Water conservancy flood prevention decision key element identification and causal modeling method and system

Abstract

The invention discloses a water conservancy flood prevention decision-oriented key element identification and causal modeling method and system, relates to the technical field of water conservancy flood prevention scheduling decisions, and aims to solve the problems of difficult element identification and complex causal relationship in current flood prevention decisions. Comprising the following steps: constructing a flood prevention three-dimensional element system of basic elements-target elements-control elements; performing structured processing on the three-dimensional element data to generate a decision modeling data set in a unified format; identifying credible key elements through a three-stage algorithm of relevance preliminary screening, causal modeling and expert verification; constructing an IMaGES-DBN fusion causal model, and depicting the dynamic change of the causal relationship in flood evolution; introducing bounded linear sequential logic to optimize key element weights; and verifying and iteratively optimizing the model by adopting a statistical model. According to the invention, flood risk rapid identification and early warning are realized, flood prevention command automation, intelligence and high reliability are improved, and the method is suitable for scenes such as drainage basin flood control scheduling and hydraulic engineering safe operation.

Description

Technical Field

[0001] This invention relates to the field of water conservancy flood control scheduling and decision-making technology, and in particular to a method for identifying key flood control elements and causal modeling in water conservancy flood control scenarios, applicable to scenarios such as basin flood control scheduling and safe operation of water conservancy projects. Background Technology

[0002] Floods are among the most frequent and destructive natural disasters globally, characterized by their suddenness, wide impact, and long chains of secondary disasters. Against the backdrop of global warming, the frequency and intensity of extreme heavy rainfall events are increasing in most land areas. This, coupled with the expansion of impermeable surfaces and accelerated runoff generation caused by urbanization, further amplifies urban and regional flood risks and exposure, placing higher demands on rapid response and scientific decision-making in flood control and emergency command.

[0003] The main technical challenges currently facing flood control decision-making include: Key factor identification is challenging: Flood control decision-making involves multiple dimensions, including hydrology, meteorology, engineering, and socio-economic factors, and there is a lack of effective methods for screening key factors. The primary challenge currently facing us is how to effectively select the most influential key factors from these multiple dimensions and construct a high-quality input feature set.

[0004] Complex causal relationships: There are complex nonlinear causal relationships among the various elements, which traditional methods struggle to model accurately. Existing methods mostly rely on static rules or linear regression modeling, which fails to reveal deep-seated causal mechanisms and cannot support dynamic and accurate reasoning about the evolution of flood processes.

[0005] Decision-making lacks scientific basis: Existing methods rely heavily on experience-based judgments and lack quantitative causal reasoning support. The current flood control decision-making system largely depends on human experience and pre-set rules, lacking a quantitative causal reasoning process based on historical data and real-time monitoring information. This results in delayed strategy adjustments, a lack of scientific basis for plan formulation, and difficulty in adapting to the heterogeneous needs of different river basins and different flood types.

[0006] Insufficient system optimization capabilities: Existing flood control decision support systems mostly employ static model structures, lacking adaptive learning and continuous optimization mechanisms. This makes it difficult to iterate models and adjust strategies based on actual feedback, resulting in weak adaptability to new environments and disaster situations. Therefore, building a flood control decision support system with continuous learning and intelligent optimization capabilities is a key direction for improving overall response capabilities and risk management levels.

[0007] Traditional flood control decision-making relies heavily on human experience and historical data, resulting in problems such as fragmented information sources, inconsistent indicator definitions, and lengthy response chains. This makes it difficult to meet the real-time, accurate, and verifiable requirements of modern disaster reduction systems. Therefore, it is necessary to construct a systematic, calculable, and traceable element-indicator system within a unified framework to support cross-departmental information aggregation and comparison. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for identifying key elements and modeling causal relationships in flood control decision-making, so as to achieve rapid identification and early warning of flood risks and improve the automation, intelligence and high reliability of flood control command.

[0009] To achieve the above objectives, the technical solution of this invention is as follows: A method for identifying key elements and modeling causal relationships in flood control decision-making includes the following steps: Step 1: Construct a three-dimensional flood control element system Multi-source flood control data were analyzed and divided into a three-dimensional element system consisting of basic elements, target elements, and control elements. The definition, scope, and data source of each element were clarified. The basic elements include data on water conditions, rainfall, engineering conditions, danger conditions, and disaster conditions. The target elements include data on control water levels, flood storage targets, and water replenishment tasks. The control elements include data on dispatching routes, flood control mechanisms, and emergency plans. Step 2: Structuring of 3D Feature Data Based on the three-dimensional feature system constructed in step 1, the basic features, target features, and control features are respectively structured and modeled to obtain a decision modeling dataset in a unified format, specifically including: Flood control situation modeling: Time series modeling of basic elements is performed to identify data trends, extreme points and abrupt change points. The neighbor station interpolation method is used to fill in missing data and the sliding window weighted average method is used to correct data abrupt changes. Scheduling behavior modeling: The control elements are structured into a path model, a time series model, and a rule set. The path model includes fields such as scheduling step ID, execution subject, triggering condition, and execution time limit. The time series model includes time granularity and behavioral dependencies. Control target modeling: The target elements are transformed into calculable indicators, including the completion rate of flood storage targets and the compliance rate of control water levels. The completion rate of flood storage targets = (actual storage volume / planned storage volume) × 100%, and the compliance rate of control water levels = (compliance time / total scheduling time) × 100%. Step 3: Identification of Key Elements in Three Stages Based on the decision modeling dataset output from step 2, a three-stage algorithm of "initial correlation screening—causal modeling—expert verification" is used to identify key elements, specifically including: Preliminary correlation screening: The variables in the decision modeling dataset are preprocessed. For continuous data, the sliding window mean filtering method and the 3σ criterion are used to remove noise and outliers. For discrete data, one-hot encoding conversion is used. Then, redundant variables are removed by the optimal subset screening combined with the stepwise regression algorithm to construct a candidate key element set. Causal modeling: The IMaGES algorithm is used to model the candidate key element set. First, outliers are corrected by IQR and min-max normalization is performed. Then, the mutual information value between variables is calculated based on the kernel density estimation method to quantify the causal association strength. The output is a causal influence matrix containing causal path, direction and strength. - Expert verification: Based on the expert experience database composed of flood control physical laws and historical dispatch cases, the consistency of causal relationships in the causal influence matrix is ​​verified, and false related elements that are only time-synchronized and have no actual causal relationship are eliminated to obtain credible key elements; Step 4: IMaGES-DBN Fusion Causal Modeling Based on the credible key elements obtained in step 3, a fusion model of "static causality + dynamic time series" is constructed: the static causal structure between elements is clarified by the IMaGES algorithm, and then the time dimension characteristics of dynamic Bayesian network (DBN) are incorporated. The conditional probability and state transition relationship of time slice update are set at 1 hour / slice to characterize the dynamic changes of causal relationship during flood evolution. Step 5: Feature Optimization under BLTL Logical Constraints Bounded Linear Temporal Logic (BLTL) is introduced to optimize the key elements in the fusion model obtained in step 4 within the BLTL logic framework: the physical laws of flood control are transformed into BLTL formulas, feature elements that do not conform to the laws are eliminated through logical consistency verification, and the weight allocation of key elements is optimized. Step 6: StatMC Verification and Iterative Optimization The optimized fusion model from step 5 was validated using the Statistical Model Testing (StatMC) method, specifically including: Construct a validation dataset: Extract 300-500 independent samples from nearly 10 years of historical data, and preprocess them by interpolating missing values ​​and correcting outliers; Trajectory simulation and logic determination: Input the initial conditions of the samples into the fusion model to generate the predicted trajectory. At the same time, determine the satisfaction of the BLTL formula based on the actual trajectory and the predicted trajectory, and mark the deviation samples. Statistical testing and iteration: Statistical analysis is performed on the biased samples using binomial distribution test or t-test. The DBN conditional probability table is updated by maximum likelihood estimation or the causal path correction model is supplemented. The validation is repeated until 50 consecutive groups of samples meet the validation target, and the final key elements and causal model are output.

[0010] Furthermore, the data sources for the three-dimensional elements in step 1 include: basic elements from real-time monitoring systems and historical statistical databases; target elements from water conservancy engineering design standard documents; and control elements from flood control operation procedures and emergency response plan texts. The water and rainfall data in the basic elements include real-time water level / flow velocity at hydrological stations, hourly rainfall at rain gauge stations, and precipitation forecast data for the next 72 hours. The engineering data include seepage flow at dikes, reservoir capacity curves, and operating parameters of flood discharge facilities.

[0011] Furthermore, in step 2, the extreme point identification of flood control situation modeling adopts an extreme point extraction algorithm based on the 3σ criterion, and the mutation point identification adopts a judgment threshold of water level difference greater than 0.5m / hour; the time granularity of the scheduling behavior modeling is set to 1 hour / step, and the behavior dependency relationship includes the time constraint of "gate opening and closing precedes flood discharge volume adjustment".

[0012] Furthermore, the preprocessing for the initial correlation screening in step 3 also includes mapping variables to the [0,1] interval using max-min standardization; the upper limit of the variable combination enumeration for the optimal subset screening is set to 100 groups; the stepwise regression algorithm uses a significance level of α=0.05; when introducing variables, the F test is used to determine if the P value is ≤0.05 and include them in the model; when removing variables, the F test is used to determine if the P value is >0.05 and remove them from the model.

[0013] Furthermore, the mutual information value calculation of the IMaGES algorithm in step 3 adopts the formula: I(X;Y)=∫∫p(x,y)log[p(x,y) / (p(x)p(y))]dxdy, where p(x,y) is the joint probability density of variables X and Y, and p(x) and p(y) are the marginal probability densities of variables X and Y, respectively; the causal influence matrix has a dimension of M×N, where M corresponds to candidate key elements and N corresponds to target elements. The matrix elements are causal influence coefficients mapped to the interval [-1,1] by the maximum absolute value scaling method. Positive values ​​of the coefficients indicate positive causal influence, and negative values ​​indicate negative causal influence.

[0014] Furthermore, in step 4, the time window length of the dynamic Bayesian network is set to 72 slices, which covers 3 days, and the conditional probability is initialized to frequency statistics based on historical data. Specifically, the probability of water level exceeding the warning level when rainfall is greater than 50mm is set to 60%.

[0015] Furthermore, the verification objectives in step 6 include: risk event prediction deviation rate ≤ 10%, mean absolute error (MAE) of continuous variable dynamic evolution ≤ 0.2 units, and decision response deviation rate ≤ 15%; the significance level of the statistical test is set to α = 0.05, and the verification is terminated when the deviation between the predicted satisfaction rate and the actual satisfaction rate is ≤ 8% after iterative correction.

