the limitations of traditional risk assessment methods that rely on empirical rules, data correlation modeling, or static indicator systems, and solve core problems such as the difficulty in integrating multi-source data, the inability to interpret causal paths, and the difficulty in quantifying and ranking risks.

[0010] To address the aforementioned technical problems, this invention provides a method for quantifying reservoir safety risk factors based on causal knowledge graphs. By constructing a structured semantic graph, introducing a causal inference algorithm, and combining graph topological attributes for risk quantification, this invention can achieve automatic identification, causal analysis, scoring, ranking, and dynamic updating of reservoir safety risk factors. The method includes the following steps: S1. Collect multi-source data from reservoirs and reservoir areas, including static structural data and dynamic operational data, and construct a set of triples including entities, relations, and attributes. The multi-source data includes, but is not limited to, dam structural parameters (such as dam type, dam height, permeability coefficient, etc.), geographical location, geological conditions, reservoir area DEM, land use type, population density, regional GDP, water level changes, reservoir capacity changes, rainfall intensity, flood events, and reservoir inundation scenarios. The data sources may include remote sensing monitoring, sensor data collection, basic geographic databases, and manual survey data.

[0011] Furthermore, based on the preprocessing and fusion of the collected multi-source data, a set of triples consisting of entities, attributes, and relationships is constructed according to the logical and causal relationships between key entities within the reservoir system. ,in, For the head entity, For tail entities, A semantic label representing a causal or dependent relationship between two things.

[0012] Furthermore, the set of triples Includes at least the following types, namely: <Rainfall, triggering, floods>; <Dam permeability, reinforcement, leakage risk>; <Population density, exposure to, flooded areas>; <Topography, exacerbation, flood propagation>; <Gate malfunction, impact on flood discharge capacity>.

[0013] S2, Set the triplet set The input graph embedding model uses the TransE algorithm or graph convolutional networks to map entities and relations into graph embedding representations in vector space, forming a representation composed of several risk factors. Structured semantic graph as graph nodes.

[0014] Furthermore, this invention employs graph embedding models such as TransE or Graph Convolutional Neural Networks (GCN) for the triple set. Vectorized modeling is employed, embedding entities and relationships into the same vector space to maintain consistency between the graph structure and semantics. The objective function is then optimized. for: in, It is a positive triplet; For negative triples generated by perturbation; This is the boundary margin hyperparameter, used to enhance the distinction between positive and negative samples; This represents the portion greater than zero. This embedding result serves as the input feature representation for subsequent causal inference modeling.

[0015] Preferably, the graph embedding model uses the TransE algorithm, whose optimization objective is to minimize the entity and relation vectors, satisfying: Used to maintain the geometric consistency of causal chains in a knowledge graph.

[0016] S3. The graph embedding is used as a covariate X, and treatment variables T (such as whether the rainstorm intensity exceeds a threshold) and response variables Y (such as casualties or property damage) are introduced. The causal effect between treatment variables T and response variables Y is then analyzed based on the Causal Forest model. Quantitative inferences were performed to estimate the average treatment effect (ATE) and individual treatment effect (ITE) for each variable.

[0017] Furthermore, this invention constructs a causal inference model with treatment variable T, covariate X, and response variable Y. Treatment variable T can represent events such as the occurrence of a specific intensity of rainstorm, water level exceeding limits, or structural anomalies. Response variable Y is a quantitative indicator of the accident's consequences, including the degree of casualties, economic losses, or ecological damage. Response variable Y can be quantified using the following disaster consequence function: Y= f (Flooding depth, population density, exposure duration, regional GDP) in, f The weighted combination function can be obtained through regression training of historical cases and is used to map physical and social exposure factors into disaster consequences.

[0018] Furthermore, the causal inference part employs the Causal Forest algorithm, which constructs multiple heterogeneous regression trees. Estimate the mean treatment effect (ATE) and individual treatment effect (ITE) for each variable. The mathematical expression of the causal inference model is as follows: Among them, the causal inference model constructs multiple heterogeneous regression trees. We performed split modeling on the covariate X and used the Bootstrap method to split each risk factor. Interval estimation was performed on the average treatment effect ATE estimate to obtain the risk factor. confidence interval This is used to enhance the credibility and statistical robustness of the inference results.

