Method for dynamic early warning and emergency scheme making of real-time flood control scheduling risk of river basin

Abstract

The application discloses a kind of basin real-time flood control scheduling risk dynamic early warning and emergency scheme making method, comprising: obtaining real-time full-element information tensor and engineering flood control design parameter set, based on safety margin distribution calculation obtains real-time hierarchical risk entropy time series section, real-time data and pre-constructed entropy flux Markov graph structure data are input to the entropy perception double-flow prediction model trained to carry out reasoning, obtain future risk prediction set.Combining preset entropy grade threshold and future failure chain probability generates dynamic risk early warning result set, and accordingly, to meet safety margin constraint and entropy reduction target as basis, generates emergency scheduling scheme set.The application realizes the dynamic early warning and decision of physical-risk strong coupling.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of smart water conservancy and flood control safety, and particularly relates to a method for dynamic early warning of real-time flood control scheduling risk and emergency scheme making in a river basin. BACKGROUND

[0002] In the face of frequent extreme rainfall events caused by global climate change, the traditional scheduling mode based on experience and static rules is increasingly severe. Therefore, researching a dynamic risk early warning and emergency decision method for real-time flood control scheduling in a river basin to realize the transition from passive response to active prediction has great theoretical and practical significance for improving the flood control resilience and emergency disposal capacity of the river basin.

[0003] Currently, research on river basin flood control scheduling mainly focuses on two aspects: one is flood evolution prediction based on a hydrodynamic model, which predicts the river water level, flow and inundation range in the future period through high-precision numerical simulation (such as HEC-RAS, MIKE, etc.); and the other is risk index prediction based on a data-driven model, which uses statistical methods or conventional time series networks (such as LSTM) to fit historical monitoring data to predict physical quantities such as water level and flow at a single or a few key sections. At the same time, in terms of risk assessment, the existing methods mostly use multi-index fuzzy evaluation, failure probability statistics or scenario analysis to post-evaluate the risk consequences of the scheduling scheme.

[0004] The existing schemes have obvious deficiencies in the physical connotation of risk measurement, the spatiotemporal propagation mechanism of risk and the physical-risk coupling of the prediction model. These problems lead to lagging early warning, black-box risk evolution process and difficulty in supporting precise emergency scheduling. First, the risk measurement is decoupled from the physical state. Traditional risk is mostly defined based on historical statistical failure probability, which has limited representation ability and is difficult to correspond to the real-time safety margin (i.e. the difference between the design bearing capacity and the real-time load) of the project, a core physical state, resulting in that the risk assessment results cannot intuitively reflect the real-time safety reserve of the project. Second, there is a lack of quantitative models for the spatiotemporal coupling propagation of risk. The existing methods regard each engineering node as an independent risk source and ignore the fact that risk flows, accumulates and dissipates in the river basin topology network (for example, from upstream reservoirs to downstream embankments) like entropy. The lack of this spatiotemporal propagation mechanism makes it impossible to accurately predict the cascading failure chain of instability at one place and emergency at the whole line. Finally, flood evolution prediction is decoupled from risk state evolution. Data-driven models (such as LSTMs) are only trained to fit physical curves such as water level and flow, and their internal states do not contain any risk information. This leads to the fact that the model cannot perceive the nonlinear impact of physical processes (such as flood peaks) on risk states (such as safety margin entropy increase), nor can it use risk constraints (such as entropy conservation) to correct its physical prediction, resulting in that the prediction results are known but the reasons are not known. SUMMARY

[0005] The application aims to provide a method for real-time flood control scheduling risk dynamic early warning and emergency scheme formulation in a river basin, so as to solve one of the above problems in the prior art.

[0006] The technical scheme is a method for real-time flood control scheduling risk dynamic early warning and emergency scheme formulation in a river basin, comprising the following steps:

[0007] Obtaining a real-time full-element information tensor and an engineering flood control design parameter set; and calculating a real-time safety margin state set and a real-time hierarchical risk entropy time sequence segment based on the real-time full-element information tensor and the engineering flood control design parameter set;

[0008] Inputting the real-time full-element information tensor, the real-time safety margin state set and a pre-constructed entropy flux Markov graph structure data into a trained entropy perception double-flow prediction model, performing reasoning, and obtaining a future risk prediction set comprising a future hierarchical risk entropy prediction result set, a future entropy flux prediction result set and a future failure chain probability distribution set;

[0009] Generating a dynamic risk early warning result set in combination with the real-time hierarchical risk entropy time sequence segment, the future risk prediction set and a preset entropy level threshold;

[0010] Generating an emergency scheduling scheme set based on the dynamic risk early warning result set, the future risk prediction set and the engineering flood control design parameter set, so as to meet the safety margin constraint and the entropy reduction target.

[0011] The beneficial effects are that the physical safety margin is introduced as an anchor point for risk quantification, the relative margin between the engineering design capacity and the real-time load is calculated, a three-level risk entropy system from a single engineering, a subsystem to a system is constructed, the risk assessment has a solid physical connotation; an entropy flux Markov graph model is constructed, the isolated node risk is converted into a dynamic propagation process coupled in time and space, the quantitative prediction of the cascading failure risk evolution path is realized; an entropy perception double-flow prediction model is designed, the physical process flow and the risk structure flow are processed in parallel, the dynamic early warning and decision-making of the physical-risk strong coupling are realized. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 The figure is a whole flowchart of the application.

[0013] Figure 2 The figure is a flowchart of the application for calculating the real-time safety margin state set and the real-time hierarchical risk entropy time sequence segment.

[0014] Figure 3 The figure is a flowchart of the application for calculating the single engineering safety margin time sequence set.

[0015] Figure 4 The figure is a flowchart of the application for obtaining the single engineering safety margin interval label set. DETAILED DESCRIPTION

[0016] In order to make the objects, technical solutions and advantages of the present application clearer, the following will further describe the present application with reference to the accompanying drawings. Figure 1 to the accompanying drawings Figure 4 and specific examples. It should be understood that the specific examples herein are only used to explain the present application and not to limit the present application.

[0017] Example 1 provides a basic framework and data preparation process for implementing the method described in the present application, which provides a unified and reliable data basis for the calculation of risk entropy, the construction of the model and the real-time deduction in all subsequent examples.

[0018] In this embodiment, first, the acquisition and full-factor information tensor of the basin multi-source data are constructed. Specifically, this process reads the basic basin multi-source monitoring original data set from the data sources such as hydrological monitoring stations, rainfall stations, remote sensing platforms and engineering automation systems within the basin. The original data set is massive and heterogeneous, and exemplarily includes real-time and historical rainfall sequences, river flow sequences, water level sequences, reservoir capacity sequences, gate opening sequences, unit output sequences and related basic meteorological elements.

[0019] After obtaining the original data set, it needs to be strictly quality controlled. The quality control operation can include identifying missing values in the data sequence, removing abnormal values exceeding a reasonable threshold, diagnosing the drift of the sensor over time, and reconstructing the missing data by interpolation algorithm. After the above operation, the quality controlled monitoring data set with controlled error range is obtained. The interpolation algorithm can use linear interpolation, Kriging interpolation or machine learning model, etc.

[0020] On this basis, the unified and assimilated fusion of the space-time reference is carried out. Since data from different sources may use different coordinate systems and time standards, it must be unified. First, according to the unified national or regional coordinate system (for example, CGCS2000), the geographical coordinates of each data source are converted to the unified spatial reference to obtain the spatially aligned observation data set. Then, according to the time zone information and observation period of the monitoring equipment, the time stamps of different sources are synchronized to a unified standard time axis (for example, UTC+8) to construct a time-aligned observation data set. In order to further improve the data quality, a data assimilation method such as ensemble Kalman filter EnKF can be used to jointly estimate the time-aligned observation data set and the output of the existing basin hydrological model (for example, SWAT model or HEC-RAS model), thereby effectively correcting the observation noise and model bias, forming the assimilated basin state data set under the unified space-time reference.

[0021] The extraction of construction engineering and exposure information and node attributes are used to obtain the disaster-bearing capacity of the engineering in the basin and the social and economic vulnerability. Specifically, the assimilated basin state data set is read, and the design specifications, as-built drawings, and historical operation archives of each engineering (such as reservoirs, dikes, gates, and pumping stations) are consulted to extract key structural parameters, design flood control standards (such as the design flood peak flow or the design allowable water level corresponding to a 100-year flood), maximum flood discharge capacity, and historical scheduling rules to form an engineering flood control design parameter set describing the disaster-bearing capacity of the engineering. At the same time, according to the latest land use type data, spatialized population density distribution map, economic output (GDP) statistical data, and the like, a set of exposure and vulnerability parameters sensitive to flood consequences is constructed. Finally, the engineering flood control design parameter set and the exposure and vulnerability parameter set are associated with the corresponding engineering nodes and spatial units in the basin topology structure, respectively, to form an engineering and exposure attribute table describing the static attributes of each engineering node and basin unit.

[0022] A full-factor information tensor is constructed. The assimilated basin state data set, the engineering and exposure attribute table, and a pre-set basin spatial grid division scheme are read. The dynamic characteristics corresponding to each time step are mapped to the corresponding basin grid and engineering node; at the same time, the engineering flood control design parameter set and the exposure and vulnerability parameter set are encoded as static characteristics and also attached to the corresponding node or grid. The dynamic characteristics include rainfall, inflow, water level, reservoir capacity, gate opening, and unit output. In this way, all information is arranged and combined into a multi-dimensional array with time dimension, space dimension, and feature dimension, i.e., a full-factor information tensor, to centrally express the hydrological state, engineering state, and exposure state of the basin at each time step, which is the unified input data for the safety margin calculation, hierarchical risk entropy construction, and prediction modeling in subsequent embodiments.