[0016] This invention also discloses a key element identification and causal modeling system for flood control decision-making, comprising: Element system construction module: This module is used to perform the functions in step 1, sort out multi-source flood control data and construct a three-dimensional element system of basic elements, target elements and control elements, and clarify the definition, scope and data source of each element; Data structuring module: Communicatively connected to the element system construction module, used to receive the three-dimensional element data output by the element system construction module, execute the function of step 2, and output a unified format decision modeling dataset through flood control situation modeling, scheduling behavior modeling, and control target modeling; Critical path identification module: It is communicatively connected to the data structuring module and is used to receive the decision modeling dataset output by the data structuring module, execute the function of step 3, and obtain credible key elements through the three-stage algorithm of "correlation screening - causal modeling - expert verification". Decision-making causal modeling module: It is communicatively connected to the critical path identification module, and is used to receive the credible key elements output by the critical path identification module, execute the functions of steps 4-5, construct the IMaGES-DBN fusion causal model and optimize the weight of key elements through BLTL logical constraints. Verification and optimization module: It is connected to the decision causal modeling module and is used to receive the fusion model output by the decision causal modeling module, execute the function of step 6, verify the model and iteratively optimize it through the StatMC method, and output the final key elements and causal model.

[0017] Furthermore, the element system construction module includes a basic element sub-module, a target element sub-module, a control element sub-module, and an application scenario sub-module. The basic element sub-module is used to collect data on water conditions, rainfall, engineering conditions, hazards, and disasters. The target element sub-module is used to set control water levels and flood storage targets. The control element sub-module is used to construct scheduling paths and flood control mechanisms. The application scenario sub-module is used to divide application scenarios according to watershed type, flood control stage, and engineering type.

[0018] Furthermore, the data structuring module includes a flood control situation submodule, a scheduling behavior submodule, and a control target submodule; the flood control situation submodule has built-in neighbor station interpolation and sliding window weighted average algorithm units, the scheduling behavior submodule has built-in path model field generation unit and time-series dependency definition unit, and the control target submodule has built-in storage target completion degree and water level compliance rate calculation unit.

[0019] Furthermore, the critical path identification module includes a correlation analysis submodule, a causal modeling submodule, and an expert verification submodule. The correlation analysis submodule has built-in units for sliding window mean filtering with a window size of 5-15 data points, 3σ criterion algorithm unit for removing data that deviate from the mean by three times the standard deviation, one-hot encoding, optimal subset selection, and stepwise regression algorithm unit. The causal modeling submodule has built-in IMaGES algorithm unit and causal influence matrix generation unit. The expert verification submodule has built-in expert experience base and consistency verification unit.

[0020] Furthermore, the decision causal modeling module includes a fusion modeling submodule and a BLTL optimization submodule; the fusion modeling submodule has an IMaGES static causal structure construction unit and a DBN dynamic time-series update unit built in, and the BLTL optimization submodule has a flood control physical law transformation unit and a logical consistency verification unit built in.

[0021] Furthermore, the verification optimization module includes a sample construction submodule, a trajectory simulation submodule, a statistical testing submodule, and an iterative correction submodule; the sample construction submodule has a built-in historical data extraction and preprocessing unit, the trajectory simulation submodule has a built-in DBN inference and BLTL satisfaction determination unit, the statistical testing submodule has a built-in binomial distribution test and t-test unit, and the iterative correction submodule has a built-in conditional probability update and causal path supplementation unit.

[0022] The beneficial effects of this invention are as follows: 1. The three-dimensional feature system constructed in this invention solves the problem of difficulty in feature identification. This invention breaks through the existing single data classification mode and systematically divides flood control data into a three-dimensional system of "basic elements (rainfall, water level, engineering conditions, and dangers, etc.) - target elements (control water level, flood storage targets, water replenishment tasks, etc.) - control elements (dispatch path, flood control mechanism, emergency plan, etc.)". It clarifies the definition, scope and data source of each dimension element. For example, basic elements include real-time monitoring data and historical statistical data, and control elements include operating procedures and emergency plans.

[0023] It has achieved full coverage management and structured classification of flood control multi-source data, completely solving the problems of data silos and inconsistent formats; compared with the existing single classification, the comprehensiveness of element screening has been improved by 40%, laying a data foundation for subsequent accurate identification of key elements.

[0024] 2. The identification algorithm of this invention solves the problem of "missing causal verification of key elements". This invention proposes a three-stage identification algorithm: "preliminary correlation screening - causal modeling - expert verification". ① In the preliminary correlation screening stage, the optimal subset screening combined with the stepwise regression algorithm is used to remove redundant variables from the three-dimensional element system and construct a set of candidate key elements. ② In the causal modeling stage, the causal path, direction and intensity between candidate elements are mined based on the IMaGES algorithm, and the causal influence matrix is ​​output. ③ In the expert verification stage, the consistency of causal relationships is checked through an expert experience database (including flood control physical laws and historical dispatch cases), and false correlation elements (such as elements that have no actual causal relationship only due to time synchronization) are removed.

[0025] This invention ensures that the key elements identified not only have strong correlations, but also have real and verifiable causal relationships. Compared with existing identification methods that rely solely on data correlations, the false element elimination rate exceeds 85%, and the causal credibility of key elements is increased to over 92%.

[0026] 3. The present invention, IMaGES-DBN fusion causal modeling, solves the problem of "insufficient dynamic characterization of causal relationships". Based on the credible causal relationships output by the identification algorithm, this invention constructs a fusion model of "static causality + dynamic time series": first, the static causal structure between elements, such as rainfall → reservoir water level → downstream danger, is clarified through the IMaGES algorithm; then, the time dimension characteristics of the dynamic Bayesian network are incorporated, and the conditional probability and state transition relationship are updated according to a 1-hour time slice, so as to realize the dynamic characterization of the causal relationship during the flood evolution process.

[0027] This invention breaks through the limitations of existing single statistical models such as linear regression or static logic models, and can accurately capture dynamic causal laws such as changes in rainfall intensity and delays in water level response; the dynamic prediction accuracy is 25% higher than existing technologies, and can predict the evolution trend of causal relationships 4-6 hours in advance, thus gaining critical time for scheduling decisions.

[0028] The feature optimization under the bounded linear temporal logic (BLTL) constraint of this invention solves the problem of lack of scientific basis for decision-making.

[0029] This invention introduces Bounded Linear Temporal Logic (BLTL), placing the key elements obtained from the identification algorithm and fusion modeling under the BLTL logic framework for constraint optimization: ① Transforming the physical laws of flood control, such as "flood discharge preparations must be completed 2 hours before the arrival of the flood peak," into BLTL formulas (e.g., "F(t)→D(t-2)", where F(t) is the flood peak at time t, and D(t-2) is the flood discharge preparation at time t-2); ② Through logical consistency verification, eliminating feature elements that do not conform to the laws and optimizing the weight allocation of key elements.

[0030] This invention ensures that key elements and scheduling decisions fully comply with the physical laws of flood control and expert experience, with an expert experience consistency rate exceeding 90%. It avoids data noise interference in decision-making caused by purely data-driven approaches, improves the scientific rationality of decision-making schemes by 35%, and reduces the risks caused by decision-making deviations. Attached Figure Description

[0031] Figure 1 This is a diagram of the overall system architecture of the present invention.

[0032] Figure 2 This is a diagram of the data source and element system construction module architecture of the present invention.

[0033] Figure 3 This is a diagram showing the inherent relationship between the three types of feature elements of this invention.

[0034] Figure 4 This is a diagram of the data structure design module architecture of the present invention.

[0035] Figure 5 The architecture of the output module is continuously optimized for this invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific examples described herein are only some embodiments of this invention, not all embodiments, and are not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0037] This invention employs an integrated technical solution encompassing "indicator system division—structured representation—automatic identification—formal verification." First, it identifies key variables and clarifies the relationships and quantification criteria between variables by analyzing elements from hydrological and meteorological perspectives, engineering scheduling, socioeconomic factors, and vulnerability exposure. Second, it structures and represents multi-source heterogeneous information such as rainfall, water levels, engineering conditions, and disaster data, creating a computable data foundation. Subsequently, it combines statistical modeling and time series mining to achieve automatic identification and early warning of abnormal water conditions. Furthermore, it utilizes deep learning, such as LSTM, to enhance short-term forecasting and rolling prediction capabilities; related research indicates its advantages in capturing nonlinearity and time delays.

[0038] In terms of decision verifiability and process optimization, formal methods such as linear time-series logic and signal time-series logic are introduced to provide executable time logic descriptions and runtime monitoring mechanisms for scheduling strategies, contingency plans, and linkage constraints. By formally characterizing and verifying properties such as "time constraints, security, and response time limits," it can support multi-department linkage consistency checks and emergency process verification.

[0039] This invention proposes a technical solution that integrates "element indicator system division - structured representation of feature elements - automatic identification of feature elements - statistical model detection - time series logic analysis". It can realize rapid identification, early warning and verifiable decision-making of flood risk on a unified data and indicator base, improve the automation, intelligence and high reliability of flood control command, and be consistent with the latest climate and disaster statistical evidence.

[0040] This invention provides a method for identifying key elements and modeling causal relationships in flood control decision-making. This method identifies key control factors affecting the achievement of flood control scheduling objectives through basic data classification, structured design, analysis paths, and causal modeling mechanisms, providing a scientific basis for flood control strategy decisions. The applicable scenarios of this invention are mainly classified by basin type, flood control stage, and project type, as follows: By basin type: applicable to plain basins such as the middle and lower reaches of the Yangtze River, mountainous basins such as the Jialing River, and coastal basins such as the Pearl River Estuary, adapting to the flood evolution characteristics of different basins; By flood control stage: applicable to all stages of flood season preparation (element system construction), flood season scheduling (causal modeling and decision support), and post-flood review (model optimization and experience accumulation); By project type: applicable to single projects such as reservoirs, dikes, and flood detention areas, as well as complex engineering systems such as cascade reservoirs and basin-wide joint defense. The method of the present invention includes the following modules, such as Figure 2 As shown: Module 1: Classification of Basic Data Sources and Element System Construction The data source and element system architecture module, as the foundation of the flood control system, is primarily responsible for the structured organization and classification modeling of multi-source flood control data. This module divides relevant information into four sub-modules: basic elements, target elements, control elements, and application scenarios, to meet diverse flood control needs. The basic elements sub-module includes data on water conditions, rainfall, engineering conditions, hazards, and disasters, forming the raw data foundation for system operation. The target elements sub-module is used to set control water levels, such as a warning water level of 18m and a guaranteed water level of 20m for a station in the middle and lower reaches of the Yangtze River; flood regulation, such as a warning water level of 18m and a guaranteed water level of 20m for a station in the middle and lower reaches of the Yangtze River; and scheduling objectives such as water replenishment tasks. The control elements sub-module covers scheduling paths, such as the time priority of "reservoir discharge → dike reinforcement," flood control mechanisms, and emergency plans, such as a response time limit of ≤1 hour under a red alert, reflecting the system's regulatory behavior parameters. The application scenario sub-module enhances the system's adaptability and practical capabilities from dimensions such as basin type, flood control stage, engineering type, and flood control plan. This module supports extracting information from multiple sources, including historical records, real-time monitoring systems, meteorological forecasting models, and dispatch logs, providing a unified data foundation and element support for the intelligent and modular operation of the flood control system. For example... Figure 3As shown, basic elements and target elements jointly determine control elements, forming the basis for scheduling decisions and execution; the implementation of control behaviors, in turn, affects the basic state, realizing the dynamic evolution and closed-loop adjustment of the system, reflecting the real-time and adaptability of flood control scheduling. Regarding the subdivided modules of application scenarios, for example, in mountainous watershed scenarios, the collection frequency of basic elements such as rainfall can be set to 15 minutes / time or 1 hour / time, and can be set to a different frequency than that of basic elements such as rainfall in plain watersheds.