[0019] S4. Calculate each risk factor in the structured semantic graph. The centrality index is used to obtain its structural weight. Its causal effect Together, they constitute a risk scoring function used to quantitatively identify and rank key hidden danger factors.

[0020] To comprehensively reflect the causal strength of variables and their importance in the structured semantic graph, this embodiment introduces a weighted scoring mechanism based on graph topology. This is achieved by calculating risk factors. The structural weights are obtained from centrality metrics such as PageRank or DegreeCentrality. Its causal effect Together they constitute the final risk scoring function. ,Right now: Furthermore, the scoring system supports interpretable ranking of risk factors and prioritization of governance.

[0021] Furthermore, to enhance the explanatory power of causal paths, this invention can also estimate the causal strength score of a variable A on the response variable Y using the conditional mutual information method. , expressed as: Used to assess the strength of the causal influence of variable A on response variable Y (outcome) given covariate X.

[0022] Among them, the causal strength score It can complement the average treatment effect ATE and be used to identify the higher-order coupling strength of key factors in nonlinear influence paths.

[0023] S5. Output a list of key risk factors and their quantitative indicators, including: for each risk factor The average treatment effect (ATE), confidence interval, contribution ranking, and risk level classification are used for the identification and management recommendations of reservoir safety early warning systems.

[0024] This invention supports a dynamic update mechanism for knowledge-structured semantic graphs. When a new round of data (such as the latest rainfall, water level anomalies, dam seepage, etc.) is collected, the system can automatically generate a new set of triples. The system can automatically insert data such as "current water level, exceeding limit, warning water level" into the graph structure, triggering the reconstruction or fine-tuning of the causal inference model, thereby achieving a closed-loop response of graph-reasoning-scoring.

[0025] Furthermore, to verify the accuracy and practicality of the causal inference model, this invention introduces a verification module to compare the identified high-risk factors with historical accident data. Indicators such as ROC-AUC, mean squared error (MSE), and F1-score are used to evaluate the model's prediction accuracy and risk level classification accuracy, ensuring that the model results have engineering credibility.

[0026] Furthermore, based on the numerical distribution characteristics of the scoring results, the system can automatically classify all risk factors into four levels—high, relatively high, medium, and low—using quantile methods, K-Means clustering, or Gaussian mixture models, and attach governance suggestion templates for each risk factor category. The output results are visualized in the form of graphs and linked to reservoir management platforms or early warning systems to provide intuitive and usable auxiliary decision support.

[0027] This invention also proposes a system for quantifying reservoir safety risk factors, comprising a data acquisition module, a graph construction module, a graph embedding and modeling module, a causal inference module, a risk scoring and output module, and a dynamic update module. These modules work collaboratively to achieve structured knowledge modeling, causal reasoning, and intelligent risk assessment of the reservoir system. Specifically: The data acquisition module is used to acquire and format static structural data and dynamic operational data of the reservoir; the graph construction module is used to generate a set of triples of <entity, relation, entity>; the graph embedding and modeling module is used to vectorize the triple set to form a structured semantic graph; the causal inference module is used to model the average treatment effect ATE, the individual treatment effect ITE, and the causal strength; the risk scoring and output module is used to output key hazard factors and quantitative scoring results; and the dynamic update module supports the dynamic supplementation of new data and the updating of inference results over time.

[0028] By employing the above technical solution, the present invention provides a method for quantifying reservoir safety risk factors based on causal knowledge graphs, which has at least the following beneficial effects: 1. The method proposed in this invention has significant advantages such as causal-driven modeling, structured semantic expression, interpretable and rankable risk scoring, and support for dynamic updates and online reasoning. It can realize the automatic identification, causal analysis, scoring, ranking and dynamic updates of reservoir safety risk factors.