[0023] Embodiment 2, the embodiment provides a general flow of a real-time flood control scheduling risk dynamic early warning and emergency scheme making method for a basin.

[0024] Step 201, obtaining a real-time full-factor information tensor and an engineering flood control design parameter set; and calculating a real-time safety margin state set and a real-time hierarchical risk entropy time sequence front segment based on the real-time full-factor information tensor and the engineering flood control design parameter set.

[0025] Specifically, the acquisition and construction of the real-time full-factor information tensor and the engineering flood control design parameter set have been described in detail in Embodiment 1. In the real-time running phase, the system will continuously receive the latest monitoring data, and update the real-time full-factor information tensor according to the data processing flow described in Embodiment 1.

[0026] On this basis, the hierarchical risk entropy construction method described in subsequent embodiments is called.

[0027] Specifically, based on the real-time full-element information tensor and the engineering flood control design parameter set, a single-project safety margin time series reflecting the relative safety level of each project is calculated; the risk entropy of a single project is defined and calculated based on the probability distribution of the safety margin; and the risk entropy of subsystems and the system are constructed level by level, combining the topological coupling relationship and exposed vulnerability between projects. The calculated safety margin-related data (e.g., safety margin values ​​and their discrete interval labels) are collected into the real-time safety margin state set, and the risk entropy sequences of each level (single project, subsystem, system) are collected into the first segment of the real-time hierarchical risk entropy time series.

[0028] Step 202: Input the real-time full-element information tensor, the real-time safety margin state set, and the pre-constructed entropy flux Markov graph structure data into the trained entropy-aware dual-flow prediction model to perform inference and obtain a future risk prediction set containing a future hierarchical risk entropy prediction result set, a future entropy flux prediction result set, and a future failure chain probability distribution set.

[0029] Specifically, the entropy flux Markov graph structure data is pre-constructed based on historical data before model training. It describes the pattern of risk entropy transmission and the probability of state transitions in the watershed topology network. The entropy-aware dual-flow prediction model that has been trained is also pre-trained.

[0030] During the real-time operation phase, the real-time full-element information tensor (reflecting the physical process) and the real-time safety margin state set (reflecting the risk state) within the current moment and the near historical window are used as inputs, while the entropy flux Markov graph structure data is loaded simultaneously. After receiving these inputs, the entropy-aware dual-stream prediction model performs a forward inference and outputs predictions for several future time steps (e.g., the next 1 hour, 3 hours, and 6 hours). These prediction results are organized into the future risk prediction set, which includes predictions of future hierarchical risk entropy (single project, subsystem, system), predictions of entropy flux on the critical path, and predictions of the probability distribution of failure of critical project combinations.

[0031] Step 203: Combine the real-time hierarchical risk entropy time series, the future risk prediction set, and the preset entropy level threshold to generate a dynamic risk warning result set.

[0032] Specifically, the future hierarchical risk entropy prediction result set and the future failure chain probability distribution set are extracted from the future risk prediction set. Then, the predicted risk entropy values are compared with the preset entropy level threshold. The entropy level threshold is determined based on historical extreme events and flood control targets of the basin, for example, the risk entropy is divided into four levels: reliable zone, alert zone, crisis zone and critical zone. Through comparison, it can be judged that the risk entropy of each time point and each engineering node in the future will be in which level. In combination with the failure risk of key projects (such as upstream reservoirs) indicated in the future failure chain probability distribution set, the warning level is corrected or confirmed. A dynamic risk warning result set including time dimension (future time) and space dimension (each engineering node and subsystem) is generated to issue clear risk prompts to dispatch personnel.

[0033] Step 204, based on the dynamic risk warning result set, the future risk prediction set and the engineering flood control design parameter set, an emergency dispatch scheme set is generated to meet the safety margin constraint and the entropy reduction target.

[0034] Specifically, when the dynamic risk warning result set indicates that the risk level is rising (for example, entering the alert zone or crisis zone), the system automatically triggers the generation of the emergency dispatch scheme. The system will first deduce a series of candidate emergency dispatch operations that meet the constraints of the engineering flood control design parameter set (for example, the reservoir cannot exceed the flood limit water level, the gate opening speed is limited, etc.), such as pre-discharge, peak shaving, flood diversion or peak shifting operation combination.

[0035] The system evaluates the effects of these candidate schemes. The evaluation is based on the following:

[0036] Safety margin constraint, that is, the dispatch operation cannot cause the safety margin of other projects to be lower than the set lower limit;

[0037] Entropy reduction target, that is, the dispatch operation should effectively reduce the predicted future risk entropy level.

[0038] The entropy flux prediction result in the future risk prediction set can be used to simulate the change trend of future entropy flux and hierarchical risk entropy after executing different candidate schemes. In this way, the optimal dispatch scheme that can achieve the maximum entropy reduction while ensuring the safety of all projects is selected, forming the final emergency dispatch scheme set for the dispatch personnel to consult and confirm and issue for execution.

[0039] Embodiment 3, this embodiment describes in detail the specific solving process of the single engineering safety margin time series set in step 201 of embodiment 2, which is the physical basis for all subsequent risk entropy calculations.

[0040] Step 301, for each engineering node, the engineering design flood control capacity thereof is extracted from the engineering flood control design parameter set.

[0041] Specifically, this step corresponds to the process of constructing the engineering flood control design parameter set in Embodiment 1. In this embodiment, for each monitored and scheduled engineering node (e.g., a reservoir dam, a section of dike, a floodgate), it is necessary to query and extract its key design flood control capacity parameters from the engineering flood control design parameter set.

[0042] The engineering design flood control capacity C_design_e is a comprehensive index used to characterize the ultimate bearing capacity of an engineering project. Its specific meaning varies depending on the type of project. For example, for reservoir projects, C_design_e can be its design flood control level (e.g., 80.0 meters) or the reservoir capacity corresponding to the check flood level; for dike projects, C_design_e can be its design allowable water level (e.g., 75.0 meters); for spillway projects, C_design_e can be its design maximum flood discharge capacity (e.g., 8000 cubic meters per second). These extracted capacity parameter values ​​are associated with the unique identifiers of the engineering nodes to form an engineering flood control capacity parameter table for subsequent calculations.

[0043] Step 302: Extract the real-time actual flood load of the corresponding engineering node from the real-time full-element information tensor.

[0044] Specifically, the real-time actual flood load L_load_e(t) is a dynamic quantity that changes with time t, used to characterize the actual load borne by the project at the current moment. This load data needs to be extracted from the real-time full-element information tensor constructed in Example 1, based on the identifier of the project node and the current time step t.

[0045] It should be noted that the real-time actual flood load L_load_e(t) is not a single observation value, but an equivalent load defined according to the engineering failure mechanism. For example:

[0046] For reservoir-type projects, L_load_e(t) may not only be the real-time water level, but also the superposition value of the real-time water level and the additional water level caused by wind and waves.

[0047] For levee projects, L_load_e(t) may be the equivalent external load calculated after comprehensively considering the real-time water level, river flow velocity and wave impact pressure.

[0048] For dam and gate projects, L_load_e(t) may be the total stress on the gate leaf and foundation calculated based on the difference in water levels between upstream and downstream and the degree of gate opening.

[0049] The refined definition of L_load_e(t) overcomes the limitation of using only a single water level or flow rate indicator in traditional methods, making the calculation of safety margin more in line with the actual stress state of the project.

[0050] At step 303, based on the engineering design flood control capacity and the real-time actual flood load, a safety margin value is calculated and obtained, and the safety margin value is collected as the single engineering safety margin time series set.

[0051] Specifically, after obtaining the engineering design flood control capacity C_design_e and the real-time actual flood load L_load_e(t) of the project e at time step t, the current safety margin value M_safe_e(t) can be calculated by a safety margin definition formula.

[0052] In this embodiment, the calculation formula is preferably in the following form:

[0053] M_safe_e(t)=(C_design_e-L_load_e(t)) / C_design_e;

[0054] The safety margin M_safe_e(t) represents the relative margin of the current load L_load_e(t) from the design capacity C_design_e.

[0055] When L_load_e(t)<C_design_e, M_safe_e(t)>0, indicating that the project is in a safe state, and the larger the value, the safer.

[0056] When L_load_e(t)=C_design_e, M_safe_e(t)=0, indicating that the project is in a critical state of design.

[0057] When L_load_e(t)>C_design_e, M_safe_e(t)<0, indicating that the project is in an overload state, i.e., the margin is negative, facing the risk of failure.

[0058] As an example, assume that the design flood control water level C_design_e of a reservoir is 80.0 meters. At time step t1, the equivalent flood load (real-time water level + wind wave addition) L_load_e(t1) is 78.5 meters. Then the safety margin at this time is:

[0059] M_safe_e(t1)=(80.0-78.5) / 80.0=1.5 / 80.0=0.01875;

[0060] At time step t2, the flood rises and the equivalent flood load L_load_e(t2) reaches 80.2 meters. The safety margin at this moment is: M_safe_e(t2) = (80.0 - 80.2) / 80.0 = -0.2 / 80.0 = -0.0025;

[0061] By repeating the above calculation for each engineering node e at all real-time time steps t, a series of safety margin values varying with time can be obtained. Organizing these values in engineering and time order, a single-engineering safety margin time series set is formed. This time series set will be directly input for the subsequent calculation of the single-engineering risk entropy.

[0062] In some optional embodiments, the calculated safety margin time series can also be corrected. For example, check if the safety margin value is within a reasonable range, for example, whether it is -0.1 to 0.5. If the value is far beyond the expected threshold, it may be caused by sensor abnormalities, at which time it can be marked as a suspected abnormal value, and corrected by interpolation of adjacent time steps or truncated to a preset boundary value, in order to enhance the robustness of the subsequent entropy calculation.