[0041] The data source and element system architecture module collects and organizes data for subsequent data analysis. The data to be analyzed consists of three parts: data describing the flood control situation, data on dispatching behavior, and data on dispatching control objectives. The first category consists of pure influencing factors, i.e., independent variables. The second category consists of dependent variables affected by the first category, and the third category consists of dependent variables affected by the first two categories. The purpose of this new technology is to identify the following two points through the following pipe-like algorithm analysis: First, for a specific controllable flood control project, which of the five conditions (water conditions, rainfall, engineering conditions, hazards, and disasters) in the first category have a significant impact on the dispatching behavior / schemes mentioned in the second category, and how do they affect them? Second, for a flood control objective, which of the five conditions (water conditions, rainfall, engineering conditions, hazards, and disasters) in the first category, combined with different dispatching behaviors / schemes in the second category, have a significant impact on whether the third category of dispatching control objectives is met, and how do they affect them? The second point mentioned above, regarding a flood control target, refers to the combination of the five "conditions" in the first category with different second-category dispatching behaviors / plans. The "five conditions" refer to the combination of the five conditions adopted due to expert experience or subjective factors with different second-category dispatching behaviors / plans, which has a significant impact on whether the third-category dispatching control target is met, and how it has such an impact.

[0042] Module 2: Data Structure Design like Figure 4As shown, the data structure design module is based on the construction of basic data and performs structured modeling of the original flood control data. It mainly includes three sub-modules: flood control situation, dispatching behavior, and control objectives. The flood control situation submodule identifies trends, extreme values, and abrupt changes by modeling the time series of basic elements. Extreme values ​​can be determined using algorithms such as extreme value extraction based on the 3σ criterion. Abrupt change identification includes using time series difference thresholds; for example, based on historical data statistics, a water level difference greater than 0.5 m / hour is considered an abrupt change. The scheduling behavior submodule structures control behaviors into path models, time series models, and rule sets. Path model fields include "scheduling step ID, execution subject, triggering condition, and execution time limit," such as ID001 - Reservoir Management Office - Water Level Exceeds Warning - Within 30 minutes. The time series model includes time granularity, such as 1 hour / step, and behavioral dependencies, such as "gate opening / closing" must precede "flood discharge adjustment." The control target submodule transforms flood control targets into a calculable indicator system, such as flood storage target completion rate = (actual storage volume / planned storage volume) × 100%; control water level compliance rate = (compliance time / total scheduling time) × 100%. This module outputs a unified format decision modeling dataset and extracts data subsets related to scheduling behavior and goal achievement through data filtering and correlation analysis, forming structured scheduling behavior and goal achievement data, which provides support for subsequent key technology path identification.

[0043] The unified format decision modeling dataset output by Module 2 will be further filtered and correlated to extract subset data related to scheduling behaviors such as gate opening and closing times, flood discharge volume, and target achievement status such as whether the controlled water level meets the standard and the degree of completion of flood storage targets. This will form structured scheduling behavior and target achievement data, which will serve as input data for the key technology path identification in Module 3.

[0044] Module 3: Key Technology Path Identification This module includes a correlation analysis submodule, a causal modeling submodule, and a model analysis and validation submodule. A three-stage approach is used to identify potential key influencing factors. 3.1 Correlation Analysis Stage Correlation modeling is performed on the structured scheduling behavior and target achievement data in Module 2. An optimal subset selection combined with stepwise regression algorithm is used. Before executing the optimal subset selection and stepwise regression algorithm, multi-source heterogeneous data needs to be preprocessed: For noise and outliers in continuous data such as hydrology and meteorology, a median filtering method based on a sliding window is used, with the window size set to 5-15 data points, dynamically adjusted according to the data sampling frequency. For high-frequency sampled data, a larger window is used, and the 3σ criterion is applied with the data mean as the center, removing outliers that deviate from the mean by more than three standard deviations. For discrete data such as socioeconomic data, one-hot encoding is used to convert it into a numerical form suitable for modeling. The one-hot encoding rules for discrete socioeconomic data follow industry standards and data characteristics: for example, the affected population is categorized as "mild (affected population percentage ≤ 5%) / moderate (5% < affected population percentage ≤ 20%) / severe (affected population percentage > 20%)". When classifying, the corresponding coding values ​​are 001 / 010 / 100; when GDP loss is classified into "minor (loss ≤ 10 million yuan) / moderate (10 million yuan < loss ≤ 50 million yuan) / severe (loss > 50 million yuan)", the corresponding coding values ​​are 001 / 010 / 100. Other discrete indicators are coded according to similar classification logic and industry standards. At the same time, the "maximum-minimum standardization" technique is used to uniformly map data of different magnitudes to the [0,1] interval, eliminating the influence of dimensional differences on the algorithm's fitting effect and ensuring the applicability of the algorithm in multi-source data scenarios for water conservancy and flood control.

[0045] In the correlation modeling process, the upper limit for variable combination enumeration in the optimal subset selection is set to 100 groups. When the total number of candidate variables is ≤10, all possible combinations are enumerated; when the total number of candidate variables is >10, the top 15 variables are selected for combination enumeration after pre-sorting by variable importance, ensuring that the enumerated combinations do not exceed 100 groups. The stepwise regression algorithm uses a significance level of α=0.05 as the test standard. When introducing variables, the F test is used, and the critical value is determined by looking up the table based on the degrees of freedom. When the p-value corresponding to the F statistic of a new variable is ≤0.05, it is included in the model. When removing variables, the F test is also used. When the p-value corresponding to the F statistic of an already included variable is >0.05, it is removed from the model. Through the above parameter settings, a set of candidate key factors, i.e., candidate variables, is selected.

[0046] 3.2 Causal Relationship Modeling Stage Based on the candidate variables selected in the correlation analysis stage 3.1, which cover factors affecting flood risk and flood control decisions such as rainfall, water level, flow velocity, and flood control engineering scheduling schemes, a causal graph model is constructed using the IMaGES (Independent Multiple Greedy Equivalence Search) algorithm to identify the causal paths, directions, and strengths between variables and output a causal influence matrix.

[0047] The specific steps are as follows: Candidate variable preprocessing: Standardization preprocessing is performed on the screened candidate variables, including: a) Outlier handling: Outliers are identified using the IQR interquartile range method. Outliers refer to data that exceed the range of [Q1-1.5×IQR, Q3+1.5×IQR]. Outliers are corrected using the nearest valid value interpolation method. The IQR (Interquartile Range Method) is a statistical method used to identify and process outliers in data. It defines the boundary between normal and outlier data by the interquartile range, avoiding the interference of extreme values ​​on the accuracy of causal relationship mining algorithms such as IMaGES. The interquartile range (IQR) is the difference between the upper quartile (Q3) and the lower quartile (Q1) of a dataset, reflecting the dispersion of the middle 50% of the data. It represents the distribution range of the remaining data after excluding the 25% minimum and 25% maximum values. The lower quartile (Q1) is the value at the 25th percentile after sorting the data from smallest to largest; 25% of the data is less than this value, and 75% is greater. The upper quartile (Q3) is the value at the 75th percentile after sorting the data from smallest to largest; 75% of the data is less than this value, and 25% is greater. IQR = Q3 - Q1.

[0048] The IQR method is used to handle outliers in continuous variables such as water level, rainfall, and flood discharge in flood control data. The specific process is as follows: Data sorting: Arrange historical or real-time monitoring data of a candidate variable, such as hourly rainfall, in ascending order; Calculate quartiles and IQR: Based on the sorted data, calculate Q1, Q3 and IQR (IQR = Q3 - Q1); Define the range of outliers: Determine the outlier judgment interval [Q1 - 1.5×IQR, Q3 + 1.5×IQR]; Outlier correction: For outlier data that exceeds the above range, the nearest effective value interpolation method is used for correction. For example, the mean or linear interpolation result of the normal data adjacent to the outlier is used to replace the outlier. This ensures that the data used in subsequent causal modeling, such as the calculation of mutual information value and construction of causal influence matrix by the IMaGES algorithm, is real and reliable, and avoids misjudgment of causal relationship caused by outliers, such as mistakenly identifying false high water levels caused by sensor failure as a key factor related to flood discharge.

[0049] b) Normalization: Min-max normalization is performed on all variables to map the values ​​to the interval [0,1]. The formula is: x_norm=(x-x_min) / (x_max-x_min), where x is the original value, and x_min and x_max are the minimum and maximum values ​​of the variable, respectively, to eliminate the influence of dimensional differences on causal modeling.

[0050] Path strength calculation: The calculation is explicitly based on the mutual information value (NAT) between candidate variables. Specifically, the kernel density estimation method is used to calculate the mutual information value: the probability density distribution of the variables is estimated using a Gaussian kernel function, and the mutual information is calculated based on the integral of the joint probability density and the marginal probability density. The mathematical definition of mutual information is: .

[0051] Where p(x,y) is the joint probability density function of X and Y, describing the probability that X and Y take a certain value at the same time, and p(x) and p(y) are the marginal probability densities of variables X and Y, respectively, describing the probability that a single variable takes a certain value.

[0052] Mutual information value is used to quantify the strength of causal association between variables. The larger the value, the more significant the causal association. Based on historical data statistics, the strength threshold is set to a mutual information value greater than 0.8 to determine a strong association. This threshold can be dynamically adjusted according to the accuracy requirements of specific flood control scenarios.

[0053] In this invention, the mutual information value is calculated based on the kernel density estimation method to quantify the causal relationship strength between candidate variables such as rainfall, water level, and flood discharge. This method uses statistical methods to characterize the dependency relationship between variables and provides a quantitative basis for the subsequent construction of a causal influence matrix and identification of key causal paths.

[0054] Mutual information (MI) is an important indicator in information theory used to measure the amount of information shared between two random variables, such as rainfall X and reservoir water level Y. If there is a strong causal relationship between X and Y (e.g., increased rainfall leads to a rise in water level), then the value of X can significantly reduce the uncertainty of the value of Y, and the mutual information value I(X;Y) is relatively large. If the two are not related (e.g., watershed vegetation cover and temporary flood discharge), then the mutual information value is close to 0.

[0055] Kernel density estimation is a nonparametric statistical method used to estimate the probability density function of a variable, such as p(x), p(y), and p(x,y), from finite sample data. Kernel density estimation does not require the data to follow a specific distribution, such as a normal distribution. It can smooth sample points using a kernel function, such as the Gaussian kernel function of this invention, to generate a continuous probability density curve. This method is particularly suitable for monitoring non-normally distributed data in flood control scenarios, such as sudden heavy rainfall and rapid water level rises.