[0029] 2. This invention reconstructs the causal chain of "disaster-causing factor - disaster-causing environment - disaster consequences" by integrating the graph structure and causal inference mechanism. It can accurately identify risk factors and their action paths, and quantify the causal effects and structural weights of each factor, so as to achieve the scientific classification of risk levels. It has the advantages of clear structural expression, clear causal mechanism, strong interpretability of results and good dynamic adaptability, and is applicable to risk identification and early warning management of various types of reservoirs.

[0030] 3. The method proposed in this invention has good dynamic response capability and scalability. It can automatically update the map structure and inference results with real-time data. It is applicable to different types of reservoirs (such as small and medium-sized reservoirs, hilly areas, and river reservoirs) and various operating environments. It can also be efficiently integrated with existing safety monitoring systems, early warning platforms and digital twin systems. It has broad application value and promotion prospects in reservoir engineering safety evaluation, operation management and disaster early warning. Attached Figure Description

[0031] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the method for quantifying reservoir safety risk factors in this invention. Detailed Implementation

[0032] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0033] Example 1: Construction of causal graph and risk scoring identification based on medium-sized hilly reservoir.

[0034] This embodiment uses a typical medium-sized hilly reservoir (hereinafter referred to as Reservoir A) as an example to verify the feasibility and effectiveness of the reservoir safety hazard identification and risk factor quantification method based on causal knowledge graph proposed in this invention. The reservoir was built in the 1990s, with a homogeneous earth dam structure, a maximum dam height of 32.5m, and a total reservoir capacity of 14.2 million m³. 3 It is a Class III reservoir, serving a population of approximately 30,000.

[0035] Reservoir A has relatively complete operational data in recent years. Its dynamic monitoring data and structural data from June 2020 to June 2023 were selected as the basic data source, as shown in Table 1.

[0036] Table 1 Static Structure Data and Dynamic Operation Data Based on the above data, and combining knowledge of reservoir safety with statistical mining techniques, a set of triples containing 42 types of entities and 71 types of relations was constructed, generating a total of 812 valid triples. Some typical triples are shown in Table 2.

[0037] Table 2. Set of triples containing entities and relations The set of triples was input into the TransE embedding model for training. The vector dimension was set to 128, the learning rate to 0.01, and the boundary margin γ to 1.0. The Adam optimizer was used for 100 iterations. The convergence of the embedding distance difference between positive and negative samples was monitored during training. The embedding results show that the triples with the relationships "inducing," "exacerbating," and "causing" have obvious directionality and semantic aggregation.

[0038] Taking <Rainstorm intensity, triggering factors, and sudden rise in water level> as an example, its embedding distance is: The distance between random negative samples (such as <rainfall intensity, causing, gate failure>) is 0.894, indicating that the structure training is effective.

[0039] Using the rainfall event in July 2021 as the primary analysis window, the following definition is provided: Treatment variable T: Whether "daily rainfall intensity > 80mm" occurs; Response variable Y: Estimated economic loss, unit: 10,000 yuan (weighted based on the distribution of inundated area × exposed population × GDP). Covariate X: Structural features output by the embedded model (including 128-dimensional vectors such as dam type, slope, permeability coefficient, and land use).

[0040] The CausalForest model was used for training, with 500 trees and a minimum number of leaf node samples of 30. Cross-validation was used to select the sample split depth. The calculation results are shown in Table 3.

[0041] Table 3. Calculation results of ATE and confidence interval for each risk factor Each risk factor in the computational graph PageRank values ​​(normalized and used as weights) ), and combine with ATE to calculate the risk scoring function. Some of the results are shown in Table 4. That is: Table 4. Risk score calculation results for each risk factor Based on the scoring results, risk factors are divided into four levels (high / relatively high / medium / low) and causal paths are output in the form of a graph structure: for example, the path "rainstorm intensity → sudden rise in water level → flooded area → economic loss" has the highest weight in the scoring and is identified as the priority intervention path.