[0063] Embodiment 4, this embodiment describes how to generate a single-engineering risk entropy time series set and a single-engineering safety margin interval label set based on the single-engineering safety margin time series set calculated in embodiment 3, by estimating the safety margin interval probability distribution.

[0064] Step 401, according to the preset safety margin interval division rule, the safety margin values in the single-engineering safety margin time series set are discretized. The continuous safety margin value M_safe_e(t) is mapped to a limited, discrete interval with clear physical meaning.

[0065] The safety margin interval division rule set is pre-set, which is used to define the level of risk.

[0066] Exemplarily, the rule set can divide the value range of the safety margin M_safe_e(t) into four intervals k = 1, 2, 3, 4, for example:

[0067] Interval 1, the margin is a negative interval: M_safe_e(t) <= 0. Corresponding to the engineering has been overloaded, in a high-risk state.

[0068] Interval 2, the margin critical interval: 0 < M_safe_e(t) <= 0.02. Corresponding to the engineering margin is extremely low, close to the critical state.

[0069] Interval 3, the margin is low interval: 0.02 < M_safe_e(t) <= 0.05. Corresponding to the engineering margin is insufficient, in a state of alert.

[0070] Interval 4, sufficient margin interval: M_safe_e(t) > 0.05. Corresponding to engineering safety, in a reliable state.

[0071] These thresholds (such as 0.02, 0.05) are determined according to the requirements of flood control safety management of the basin and historical experience.

[0072] In some optional embodiments, the interval division rule can not be fixed. For example, different thresholds can be set according to the category of different projects (such as reservoirs, embankments). Or, a data-driven approach can be used, such as using statistical methods (such as percentile method) or clustering algorithms (such as K-Means algorithm) in machine learning based on historical safety margin time series set to automatically determine the optimal interval boundary, so that the interval division can better reflect the actual distribution characteristics of the data.

[0073] Step 402, at each time step, the conditional probability of the safety margin value falling into each safety margin interval is counted to obtain the safety margin interval probability distribution.

[0074] Specifically, the purpose of this step is to estimate the uncertainty of the safety margin M_safe_e(t) around the current time step t. Directly using the instantaneous value of M_safe_e(t) (such as 0.01875) is information loss, because it does not reflect whether the value is stable at 0.01875 or fluctuates sharply between 0 and 0.04.

[0075] To estimate this uncertainty, this embodiment adopts the method of sliding time window. Specifically, a sliding time window is constructed for each project node e, for example, the window length is set to W (such as W = 60, representing the past 60 time steps, for example, 60 minutes). At each current time step t, W safety margin samples of the project in the [t-W+1, t] time window are extracted.

[0076] According to the interval division rule defined in step 401, the frequency N_k of the W samples falling into each interval k = 1, 2, 3, 4 is counted. The conditional probability p_e_k(t) in this time window is obtained by normalizing the frequency:

[0077] p_e_k(t) = N_k / W;

[0078] For example, at time step t, W = 60 samples of a certain project e, 6 of which fall into interval 1, 30 of which fall into interval 2, 24 of which fall into interval 3, and 0 of which fall into interval 4. The safety margin interval probability distribution of e at time t is:

[0079] p_e_k(t)=[p_e_1(t),p_e_2(t),p_e_3(t),p_e_4(t)]=[6 / 60,30 / 60,24 / 60,0 / 60]=[0.1,0.5,0.4,0.0];

[0080] In some alternative implementations, to avoid overly extreme probability estimates (e.g., 0.0) when the sample size is small (e.g., W is small or some intervals have 0 samples), smoothing techniques can be used, such as adding a small pseudo-count (e.g., 1) to the frequency N_k of each interval, i.e., Laplace smoothing.

[0081] Step 403: Based on the probability distribution of the safety margin interval, calculate the risk entropy value of a single project using the information entropy formula, and collect them into a time series set of risk entropy for the single project.

[0082] Specifically, after obtaining the probability distribution p_e_k(t) of the safety margin interval at time step t, the definition of information entropy (Shannon entropy) is applied to quantify the uncertainty of this distribution. The formula for calculating the single-project risk entropy H_single_e(t) is: H_single_e(t) = -Σ_k(p_e_k(t) * log(p_e_k(t)));

[0083] Here, Σ_k represents the summation over all safety margin intervals k=1,2,3,4... The logarithm can be base 2 or base e; in this invention, a consistent base is sufficient. To prevent the logarithmic calculation from being meaningless when the probability p_e_k(t) is 0, it is agreed that when p_e_k(t)=0, p_e_k(t)*log(p_e_k(t))=0.

[0084] The higher the value of the risk entropy H_single_e(t), the greater the uncertainty of the safety margin (i.e., the margin value is dispersed across multiple intervals), which in itself represents a risk. Conversely, the lower the value, the more certain the safety margin falls within a single interval.

[0085] Continuing with the example from step 402, whose probability distribution is [0.1, 0.5, 0.4, 0.0], the single-project risk entropy (base 2) at that moment is calculated as follows:

[0086] H_single_e(t)=-(0.1*log2(0.1)+0.5*log2(0.5)+0.4*log2(0.4)+0*log2(0));

[0087] H_single_e(t)=-(0.1*(-3.32)+0.5*(-1.0)+0.4*(-1.32));

[0088] H_single_e(t) = -(-0.332 - 0.5 - 0.528) = 1.36 bits;

[0089] By repeating this calculation for each project e at all time steps t, the single-project risk entropy time series set is obtained.

[0090] Step 404, meanwhile, mark the safety margin interval corresponding to the safety margin value of each time step as the single-project safety margin interval label set.

[0091] Specifically, this step is a parallel marking process. In step 401, the instantaneous value of M_safe_e(t) is discretized. This step is to save the result of this discretization.

[0092] For example, at time step t, the M_safe_e(t) of a project is 0.01875. According to the rule in step 401, this value falls into interval 2 (the margin critical interval). Then the interval label of this project at this time is interval 2 (or its numerical index 2).

[0093] Collecting the interval labels of all projects at all time steps, we get the single-project safety margin interval label set. This label set will be used to construct discrete risk states and build model input features in subsequent examples.

[0094] Example 5, this example describes how to combine the single-project risk entropy time series set generated in Example 4 with the pre-constructed graph structure data to construct the subsystem risk entropy time series set.

[0095] Step 501, extract the project topological relationship and hydraulic connection from the pre-constructed entropy flux Markov graph structure data.

[0096] Specifically, the entropy flux Markov graph structure data is a data structure pre-established to describe the risk propagation network of the basin. In this step, we need to query and extract the static relationship between projects from this structure data.

[0097] The topological relationship of the project is the spatial layout and connection relationship of the project in the river system network of the basin, for example, project A is located upstream of project B, projects C and D are connected in series, projects E and F are connected in parallel for flood discharge, etc.

[0098] The hydraulic connection represents the mutual influence in water flow dynamics corresponding to the topological relationship, for example, the influence degree of the flood discharge of the upstream reservoir on the water level of the downstream embankment, the propagation time lag of the confluence of two sub-basins, etc.

[0099] This step also includes an implicit process, i.e. to divide all the engineering nodes in the basin into several subsystems according to the topological relations and hydraulic connections described above. A subsystem is a collection of a group of projects with close hydraulic connections or sharing common flood control tasks. For example, a group of reservoirs on the same main stream can be a subsystem, and a group of dikes and pumping stations around the same protected area (e.g. a city) can be another subsystem.

[0100] Step 502, based on the engineering topological relations, determine the topological weight of each engineering in the subsystem.

[0101] Specifically, after the engineering is assigned to the subsystem s, it is necessary to determine the topological weight w_topo_s_e corresponding to each engineering e in the subsystem s. This weight reflects the importance or risk contribution of the engineering e in the subsystem s.

[0102] The determination of this topological weight can be based on various factors. For example:

[0103] Based on engineering location: the weight w_topo_s_e of the engineering located at the key nodes of the subsystem (such as confluence, downstream outlet) may be higher.

[0104] Based on engineering scale: the weight w_topo_s_e of the engineering with larger design flood control capacity C_design_e and wider protection range may be higher.

[0105] Based on failure consequences: the weight w_topo_s_e of the engineering whose failure has more serious impact on the downstream or protected area may be higher.

[0106] All weights are normalized in the subsystem, i.e. Σ_ew_topo_s_e=1.

[0107] Step 503, based on the hydraulic connection, determine the coupling strength coefficient between the engineering in the subsystem.

[0108] Specifically, this step is used to quantify the mutual influence strength between the engineering pairs (e.g. e1 and e2) in the subsystem s that have direct hydraulic or scheduling coupling relations. This strength is defined as the coupling strength coefficient c_s_e1_e2.

[0109] For example, if the flood discharge operation of engineering e1 (upstream reservoir) will immediately and significantly affect the water level of engineering e2 (downstream dike), then the coupling strength coefficient c_s_e1_e2 between them is higher. If two engineering e3 and e4 are located in different tributaries of the same subsystem, the mutual influence is small, then c_s_e3_e4 is lower or 0. This coefficient can be calibrated based on hydrodynamic model simulation or historical data statistics.

[0110] Step 504, the single-engine risk entropy time series set is weighted according to the topology weight, and the coupling strength coefficient between the engines and the single-engine risk entropy difference are fused to jointly construct the subsystem risk entropy time series set.

[0111] Specifically, this step is the core calculation process of the subsystem risk entropy H_sub_s(t). In this embodiment, a calculation formula that fuses weighted average and coupling difference is preferably adopted:

[0112] H_sub_s(t)=Σ_e(w_topo_s_e*H_single_e(t))+λ_s*Σ_(e1,e2)(c_s_e1_e2*|H_single_e1(t)-H_single_e2(t)|);

[0113] Wherein: H_sub_s(t) is the risk entropy of the subsystem s at the time step t.