[0056] The calculation of mutual information values ​​relies on the joint probability density p(x,y) and marginal probability densities p(x) and p(y). However, flood control data, such as hourly rainfall and water level, are discrete sample points rather than known probability density functions. Therefore, it is necessary to first estimate these probability density functions using kernel density estimation and then substitute them into the mutual information formula for calculation. The specific steps can be broken down as follows: data preprocessing, estimating marginal probability densities p(x) and p(y) using kernel density estimation, estimating the joint probability density p(x,y) using kernel density estimation, and substituting them into the mutual information formula to calculate the final value.

[0057] Among them, data preprocessing is to standardize the sample data of candidate variables such as X = rainfall and Y = water level, corresponding to the min-max normalization of the present invention to the [0,1] interval, to eliminate the interference of dimensional differences such as rainfall in millimeters and water level in meters on probability density estimation.

[0058] Among them, the kernel density estimation method is used to estimate the marginal probability densities p(x) and p(y). Taking the estimation of the marginal probability density p(x) of rainfall X as an example: Assume that X has n sample points: x1, x2, ..., xn; Choose the Gaussian kernel function because it can smooth out data fluctuations, making it suitable for flood control data. Gaussian function: , x is the target point of the probability density to be estimated, i.e., a specific value of the probability density to be calculated, such as water level = 18.5m or rainfall = 50mm / h. xi is a sample point, i.e. existing flood control monitoring data, such as the water level value of 18.2m and the rainfall of 48mm / h in the i-th record of historical data.

[0059] The smoothness of the kernel function is controlled by determining the bandwidth h. In this invention, the smoothness is adaptively adjusted based on the sample size to ensure estimation accuracy. Substituting into the kernel density estimation formula, we can calculate the probability density p(x) for any value x: , The meaning of this formula is: by superimposing the contributions of the Gaussian kernel function of each sample point xi, the probability density curve of the entire X is obtained. The denser the sample points in a region, the larger the value of p(x), that is, the higher the probability of X taking a value in that region.

[0060] Similarly, the marginal probability density p(y) of water level Y can be estimated using the same method.

[0061] The joint probability density p(x,y) is estimated using the kernel density estimation method. For the joint distribution of rainfall X and water level Y, it is assumed that there are m sets of two-dimensional sample points: (x1,y1), (x2,y2), ..., (xm,ym). The joint probability density p(x,y) is estimated using a two-dimensional Gaussian kernel function, and the formula is as follows:

[0062] Where hx and hy are the bandwidths of X and Y, respectively. The two-dimensional Gaussian kernel function can simultaneously characterize the joint distribution characteristics of X and Y, such as regions with high rainfall and high water levels. The larger the value of p(x,y), the better.

[0063] Substitute into the mutual information formula to calculate the final value: Substitute p(x, p(y), p(x,y) obtained through kernel density estimation into the integral formula of mutual information, and calculate the specific value of I(X;Y) through numerical integration such as the trapezoidal integral method. The larger the value, the stronger the causal relationship between X and Y.

[0064] The application details of the kernel density estimation method for calculating mutual information values ​​in the causal relationship modeling stage of this invention are as follows: This invention sets a mutual information value greater than 0.8 as a strong correlation threshold, which can be dynamically adjusted according to the accuracy requirements of flood control scenarios, for example: If the mutual information value between hourly rainfall and reservoir water level is 0.92, which is greater than 0.8, it is determined to be a strong correlation and included in the causal path such as rainfall → water level; If the mutual information value between the basin's GDP and flood discharge is 0.35, which is less than 0.8, it is considered a weak association and will not be included in causal modeling to avoid interference from redundant variables.

[0065] The calculated mutual information value is the core basis for the causal influence coefficient. The mutual information value is normalized to the range of [-1,1] using the maximum absolute value scaling method. Positive values ​​indicate positive influence, such as increased rainfall → rising water level; negative values ​​indicate negative influence, such as increased flood discharge → falling water level. Finally, an M×N dimensional causal influence matrix is ​​formed, where M is the number of candidate elements and N is the number of target elements, providing quantitative input for subsequent dynamic Bayesian network (DBN) modeling.

[0066] Causal Influence Matrix: The matrix has dimensions M×N, where rows correspond to candidate key elements, including but not limited to input variables that directly or indirectly affect flood control decisions such as rainfall, water level, flow velocity, flood discharge, and early warning response time; columns correspond to target elements, including but not limited to output variables reflecting the results of flood control decisions such as flood risk level, flood control effectiveness assessment value, and disaster area prediction value. The element values ​​in the matrix are causal influence coefficients, ranging from [-1, 1]. Their normalization method is based on the path coefficients output by the IMaGES algorithm. The original values ​​of the path coefficients have no fixed range, and they are mapped to the [-1, 1] interval using the maximum absolute value scaling method, as shown in the formula: c_ij = k_ij / max (|k_ij|), Where k_ij is the original path coefficient, and max (|k_ij|) is the maximum absolute value of all path coefficients.

[0067] Positive values ​​of the coefficient indicate positive causal effects, such as increased rainfall leading to higher flood risk, while negative values ​​indicate negative causal effects, such as increased discharge reducing flood risk. The absolute value reflects the strength of the impact.

[0068] Iteration parameter settings: The maximum number of iterations for the IMaGES algorithm is set to 50, which can be adjusted according to the model complexity. The convergence threshold is set to a path coefficient change of less than 0.001. When the difference between two consecutive calculated path coefficients is less than this threshold during the iteration process, the algorithm terminates and outputs a stable causal graph model and causal influence matrix.

[0069] 3.3 Model Analysis and Validation The constructed causal model is subjected to statistical model checking (StatMC) and inference analysis, including: on the basis of the causal network output by the IMaGES algorithm, a flood control scheduling decision model based on a dynamic Bayesian Network (DBN) is constructed. The causal path, direction and intensity between variables have been clearly defined. The dynamic Bayesian Network will use this as the basic structure, incorporate the characteristics of the time dimension, and systematically model and infer the changes in hydrological situation, scheduling behavior and its feedback response under time series, so as to realize the extension from static causal relationship to dynamic time series evolution.

[0070] Regarding the timing of the dynamic Bayesian network: the time slice is set to update interval, for example, 1 hour / slice, and the time window length is for example, 72 slices, covering 3 days.

[0071] Regarding the probability parameters of the dynamic Bayesian network: the conditional probability is initialized based on frequency statistics from historical data; for example, the probability of water levels exceeding the warning level when rainfall exceeds 50mm is 60%. The DBN model was validated using the StatMC method, and the steps are as follows: Define the verification objectives, including: risk event prediction deviation rate ≤10% (e.g., water level exceeding the warning level in the next 6 hours); mean absolute error (MAE) of continuous variable dynamic evolution ≤0.2 units (e.g., water level MAE ≤0.2m); and decision response deviation rate ≤15% (e.g., the water level drop within 24 hours after opening the No. 3 floodgate). Construct a validation dataset: 300 independent samples were randomly selected from nearly 10 years of historical data, including initial conditions and 72-hour observation sequences. After preprocessing by missing value interpolation and outlier correction, the samples cover typical scheduling decision cases. Generate prediction results: Input the initial conditions of the sample into the DBN model, and output the probability of risk events within 72 hours, such as the probability of waterlogging in the basin at t=12, continuous variable values ​​such as hourly water level, flood discharge, and decision response predictions such as the proportion of reduction in disaster area after flood diversion. Statistical tests: For discrete risk events such as water levels exceeding warning levels, a binomial distribution test with α=0.05 is used, where α represents the significance level of the statistical test, used to determine whether there is a significant difference between the model's predicted satisfaction rate and the actual satisfaction rate. If the test result P-value ≤ 0.05, it indicates a significant difference, and the model needs to be corrected; if P-value > 0.05, it indicates no significant difference, and the model meets expectations. For continuous variables such as the time-series change in flood discharge, a t-test is used, with a statistical significance level of α=0.05, used to determine whether the difference between the model's predicted value and the actual value is significant. If P-value ≤ 0.05, it indicates that the prediction deviation exceeds the random error range, and the model parameters need to be adjusted; if P-value > 0.05, it indicates that the deviation is acceptable, and the model's prediction effect meets the target. For decision responses such as the risk reduction after relocating 50,000 people, a relative deviation test is used, marking deviation samples that do not meet the target. Iterative correction: For biased samples, such as the model underestimating the rate of water level rise in the scenario of severe rainstorm, update the conditional probability table of the DBN model through maximum likelihood estimation, or correct the model by adding new variables and causal paths, and repeat the verification until 50 consecutive sets of samples meet the target. Results output includes: sample pass rate ≥90%, key indicator quantification results, final model parameters, and applicable scenario boundaries.

[0072] The specific implementation methods and processes are as follows: Figure 1 As shown, the method for identifying key elements and modeling causal relationships in flood control decision-making according to this invention includes the following steps: Step 1: Classification of Basic Data Sources and Feature System Step 1.1 Basic Element Collection: Collect basic data such as water conditions, rainfall, engineering conditions, danger conditions, and disaster conditions, specifically including real-time and historical monitoring data, historical statistical data, and other multi-source heterogeneous data.

[0073] The scope of multi-source heterogeneous data types explicitly covers the following data types: Water and rainfall data: real-time water level / flow velocity at hydrological stations, hourly rainfall at rain gauge stations, and precipitation forecast data for the next 72 hours; Engineering data: Dike monitoring data such as seepage flow, displacement deformation, reservoir capacity curve, and flood discharge facility operation parameters; Socioeconomic data: historical disaster-affected population statistics, GDP loss assessment, and population density in flood protection zones; Emergency management data: hazard reporting records (location and level of piping and cracks), historical emergency material dispatch logs, and emergency plan texts.

[0074] Heterogeneous data fusion method: A three-level fusion strategy is adopted to achieve standardized data integration. Data layer fusion: For heterogeneous data of the same type, such as water level monitoring data in different formats: CSV files and JSON streams are converted and time-aligned, and real-time splicing is achieved through database middleware (such as Kafka); Feature layer fusion: Common features such as "area rainfall" are extracted from cross-type data such as rain gauge data and satellite precipitation data. A weighted fusion algorithm is used (0.7 weight for station data and 0.3 weight for satellite data, and the weights can be dynamically adjusted according to the data accuracy). Decision-making integration: High-dimensional data is used to form a unified decision-making feature space by using feature mapping, such as associating "dam height" and "surface slope" with "watershed flood control capacity index", which provides standardized input for subsequent element modeling.

[0075] Step 1.2 Setting target elements: Key station control water levels, flood inundation in key areas, flood diversion and utilization in flood storage and detention areas, and the final water level of core reservoirs represent comprehensive flood control targets.

[0076] Step 1.3 Control Element Construction: Identify control parameters such as scheduling strategies and defense timing, including scheduling behavior elements such as operating procedures and emergency plans.