[0042] In this embodiment, the model's predicted economic losses from the 2021 rainstorm event differed from the government's published estimates by less than 10%, demonstrating the model's good accuracy in causal reasoning. The model not only identified the explicit disaster-causing factor of "rainstorm intensity" but also quantitatively revealed the secondary impacts of structural conditions such as "dam permeability" and "topographic slope" on the risk outcome. Through interpretable scoring outputs, the system can provide local water resources departments with governance suggestions, such as strengthening dam seepage prevention projects and optimizing the control of high-risk slope development, demonstrating strong engineering practicality.

[0043] Example 2: Instance of map update and risk score adjustment driven by dynamic data of rainstorm process.

[0044] This embodiment uses a five-day continuous rainstorm event at Reservoir A in July 2022 as the application background to verify the ability of the system of the present invention to automatically update the map, reconstruct the causal model, and adjust the scoring results in real time under dynamic data-driven conditions. Through comparison with historical events and error evaluation indicators, the system verifies the accuracy and stability of the inference results, fully demonstrating the engineering adaptability and causal inference enhancement effect of the present invention in dynamic risk response scenarios.

[0045] From July 18 to 22, 2022, the A reservoir basin experienced continuous heavy rainfall, with a cumulative rainfall of 270 mm. During this period, the system received the following dynamic data every 15 minutes: Real-time water level and reservoir capacity data (automatic water level gauge); Rainfall intensity and distribution (radar station + rain gauge); Data from dam seepage sensors; Simulated images of river network inflow and reservoir inundation.

[0046] When the system detects the following abnormal event triggering conditions, it will automatically perform incremental updates to the graph structure: Daily rainfall intensity ≥90mm; Water level at any cross-section is greater than or equal to the warning level; The seepage rate at monitoring point No. 2 of the dam body exceeded the benchmark value for five consecutive times. Newly covered village roads and farmland areas in the flooded area (extracted and identified from remote sensing images).

[0047] Based on the above real-time triggering mechanism, the system automatically generates and inputs the following newly added triples, as shown in Table 5.

[0048] Table 5. Newly added triples containing entities and relations A total of 36 new triples were added. The system performs incremental embedding learning on the basis of the original graph structure, only updating the local vector representation associated with the new nodes / edges, reducing the computational burden of the system and avoiding overall reconstruction.

[0049] The updated map embedding results are automatically fed into the CausalForest model for local retraining. The processing variable T is set as "whether a red rainstorm warning is triggered", the response variable Y is set as "the percentage increase in flooded area", and the covariate X comes from the structural features after map embedding.

[0050] A forest was constructed using 500 regression trees. All updates only involved local splitting and relearning of subtrees where the factor influence paths changed, while preserving stable paths and weight configurations from the original model. The system recorded changes in ATE estimates during the rainstorm process, as shown in Table 6.

[0051] Table 6. Changes in ATE estimates during the rainstorm process. It can be seen that after the introduction of new flooding and seepage data, the causal strength of "rainstorm intensity" on flooding consequences is significantly amplified, and the model's response capability to extreme processes is enhanced. After each update, the system outputs the ATE confidence interval in real time, allowing engineers to understand the range of uncertainty.

[0052] To further identify potentially underestimated structural disaster-causing factors during rainstorms, the system introduces a conditional mutual information index to calculate causal strength, namely: Before the update, the conditional mutual information of the "dam permeability" variable was 0.126. After the update, it increased to 0.291 under the activation of the new data, indicating that it acts as a "hidden amplifier" and has a significant impact on risk outcomes during extreme rainfall. The system automatically added a new causal path to the spectral path: "Rainstorm intensity → Dam permeability → Abnormal leakage → Risk of local collapse → Economic loss".

[0053] This path is labeled with a high causal weight for subsequent scoring weight adjustments.

[0054] The system updates the ATE value and the structural centrality weight in each round. Calculate risk score The risk factor classification is dynamically adjusted based on the value, as shown in Table 7.

[0055] Table 7. Ranking Results of Each Risk Factor The "dam permeability" factor, which was originally at the medium level, jumped to a risk score of 15.91 due to the updated causal strength. The level was automatically adjusted to "higher" by the system and pushed to the governance priority list.