[0114] The first term Σ_e(w_topo_s_e*H_single_e(t)) is the topology weighted average value of the single-engine risk entropy (from embodiment 4) of all engines in the subsystem, which reflects the overall average risk level of the subsystem.

[0115] The second term is the coupling difference term. Σ_(e1,e2) represents the summation of all coupled engineering pairs (e1, e2) in the subsystem s. |H_single_e1(t)-H_single_e2(t)| is the absolute difference of the risk entropy between the engines e1 and e2. c_s_e1_e2 is the coupling strength coefficient between them. λ_s is the coupling strength adjustment coefficient of the subsystem s (a hyperparameter, used to balance the dimensions and importance of the first and second terms).

[0116] This formula not only considers the average risk of the subsystem (the first term), but also introduces the risk imbalance as an independent risk source (the second term). For example, if the entropy of an engine e1 in a subsystem is very high (such as 0.9, close to failure), and its closely coupled engine e2 has very low entropy (such as 0.1, very safe), this huge entropy difference |H_single_e1(t)-H_single_e2(t)|=0.8 itself represents the instability and potential cascade failure risk of the system. This formula quantifies it through the second term and includes it in the total entropy of the subsystem, thereby more comprehensively measuring the risk.

[0117] As an example, assume that the subsystem s contains only two engines e1 and e2. Its topology weights w_topo_s_1=0.6, w_topo_s_2=0.4. The coupling coefficient between them c_s_e1_e2=0.8. The adjustment coefficient λ_s=0.1.

[0118] At time step t, H_single_e1(t) = 1.36, H_single_e2(t) = 0.5.

[0119] Then H_sub_s(t) = (0.6*1.36 + 0.4*0.5) + 0.1*(0.8*|1.36-0.5|) ;

[0120] H_sub_s(t) = (0.816 + 0.2) + 0.1*(0.8*0.86) ;

[0121] H_sub_s(t) = 1.016 + 0.1*(0.688) = 1.016 + 0.0688 = 1.0848;

[0122] By repeating this calculation for all subsystems s and all time steps t, the subsystem risk entropy time series set is obtained.

[0123] In some alternative embodiments, the calculation of subsystem risk entropy can also take other forms. For example, a simplified way is to use only the first term, i.e. H_sub_s(t) =∑_e(w_topo_s_e*H_single_e(t)). Another alternative way is not to use linear weighting, but to use a nonlinear function, for example H_sub_s(t) = max_e(H_single_e(t)), that is, the risk of the subsystem is determined by the highest risk of its internal engineering.

[0124] Embodiment 6, this embodiment describes how to construct the system risk entropy time series, and how to integrate all intermediate results generated in the calculation process of the present application into the real-time safety margin state set and the front segment of the real-time hierarchical risk entropy time series required for the subsequent steps.

[0125] Step 601, fuse the subsystem risk entropy time series set and the engineering flood control design parameter set to construct the system risk entropy time series.

[0126] Specifically, this step is to further aggregate upwards on the basis of the subsystem risk entropy H_sub_s(t) to construct the system risk entropy H_sys(t) at the whole basin level.

[0127] This construction process needs to introduce the exposure and vulnerability parameter set and the engineering flood control design parameter set extracted in Embodiment 1 to determine the weight of each subsystem s. In this embodiment, it is preferred to set two weights for each subsystem s:

[0128] Exposure weight w_exp_s: This weight is determined based on the exposure and vulnerability parameter set, and is used to quantify the severity of consequences that can be caused by the failure of subsystem s. For example, a subsystem s1 protects important cities and a large population, and its exposure weight w_exp_s1 is high (e.g., 0.5); while another subsystem s2 protects farmland, and its exposure weight w_exp_s2 is lower (e.g., 0.1).

[0129] Control weight w_ctrl_s: This weight is determined based on the engineering flood control design parameter set, and is used to quantify the flood control ability and control importance of subsystem s. For example, it is determined according to the scheduling flexibility, engineering redundancy, and design flood control level within the subsystem. A subsystem with strong scheduling ability, high redundancy, and high flood control level may also have a corresponding control weight w_ctrl_s set (e.g., set to 0.8).

[0130] The calculation formula of system risk entropy H_sys(t) can be: H_sys(t)=Σ_s(w_exp_s*w_ctrl_s*H_sub_s(t)).

[0131] Where Σ_s represents the summation of all subsystems s in the basin. The meaning of this formula is that the total risk entropy of the basin is the weighted sum of all subsystem risk entropies H_sub_s(t), and this weighting also considers the consequences of failure of the subsystem (exposure weight) and its own importance (control weight). A subsystem with a high H_sub_s(t), a high exposure weight w_exp_s, and a high control weight w_ctrl_s will be the main contributor to the total system risk entropy H_sys(t).

[0132] By repeating this calculation at all time steps t, the system risk entropy time series can be obtained.

[0133] In some optional embodiments, the system risk entropy H_sys(t) can also be in a non-weighted manner, for example, H_sys(t)=max_s(H_sub_s(t)), that is, the total system risk is determined by the riskiest subsystem.

[0134] Step 602, combining the set of single-engineering safety margin time series and the set of single-engineering safety margin interval labels to form the set of real-time safety margin states; and collecting the set of single-engineering risk entropy time series, the set of subsystem risk entropy time series, and the set of system risk entropy time series to form the real-time hierarchical risk entropy time series front section.

[0135] Specifically, this step is a standardized data integration of all results produced by embodiments 3 to 6, so as to be input into the subsequent prediction model (such as embodiment 8) as a unified data structure.

[0136] The real-time safety margin state set is a data structure. In this structure, for each project e and each time step t, two key information are stored:

[0137] The original safety margin value M_safe_e(t) from example 3 (e.g. 0.01875);

[0138] The discrete interval label from example 4 (e.g. interval 2).

[0139] This combination allows the model to obtain both accurate continuous values and discrete features with clear risk level implications.

[0140] The real-time hierarchical risk entropy time series prefix is another data structure. Hierarchical refers to the fact that it contains the three risk levels built by the invention; prefix refers to the fact that it contains all data from the historical starting point to the current time step t, which is used as an input sequence to predict the future. This data structure collects: all single-project risk entropies H_single_e(t); all subsystem risk entropies H_sub_s(t); system risk entropy H_sys(t).

[0141] In summary, examples 3, 4, 5, and 6 fully describe the entire process of calculating the real-time safety margin state set and the real-time hierarchical risk entropy time series prefix from the original data.

[0142] Example 7, this embodiment details the specific construction method of the pre-constructed entropy flux Markov graph structure data cited in example 2. This data structure is pre-established based on historical data and basin physical characteristics before model training and real-time operation.

[0143] Step 701, abstract the basin topology and establish the graph structure. Before constructing the entropy flux Markov graph structure data, the physical structure of the basin needs to be abstracted first. Specifically, read the basin basic geographic information data and full-element information tensor obtained in example 1, and according to the river connectivity relationship, upstream and downstream water level control relationship, and engineering dispatching path, abstract the basin into a basin engineering topology graph structure data composed of engineering nodes and river section edges.

[0144] Among them, the engineering node can represent a single project (such as a reservoir dam, a critical dike, or a flood discharge gate group) or a local confluence area. The river section edge represents the hydraulic transmission path or dispatching control path between nodes. Preferably, an attribute describing its physical characteristics is added to each edge, such as average flow direction, average propagation time lag (e.g. 30 minutes) and typical flow range.

[0145] Step 702, obtaining an engineering node from an engineering flood control design parameter set, and establishing an entropy balance equation for any engineering node. This step is the theoretical basis for constructing the Markov graph.

[0146] Specifically, the present application regards the single engineering risk entropy H_single_j(t) calculated in Example 4 as a quantity that can flow and dissipate between nodes j of the topology graph, thereby establishing the transmission relationship of risk entropy between nodes.

[0147] It should be noted that the risk entropy (a concept of information theory) is analogous to a flowable physical quantity here, which is an analog modeling method constructed to solve the problem of risk spatiotemporal propagation in flood control scheduling. It is not equivalent to the energy conservation in thermodynamics or the mass conservation in fluid mechanics. Under certain assumptions (for example, the transmission relationship of risk between adjacent projects can be quantified by an entropy flux function), the model can be used as an effective approximation to describe the spatiotemporal evolution of risk.

[0148] Based on this, for any engineering node j, its entropy balance equation can be constructed as:

[0149] H_single_j(t+1)=H_single_j(t)+Σ_iF_entropy_i_j(t)-Σ_kF_entropy_j_k(t)-D_entropy_j(t);

[0150] Where H_single_j(t+1) and H_single_j(t) are the single engineering risk entropies of node j at adjacent time steps (t+1 and t) (historical sequences from Example 4). Σ_iF_entropy_i_j(t) represents the aggregated entropy flux flowing into node j from all upstream nodes i. Σ_kF_entropy_j_k(t) represents the aggregated entropy flux flowing out of node j to all downstream nodes k. D_entropy_j(t) represents the local entropy dissipation term at node j at time step t.

[0151] The local entropy dissipation term D_entropy_j(t) has a clear physical meaning, which represents the reduction of local risk uncertainty (i.e., entropy reduction) due to local proactive flood control measures (such as temporary reinforcement, opening of pump station for strong drainage) or scheduling operations (such as opening of flood diversion channel). This term can be fitted as a function of risk entropy level and scheduling operation instructions through historical data, forming an entropy dissipation parameter set.

[0152] Step 703, calibrating the entropy flux coefficient. In order to make the above entropy balance equation set (established for all nodes j) solvable, it is necessary to first calibrate the aggregated entropy flux inflow F_entropy_i_j(t).