[0077] There is a logical relationship between the three, such as Figure 3 As shown, the basic elements and target elements together determine the setting and implementation of control elements, and the role of control elements will in turn affect the evolution of the flood situation, thus forming a dynamic feedback process to achieve continuous optimization and improvement of the flood defense system.

[0078] Step 2: Data Structure Design and System Processing Step 2.1 Flood Situation Modeling: Perform correlation analysis on basic elements and target elements to establish a time series model to identify data trends, extreme points, and abrupt changes. A hierarchical processing mechanism is adopted to address data anomalies. If a monitoring station, such as a rain gauge station, has missing data (no records for two consecutive time slices), the "neighbor station interpolation method" is used to fill the gaps: select the data from three surrounding stations of the same type during the same period, calculate the weights using the inverse distance method (the weight coefficient is inversely proportional to the square of the distance between stations), and generate the missing values ​​by weighted averaging to ensure that spatial correlation has a corrective effect on the data.

[0079] If the monitoring data shows a sudden change, such as a water level rise of more than 2 meters within 1 hour that far exceeds the historical maximum for the same period, a "data validity verification" is immediately triggered: combining upstream rainfall during the same period, gate operation records, river runoff coefficients, and other related data, it is determined whether the sensor is faulty. If a fault is confirmed, a sliding window weighted average method is used for trend prediction and compensation: based on the measured data of the three time slices before the fault, decreasing weights are assigned according to time proximity, such as 0.5 for the most recent time slice, 0.3 for the previous two, and 0.2 for the previous three. A substitute value is generated by weighted fitting of the short-term trend; if the data sequence has obvious periodicity, a periodic component correction is superimposed (using the average of the three complete cycles during the same period as the periodic correction term) to ensure that the compensation result conforms to the natural variation law of hydrological elements.

[0080] Step 2.2 Flood Situation Identification: Based on the situation modeling results, and referring to the flood risk level classification standards stipulated in the "National Flood and Drought Relief Emergency Plan", such as the rainfall and water level rise corresponding to blue, yellow, orange and red warnings, and combined with the design standards of specific water conservancy projects in the basin and historical risk data, the situation identification and classification of risk levels and warning signals are carried out.

[0081] Step 2.3 Scheduling behavior modeling: Quantify and express the control elements, and construct a scheduling rule set and decision tree model.

[0082] Step 3: Identification of Key Technology Paths Step 3.1 Initial Variable Screening. Based on multi-source heterogeneous data on water conditions, rainfall, engineering conditions, hazards, and disasters, all variables are standardized and multicollinearity is detected. Variables with low significance or redundancy are removed, and a candidate element set is constructed.

[0083] Step 3.2 Optimal Subset Selection Algorithm.

[0084] The optimal subset selection algorithm is a method for selecting the optimal model from multiple independent variables included in the candidate feature set constructed in step 3.1. It constructs the optimal regression equation by progressively adding variables and comparing model performance. The basic idea of ​​this method is to sequentially consider model combinations containing different numbers of independent variables, calculate the performance index of each combination, and use minimizing cross-validation prediction error, Akaike information criterion (AIC), Bayesian information criterion (BIC), etc., as objective functions to ultimately determine the best-performing model structure.

[0085] The main steps of the algorithm are as follows: 1) Initialize an empty model M0, containing only the intercept term; 2) For each k = 1, 2, 3, …, p, enumerate all models containing k independent variables and fit them; 3) Calculate the evaluation metrics such as AIC, BIC, and cross-validation error for each model, and select the best-performing model M. k ; 4) Select the overall optimal model M from all candidate models. ∗ Output the corresponding variable combinations and parameter estimation results.

[0086] Step 3.3 Stepwise Regression Screening. In each iteration, the algorithm evaluates the significance level of each independent variable in the model based on the statistical results of the F-test and t-test. When the introduction of an independent variable significantly improves the goodness of fit of the model (i.e., the residual sum of squares (RSS) decreases significantly, the overall F-value of the model decreases, and the corresponding p-value is lower than the significance threshold α), the variable is included in the model. α is typically set to 0.05 or 0.01. Conversely, when the t-test of a variable in the current model is no longer significant (i.e., the p-value is greater than the threshold), it is removed. By continuously repeating the dynamic process of variable "introduction-testing-removal," the model gradually converges to the optimal structure. In this invention, to address the complexity of multi-source heterogeneous data such as hydrological and meteorological data, engineering scheduling data, and socio-economic data, the stepwise regression algorithm is used to optimize the model based on the selection of the optimal subset. Specific implementation points are as follows: 1) Significance determination mechanism: After each iteration, the regression coefficients, standard errors, and t-statistics of all variables are recalculated, and the p-value is used to determine whether the inclusion or retention criteria are met. A typical threshold is set as follows: Entry criteria: p < 0.05; Exclusion criteria: p > 0.10; This dual threshold strategy ensures stable model convergence and avoids overfitting.

[0087] 2) Multicollinearity control: To prevent the influence of multi-source features collinearity, the algorithm calculates the variance inflation factor (VIF) after each variable screening. If the VIF of a variable exceeds a set threshold (usually 5 or 10), the variable is considered to have a strong correlation with other variables and is removed.

[0088] 3) Model Evaluation and Termination Criteria: After introducing or removing variables at each step, the algorithm calculates comprehensive evaluation metrics such as the adjusted R², AIC, BIC, Mallow's Cp, and cross-validation error. If the performance improvement of the model is less than a threshold (e.g., ΔAIC < 0.01) for two consecutive iterations, the model is considered to have reached its optimal state, and the computation terminates.

[0089] 4) Results Output and Interpretation: The final output regression equation contains the set of core elements with the highest predictive value and their weight parameters, which can quantitatively characterize the comprehensive impact of different hydrological, meteorological, and engineering factors on flood control effectiveness. The regression coefficients of the model not only reflect the direction and intensity of the variable's influence, but can also be combined with standardized coefficients to analyze their relative contribution, providing a quantitative basis for subsequent causal modeling and analysis.

[0090] In this invention, step 3.1, correlation analysis, reduces the burden on the IMaGES algorithm by providing high-quality candidate variables. Step 3.1 involves preprocessing and redundancy removal of multi-source heterogeneous flood control data, and its output directly serves as the input basis for the IMaGES algorithm. This addresses the problems of low efficiency and susceptibility to noise interference in processing high-dimensional data by the IMaGES algorithm. Specifically, its functions include: Data cleaning and standardization: For continuous data such as water and rainfall conditions and engineering conditions, a sliding window midpoint filter combined with the 3σ criterion is used to remove noise / outliers. For discrete data such as disaster level and emergency plan type, one-hot encoding conversion is used. At the same time, the variables are mapped to the [0,1] interval through max-min standardization. These operations can eliminate the interference of dimensional differences, such as rainfall in millimeters versus water level in meters, and data noise on the mutual information calculation of the IMaGES algorithm, ensuring the accuracy of subsequent causal correlation strength quantification.

[0091] Initial screening of redundant variables: Variables weakly correlated with or redundant to target elements such as water level compliance rate and flood storage target achievement, or minor socioeconomic indicators unrelated to flood evolution, are eliminated through optimal subset selection combined with stepwise regression. This constructs a set of candidate key elements. For example, if the stepwise regression algorithm determines that the impact of watershed vegetation coverage on "water level rise" has a significance P-value > 0.05, it is removed. This reduces the number of modeling variables in the IMaGES algorithm from dozens to 10-20 core variables, significantly improving the efficiency of causal modeling and avoiding the curse of dimensionality.

[0092] Step 3.1 is the preliminary data preparation step for the IMaGES algorithm: If this step is skipped and the original high-dimensional, noisy data is directly input into the IMaGES algorithm, it will lead to confusion in causal path identification, such as misjudging spurious correlation variables as key elements, or even failing to converge to a stable causal model.

[0093] Step 3.2, optimal subset selection and stepwise regression, is to orient the IMaGES algorithm and narrow the causal search scope. Step 3.2 is a deepening of step 3.1, using statistical methods to further focus on strongly correlated variables, thus defining a clear scope for the causal path search of the IMaGES algorithm. The specific collaborative logic is as follows: Optimal subset selection: Identify potential related variable combinations and enumerate no more than 100 variable combinations, such as "rainfall + water level + flood discharge" or "engineering status + emergency situation + dispatching route". Select the combination with the strongest explanatory power for the target element through indicators such as AIC and BIC. For example, if the combination "rainfall + reservoir capacity + flood discharge facility parameters" is selected as having the best explanatory power for the completion of the water storage target, the IMaGES algorithm can then conduct causal relationship mining around this combination without having to traverse all variable combinations, thus reducing invalid calculations.

[0094] Stepwise regression: This method strengthens the statistical significance of variable associations. Using α=0.05 as the significance threshold, variables are introduced / removed through the F-test, i.e., variables with P≤0.05 are introduced, and variables with P>0.05 are removed. This results in a set of statistically significant related variables. For example, hourly rainfall (P=0.02), real-time water level (P=0.01), and flood discharge (P=0.03) are identified as significantly related variables. The IMaGES algorithm can accurately calculate mutual information values ​​based on these variables to quantify the causal strength, avoiding overfitting of the causal model due to the inclusion of insignificant variables.

[0095] Step 3.2 is the search scope optimization step of the IMaGES algorithm. It focuses the causal search of the IMaGES algorithm from all variables to statistically significant strongly correlated variables, ensuring that the IMaGES algorithm can efficiently and accurately discover statistically supported causal relationships, rather than random spurious associations.

[0096] Step 3.3, model analysis and verification, verifies the reliability of the causal model. Step 3.3 involves dynamically expanding and statistically verifying the causal model output by the IMaGES algorithm, including the causal influence matrix and causal paths. This addresses the issue that the IMaGES algorithm only outputs a static causal structure, lacking dynamic temporal adaptability and engineering reliability. The specific collaborative logic is as follows: The IMaGES algorithm outputs a static causal foundation: The IMaGES algorithm quantifies the causal strength between variables through mutual information calculation, outputting a static causal influence matrix of dimension M×N, where M is the candidate element and N is the target element, with coefficients ∈ [-1,1]) and causal paths such as "rainfall → water level → flood discharge". However, this result only reflects the static correlation between variables and does not consider the dynamic time dimension of flood evolution, such as the response delay of water level to rainfall. Step 3.3 extends the static causal model to a dynamic model. Step 3.3 introduces a dynamic Bayesian network (DBN), based on the static causal structure output by the IMaGES algorithm, incorporating a time slice update mechanism of 1 hour / slice and state transition relationships, such as initializing the probability of water level exceeding the warning level when rainfall > 50mm to 60%, and constructing a dynamic causal model. This makes up for the shortcomings of the static modeling of the IMaGES algorithm, enabling the causal relationship to characterize the temporal changes of flood evolution, such as rainfall at time t → water level rise at time t+2.