[0056] The system compares the predicted inundation area (map + inference output) with the measured inundation area extracted from UAV remote sensing images, and the evaluation is as follows: F1 score: 0.872; Predicted area deviation rate: 7.4%; The model score's maximum ranking change across 5 rounds of updates is ≤2 places; The consistency rate of high-risk factor identification (compared with historical events) was 92%.

[0057] Furthermore, the updated ATE valuation has gradually stabilized over the last three rounds of changes, indicating that the system has good dynamic stability and robustness, avoiding problems such as "overresponse" or drastic score fluctuations.

[0058] Example 3: Causal graph modeling and risk factor comparative analysis of typical reservoirs across regions.

[0059] This embodiment selects three typical reservoirs with significant differences in geographical location, structural characteristics, and operating environment. Based on the method of this invention, causal knowledge graphs and causal inference models are constructed for each reservoir. Risk factor comparison and hierarchical analysis are then conducted under a unified scoring mechanism to verify the universality, transferability, and consistency of the output results of the system. The three reservoirs are located in the mountainous areas of East China, the plains of Central China, and the river valleys of Southwest China, respectively, and are referred to as Reservoirs A, B, and C. The collected multi-source data are shown in Table 8.

[0060] Table 8. Results of multi-source data collected from Reservoirs A, B, and C Static structural data (dam type, permeability coefficient, DEM, land use, population density, regional GDP, etc.) and dynamic operational data (water level, reservoir capacity, rainfall, flow rate, leakage, inundation scenarios, etc.) were collected for all three reservoirs. The resulting ternary scales were formed during the structural modeling process, as shown in Table 9.

[0061] Table 9. Triple Scale of Each Reservoir Including Entities and Relationships Due to differences in reservoir functions, Reservoir C has more related entities such as transportation and water ecology, while Reservoir A highlights mountain risk relationships such as "slope, landslide, and seepage", and Reservoir B focuses on the relationship structure of "population exposure-asset density".

[0062] All three database maps were embedded using TransE embedding, with a uniform dimension (128) and boundary margin γ=1.0. The number of training epochs was determined based on the convergence of training error (approximately 100 epochs). After map embedding, a causal forest model was constructed for each reservoir, with "heavy rainfall" as the treatment variable and "economic loss estimate" as the response variable. The covariate was the map embedding structure feature vector. The causal forest parameters were kept consistent (500 trees, 30 leaf nodes). The risk scoring results in the inference results of the three databases are shown in Table 10.

[0063] Table 10 Risk score calculation results for each risk factor visible: Due to its steep terrain, Library A has a higher structural risk and a higher structural factor score. Due to its densely populated residential areas and high asset density, B-type risk factors score highly. Library C primarily focuses on operational facility risks, with an overall low factor score, indicating that its disaster paths are relatively short.

[0064] After scoring, the system automatically classifies each database into risk factor levels (high, relatively high, medium, and low) and generates a priority list of governance recommendations. This consistent scoring strategy across regions demonstrates the system's ability for comparative analysis and generalized deployment.

[0065] By linking structural graphs with model inference, the system automatically identifies the most representative risk paths in each database, such as: Reservoir A: "Rainstorm intensity → Topographic slope → Leakage risk → Flooded area → Economic loss"; B database: "Rainstorm intensity → Flooded area → Population density → Direct economic loss"; C Reservoir: "Rainstorm intensity → rubber dam elasticity attenuation → gate discharge efficiency reduction → water level accumulation"; Meanwhile, the system calculates the conditional mutual information CS(A→Y) between key hidden danger factors and response variables, and finds that the mutual information of the "topography slope" factor in database A is the highest, while that in database B is "population density" and that in database C is "gate aging level", further confirming the differences in the main controlling factors of causal mechanisms in different regions.

[0066] The core of this invention lies in transforming multi-source reservoir operation data into a structured semantic graph and integrating graph embedding modeling with causal inference algorithms (such as causal forests) to achieve causal path modeling, risk scoring, and dynamic updating between disaster-causing factors, disaster-inducing environments, and disaster consequences, thereby improving the interpretability, accuracy, and responsiveness of reservoir risk assessment.

