Method for identifying risk sources and generating scenario set of reservoir group flood resource utilization

Abstract

The application discloses a reservoir group flood resource utilization risk source identification and scene set generation method. It comprises the following steps: constructing a risk condition hydrology-risk joint manifold, taking the target risk level as the front input condition, using a parameter mapping model to automatically generate a flood process, and through simulation deviation feedback closed loop iteration correction manifold parameters to generate a sample set with controllable risk distribution; inputting the sample set into a risk topology self-organizing mapping network, clustering based on hydrology-risk joint distance measurement and risk sensitive neighborhood rules, and constructing a hierarchical scene tree with risk level semantics; based on the scene tree, multi-stage risk budget and discount parameter optimization under the fairness constraint are carried out to generate a risk control scene set and identify the risk source. The application solves the problems of uncontrollable generation process, easy loss of high risk scene and uneven stage risk allocation in the traditional method, and realizes efficient scene generation and scientific scheduling in a risk-oriented manner.

Description

Technical Field

[0001] This invention pertains to flood risk identification technology, particularly a method for identifying risk sources and generating scenario sets for flood resource utilization in reservoir groups. Background Technology

[0002] Utilizing flood resources from reservoir groups is a key means of alleviating the conflict between water scarcity and flood control safety. Through scientific dynamic control of water levels during the flood season and pre-storage and pre-release scheduling, the beneficial potential of flood resources can be maximized while ensuring flood control safety. However, the utilization of flood resources involves a balance between risk and benefit. As the scale of reservoir groups expands and scheduling rules become more complex, potential risk sources (such as insufficient pre-release capacity and misjudgment of the timing of recharge) exhibit characteristics of high concealment and complex transmission paths. Therefore, accurately identifying risk sources and constructing a highly representative set of flood scenarios is the fundamental guarantee for formulating robust scheduling strategies and maximizing the benefits of reservoir groups.

[0003] Currently, in the field of flood scenario generation and risk analysis, mainstream methods typically adopt a pattern of first generating scenarios randomly and then conducting simulation evaluation. Specifically, this often relies on historical hydrological data, using Copula functions to construct multidimensional joint distributions or employing Monte Carlo methods for random simulation to generate a large number of flood processes that conform to historical statistical patterns. These flood processes are then input into a scheduling model for simulation to calculate risk indicators. Regarding scenario reduction and clustering, existing technologies widely employ K-means clustering or standard self-organizing map (SOM) networks. These methods primarily classify a large number of samples based on the Euclidean distance of hydrological characteristics such as flood peak and flood volume, selecting the sample closest to the center as a typical scenario. This forms the basis for constructing a scenario tree or representative scenario set, which is then used for subsequent stochastic programming or robust optimization solutions.

[0004] However, the aforementioned existing technologies still have some problems when facing risk management of complex reservoir groups: the generation mechanism has risk blind spots and lacks the ability to proactively construct extreme risk scenarios. Existing methods are based on blind generation according to hydrological statistical laws, with risk only serving as a post-event evaluation label. This results in a small number of truly destructive high-risk samples among the tens of thousands of generated samples, making it difficult to cover the long tail of risk distribution and failing to meet the need for targeted stress testing of certain high-risk levels (such as extreme risks). In addition, there is a hydrological-risk topology mismatch problem in the scenario reduction process. Traditional clustering algorithms only maintain the topological structure of the hydrological feature space, ignoring nonlinear features that are similar in shape but have different risks or similar in risk but different in shape. This causes high-risk scenarios to be easily averaged or submerged by a large number of medium- and low-risk scenarios during the clustering process, and the constructed scenario tree cannot truly reflect the hierarchical structure of risks, resulting in the loss of key risk information. The scheduling optimization lacks stage risk fairness constraints, leading to an imbalance in the allocation of risk responsibility. Existing optimization methods often focus on minimizing total risk throughout the entire period, neglecting the cumulative differences in risk across different scheduling stages, such as pre-storage without rain, pre-release for beneficial purposes, and pre-release for flood control. This often leads to optimized strategies that concentrate excessive risk pressure on a particular stage, such as overdrawing flood control capacity in exchange for beneficial benefits, making the scheduling plan unfeasible in practice due to violations of the stage's safety baseline. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying risk sources and generating scene sets for the utilization of flood resources in reservoir groups, in order to solve one of the problems existing in the prior art.