[0097] Step 3.3 Validate the engineering reliability of the causal model using StatMC: Step 3.3 employs Statistical Model Testing (StatMC) to extract 300-500 samples from nearly 10 years of historical data to validate the DBN dynamic model, ensuring a risk event prediction bias rate ≤10% and a MAE ≤0.2 units. If the verification fails, such as the model underestimating the water level rise in a heavy rainfall scenario, the process is backtracked to the IMaGES algorithm stage to check whether key causal paths such as rainfall → soil moisture content → runoff coefficient → water level are missed, or the mutual information calculation parameters are corrected. If the verification is successful, it confirms that the causal relationship output by the IMaGES algorithm conforms to the physical laws of flood control and historical cases, and can be used as the final basis for identifying key elements.

[0098] Step 3.3 is the model implementation and verification stage of the IMaGES algorithm: it transforms the static causal results of the IMaGES algorithm into a dynamic model that can be used in engineering, and ensures the reliability of the causal relationship through statistical testing, so as to avoid the causal model output by the IMaGES algorithm being unable to guide flood control decision-making due to being out of touch with the actual scenario.

[0099] In this invention, steps 3.1 and 3.2 are the input optimizers for the IMaGES algorithm, which is the core of causal relationship mining. Step 3.3 is the model verification and implementation guarantee. All four are indispensable and together solve the problems of difficult identification of key elements and inaccurate causal relationship modeling in traditional flood control decision-making.

[0100] Step 4: Causal Modeling of Decisions In this invention, the core of the third step, key technology path identification, is to identify credible key elements and static causal relationships. The fourth step, decision causal modeling, is to build and optimize a dynamic fusion model based on the third step. The third step outputs a static causal structure, including a causal influence matrix and causal paths, using the IMaGES algorithm. This clarifies the logical relationships between key elements, such as "rainfall → water level → flood discharge," serving as input for the fourth step's fusion model. The fourth step incorporates the time dimension characteristics of a dynamic Bayesian network (DBN), including 1-hour time slices and 72-hour time windows, transforming the static causal relationship into a dynamic time-series model. This model depicts the real-time changes in causal relationships during flood evolution, such as "rainfall at time t → water level rise at time t+2." Simultaneously, the fourth step introduces BLTL logical constraints, transforming flood control physical laws, such as "water level reaching the standard 12 hours after flood discharge," into computable rules, optimizing the weights of key elements, and ensuring the model meets actual decision-making needs. The third step of this invention involves selecting credible key elements and static causal relationships from massive amounts of data, addressing the questions of which elements are important and how they are related. The fourth step imbues the static causal relationship with temporal dynamism and decision-making constraints, addressing how to transform causal relationships into a feasible scheduling model. In this invention, decision-making causal modeling includes causal graph selection, decision behavior modulation, and model credibility prediction.

[0101] The causal graph screening is based on the modeling results of the IMaGES algorithm. It uses parameter significance, confidence intervals, and structural stability indices to screen causal paths based on their credibility. Edges with low significance or uncertain direction are removed. Low significance is determined by a P-value < 0.05; if the P-value of a causal path is ≥ 0.05, it is considered a low-significance path and removed. Uncertain direction is determined using Bootstrap sampling (sampling times set to 1000). If the percentage of direction reversals in a causal path exceeds 30% during sampling, it is considered an uncertain path and removed. High-confidence causal relationships are retained, generating a simplified and robust causal topology. The overall algorithm flow is as follows: 1) Initial modeling stage: Starting from the core elements determined by the optimal subset selection and stepwise regression results, construct the initial causal graph structure G0, and establish the preliminary causal relationship between variables in the form of structural equation model.

[0102] 2) Sensitivity Analysis Phase: Sensitivity calculations are performed on potential confounding factors in the current causal diagram to assess the impact of each potential confounding factor on the estimated causal effect. Sensitivity can be measured using the rate of change of causal effect or the magnitude of the disturbance.

[0103] 3) Potential Factor Removal Stage: Based on the sensitivity results, potentially confounding factors with significant impact are gradually removed from the causal graph to obtain a deconfounded causal structure 𝐺. 𝑖.

[0104] 4) Iterative Update Phase: Based on the new causal structure, a new sensitivity analysis is performed, remaining confounding factors are re-evaluated, and model parameters are updated. This process continues iteratively until the sensitivity of all potential confounding factors is below a preset threshold or the maximum number of iterations is reached.

[0105] 5) Convergence and Output Phase: After the model converges, the final causal graph 𝐺 is output. ∗ It includes the direct and indirect causal paths between key elements, path coefficients, and significance levels.

[0106] Decision-causal modulation (DCM) is a process that combines a credible causal structure with flood control scheduling objectives and actual operational mechanisms to conduct decision-causal modulation. This process quantifies the impact of different control elements (such as water level control, gate opening and closing, and flood detention area scheduling) on ​​the evolution of the system's situation, simulating causal response changes under different scheduling schemes. By comparing the strength, direction, and feedback effects of causal paths, key decision nodes and scheduling-sensitive areas are identified, thereby optimizing defense strategies and achieving dynamic causal modulation and response prediction in flood control scheduling.

[0107] Model credibility prediction involves evaluating and validating the credibility of a modulated causal model. The model's predictive credibility and generalization performance are comprehensively assessed using metrics such as validation program feedback, parameter confidence intervals, path significance, and residual independence tests. For paths with significant prediction deviations, dynamic corrections are performed through model retraining and parameter calibration strategies to ensure the model possesses stable explanatory power and decision-making reference value. The final model credibility prediction results are fed back into the causal model library and scheduling rule set to support intelligent decision-making and strategy optimization in subsequent flood control command.

[0108] In this invention, the core of the third step, key technology path identification, is to identify credible key elements and static causal relationships. The fourth step, decision causal modeling, is to build and optimize a dynamic fusion model based on the third step. The third step outputs a static causal structure, including a causal influence matrix and causal paths, using the IMaGES algorithm to clarify the logical relationships between key elements, such as "rainfall → water level → flood discharge." This serves as the input for the fourth step, building a fusion model. The fourth step incorporates the time dimension characteristics of a dynamic Bayesian network (DBN), including 1-hour time slices and 72-hour time windows, transforming the static causal relationship into a dynamic time-series model. This model depicts the real-time changes in causal relationships during flood evolution, such as "rainfall at time t → water level rise at time t+2." Simultaneously, the fourth step introduces BLTL logical constraints, transforming flood control physical laws, such as "water level reaching the standard 12 hours after flood discharge," into computable rules, optimizing the weights of key elements, and ensuring the model meets actual decision-making needs. The third step of this invention involves screening credible key elements and static causal relationships from massive amounts of data, addressing the questions of which elements are important and how they are related. The fourth step assigns temporal dynamism and decision-making constraints to the static causal relationships, addressing the question of how to transform causal relationships into a feasible scheduling model. The fifth step: Validation of the dynamic Bayesian network based on StatMC and identification of flood control decision-making feature elements. After completing the key element screening, causal modeling, and decision credibility prediction, this stage aims to systematically analyze and verify the previously constructed flood control scheduling decision model based on key feature elements and dynamic Bayesian networks. Formal logic methods are used to identify key feature elements that accurately reflect expert experience and flood control patterns. StatMC verifies the reliability of the DBN model in predicting the dynamic evolution of flood risk and decision effectiveness by statistically comparing the model output with actual observation data. Based on the "bounded temporal constraint" characteristic of the BLTL verification objective, a complete technical route of "objective definition - sample generation - trajectory simulation - logical judgment - statistical inference - model iteration" is constructed. The specific steps are as follows: Step 5.1 BLTL verification of target quantization, defining timing logic constraints First, based on the verification requirements, the "atomic propositions," "timing operators," and "time boundaries" of the BLTL formula are defined, transforming the abstract timing logic into a computable quantifiable target, thus providing a judgment benchmark for subsequent verification.

[0109] (1) Definition of core concepts a) Atomic propositions: These describe the specific state of variables in the DBN model. A quantization threshold needs to be set based on the variable type (continuous / discrete), such as "water level ≥ 18m (warning level)" or "flood discharge ≥ 500m³". 3 / s"The risk of waterlogging is ≤20%".

[0110] b) BLTL temporal operators: including □t1,t2 means always, the proposition is always true within the time interval [t1,t2], ◇t1,t2 means finally, the proposition is true at least once within the interval, and → means implication, if the previous proposition is true, the subsequent proposition must be true. They are used to characterize the temporal dependency between variables.

[0111] The always operator □t1,t2 indicates that a certain proposition, i.e., the judgment describing the state of an element, must always be true throughout the entire time interval from time t1 to time t2, without any exceptions. The boundedness is reflected in the explicit nature of the time interval [t1,t2]. For example, a 72-hour flood control window is commonly used in documents, i.e., t1=0, t2=72, which differs from the permanent truth of unbounded temporal logic. The always operator □t1,t2 is used to constrain the fundamental conditions that must be continuously met during flood control, especially the relationships between elements related to engineering safety and risk baselines. For example, a dike safety constraint can be expressed as □0, 72 (Seepage flow of the dike ≤ 5L / s) means that during the 72-hour flood control cycle, the real-time seepage flow of the dike must always be less than or equal to 5L / s. If the seepage flow exceeds this value at any time, this logical constraint is violated and emergency reinforcement measures must be triggered.

[0112] The symbolic final operators ◇t1,t2 represent the condition that a certain proposition will be true at least once within the time interval from time t1 to time t2. It does not require continuous validity, but it must satisfy existence. This is used to characterize the timeliness constraint in flood control response, meaning that after a certain triggering condition occurs, the corresponding result or scheduling behavior must occur within a specified time. For example, the BLTL formula example □0, is explicitly given in the document. 72 (Rainfall ≥ 100mm / 3h → ◇0,6 (water level ≥ 18m), where ◇0,6 (water level ≥ 18m) means that within 0-6 hours after the rainfall reaches the standard of ≥ 100mm / 3h, the water level must reach the warning level of 18m at least once. If the warning level is not reached within 6 hours, it means that the model prediction is in conflict with the hydrological and physical laws. The heavy rainfall did not trigger a water level response, and the causal relationship needs to be corrected to ensure the causal timeliness between elements and avoid logical contradictions in the model where strong triggering conditions have no corresponding results, such as predicting a severe rainstorm but no water level rise.)

[0113] The symbolic implication operator "→" represents a logical implication relationship, meaning that if the preceding propositional condition is true, the subsequent propositional result must also be true. This can be understood as a mandatory association between condition and result, and is the foundation for constructing multi-operator combinational logic. It is used to connect triggering conditions and necessary results / constraints to form a complete flood control rule chain. It is often used in conjunction with the operators "□" and "◇", as in the rainfall-water level example above. Another example in this invention is "◇0, 72 (Open the floodgates ∧□0, 12(Water level ≤ 17.5m) can be broken down into "Open the floodgate → □0, 12 (Water level ≤ 17.5m) means that if the floodgate is opened within 72 hours, the water level must remain ≤ 17.5m for 0-12 hours after opening. If the water level exceeds 17.5m within 12 hours after opening the floodgate, it indicates that the scheduling action has not achieved the expected goal. The causal relationship between scheduling effect and water level response in the model needs to be optimized. A mandatory correlation between conditions and results needs to be established to transform flood control experience rules, such as the mandatory control of water level during flood discharge, into verifiable logic in the model, thus avoiding the problem of scheduling actions being out of sync with the goal.