[0153] In the present embodiment, the entropy flux from node i to node j is defined to be related not only to the risk entropy H_single_i(t) of the upstream node i, but also to the hydraulic connection (e.g. flow) and the operation of scheduling (e.g. gate) between them. A preferred formula for the entropy flux is:

[0154] F_entropy_i_j(t) = a_ij * H_single_i(t) * g_flow(Q_ij(t), U_ij(t))

[0155] where a_ij is the entropy transfer coefficient from node i to node j, reflecting the coupling strength of the hydraulic or scheduling, which is a core parameter to be calibrated by historical data. H_single_i(t) is the single-engine risk entropy of the upstream node i. g_flow(...) is a normalization function to map the actual flow Q_ij(t) on the edge i->j and the associated scheduling operation strength U_ij(t) (e.g. gate opening) to the contribution to the entropy transfer.

[0156] By reading the hierarchical risk entropy time series set and the full-element information tensor in the historical flood cases, sample periods containing typical flood processes can be screened. Using the entropy change and flow change relationship in these sample periods, the entropy transfer coefficients a_ij on all edges are fitted and estimated by least squares method or other optimization algorithms, and finally the entropy flux coefficient matrix set is formed.

[0157] Specifically, the calibration process of the entropy transfer coefficient a_ij can use the following method:

[0158] Select at least 30 complete flood processes in the historical flood cases as training samples. For each sample, extract the single-engine risk entropy change amount AH_single_j(t), the risk entropy H_single_i(t) of the upstream node, the actual flow Q_ij(t) on the edge, and the scheduling operation strength U_ij(t) of the adjacent time step.

[0159] Construct an optimization objective function, for example: minimize Σ_samples Σ_t Σ_j (AH_single_j(t) - Σ_i(a_ij * H_single_i(t) * g_flow(Q_ij(t), U_ij(t))) + D_entropy_j(t))^2;

[0160] The least square method or gradient descent algorithm is used to solve the above optimization problem to obtain the entropy transfer coefficient a_ij that minimizes the error between the predicted entropy change and the actual entropy change. In a preferred embodiment, the entropy transfer coefficient a_ij is usually in the range of 0 to 1, where the edges with strong hydraulic coupling have larger a_ij values (e.g. 0.6 to 0.9), and the edges with weak hydraulic coupling have smaller a_ij values (e.g. 0.1 to 0.3).

[0161] At step 704, the entropy increment is calculated and the entropy flux Markov transition matrix is derived.

[0162] After the entropy balance equation is established (step 702) and the entropy flux coefficients are calibrated (step 703), the theoretical entropy increment of any node j at any time step t, ΔH_single_j(t) = H_single_j(t+1) - H_single_j(t), can be calculated using historical data.

[0163] This entropy increment ΔH_single_j(t) and the current safety margin state (M_safe_j(t) from Example 6) jointly drive the evolution of the engineering node risk state. This evolution process is modeled as a Markov process at this step.

[0164] First, a finite number of discrete risk states need to be defined. These states can be defined based on the safety margin interval labels in Example 4, for example, defining sufficient margin, low margin, critical margin, and negative margin as four discrete states S1, S2, S3, S4.

[0165] The entropy flux Markov transition matrix P_j has elements P_j(m,n) representing the probability of node j transitioning from the current state Sm to the next state Sn. This probability P_j(m,n) is derived based on the entropy increment ΔH_single_j(t) and the safety margin M_safe_j(t), rather than simply historical frequency statistics.

[0166] For example, a probability function P_j(m,n) = f_markov(ΔH_single_j(t), M_safe_j(t)) can be constructed. Through historical data, the parameters of this function f_markov are solved using maximum likelihood estimation or Bayesian estimation methods. For example, when the entropy increment ΔH_single_j(t) is significantly positive (entropy increases sharply), the model-derived probability P_j(2,3) of transitioning from low margin (S2) to critical margin (S3) will be significantly increased.

[0167] This derivation is performed for all nodes j, and the set of entropy flux Markov transition matrices P_j is obtained.

[0168] Step 705, integrate the graph structure data. Finally, the data generated in the above steps is unified coded and integrated. The watershed engineering topology graph structure data (nodes and edges) generated in step 701, the calibrated entropy flux coefficient matrix set in step 703, the determined entropy dissipation parameter set in step 702, and the derived entropy flux Markov transition matrix set in step 704 are unified coded and stored.

[0169] The finally formed data packet is the pre-constructed entropy flux Markov graph structure data. The data packet will be used as an input of the prediction model in subsequent embodiments to express the propagation mode of entropy on the topology network and deduce the failure chain.

[0170] Embodiment 8, this embodiment describes the construction and training method of the trained entropy-aware double-flow prediction model for step 202 in embodiment 2.

[0171] Step 801, construct the entropy-aware double-flow network structure. The network structure is the core of the prediction model of the present application. The structure preferably includes two parallel but interactive data flows, namely a physical process flow and a risk structure flow.

[0172] The physical process flow aims to extract the dynamic characteristics of flood evolution and engineering response. Specifically, this flow takes the dynamic time series in the full-element information tensor generated in embodiment 1 as input, such as rainfall, inflow, water level, reservoir capacity, gate opening, etc. The network structure of this flow can preferably use deep learning networks that are good at processing time series, such as multi-layer stacking of long short-term memory networks LSTM, gated recurrent units GRU or time convolution networks TCN.

[0173] The risk structure flow aims to express the propagation mode of entropy on the topology network. Specifically, this flow takes the hierarchical risk entropy time series set generated in embodiment 6, the safety margin state set, and the entropy flux Markov graph structure data generated in embodiment 7 as input. Since the input naturally has a graph structure (engineering nodes and river section edges), the network structure of this flow can preferably use graph neural networks (GNN), such as graph attention networks (GAT) or graph convolution networks (GCN). These networks can capture the spatial dependence and propagation characteristics of risk entropy on the topology structure of the watershed.

[0174] Step 802, set an explicit entropy channel between the physical process flow and the risk structure flow to achieve controlled double-flow information interaction. The detailed implementation of this step is further described in embodiment 9. The purpose is to ensure that the evolution of physical processes (such as a rainstorm) can be perceived by the risk structure, while the changes in risk state (such as a sudden increase in entropy of an engineering) can also constrain the prediction of physical processes.

[0175] Step 803, training the entropy-aware dual-flow network structure with the joint loss function to obtain the trained entropy-aware dual-flow prediction model. The detailed implementation of this step is further described in Embodiment 10.

[0176] In the training phase, first, a training dataset and a validation dataset need to be constructed. Specifically, a plurality of complete flood events are selected from historical data, processed according to the methods of Embodiments 1 to 7, and corresponding input features (full-factor information tensor, hierarchical risk entropy time series front section, etc.) and supervision labels (future time step hierarchical risk entropy true value, entropy balance equation constraint, historical failure chain sample, etc.) are generated. These data pairs are divided into training set and validation set according to time.

[0177] Subsequently, an optimization algorithm (such as Adam optimizer) is configured, and the initial learning rate, batch size, and upper limit of training rounds are set.

[0178] In a preferred embodiment, the initial learning rate can be set in the range of 0.0001 to 0.001, for example 0.0005; the batch size can be set in the range of 16 to 128, for example 32 or 64; the upper limit of training rounds can be set in the range of 100 to 500 rounds. For Adam optimizer, β1 can adopt the default value of 0.9, and β2 can adopt the default value of 0.999. If an early stopping strategy is adopted, when the validation loss does not decrease for 10 to 20 consecutive rounds, the training is terminated and the current optimal model parameters are saved.

[0179] In an optional embodiment, for the time series network (such as LSTM or GRU) in the physical process flow, the hidden layer dimension can be set in the range of 64 to 256, and the number of layers can be set to 2 to 4 layers; for the graph neural network (such as GAT or GCN) in the risk structure flow, the hidden dimension can be set in the range of 32 to 128, and the number of attention heads can be set to 4 to 8.

[0180] In each training round, the training dataset is input into the entropy-aware dual-flow network in batches, forward propagation is performed, and the joint loss is calculated. Then, according to the joint loss, the network parameters are updated by backpropagation. At the same time, the validation loss is evaluated on the validation dataset regularly to monitor overfitting. When the validation loss no longer significantly decreases or the early stopping strategy is triggered, the current network parameters are saved, and the trained entropy-aware dual-flow prediction model is obtained.

[0181] Embodiment 9, as a preferred implementation of step 802 in Embodiment 8, details how to set an explicit entropy channel to achieve controlled dual-flow information interaction.

[0182] Step 901, for each node, construct an entropy channel feature vector containing risk entropy, safety margin and entropy flux indicators.

[0183] Specifically, at each time step t of model training and inference, for each engineering node e in the topology graph, a set of key risk indicators are extracted from the data of embodiments 6 and 7. The feature vector exemplarily can include:

[0184] Current single-engine risk entropy H_single_e(t);

[0185] Current single-engine safety margin M_safe_e(t);

[0186] Current subsystem risk entropy H_sub_s(t) (where e belongs to s);

[0187] Current system risk entropy H_sys(t);

[0188] Entropy increment ΔH_single_e(t) within the near history window;

[0189] Entropy flux inflow / outflow ΣF_in, ΣF_out within the near history window.

[0190] Splice these indicators to form a high-dimensional entropy channel feature vector V_entropy_e(t). This vector collectively summarizes the full-level risk state of node e at time t.

[0191] Step 902, embed the entropy channel feature vector as an independent entropy feature token into the physical process flow and the risk structure flow.

[0192] Specifically, this step is to introduce V_entropy_e(t) as a special token into the dual-flow network. In deep learning implementation, this can be regarded as an additional splicing of this V_entropy_e(t) vector in the input sequence of the physical process flow (such as LSTM) and the node features of the risk structure flow (such as GNN).