[0067] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0068] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Since the above embodiments are substantially similar to the method embodiments, their descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0069] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for quantifying reservoir safety risk factors based on causal knowledge graphs, characterized in that, The method includes the following steps: S1. Collect multi-source data on the reservoir and its surrounding area, including static structural data and dynamic operational data. Based on the logical and causal relationships between key entities within the reservoir system, construct a set of triples including entities, relationships, and attributes. The set of triples Includes at least the following types, namely: <Rainfall, triggering, floods>; <Dam permeability, reinforcement, leakage risk>; <Population density, exposure to, flooded areas>; <Topography, exacerbation, flood propagation>; <Gate malfunction, impact on flood discharge capacity>; in, For the head entity, For tail entities, Semantic tags used to represent the causal or dependency relationship between the two; S2, Set the triplet set The input graph embedding model uses the TransE algorithm or graph convolutional networks to map entities and relations into graph embedding representations in vector space, forming a representation composed of several risk factors. A structured semantic graph as graph nodes; S3. The graph embedding is used as a covariate X, and treatment variables T and response variables Y are introduced. Based on a causal inference model, the causal effect between treatment variables T and response variables Y is analyzed. Quantitative inference is performed to estimate the average treatment effect ATE and the individual treatment effect ITE for each variable; the treatment variable T is whether a specific intensity of rainstorm, water level exceeding the limit, or structural anomaly occurs; the causal inference model adopts the causal forest algorithm. The response variable Y is quantified by the following disaster consequence function, namely: Y= f (Flooding depth, population density, exposure duration, regional GDP); in, f This is a weighted combination function used to map physical and social exposure factors to disaster consequences; S4. Calculate each risk factor in the structured semantic graph. The centrality index is used to obtain its structural weight. and its causal effects Together, they form a risk scoring function for the quantitative identification and ranking of key hazard factors. The centrality metric is expressed as PageRank or DegreeCentrality. The risk scoring function The calculation formula is: ； in, Representing risk factors in structured semantic graphs The centrality score; S5. Based on risk scoring function Output a list of key hidden danger factors and quantitative indicators for identification and management recommendations in reservoir safety early warning systems.

2. The method for quantifying reservoir safety risk factors according to claim 1, characterized in that, The multi-source data includes dam structure parameters, geographical location, geological conditions, reservoir area DEM, land use type, population density, regional GDP, water level changes, reservoir capacity changes, rainfall intensity, flood process, and reservoir inundation scenarios. The sources of the multi-source data include remote sensing monitoring, sensor data collection, basic geographic databases, and manual survey data.

3. The method for quantifying reservoir safety risk factors according to claim 1, characterized in that, The graph embedding model employs the TransE algorithm, whose optimization objective is to minimize the entity and relation vectors, satisfying: ； Used to maintain the geometric consistency of causal chains in knowledge graphs; Optimize objective function for: ； in, It is a positive triplet; For negative triples generated by perturbation; This is the boundary margin hyperparameter, used to enhance the distinction between positive and negative samples; It represents the part that is greater than zero.

4. The method for quantifying reservoir safety risk factors according to claim 1, characterized in that, The mathematical expression for the causal inference model is: ； The causal inference model constructs multiple heterogeneous regression trees. We performed split modeling on the covariate X and used the Bootstrap method to split each risk factor. Interval estimation was performed on the average treatment effect ATE estimate to obtain the risk factor. confidence interval This is used to enhance the credibility and statistical robustness of the inference results.

5. The method for quantifying reservoir safety risk factors according to claim 1, characterized in that, It also includes estimating the causal strength score of a variable A on the response variable Y using conditional mutual information methods. , expressed as: ； Used to assess the strength of the causal influence of variable A on response variable Y given covariate X; The causal strength score It can complement the average treatment effect ATE and be used to identify the higher-order coupling strength of key factors in nonlinear influence paths.

6. The method for quantifying reservoir safety risk factors according to claim 1, characterized in that, The quantitative indicators include each risk factor. The average treatment effect (ATE), confidence interval, contribution ranking, and risk level classification.

Images