[0006] According to one aspect of this application, a method for identifying risk sources and generating scene sets for flood resource utilization in reservoir groups includes:

[0007] Based on basic hydrological operation data, a risk-condition hydrological-risk joint manifold is constructed. By using the preset target risk level as the input condition for morphological transformation, the risk-condition hydrological-risk joint manifold is iteratively corrected using the pre-calculated simulation risk deviation, thereby generating a risk-condition flood process sample set.

[0008] Input the sample set of flood process under risk conditions into the risk topology self-organizing mapping network, perform topology mapping based on hydrology-risk joint distance metric and risk-sensitive neighborhood update rule, and construct a risk hierarchical scenario tree;

[0009] Based on a risk-layered scenario tree, discount parameters are optimized under the constraints of multi-stage risk budgets. Scheduling strategy combinations that meet the stage risk fairness requirements are selected, a risk control scenario set is generated, and risk attribution analysis is performed to identify risk sources.

[0010] Beneficial effects: This invention solves the problems of uncontrollable generation process, easy loss of high-risk scenes and uneven distribution of stage risks in traditional methods, and realizes risk-oriented efficient scene generation and scientific scheduling. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall process of the method for identifying risk sources and generating scene sets for the utilization of flood resources in reservoir groups, provided in an embodiment of this application.

[0012] Figure 2 This is a schematic diagram illustrating the process of generating a sample set of flood processes under risk conditions, provided in an embodiment of this application.

[0013] Figure 3 This is a schematic diagram illustrating the process of constructing a risk feedback-driven manifold correction closed loop and performing physical simulation verification and adaptive correction on a sample set of flood processes under risk conditions, as provided in the embodiments of this application.

[0014] Figure 4 This is a schematic diagram illustrating the process of constructing a risk-layered scenario tree, as provided in an embodiment of this application.

[0015] Figure 5 This is a schematic diagram illustrating the principle of a risk topology self-organizing mapping network provided in an embodiment of this application.

[0016] Figure 6 Risk distribution diagrams under different combinations of discount parameter variations provided in embodiments of this application.

[0017] Figure 7 This is a schematic diagram of an optional scheme for a method of identifying risk sources and generating scene sets for the utilization of flood resources in a reservoir group, provided in an embodiment of this application. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0019] Example 1 describes a method for identifying risk sources and generating scenario sets for flood resource utilization in reservoir groups based on risk pre-generation and feedback closed loop, such as... Figure 1 As shown, by constructing a risk-condition hydrological-risk joint manifold, the active generation from the target risk level to the flood process is realized, and a risk control scenario set is constructed through risk topology mapping and multi-stage optimization.

[0020] Step 101: Based on basic hydrological operation data, construct a risk-condition hydrological-risk joint manifold. By using the preset target risk level as the input condition for morphological transformation, and using the pre-calculated simulation risk deviation to perform iterative correction on the risk-condition hydrological-risk joint manifold, a risk-condition flood process sample set is generated.

[0021] In this embodiment, the basic hydrological operational data is not merely a simple numerical record, but rather a multidimensional time-series dataset that has undergone cleaning and standardization. Specifically, this dataset includes historical rainfall data, historical inflow data, historical water level data, historical outflow data, and historical scheduling operation records. After missing value imputation, outlier removal, and time benchmark alignment, the above data forms a standardized sequence with a consistent structure. For flood event identification, the system jointly determines the event based on the inflow threshold and the water level rise rate. For example, when the inflow exceeds a set flow threshold and the water level rise rate exceeds a set rate threshold, it is determined as the start of a flood; when the flow drops to a termination threshold, it is determined as the end of a flood. Based on this rule, multiple flood events are extracted from the long sequence and further divided into multiple operational stages, such as a rainless pre-storage stage, a beneficial pre-release stage, and a flood control pre-release stage, in conjunction with scheduling rules, providing fine-grained sample support for subsequent modeling.

[0022] A risk-conditional hydrological-risk joint manifold refers to a mathematical structure established in the feature space that describes the nonlinear mapping relationship between hydrological morphology, risk level, and morphological transformation parameters. Unlike traditional methods that first generate a flood process and then calculate the risk in a posterior model, this step explicitly uses the target risk level as the input condition for the morphological transformation model. The system can directly deduce the corresponding morphological transformation parameters based on a preset risk target, such as generating a flood process with a high risk level, thus generating a flood process that conforms to that risk characteristic.