[0193] In this way, when the physical process flow is processing the feature of water level 78.5 meters, it can simultaneously see an entropy feature token that tells it the current risk entropy 1.36, which is in a critical state. Correspondingly, when the risk structure flow is processing the feature of risk entropy 1.36, it can also see the information of water level 78.5 meters from the physical process flow.

[0194] Step 903, and by configuring a limited attention mechanism in the entropy-aware dual-flow network structure, the dual-flow information interaction is constrained by the entropy feature token.

[0195] Specifically, only feature concatenation (as step 902) is a weak interaction. To achieve a controlled strong interaction, the present embodiment preferably employs a constrained attention mechanism.

[0196] In modern deep learning networks (such as Transformer or graph attention network), information interaction is achieved through attention heads. In the present embodiment, constraints are imposed on some attention heads in the network, so that they can only calculate attention weights between the entropy feature label and other node features.

[0197] For example, in the physical process flow, an entropy-guided attention head is configured, which must forcibly (or with high weight) pay attention to the entropy feature label V_entropy_e(t) when calculating the future water level.

[0198] In the risk structure flow, a physics-aware attention head is configured, which must forcibly pay attention to the features from the physical process flow (such as the predicted future flow) when calculating the future risk entropy.

[0199] Through this explicit and constrained attention design, the present invention builds a dedicated entropy channel. This channel ensures that the prediction of physical flood evolution is always constrained by the current and future risk entropy state; at the same time, the propagation prediction of risk entropy can always perceive the changes of real hydrological processes, thereby realizing the deep and controlled fusion of the two flows.

[0200] Embodiment 10, the present embodiment describes a preferred implementation of step 803 in embodiment 8, which details the specific composition and implementation of the joint loss function. The joint loss function is the optimization objective for training the entropy-aware dual-flow prediction model.

[0201] In the present embodiment, the joint loss function L_total is preferably composed of three loss terms with different physical or statistical meanings, which are combined by weights. L_total=β1*L_entropy+β2*L_flux+β3*L_chain, where β1, β2, β3 are weight coefficients for balancing the influence degree of each loss term.

[0202] Hierarchical risk entropy prediction error loss term L_entropy. This loss term is the main component of the joint loss function, which is used to quantify the deviation between the future hierarchical risk entropy prediction result set output by the model and the historical hierarchical risk entropy time series (i.e. the true label).

[0203] Specifically, this loss term L_entropy preferably uses Mean Squared Error (MSE) or its variants (such as Smooth L1 loss) to calculate. An example of its pure text calculation formula is:

[0204] L_entropy =∑_t,e (H_single_e_pred(t) - H_single_e_true(t))^2 +∑_t,s (H_sub_s_pred(t) - H_sub_s_true(t))^2 +∑_t (H_sys_pred(t) - H_sys_true(t))^2;

[0205] where H_single_e_pred(t) is the model-predicted single-engine risk entropy at future time step t, H_single_e_true(t) is the corresponding true single-engine risk entropy in the training dataset (calculated from Example 6). H_sub_s and H_sys correspond to the predicted and true values of subsystem and system, respectively.∑_t,e,∑_t,s,∑_t represent the summation over all predicted time steps and all engines, subsystems, and systems, respectively.

[0206] The role of this L_entropy loss term is to drive the model parameters to update in the direction that can more accurately regress the predicted future three-level risk entropy true values through backpropagation.

[0207] The entropy budget conservation constraint loss term L_flux. This loss term is used to penalize the deviation of the future-level risk entropy prediction result set and the future entropy flux prediction result set from the pre-established entropy budget equation set (from Example 7).

[0208] Specifically, this loss term is a Physics-informed Constraint. It first obtains the model-predicted single-engine risk entropy values H_single_j_pred(t+1) and H_single_j_pred(t) at the future adjacent time step, and the entropy flux inflow prediction value∑_iF_entropy_i_j_pred(t) and outflow prediction value∑_kF_entropy_j_k_pred(t) at the same time step, and combines the known local entropy dissipation term D_entropy_j(t) (from Example 7).

[0209] Calculate the residual of the entropy budget equation Residual_j(t):

[0210] Residual_j(t) = H_single_j_pred(t+1) - H_single_j_pred(t) - (∑_iF_entropy_i_j_pred(t) -∑_kF_entropy_j_k_pred(t) - D_entropy_j(t));

[0211] Ideally, the prediction of the model should make this residual zero. Thus, the L flux loss term is the sum of the square of the residual for all nodes j and time steps t: L flux =∑ t,j (Residual j(t)) 2;

[0212] The role of this L flux loss term is to make the model learn to predict the entropy and the flux of entropy while also having to learn to obey the conservation relationship between them on the topological network, so as to avoid the model producing prediction results that are obviously contrary to the law of risk propagation, thereby improving the generalization ability and physical consistency of the model.

[0213] The L chain loss term for failure chain consistency. This loss term is used to calibrate the consistency between the set of future failure chain probability distributions and the failure path (or historical disaster sample) deduced based on the pre-constructed entropy flux Markov graph structure data.

[0214] Specifically, this loss term L chain preferably uses Cross-Entropy loss or Kullback-Leibler Divergence to calculate. One output head of the model (as described in Embodiment 8) outputs a probability distribution vector P_pred(t), representing the probability of occurrence of each pre-defined failure chain (for example, chain 1: A dam failure, chain 2: A dam failure leading to B levee breach) at future time step t.

[0215] At the same time, a true probability distribution P_true(t) is obtained from the training data (for example, historical real disaster samples or steady-state probabilities deduced by Markov graph).

[0216] L chain is the cross-entropy between P_pred(t) and P_true(t):

[0217] L chain = -∑ t,c (P_true_c(t) * log(P_pred_c(t))); where c represents traversing all pre-defined failure chains.

[0218] The role of this L chain loss term is to ensure that the model can not only predict the total amount of risk (L_entropy), but also accurately predict the pattern of risk (i.e. the most likely way of cascading failure to break out), making the prediction results more meaningful for decision-making.

[0219] During the training of the model, the weight coefficients β1, β2, β3 are optimized on the validation set (for example, by grid search or Bayesian optimization) to obtain the optimal combination. For example, one possible setting is β1=0.6, β2=0.2, β3=0.2, indicating that the prediction accuracy of risk entropy is the main consideration, supplemented by physical constraints and failure chain constraints.

[0220] Embodiment 11, this embodiment describes how to use the model trained in embodiment 10 to make inference on real-time data to obtain the future risk prediction set in real-time operation phase.

[0221] Step 1101, real-time full-factor information tensor and engineering parameter acquisition. In the real-time flood control operation phase, the method of the present application will continue to run. The system will read the latest monitoring data stream from the monitoring systems (such as hydrological stations, rainfall stations, engineering automation systems) described in embodiment 1 in real time, and obtain the latest weather forecast data.

[0222] The system will adopt exactly the same data processing procedure as in embodiment 1, including quality control, time and space reference unification, assimilation fusion and attribute mapping, to process these new data, and combine them with the near-historical data to form a real-time full-factor information tensor that is updated in time. At the same time, the static stored engineering flood control design parameter set is loaded.

[0223] Step 1102, real-time safety margin and real-time hierarchical risk entropy calculation. This step is to prepare the current state input for the prediction model. The system will obtain the real-time full-factor information tensor in step 1101 and the engineering flood control design parameter set.

[0224] It is emphasized that this step is a calculation, not a prediction. It quickly calculates the current and near-historical period's:

[0225] Real-time single-engineering safety margin time series; real-time single-engineering risk entropy time series; real-time subsystem risk entropy time series; real-time system risk entropy time series.

[0226] These calculation results are integrated into the real-time safety margin state set and the real-time hierarchical risk entropy time series front segment defined in embodiment 6.

[0227] Step 1103, entropy-aware dual-flow model forward inference. This is the core of this embodiment, that is, to perform real-time prediction of the model.

[0228] Load three key data:

[0229] The entropy-aware dual-flow prediction model trained in embodiment 10 (including all its network parameters);

[0230] The entropy flux Markov graph structure data pre-constructed in embodiment 7 (including topology and transition matrix);

[0231] The real-time input data generated in step 1102 (real-time safety margin state set and real-time hierarchical risk entropy time series front segment).

[0232] The system packs the real-time input data into the required input format for the model, according to the entropy channel construction defined in Example 9.

[0233] A forward pass is performed. That is, the real-time input data is fed into the trained prediction model, and the model directly calculates the output results according to the parameters it has learned, without backpropagation and parameter updating. This calculation process outputs the results of the three prediction heads defined in Example 8:

[0234] A future hierarchical risk entropy prediction result set (for example, H_single_e, H_sub_s, H_sys prediction values every 15 minutes within the next 6 hours);

[0235] A future entropy flux prediction result set (for example, F_entropy_i_j prediction values on the critical path within the next 6 hours);

[0236] A future failure chain probability distribution set (for example, probabilities of P(chain1), P(chain2), etc. within the next 6 hours).

[0237] The system time-aligns and calibrates these three sets of prediction results (for example, to specific prediction timestamps), and encapsulates them into a data structure, namely a future risk prediction set. This prediction set will be directly input into the downstream warning and decision-making steps.

[0238] In some optional embodiments, in order to evaluate the reliability of the prediction results, a prediction uncertainty evaluation can be added in this step. For example, the Monte Carlo Dropout technique is used to run the forward propagation multiple times (for example, 30 times) and retain Dropout during the inference phase, so as to obtain a distribution of a series of prediction results, according to which the confidence interval (for example, 95% confidence interval) of each prediction value (for example, H_sub_s_pred(t+1)) can be calculated, and the uncertainty evaluation result is output together.