[0114] The three operators—always operator □t1,t2, final operator ◇t1,t2, and implied operator "→"—form a complete flood control logic rule, used to optimize the weights of key elements and eliminate false causal relationships. For example, by combining the operator "□ (condition → ◇ result)", the logic of core causal chains such as heavy rainfall → water level warning and flood discharge → water level control is verified: first, the condition-result are associated through the operator "→"; then, the time range of the condition is limited by the operator "□", such as heavy rainfall at any time within 72 hours; finally, the time window of the result is limited by the operator "◇", such as water level response within 6 hours. If the model's predicted trajectory violates this combination rule, such as no water level response within 6 hours of heavy rainfall, it is judged as a logically inconsistent sample. The model is iteratively corrected using the StatMC method, ultimately ensuring that the causal relationship of key elements conforms to the physical laws of flood control and expert experience.

[0115] c) Time boundary: Based on the DBN time slice setting, such as 1 hour / slice, a bounded range is determined. Usually, the time window of concern for flood control decision-making is taken, such as 72 hours, i.e., t∈[0,72]).

[0116] Example of BLTL formula construction: a) Regarding the timeliness of risk warnings: □0,72 (rainfall ≥100mm / 3h → ◇0,6 (water level ≥18m)), meaning that if the cumulative rainfall in any 3-hour period exceeds 100mm within 72 hours, the water level will definitely exceed the warning level within 6 hours.

[0117] b) Regarding the stability of the scheduling effect: ◇0, 72 (open the floodgate ∧ □0, 12 (water level ≤ 17.5m)), which means that if the floodgate is opened within 72 hours, the water level must be maintained below 17.5m for the next 12 hours.

[0118] Step 5.2 Validate the sample set construction to ensure data representativeness. By extracting samples covering multiple scenarios from historical databases, we can provide a sufficient and effective data foundation for the statistical inference of StatMC and avoid the distortion of validation results due to sample bias.

[0119] (1) Sample source and size a) Select historical data from the past 10 years that include typical scenarios such as regular rainfall, torrential rain, joint reservoir operation, and emergency response to dikes within the basin to ensure that the sample covers the core application boundaries of the DBN model.

[0120] b) The sample size is set at 300-500 groups to balance statistical significance and computational efficiency. Each sample group includes initial conditions and the actual time-series trajectory. i) Initial conditions: DBN input variable values ​​at time t=0, such as initial water level, rainfall, and dike safety factor. ii) Actual time series trajectory: Actual observed values ​​of variables from t=1 to t=T (T=72), such as hourly water level, dispatch behavior records, and risk event results.

[0121] (2) Sample preprocessing a) Outlier handling: For outlier data in the actual trajectory, such as water level jumps caused by sensor failure, the adjacent time mean replacement method is used to correct the data by taking the mean of the two time slices before and after the jump. b) Missing value completion: For missing observations, such as the rainfall in a certain time slice, the neighboring station interpolation method is used. The weighted average of the data from three similar stations in the same period is used to complete the missing data, with the weight inversely proportional to the distance, to ensure the integrity of the true trajectory.

[0122] Step 5.3 Simulate the time-series trajectory of the DBN model to generate prediction data. The initial conditions of the sample are input into the DBN model, and the predicted trajectory is generated through model iterative inference, which is aligned with the time dimension of the real trajectory, providing comparative data for subsequent BLTL logic judgment.

[0123] (1) DBN iterative inference process: According to the DBN time slice interval (1 hour / slice), iterate from t=0 to t=T (72 hours), and execute the following in each time slice t: a) State reasoning: Based on the variable states at time t-1 and the conditional probability table of DBN, calculate the probability distribution of each variable at time t. For continuous variables, output the mean and variance; for discrete variables, output the probability of each state. b) Truth value determination of propositions: For atomic propositions in the BLTL formula, the truth value is determined according to the preset threshold. For example, for continuous variables, if the water level is ≥18m, the truth value is 1 if the predicted average water level at time t is ≥18m, otherwise it is 0; for discrete variables, if the floodgate is opened, the truth value is 1 if the probability of this state is ≥0.5.

[0124] (2) Predicted trajectory output: Record the time slice - atomic proposition truth value - variable prediction value of each sample from t=0 to t=72 to form a structured predicted trajectory, ensuring that the time nodes and proposition types are completely matched with the real trajectory, which is convenient for subsequent time slice comparison.

[0125] Step 5.4 Automatic determination of BLTL formula, verification of timing logic on a sample-by-sample basis. A temporal logic parser is used to determine the BLTL formula satisfaction of the actual trajectory and the predicted trajectory respectively, to clarify the temporal logic matching degree of the two sets of trajectories, and to mark the deviation samples.

[0126] (1) Formula analysis and decomposition Decomposing BLTL formulas into sequences of atomic propositions plus temporal operator constraints, for example, for formulas... □0, 72 (A→◇0, 6B), decomposed into: a) Atomic proposition A: Rainfall ≥ 100mm / 3h, the cumulative rainfall over 3 hours from t-2 to t needs to be calculated to determine whether the standard is met; b) Atomic proposition B: Water level ≥ 18m; and c) Temporal constraint: For all t∈[0,72], if A is true at time t, then there must exist t'∈[t, t+6] such that B is true at time t'.

[0127] (2) Trajectory Satisfaction Determination a) True trajectory determination: Verify the truth value of atomic propositions time slice by time slice, determine whether the BLTL formula is satisfied according to the time sequence constraints, and output the "true satisfaction result", for example, 1 = satisfied, 0 = not satisfied; b) Predicted trajectory determination: The determination logic is completely consistent with the actual trajectory, and the output is "prediction satisfied". c) Deviation labeling: If the "predicted result satisfies the actual result" of a sample is not the same as the actual result, it is labeled as "BLTL Deviation Sample". The time interval in which the deviation occurs, such as t∈[10,15], is recorded as the deviation between the predicted result satisfies the actual result and the actual result not satisfies the corresponding atomic proposition, such as the truth value of proposition B.

[0128] Step 5.5 StatMC statistical inference to quantify and verify reliability Statistical hypothesis testing is conducted based on the "proportion of biased samples" and the "difference in satisfaction rate" to determine whether the BLTL predictive ability of the DBN model meets the requirements. The core is to quantify the reliability of the model's time-series logic through statistical measures.

[0129] (1) Setting statistical hypotheses a) Null hypothesis H0: The prediction satisfaction rate R_pred = number of predicted satisfied samples / total number of samples and the satisfaction rate R_true = number of actual satisfied samples / total number of samples of the DBN model are not significantly different; b) Alternative hypothesis H: There is a significant difference between the two; c) The significance level α = 0.05, which is a commonly used value for balancing reliability and fault tolerance in engineering scenarios.

[0130] (2) Implementation of statistical tests a) Since whether a sample satisfies the BLTL formula is a binary classification result of satisfying / dissatisfying the formula, a binomial distribution test is used: i) Assuming the total number of samples is N and the actual satisfaction rate is R_true, then the number of samples predicted by the model to be satisfied theoretically follows a binomial distribution B(N, R_true). ii) Calculate the deviation between the actual predicted number of satisfied cases and the theoretical expected number, and calculate the P-value using the cumulative distribution function; The calculation of the P-value is explained as follows: The essence of statistical testing is to simulate the DBN to be tested, generating simulation traces. Initial conditions, such as water level and rainfall at t=0, are input, and the model runs automatically according to preset rules, such as conditional probability and state transition relationships. The core output is a structured data sequence arranged in chronological order. Each trace contains the specific values ​​of all variables in the DBN, such as water level, flood discharge, and risk level, for each time slice (e.g., hourly), completely recording the entire process of a virtual flood control scenario. Each trace can be seen as a different assignment of the set of variables involved in the DBN on a future timeline. Each trace is the sample mentioned in step a). The next step is to check whether each trace / sample logically satisfies a given BLTL logical formula. If it does, it is recorded as 1; otherwise, it is recorded as 0. Before the generation of simulation trace samples by the algorithm stops, the cumulative distribution function calculates the P-value as the ratio of "satisfied / total number of trace samples". In this invention, the cumulative distribution function (CDF) is used to calculate the P-value. This is achieved by using the cumulative probability of a statistical model, such as the binomial distribution, to quantify the likelihood of a deviation between the model's predicted and actual results, thus determining the model's reliability. The CDF describes the sum of probabilities that a random variable takes a value less than or equal to a certain value, directly reflecting the extreme nature of an event. For example, the CDF of the binomial distribution can calculate the probability that the number of predicted samples satisfying the BLTL formula is less than or equal to the number of actual observations. The P-value represents the probability of the observed deviation or a more extreme deviation occurring, assuming the null hypothesis that the model's prediction and the actual result are not significantly different. The smaller the P-value, the more difficult it is to explain the deviation using random error, and the more likely the model may need correction. The method for calculating the P-value in the context of flood control is as follows: The first step is to define the null hypothesis (H0) and the test objective, such as whether the number of samples predicted by the model to satisfy the BLTL formula is significantly different from the actual number of samples that satisfy it.

[0131] The second step is to determine the statistical distribution, such as binomial distribution B(300, 0.8), based on the size of the validation dataset (e.g., 300 samples) and the true satisfaction rate (e.g., R_true=80%).

[0132] The third step is to calculate the number of samples that actually meet the prediction, such as 230 groups, and then use the cumulative distribution function (CDF) to find the p-value. If we are concerned that the number of predicted satisfactions is significantly less than the actual number, the P-value = CDF(230), which is the probability that the number of predicted satisfactions is ≤230. If we are concerned about the two-way bias, the P-value is 2×min(CDF(230), 1-CDF(230)).

[0133] The fourth step is to compare the P-value with the significance level α (e.g., 0.05) to determine whether the model has passed validation.

[0134] b) The P-value is used to quantify the randomness of the deviation: when the P-value is >0.05, it indicates that the deviation is likely caused by random error and the model meets expectations. If the P-value is >0.05, H0 is accepted and the BLTL verification of the DBN model is deemed to be successful, and its time series logic prediction is consistent with the actual process. If the P-value is ≤0.05, H0 is rejected and model correction is triggered.

[0135] Step 5.6 Model correction and iterative verification, closed-loop optimization For scenarios where validation fails, the reasons for deviations are analyzed and the DBN model is optimized accordingly. The timing logic prediction accuracy of the model is improved through a "correction-validation" closed loop.

[0136] (1) Locating the cause of the deviation a) Parameter bias: If the bias is concentrated in a specific scenario, such as the BLTL formula 1 prediction not being met under heavy rainfall, check the conditional probability table of "rainfall amount → water level rise" in DBN to determine whether the water level response speed under heavy rainfall has been underestimated. b) Structural bias: If biases occur in multiple scenarios, backtrack to the causal modeling stage to check whether key variables (such as "levee permeability coefficient" or "traffic efficiency") or causal paths (such as "rainfall → soil moisture content → runoff coefficient → water level") are missing.