[0023] To ensure that the generated flood events accurately reflect the expected risk levels in actual physical simulations, this step introduces an iterative correction mechanism based on simulation risk bias. Specifically, the generated flood events are input into the reservoir group scheduling simulation model to calculate their actual risk indicators (such as the probability of exceeding flood limits and tail losses), and compared with the target risk level. If systematic biases exist, such as the generated high-risk samples showing lower risk in actual simulations, the parameters of the joint manifold, such as the risk level boundary or the template parameters of the morphological transformation, are adjusted through a feedback loop until the bias converges. The final generated sample set of risk-conditional flood events is not only reasonable in terms of hydrological statistical characteristics, but also accurately covers all levels from low to high in terms of risk distribution, with a significantly enhanced coverage of extreme high-risk scenarios.

[0024] In some alternative implementations, the basic hydrological operational data may also include soil moisture content data within the watershed, meteorological forecast data, and engineering characteristic parameters of the reservoir group (such as reservoir capacity curves and discharge capacity curves). Auxiliary data helps to more accurately construct physical constraints and improve the physical reliability of the generation process. The iterative correction process can employ automated optimization algorithms, such as gradient-based adjustment or heuristic search algorithms, to improve correction efficiency.

[0025] Step 102: Input the sample set of flood process under risk conditions into the risk topology self-organizing mapping network, perform topology mapping based on the hydrology-risk joint distance metric and risk-sensitive neighborhood update rules, and construct a risk-layered scenario tree.

[0026] The Risk Topology Self-Organizing Map (SOM) network is an improvement on the traditional SOM network. Traditional SOMs typically cluster samples solely based on the Euclidean distance in the feature space (i.e., hydrological morphology), leading to samples with similar risks but significant morphological differences being mapped to distant nodes, failing to reflect the continuity of risk in the topological structure. In this embodiment, to address this issue, a joint hydrological-risk distance metric is defined. This metric considers not only the distance between hydrological feature vectors but also the distance between risk indicators, balancing the influence of both through weighted coefficients. Using this metric, the network will cluster samples with similar hydrological morphology and similar risk levels together during training.

[0027] Furthermore, to enhance the representation of high-risk samples in the network structure, this step employs a risk-sensitive neighborhood update rule. Specifically, during training, when a high-risk sample is input, the update magnitude of the weight vectors of the winning node and its neighboring nodes is amplified. This mechanism essentially gives high-risk samples greater attraction, causing the network to allocate more node resources to high-risk regions in its topology, thereby improving the resolution of extreme risk scenarios.

[0028] Based on the trained SOM network, this step further constructs a risk-stratified scenario tree. This scenario tree adopts a hierarchical structure, dividing mapping layer nodes into different risk levels, such as high-risk, medium-risk, and low-risk, according to their risk attributes. Within each risk level, secondary clustering is performed based on hydrological morphological characteristics to form specific morphological clusters. This structure of risk stratification followed by morphological clustering allows the scenario tree to clearly display typical hydrological morphologies at different risk levels, providing structured scenario support for subsequent risk decision-making.

[0029] In some alternative implementations, the weighting coefficient α in the hydrological-risk joint distance metric can be dynamically varied. For example, in the early stages of training, the focus can be on learning hydrological morphology, and as the number of iterations increases, the weight of risk distance can be gradually increased to optimize the risk topology. The hierarchical division of the risk stratification scenario tree can also be adjusted according to actual needs, such as adding more intermediate risk levels, or further subdividing specific flood characteristics within the morphological clusters, such as peak occurrence time and flood volume concentration.

[0030] Step 103: Based on the risk-layered scenario tree, discount parameter optimization is performed under the constraint of multi-stage risk budget, scheduling strategy combinations that meet the stage risk fairness requirements are selected, risk control scenario set is generated, and risk attribution analysis is performed to identify risk sources.

[0031] This step treats risk as an allocable and manageable resource, introducing the concept of a multi-stage risk budget. The total risk budget refers to the maximum amount of risk that managers can tolerate throughout the entire scheduling cycle, such as the upper limit of the probability of exceeding the flood limit. The total budget is further decomposed into risk quotas for each scheduling stage (e.g., pre-storage stage, beneficial development stage, and flood control stage). Discount parameters are coefficients used to adjust flood control limit water levels or pre-release flows in the pre-storage and pre-release scheduling rules (e.g., beneficial development discount coefficient, flood control discount coefficient). The goal of the optimization process is to find the combination of discount parameters that maximizes comprehensive benefits (e.g., power generation, water supply guarantee rate) while satisfying the total risk budget and the risk quotas for each stage.