[0239] If it is detected that there is missing or abnormal data in the real-time input data, for example, if the data interruption of a certain monitoring station exceeds the preset time length, such as 30 minutes, the system can use one of the following strategies: (1) fill in with the latest valid data; (2) interpolate and estimate using adjacent monitoring station data; (3) mark "data is incomplete, the prediction result is for reference only" in the warning result.

[0240] If an abnormality occurs in the model inference process, for example, numerical overflow or model output exceeding a reasonable range, the system can trigger a backup plan: (1) fall back to the valid prediction result at the last time; (2) switch to a simplified rule-based warning mechanism; (3) send a system abnormality alarm to the dispatcher, prompting manual intervention.

[0241] Embodiment 12. This embodiment describes how to generate a dynamic risk warning result set, in combination with the future risk prediction set generated in Embodiment 11 and the preset threshold.

[0242] Step 1201, extract future risk prediction data. The system first reads the future risk prediction set generated in Embodiment 11. From this data set, two key pieces of information are extracted for this step:

[0243] a future hierarchical risk entropy prediction result set (e.g., the H_single_e_pred(t) and H_sub_s_pred(t) sequences for all projects e and subsystems s in the next 6 hours);

[0244] a future failure chain probability distribution set (e.g., the P_pred_c(t) sequence for the next 6 hours).

[0245] The future entropy flux prediction result set is mainly reserved for Embodiment 13 for scenario evaluation.

[0246] Step 1202, preliminary risk level discrimination based on entropy level thresholds. The system compares the future hierarchical risk entropy prediction result set with the preset entropy level thresholds.

[0247] The entropy level thresholds are determined in advance based on historical extreme events and basin flood control safety management requirements. These thresholds map continuous risk entropy values into discrete, easily understood warning levels for dispatch personnel. Exemplarily, a set of thresholds for the subsystem risk entropy H_sub_s can be defined as follows:

[0248] Level 4 (reliable zone): H_sub_s < 0.5;

[0249] Level 3 (alert zone): 0.5 <= H_sub_s < 1.0;

[0250] Level 2 (crisis zone): 1.0 <= H_sub_s < 1.5;

[0251] Level 1 (critical zone): H_sub_s >= 1.5;

[0252] The system compares the predicted risk entropy H_pred(t) of each project node e and subsystem s at each future time step t (e.g., t+1h, t+2h,...) with the above thresholds.

[0253] For example, if the model predicts that a subsystem s1 has a risk entropy H_sub_s1_pred(t+2h) = 1.15 in 2 hours, the preliminary risk level of this subsystem at that time is judged to be level 2 (crisis zone).

[0254] Step 1203 involves correcting and confirming the risk level by incorporating the failure chain probability. This step is crucial to the invention, as it avoids potential misjudgments or insufficient information arising from relying solely on entropy thresholds. The system will use the future failure chain probability distribution set extracted in step 1201 to correct or confirm the preliminary judgment result from step 1202.

[0255] Specifically, the combination methods include, but are not limited to:

[0256] Risk Confirmation and Escalation: Assume that step 1202 identifies subsystem s1 as a crisis zone. Simultaneously, the system discovers in the failure chain probability set that the predicted probability P_pred_c(t+2h) of the critical failure chain related to s1 (e.g., failure of dam A within s1) is also significantly increased (e.g., >30%). This strongly confirms the crisis zone identification, and the warning result will be marked as a crisis zone (high risk: dam A failure chain).

[0257] Risk downgrading or identification: Suppose step 1202 determines that subsystem s2 is a crisis zone, but the failure chain probability set shows that the probabilities of all known critical failure chains related to s2 are low (e.g., <5%). This may indicate that although the system uncertainty (entropy) is high, no clear, high-probability failure mode has formed. The warning result may be marked as a crisis zone, i.e., high uncertainty, but no clear failure chain, prompting dispatchers to conduct a comprehensive investigation rather than targeting a specific path.

[0258] Risk Coverage and Supplementation: Assume that the discrimination subsystem s3 in step 1202 is only in the alert zone (with low entropy), but the failure chain probability set shows that the probability P_pred_c(t+2h) of a black swan failure chain (e.g., the cascading failure of reservoirs B and C) has increased abnormally and slightly (e.g., from 0.1% to 2%). The system can specifically address this situation by adding a special attention marker to the alert zone's warning: an abnormal probability marker for the BC cascading failure chain.

[0259] Through the above combined judgment, the system finally forms the dynamic risk warning result set. This result set is a multi-dimensional information table that not only indicates the warning level (such as crisis zone) for each future time step and each project node, but may also include the specific failure mode that leads to that level (such as the failure chain of Dam A), thus providing dispatchers with richer and more interpretable warning information.

[0260] Example 13: This example describes how to generate an emergency dispatch scheme set based on early warning results and prediction sets, with specific constraints and objectives.

[0261] Step 1301: In response to the dynamic risk warning results, simulate candidate emergency dispatch operations.

[0262] In particular, the triggering condition of this step is that the dynamic risk warning result set generated in embodiment 12 indicates an increase in risk level, for example, a certain subsystem or a certain piece of subsystem will enter the alert zone, crisis zone or critical zone within a future time step (for example, 3 hours later).

[0263] Once triggered, the system will read three key input data: 1) the dynamic risk warning result set generated in embodiment 12 (for determining the severity, location and possible failure mode of the risk); 2) the future risk prediction set generated in embodiment 11 (in particular, the predicted sequence of future entropy flux and future hierarchical risk entropy); 3) the engineering flood control design parameter set constructed in embodiment 1 (for explicitly defining the physical constraints of operation).

[0264] Based on the above inputs, the system begins to deduce candidate emergency dispatch operations. These operations must first meet the hard constraints in the engineering flood control design parameter set. For example, the pre-discharge operation of a reservoir, whose discharge flow cannot exceed the safe discharge capacity of the downstream river channel or its own structural limit; the opening speed of the gate must be within its physical limit; the dispatch instruction must comply with the basic provisions in the flood control plan of the basin.

[0265] The candidate emergency dispatch operation can be one or a sequence. For example: within 0-2 hours, the discharge flow of A reservoir is increased to 500 cubic meters per second; or scheme B: within 1-3 hours, B gate is opened to divert flood, while C pump station is operated at full load. These candidate schemes can be matched from a pre-defined dispatch rule library, or obtained from a historical successful dispatch case library through case-based reasoning, or initially generated through optimization algorithms (such as genetic algorithm or particle swarm algorithm).

[0266] Step 1302, evaluate the candidate emergency dispatch operation to screen the emergency dispatch scheme set.

[0267] In particular, this step is to evaluate, sort and screen the multiple (for example, N) candidate dispatch operation schemes generated in step 1301. The core basis of evaluation is two objectives:

[0268] Meet the safety margin constraint: This is a bottom line or hard constraint objective. That is, the execution of a dispatch scheme (such as A reservoir pre-discharge) cannot be at the expense of the safety of another project (such as C embankment). The system will evaluate whether the safety margin M_safe_e(t) (its definition and calculation are shown in embodiment 3) of all related projects in the basin within the future predicted time after the execution of the scheme remains above a preset safety lower limit (for example, M_safe_e(t)>0.01). Any scheme that causes the safety margin of any project to bottom out will be directly eliminated.

[0269] Entropy reduction objective: This is an optimization or soft constraint objective. That is, the dispatching scheme should reduce the overall risk uncertainty of the system in the future most effectively, given that all safety margin constraints are satisfied. This entropy reduction can be quantified by multiple metrics, for example, the maximum reduction of the peak value of the system risk entropy H_sys(t), or the maximum reduction of the integral value of the system risk entropy H_sys(t) over the whole warning period.

[0270] To achieve the above evaluation, the present application preferably utilizes the future risk prediction set in Example 11 and the trained prediction model in Example 10. The system simulates the execution of each candidate scheme (e.g., scheme i).

[0271] A preferred simulation method is to take scheme i (e.g., the new discharge process of A reservoir at t+1, t+2... time) as a new set of operation features, update the real-time full-element information tensor in Example 11, and then call the entropy-aware double-flow prediction model again for a forward inference. The model will output a new set of future hierarchical risk entropy prediction results and future entropy flux prediction results under the intervention of scheme i.

[0272] The system then analyzes this new set of prediction results. First, check whether the safety margin of all projects (which can be indirectly reflected by the new entropy prediction or calculated by the model) meets the constraint (objective 1). Then, calculate the total entropy reduction brought by this scheme i compared to not taking any action (i.e., the original prediction) (objective 2).

[0273] For example, scheme i causes the peak value of the system total entropy H_sys(t) to decrease by 20% in the next 6 hours, and all project safety margins meet the constraints. Scheme j only reduces the entropy by 5% and causes the safety margin of D dike to be below the lower limit.

[0274] In this way, the system ranks N candidate schemes, eliminates schemes that do not meet the safety margin (such as scheme j), and selects the best one or more (k) schemes in terms of entropy reduction effect among the schemes that meet the conditions (such as scheme i).

[0275] Organize these k preferred schemes into the set of emergency dispatching schemes. This scheme set can be further visualized and displayed, for example, in the form of dispatching instruction sheets, risk entropy evolution comparison graphs under different schemes, etc., to the flood control dispatching personnel of the basin, as a scientific decision-making basis for their consultation, confirmation, and final issuance of execution, thereby completing the complete closed loop from data to warning to decision.

[0276] Suppose a river basin contains 3 reservoirs (A1, A2, A3), 2 embankments (B1, B2) and 1 flood diversion gate (C1), where A1 is located at the uppermost, A2 and A3 are in parallel at the middle reaches, B1 and B2 are in series at the downstream, and C1 is used for flood diversion to protect the downstream city.