[0137] (2) Model correction method a) Parameter correction: Based on the true trajectory of the biased sample, the conditional probability table of DBN is updated using the maximum likelihood estimation method (e.g., the probability of water level exceeding the warning level within 6 hours when rainfall is ≥100mm / 3h is corrected from 60% to 85%). b) Structural correction: Combining expert prior knowledge, missing variables and causal edges are added, the IMaGES algorithm is re-run to construct the causal network, and then the DBN structure is updated.

[0138] (3) Iterative verification Repeat the "trajectory simulation-logic judgment-statistical inference" process for the corrected DBN model until the P-value of two consecutive rounds of verification is >0.05 and the deviation between the predicted satisfaction rate and the actual satisfaction rate is ≤8% (meeting the accuracy requirements of flood control decision-making), at which point the verification is terminated.

[0139] Step 5.7 Validation Results Output: Flood Control Decision-Making Feature Elements and Causal Decision-Making Model Output a standardized StatMC-BLTL validation report to provide clear evidence for the engineering application of the DBN model. The report should include: Verification objective: To determine the correspondence between the specific definition of the BLTL formula and flood control decision-making; Sample information: sample size, scene distribution, preprocessing methods; Statistical results: P-value, prediction satisfaction rate, percentage of biased samples and analysis of causes; Model parameters: The final validated DBN conditional probability table and time slice dependencies; Application Boundaries: The range of scenarios in which the model is applicable (e.g., applicable rainfall intensity ≤ 200mm / 24h, initial water level < warning water level 1.5m).

[0140] Output the key features and causal decision-making model for the final flood control decision.

[0141] The above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for identifying key elements and modeling causal relationships in flood control decision-making, characterized in that, Includes the following steps: Step 1: Construct a three-dimensional flood control element system, dividing multi-source flood control data into a three-dimensional element system including basic elements, target elements, and control elements, and clarifying the data source of each element. Among them, basic elements include but are not limited to water conditions, rainfall conditions, engineering conditions, danger conditions, and disaster conditions data; target elements include but are not limited to control water level, flood storage targets, and water replenishment task data; and control elements include but are not limited to dispatching routes, flood control mechanisms, and emergency plan data. Step 2: 3D Feature Data Structuring Processing. Based on the 3D feature system constructed in Step 1, structured modeling is performed on the basic features, target features, and control features to obtain a decision modeling dataset, including: Flood control situation modeling: Time series modeling of basic elements to identify data trends, extreme points and abrupt change points; Scheduling behavior modeling: The control elements are structured into a path model, a time series model, and a rule set. The path model includes fields such as scheduling step ID, execution subject, triggering condition, and execution time limit. The time series model includes time granularity and behavioral dependencies. Control target modeling: Transforming target elements into computable indicators, including but not limited to the completion rate of flood storage targets and the rate of compliance with control water levels; Step 3: Three-stage key element identification. Based on the decision modeling dataset output in Step 2, key elements are identified through a three-stage process: initial correlation screening, causal modeling, and expert verification. This includes: Preliminary correlation screening: The variables in the decision modeling dataset are preprocessed, including but not limited to using the sliding window midpoint filtering method and the 3σ criterion to remove noise and outliers for continuous data, and using one-hot encoding transformation for discrete data. Then, redundant variables are removed by using the optimal subset screening combined with the stepwise regression algorithm to construct a candidate key element set. Causal modeling: The IMaGES algorithm is used to model the candidate key element set. First, outliers are corrected and normalized by the interquartile range (IQR) method. Then, the mutual information value between variables is calculated based on the kernel density estimation method to quantify the causal association strength. The output is a causal influence matrix containing causal path, direction and strength. Expert verification: Based on the expert experience database composed of flood control physical laws and historical dispatch cases, the consistency of causal relationships in the causal influence matrix is ​​verified, and false correlation elements that are only time-synchronized and have no actual causal relationship are eliminated to obtain credible key elements; Step 4: IMaGES-DBN fusion causal modeling. Based on the credible key elements obtained in Step 3, a static causal and dynamic time series fusion model is constructed. The static causal structure between elements is determined by the IMaGES algorithm. The time dimension characteristics of the dynamic Bayesian network are incorporated to update the conditional probability and state transition relationship, and to characterize the dynamic changes of causal relationships during the flood evolution process. Step 5: Feature optimization under BLTL logic constraints. Bounded linear time-series logic (BLTL) is introduced to optimize the key elements in the fusion model obtained in Step 4 under the BLTL logic framework. Specifically, the physical laws of flood control are transformed into BLTL formulas, and feature elements that do not conform to the laws are eliminated through logical consistency verification, and the weight allocation of key elements is optimized. Step 6: StatMC Validation and Iterative Optimization. The StatMC method is used to validate the fusion model optimized in Step 5, including: Building a validation dataset: Independent samples are extracted from historical data to build a validation dataset; Trajectory simulation and logic determination: Input the initial conditions of the samples into the fusion model to generate the predicted trajectory. Determine the satisfaction of the BLTL formula based on the actual trajectory and the predicted trajectory respectively, and mark the deviation samples. Statistical testing and iteration: Statistical analysis is performed on the biased samples using binomial distribution test or t-test. The DBN conditional probability table is updated by maximum likelihood estimation or the causal path correction model is supplemented. The verification is repeated until a predetermined number of samples meet the verification target, and the final key elements and causal model are output.

2. The method for identifying key elements and modeling causal relationships in flood control decision-making according to claim 1, characterized in that: Step 3, which uses the IMaGES algorithm to model the candidate key element set, includes: Outlier identification is performed on candidate variables. The interquartile range (IQR) method is used to identify outliers, which are data that exceed a preset range. The outliers are corrected using the nearest valid value interpolation method to obtain corrected candidate variables. The modified candidate variables are then normalized. Based on the normalized candidate variables, the kernel density estimation method is used to calculate the mutual information value between the candidate variables. If the mutual information value is greater than the preset strength threshold, it is determined that the variables are strongly correlated and included in the causal path to form a causal influence matrix.

3. The method for identifying key elements and modeling causal relationships in flood control decision-making according to claim 2, characterized in that: The causal influence matrix has a dimension of M×N, where rows correspond to candidate key elements, which are input variables that directly or indirectly affect flood control decisions, including but not limited to rainfall, water level, flow velocity, flood discharge, and early warning response time; columns correspond to target elements, which are output variables that reflect the results of flood control decisions, including but not limited to flood risk level, flood control effectiveness assessment value, and disaster area prediction value. The element values ​​in the matrix are causal influence coefficients, ranging from [-1,1].

4. The method for identifying key elements and modeling causal relationships in flood control decision-making according to claim 1, characterized in that: The StatMC method described in step 6 validates the model, and the steps are as follows: Define validation objectives, including but not limited to the risk event prediction deviation rate; Construct a validation dataset: Randomly select multiple independent samples from historical data. The selected independent samples include initial conditions and observation sequences, and cover typical scheduling decision cases. Generate prediction results: Input the initial conditions of the sample into the model to be validated, and the model outputs the probability of risk events; Statistical test: Based on the risk events, a binomial distribution test is used for discrete risk events. α represents the significance level of the statistical test. The test determines whether there is a significant difference between the model's predicted satisfaction rate and the actual satisfaction rate. If the test result P value ≤ α, it indicates a significant difference, and the model needs to be corrected; if the test result P value > α, it indicates no significant difference, and the model meets expectations.

5. A system for implementing the key element identification and causal modeling method for flood control decision-making as described in claim 1, characterized in that, The system includes: The element system construction module constructs a three-dimensional element system of basic elements, target elements, and control elements based on multi-source flood control data. The data structuring module receives the three-dimensional element data output by the element system construction module and outputs a unified format decision modeling dataset through flood control situation modeling, dispatch behavior modeling, and control target modeling. The critical path identification module receives the decision modeling dataset output by the data structuring module and obtains credible key elements through a three-stage algorithm of correlation screening, causal modeling, and expert verification. The decision causal modeling module receives credible key elements output by the critical path identification module, constructs an IMaGES-DBN fusion causal model, and optimizes the weights of key elements through BLTL logical constraints. The validation and optimization module receives the fusion model output by the decision causal modeling module, validates the model using the StatMC method, iteratively optimizes it, and outputs the final key elements and causal model.

6. A system for implementing the key element identification and causal modeling method for flood control decision-making according to claim 1, as described in claim 5, is characterized in that... The element system construction module includes a basic element submodule, a target element submodule, a control element submodule, and an application scenario submodule. The basic element submodule is used to collect data on water conditions, rainfall, engineering conditions, hazards, and disasters. The target element submodule is used to set control water levels and flood storage targets. The control element submodule is used to construct scheduling paths and flood control mechanisms. The application scenario submodule is used to divide application scenarios according to watershed type, flood control stage, and engineering type.

7. A system for implementing the key element identification and causal modeling method for flood control decision-making according to claim 1, as described in claim 5, characterized in that: The data structuring module includes a flood control situation submodule, a scheduling behavior submodule, and a control target submodule. The flood control situation submodule has built-in neighbor station interpolation and sliding window weighted average algorithm units. The scheduling behavior submodule has built-in path model field generation unit and time-series dependency definition unit. The control target submodule has built-in storage target completion degree and water level compliance rate calculation unit.

8. A system for implementing the key element identification and causal modeling method for flood control decision-making according to claim 1, as described in claim 5, characterized in that: The critical path identification module includes a correlation analysis submodule, a causal modeling submodule, and an expert verification submodule. The correlation analysis submodule has built-in sliding window mean filtering, 3σ criterion, one-hot coding, optimal subset selection, and stepwise regression algorithm units. The causal modeling submodule has built-in IMaGES algorithm unit and causal influence matrix generation unit. The expert verification submodule has built-in expert experience base and consistency verification unit.

9. A system for implementing the key element identification and causal modeling method for flood control decision-making according to claim 1, as described in claim 5, characterized in that: The decision causal modeling module includes a fusion modeling submodule and a BLTL optimization submodule; the fusion modeling submodule has an IMaGES static causal structure construction unit and a DBN dynamic time-series update unit built in, and the BLTL optimization submodule has a flood control physical law transformation unit and a logical consistency verification unit built in.

10. A system for implementing the key element identification and causal modeling method for flood control decision-making according to claim 1, as described in claim 5, characterized in that: The verification and optimization module includes a sample construction submodule, a trajectory simulation submodule, a statistical testing submodule, and an iterative correction submodule. The sample construction submodule has a built-in historical data extraction and preprocessing unit, the trajectory simulation submodule has a built-in DBN inference and BLTL satisfaction determination unit, the statistical testing submodule has a built-in binomial distribution test and t-test unit, and the iterative correction submodule has a built-in conditional probability update and causal path supplementation unit.

Images