[0032] Furthermore, this step introduces a stage risk fairness requirement. In traditional optimization, a significant amount of risk may be transferred to a single stage to meet the overall risk constraint (e.g., excessively sacrificing the benefits of the benefit-generating stage to ensure flood control safety, or vice versa). To avoid an imbalance in risk responsibility, this embodiment constructs a risk responsibility allocation matrix, calculates the contribution ratio of each stage to the total risk increment, and sets fairness indicators, such as the Gini coefficient or range ratio. Only when the stage risk distribution caused by a certain combination of discount parameters meets the fairness threshold is the combination considered a feasible solution. By selecting the optimal combination of discount parameters, representative scenarios in the risk stratification scenario tree are simulated to obtain a set of scenarios that meet both physical constraints and risk control standards, and fairly share risks among stages—that is, the risk control scenario set.

[0033] In some alternative implementations, the setting of multi-stage risk budgets can be dynamically adjusted by combining historical operating experience and expert judgment. Optimization algorithms can employ intelligent optimization algorithms such as genetic algorithms and particle swarm optimization to efficiently search for optimal solutions in complex parameter spaces. The specific form of risk fairness indicators can also be customized according to the manager's preferences; for example, the maximum-minimum fairness criterion can be used to ensure that the risk increment in the stage with the heaviest risk does not exceed a predetermined relative proportion.

[0034] Example 2 describes the process of constructing a risk-condition hydrological-risk joint manifold, focusing on how to construct a target hydrological-risk feature template, and how to define and train the risk-condition morphological transformation model and its parameter mapping model.

[0035] Step 201: Perform multidimensional risk index statistics on historical flood events in the basic hydrological operation data, analyze the risk distribution characteristics, and divide several non-overlapping risk level intervals.

[0036] In this step, risk indicators corresponding to each historical flood event need to be extracted from basic hydrological operational data. Risk indicators may include, but are not limited to, the probability of exceeding the flood limit level, the maximum extent of exceeding the flood limit level, and tail losses. For the extracted multidimensional risk indicator vector, its distribution characteristics in the historical sample space are analyzed using statistical methods (such as kernel density estimation and quantile analysis). Based on the distribution characteristics, the risk space is divided into K non-overlapping risk level intervals. For example, based on the cumulative probability distribution of the risk indicator R, the interval where R is less than the 30th quantile can be defined as a low-risk interval, the interval between the 30th and 70th quantiles as a medium-risk interval, and the interval greater than the 70th quantile as a high-risk interval. For each interval, the corresponding risk indicator center value or representative value is calculated as a risk reference value for that level. The above division is no longer simply labeling events, but rather establishing the target for subsequent generative models.

[0037] Step 202: Within each risk level interval, perform hydrological feature clustering analysis independently to identify typical hydrological morphological clusters existing under that risk level, and structurally bind the risk level intervals with typical hydrological morphological clusters to construct a target hydrological-risk feature template that includes the target risk level and the target typical hydrological features.

[0038] Unlike traditional methods that cluster data across the entire sample space, this step performs conditional clustering within the same risk level. Specifically, for all historical flood events belonging to the high-risk range, hydrological feature vectors are extracted, such as peak flow, total flood volume, peak time, and flood duration. K-means or hierarchical clustering algorithms are then used for cluster analysis to identify typical patterns within that risk level, such as high-risk – short-peak, high-duration patterns and high-risk – long-peak, high-duration patterns. Similarly, for low-risk ranges, their unique patterns can be identified, such as low-risk – gentle patterns. This establishes a strong coupling between risk level and hydrological pattern. The final target hydrological-risk feature template is a structured data object that includes both the target risk level (e.g., Level 4) and the corresponding target typical hydrological feature vector (e.g., peak flow Q). _max =10000m 3 / s, total flood volume V=5×10 8 m 3 The flood lasted for T days. _dur =72h). The above template will serve as input instructions for the subsequent parameter mapping model, guiding the generator to produce specific flood processes.