[0277] At 14:00 on a certain day, heavy rainfall occurs at the upper reaches, with a forecast of 120 mm of cumulative rainfall in the next 6 hours.

[0278] Step 1401, data acquisition and full-element tensor construction. The system acquires real-time data from the hydrological monitoring station: A1 reservoir current water level 78.5m (design flood control water level 80.0m), inflow 800m³ / s; A2 and A3 reservoir water levels are 65.2m and 66.8m respectively; B1 embankment monitoring water level 12.3m (design allowable water level 15.0m). The system constructs the full-element information tensor at the current time.

[0279] Step 1402, safety margin and risk entropy calculation. According to the method of Example 3, the safety margin of each project is calculated:

[0280] A1 reservoir: M_safe_A1 = (80.0-78.5) / 80.0 = 0.01875 (margin critical interval);

[0281] B1 embankment: M_safe_B1 = (15.0-12.3) / 15.0 = 0.18 (margin sufficient interval).

[0282] According to the method of Example 4, based on the safety margin sample distribution of the past 60 minutes, the single-project risk entropy is calculated:

[0283] H_single_A1(t) = 1.36 bits (higher, with risk);

[0284] H_single_B1(t) = 0.32 bits (lower, temporarily safe).

[0285] Step 1403, model inference. The current state is input into the entropy-aware double-flow prediction model, and the model outputs the prediction results for the next 6 hours:

[0286] The predicted risk entropy of A1 reservoir at t+3h will reach 1.85 bits (entering crisis zone);

[0287] The probability of "A1 failure → B1 collapse" failure chain is 35%.

[0288] Step 1404, warning generation. According to the preset threshold (crisis zone ≥ 1.0 bits), the system generates a warning:

[0289] "Warning: A1 reservoir is expected to enter crisis zone in 3 hours, associated failure chain probability 35%, recommend taking emergency measures immediately."

[0290] Step 1405, emergency scheme generation. The system evaluates candidate schemes:

[0291] Scheme 1: A1 reservoir immediately pre-discharge to 500 m³ / s → predicted entropy reduction to 1.2 bits, but B1 margin drops to 0.12 (still meets constraints);

[0292] Scheme 2: Open C1 floodgate → predicted entropy reduction to 1.0 bits, B1 margin remains 0.15.

[0293] The system recommends scheme 2 because it has better entropy reduction effect and less impact on downstream.

[0294] In summary, the invention introduces physical safety margin as an anchor point for risk quantification. By first calculating the relative margin between engineering design capacity and real-time load, and then calculating the information entropy of the probability distribution of this safety margin, a three-level risk entropy system from single engineering, subsystem to system is constructed. This way makes risk entropy directly linked to the physical safety reserve of the project, solving the problem of traditional failure probability measurement being abstract and not intuitive, making risk assessment have a solid physical connotation.

[0295] In addition, an entropy flux Markov graph model is constructed. By establishing an analogous entropy balance equation, risk entropy is regarded as a kind of flux that can flow and dissipate in the topological network of the basin, and the Markov transition matrix is derived based on this entropy balance relationship. This successfully transforms isolated node risk into a spatiotemporal coupled dynamic propagation process, realizing the quantitative prediction of the evolution path of cascading failure risk.

[0296] An entropy-aware double-flow prediction model is designed. This model processes flood and entropy information through parallel processing of physical process flow and risk structure flow, and through explicit entropy channels and restricted attention mechanisms, it forces deep interaction and constraints between physical prediction (such as flow) and risk state (such as entropy increase). At the same time, the joint loss function containing entropy balance conservation ensures that the model prediction must meet both physical laws and risk propagation laws from the training target, achieving a strong coupling prediction effect of physics and risk.

Claims

1. A method for real-time flood control scheduling risk dynamic early warning and emergency plan formulation in a river basin, characterized in that, include: Acquire real-time full-element information tensors and engineering flood control design parameter sets; Based on this, the real-time safety margin state set and the first segment of the real-time hierarchical risk entropy time series are calculated. The real-time full-element information tensor, the real-time safety margin state set, and the pre-constructed entropy flux Markov graph structure data are input into the trained entropy-aware dual-flow prediction model to perform inference and obtain a future risk prediction set that includes the future hierarchical risk entropy prediction result set, the future entropy flux prediction result set, and the future failure chain probability distribution set. By combining the real-time hierarchical risk entropy time series, the future risk prediction set, and the preset entropy level threshold, a dynamic risk warning result set is generated. Based on the dynamic risk early warning result set, the future risk prediction set, and the engineering flood control design parameter set, and in accordance with the safety margin constraints and entropy reduction objectives, an emergency dispatch scheme set is generated. Based on this, the real-time safety margin state set and the first segment of the real-time hierarchical risk entropy time series are calculated, including: For the real-time full-element information tensor and the set of flood control design parameters of the project, the time series set of safety margin of a single project is obtained by solving the problem. Based on the time series set of safety margin of a single project, the risk entropy time series set of a single project is generated by estimating the probability distribution of the safety margin interval, and the label set of the safety margin interval of a single project is obtained. By combining the risk entropy time series set of a single project with the pre-constructed entropy flux Markov diagram structure data, a risk entropy time series set of the subsystem is constructed. By integrating the risk entropy time series set of the fusion subsystem with the flood control design parameter set of the engineering, a system risk entropy time series is constructed. Combine the single-project safety margin time series set with the single-project safety margin interval label set to form a real-time safety margin status set; and gather the single-project risk entropy time series set, the subsystem risk entropy time series set, and the system risk entropy time series to form the front end of the real-time hierarchical risk entropy time series. Based on the real-time full-element information tensor and the engineering flood control design parameter set, the time series set of safety margin for a single project is calculated, including: For each engineering node, the engineering design flood control capacity is extracted from the engineering flood control design parameters. Extract the real-time actual flood load of the corresponding engineering node from the real-time full-element information tensor; Based on the flood control capacity of the engineering design and the real-time actual flood load, the safety margin value is calculated and collected into a time series set of safety margin for a single project. Based on the time series set of safety margin for a single project, a time series set of risk entropy for a single project is generated by estimating the probability distribution of the safety margin interval, and a label set of safety margin intervals for the single project is obtained, including: Based on the preset safety margin interval division rules, the safety margin values ​​in the single project safety margin time series set are discretized. At each time step, the conditional probability of the safety margin value falling into each safety margin interval is calculated to obtain the probability distribution of the safety margin interval. Based on the probability distribution of the safety margin interval, the risk entropy value of a single project is calculated using the information entropy formula and then compiled into a time series set of risk entropy for a single project. At the same time, the safety margin interval corresponding to the safety margin value at each time step is marked as a single project safety margin interval label set.

2. The method according to claim 1, characterized in that, By combining the single-project risk entropy time series set with pre-constructed entropy flux Markov graph structure data, a subsystem risk entropy time series set is constructed, including: Extract engineering topology and hydraulic connections from pre-constructed entropy flux Markov graph structure data; Based on the engineering topology relationships, determine the topological weight of each project within the subsystem; Based on hydraulic connections, determine the coupling strength coefficients between engineering components within the subsystem; The risk entropy time series of individual projects is weighted according to topological weights, and the coupling strength coefficient between projects and the risk entropy difference of individual projects are combined to jointly construct the risk entropy time series of subsystems.

3. The method according to claim 1, characterized in that, The pre-constructed entropy flux Markov graph structure data is obtained through a method that includes the following steps: For engineering nodes within the basin, an entropy budget equation set is established. The entropy budget equation set quantifies the relationship between the time change of risk entropy of a single project, upstream entropy flux inflow, downstream entropy flux outflow and local entropy dissipation. Based on the entropy budget equations, the entropy increment of each engineering node is calculated. Based on the entropy increment and the corresponding safety margin state, the entropy flux Markov transition matrix used to describe the evolution of node risk states is derived.

4. The method according to claim 3, characterized in that, Establish a system of entropy budget equations, including: For any engineering node, construct its single-engineering risk entropy in subsequent time steps; the single-engineering risk entropy in subsequent time steps is jointly determined by its single-engineering risk entropy in the current time step, the aggregated entropy flux inflow from all upstream nodes, the aggregated entropy flux outflow from all downstream nodes, and the local entropy dissipation term.

5. The method according to claim 1, characterized in that, The successfully trained entropy-aware dual-stream prediction model is obtained through a training method that includes the following steps: An entropy-aware dual-flow network structure is constructed, which includes a physical process flow for extracting flood evolution characteristics and a risk structure flow for expressing the entropy topology propagation pattern. An explicit entropy channel is set up between the physical process flow and the risk structure flow to achieve controlled two-flow information interaction; A joint loss function is used to train the entropy-aware two-stream network structure to obtain the trained entropy-aware two-stream prediction model.

6. The method according to claim 5, characterized in that, Setting up explicit entropy channels to achieve controlled two-stream information exchange includes: For each node, construct an entropy channel feature vector containing risk entropy, safety margin, and entropy flux indicators; The entropy channel feature vector is used as an independent entropy feature label and embedded into the physical process flow and risk structure flow; Furthermore, by configuring a restricted attention mechanism in the entropy-aware two-stream network structure, the interaction of two-stream information is constrained by entropy feature labels.

7. The method according to claim 5, characterized in that, The joint loss function includes: The hierarchical risk entropy prediction error loss term is used to quantify the deviation between the future hierarchical risk entropy prediction result set output by the model and the historical hierarchical risk entropy time series. The entropy balance constraint loss term is used to penalize deviations between the future hierarchical risk entropy prediction result set and the future entropy flux prediction result set and the pre-established entropy balance equation set. The failure chain consistency loss term is used to calibrate the consistency between the future failure chain probability distribution set and the failure paths inferred based on the pre-built entropy flux Markov graph structure data.

Images