[0039] Step 203: Define a risk condition morphological transformation operator that includes amplitude scaling parameters, time scale parameters, time translation parameters, and shape correction parameters. Construct a set of physical constraints with the conservation of total flood volume after transformation, non-negative inflow, and smooth water level connection as hard boundaries, and encapsulate them to form a risk condition morphological transformation model.

[0040] The risk condition morphology transformation model is the actuator for generating the flood process. Specifically, the risk condition morphology transformation operator Q is defined. _new (t) is as follows:

[0041] Q _new (t)=a*Q _base (b*t+c)+d+Δ _Q_shape (t;θ);

[0042] Among them, Q _base (t) represents the normalized baseline data entry process; t is the time point; a is the amplitude scaling parameter, used to control the magnitude of the flood; b is the time scale parameter, used to adjust the duration of the flood, b greater than 1 indicates compression, less than 1 indicates stretching; c is the time shift parameter, used to control the timing of the flood peak; d is the baseline correction parameter, used to adjust the inrush flow or baseflow level; Δ _Q_shape (t;θ) is a shape correction function controlled by the shape parameter θ, such as based on polynomial or radial basis functions, used to fine-tune the flood waveform, such as adjusting skewness or kurtosis.

[0043] To ensure that the generated flood process conforms to the physical laws of hydraulic engineering, a set of physical constraints must be constructed. Specifically, this includes:

[0044] 1. Total amount conservation constraint, that is, the integral of the transformed flood process over a specified period should be approximately equal to the target flood total amount, allowing for a small error ε;

[0045] 2. Non-negativity constraint on flow rate, i.e., for any time t, Q _new (t) must be greater than or equal to 0;

[0046] 3. Water level transition smoothness constraint, that is, the rate of change of flow dQ / dt and the corresponding rate of change of water level must be within the allowable range of the reservoir discharge facilities and the river channel bearing capacity to avoid abrupt changes in hydraulics.

[0047] The above constraints are formalized as inequalities or equality constraints and embedded into the solution process of the model.

[0048] Step 204: Construct a training sample set based on historical flood events. For each sample, construct a parameter optimization problem with the objective of minimizing the deviation of hydrological characteristics and risk indicators and with the physical constraint set as the constraint condition. Obtain the conditionally optimal morphological transformation parameters through numerical solution.

[0049] To train the parameter mapping model, training data is first required. For each historical flood event, it is used as the target, and a baseline process E is selected. _base Construct an optimization problem to find an optimal set of transformation parameters (a). _opt ,b _opt ,c _opt ,d _opt ,θ _opt ), so that the flood process E generated by the transformation operator _gen With target event E _i The optimization problem aims to minimize deviations in hydrological characteristics (such as peaks, quantities, and shapes) and also minimize deviations in risk indicators calculated in the simulation model. This optimization problem is strictly constrained by the aforementioned set of physical constraints. Numerical optimization methods such as Sequential Quadratic Programming (SQP) or Genetic Algorithms are used to solve the problem, yielding the conditionally optimal morphological transformation parameters for each pair of baseline-target combinations. These conditionally optimal morphological transformation parameters not only represent morphological approximation but also implicitly contain the correspondence of risk characteristics.

[0050] Step 205: Using the target hydrological features and target risk level as input features, and the optimal morphological transformation parameters as output labels, a supervised learning algorithm is used to train a neural network to establish a nonlinear mapping relationship from the hydrological-risk joint space to the morphological parameter space, thereby generating a risk conditional morphological transformation parameter mapping model.

[0051] Using the input (target hydrological features + target risk level) - output (optimal morphological transformation parameters) data pairs obtained in step 204, a supervised learning training set is constructed. A neural network model with strong nonlinear fitting capabilities, such as a multilayer perceptron (MLP) or radial basis function network (RBF), is selected as the parameter mapping model. The input layer of the model receives the feature vector from the target template, and the output layer directly predicts the transformation parameters (a, b, c, d, θ). The risk-conditional morphological transformation parameter mapping model is trained by minimizing the loss function, such as mean squared error (MSE), between the predicted parameters and the optimal parameters through backpropagation. The trained model can achieve end-to-end generation: given any target risk level and target hydrological features, the model can instantly output a set of morphological transformation parameters that meet physical constraints and risk standards, without the need for time-consuming iterative optimization, thus improving generation efficiency.

[0052] In some alternative implementations, to improve the model's generalization ability, a regularization term can be added during training to prevent overfitting. Furthermore, ensemble learning methods can be employed, training multiple neural networks with different structures and performing a weighted average to obtain more robust parameter prediction results. For the shape correction function Δ... _Q_shape Parametric representation can be achieved by using methods such as orthogonal polynomial expansion or wavelet transform reconstruction.

[0053] According to one aspect of this application, the risk condition morphological transformation operator structure definition can also be:

[0054] A parameterized combination of linear and nonlinear transformation formulas is employed. Specifically, the morphological transformation operator Q... _new (t) can be expressed as:

[0055] Q _new (t)=a*Q _base (b*t+c)+d+Δ _Q_shape (t,θ),

[0056] Among them, Q _base Δ represents the flow rate value during the normalized baseline data entry process; a is the amplitude scaling parameter, preferably ranging from 0.8 to 1.5, used to adjust the peak flow magnitude; b is the time scale parameter, preferably ranging from 0.8 to 1.2, used to compress or stretch the duration of the flood process; c is the time shift parameter, used to adjust the relative time of the peak flow; d is the baseline correction parameter, used to shift the overall flow rate baseline; _Q_shape (t,θ) represents the shape correction function controlled by the parameter vector θ. This function can be in the form of a weighted sum of Gaussian radial basis functions or orthogonal polynomials (such as Legendre polynomials) to nonlinearly fine-tune the local characteristics (such as plumpness and skewness) of the flood waveform.

[0057] For example, the formula for calculating the loss function during parameter mapping model training is:

[0058] Loss=w _1 *||Q _feat -Q _target ||+w _2 *||R _sim -R _target ||;

[0059] Where Loss represents the total loss value; w _1 and w _2 Q represents the weighting coefficients of the hydrological characteristic deviation term and the risk index deviation term, respectively; _feat Q represents the hydrological feature vector of the flood process calculated based on the generated morphological transformation parameters; _target R represents the target hydrological feature vector in the target hydrological-risk feature template; _sim This represents the risk index obtained from simulating the generated flood process under the baseline scheduling; R _target The reference value of the risk indicator corresponding to the target risk level is represented by ||...||; the norm operation of the vector is represented by L2 norm.

[0060] Example 3 details how to use the model constructed in Example 2 to generate a flood process under risk conditions, and how to adaptively correct the generated manifold through active learning closed loop to solve the problem of superficial resemblance in traditional generation methods. Figure 2 , Figure 3 As shown.

[0061] Step 301: Retrieve historical inflow sequences from basic hydrological operational data that match the preset target hydrological-risk characteristic template, and perform time-scale normalization as the normalization benchmark for inflow.

[0062] In the generation phase, the first step is to use the currently selected target hydrological-risk characteristic template, including the target risk level R. _target and target hydrological characteristics F _target The most suitable seed is selected from the historical database as the starting point. The matching principle is usually based on the similarity of hydrological morphology, such as selecting a seed that matches the target feature F. _target The closest historical flood event in Euclidean distance is selected. Once selected, this historical event is time-normalized and mapped to a unified time axis, such as the [0,1] interval, to eliminate the influence of differences in the original flood duration, resulting in the normalized baseline data entry process Q. _base This step ensures that subsequent morphological transformations are performed on a standardized benchmark.

[0063] Step 302 involves invoking a pre-constructed risk condition morphological transformation parameter mapping model to map the target hydrological features and target risk levels contained in the target hydrological-risk feature template into morphological transformation parameters. The risk condition morphological transformation parameter mapping model represents the nonlinear correspondence from the hydrological-risk joint feature space to the morphological transformation parameter space.

[0064] The target feature F determined in step 301 _target and target risk level R _target The input is fed into the trained parameter mapping model. The model, through its internally learned nonlinear mapping relationships, directly outputs a set of predicted morphological transformation parameters P. _pred =(a _pred ,b _pred ,c _pred ,d _pred ,θ _pred This process is an instantaneous forward computation, which can quickly respond to requests for generating large-scale scenes.

[0065] Step 303: Input the morphological transformation parameters and the normalized benchmark into the risk condition morphological transformation model, drive the parameterized transformation operator containing physical constraints to perform morphological reconstruction, and generate a risk condition flood process sample set that combines the target hydrological statistical characteristics and the target risk level structure.

[0066] Using the shape transformation parameters P obtained in step 302 _pred Driving morphological transformation operator Q _new For the normalized baseline process Q _base The transformation is performed. The calculated new flood process Q is obtained. _gen =Q _new (t;P _pred ), which are the generated candidate samples.

[0067] Because the transformation operator incorporates physical constraint mechanisms, such as non-negative flow correction, the generated flood process is physically plausible. Repeating the above process for all target templates yields the initial risk-condition flood process sample set.

[0068] Step 304: Input the sample set of flood process under risk conditions into the reservoir group scheduling simulation model, analyze the actual risk indicators under the preset scheduling rules, measure the difference between the actual risk indicators and the target risk level corresponding to the target hydrological-risk characteristic template, and generate a risk deviation statistical set.

[0069] To verify whether the generated samples truly meet the expected risk level, physical simulation is necessary. The generated sample set is input into an actual reservoir group scheduling model, and calculations are performed according to the current scheduling rules. The model outputs corresponding process data such as flood control high water level and discharge flow, and based on this, the actual risk index R is calculated._actual Calculate the actual risk R. _actual Compared with the target risk R set when the sample was generated _target The deviation Δ between _R =R _actual -R _target The deviations of all samples under the same risk level are statistically analyzed, and the mean deviation (Mean) is calculated. _Δ_R Sum of deviations and standard deviations (Std) _Δ_R This forms a risk deviation statistics set. If Mean _Δ_R If the value is significantly non-zero, it indicates that there is a systematic bias in the generative model. For example, the actual risk of the generated high-risk samples is generally lower than expected.

[0070] Step 305: Based on the risk deviation statistics set, perform adaptive correction operations on the data parameters constituting the risk condition hydrological-risk joint manifold. This includes shifting the boundary thresholds of the risk level intervals to calibrate the systematic shift in risk distribution, and adjusting the characteristic parameters of the target hydrological-risk characteristic template to compensate for the response error of hydrological morphology to risk indicators, thereby forming an updated risk condition hydrological-risk joint manifold.

[0071] For the deviations identified in step 304, adaptive correction of the manifold parameters is performed. The specific operation involves two dimensions:

[0072] First, the risk level range can be corrected. If the actual risk of a certain level is too low, it means that the entry threshold for that level is too high or the range definition is incorrect. The distribution can be calibrated by shifting the risk boundary threshold of that level downward (for example, lowering the threshold for judging high risk from 0.8 to 0.75).

[0073] Second, the target template parameters are corrected. If the boundaries are confirmed to be correct but the deviation persists, it indicates an error in the correspondence between the target hydrological characteristics and the risk. For example, a peak value of 10,000 was initially thought to correspond to high risk, but simulations show that 12,000 is required. In this case, the target characteristic parameters in the template are adjusted, such as increasing the target flood peak value to 12,000, to compensate for the response error. These two operations work together to update the joint manifold.

[0074] Optionally, the adaptive correction operation performed on the data parameters constituting the risk condition hydrological-risk joint manifold can also be: analyzing the level offset in the risk deviation statistical set and performing a translation operation on the boundary threshold of the risk level interval to calibrate the systematic deviation of the risk distribution; analyzing the response error in the risk deviation statistical set and performing an adjustment operation on the target hydrological feature values ​​in the target hydrological-risk feature template to compensate for the nonlinear mapping error of hydrological morphology to risk indicators; and reconstructing the correspondence between the target risk level and the target hydrological feature through translation and adjustment operations to complete the update of the risk condition hydrological-risk joint manifold.

[0075] Step 306: Based on the updated risk-condition hydrological-risk joint manifold, trigger the iterative generation mechanism, re-call the risk-condition morphological transformation parameter mapping model to perform parameter mapping and morphological transformation until the risk deviation statistics set meets the preset convergence condition, and output the converged risk-condition flood process sample set.

[0076] Using the updated manifold parameters from step 305, the new risk boundary, and the new target template, re-execute steps 302 and 303 to generate a new sample set. Perform the simulation and bias calculation again to check if the bias has decreased.

[0077] Preset convergence conditions may include: in N consecutive iterations, the rate of change of the average risk deviation for each risk level is less than a threshold ε1 (e.g., 1%), and the rate of change of the standard deviation of the risk deviation distribution is less than a threshold ε2. When the convergence conditions are met, the iteration stops, and the final converged sample set is output.

[0078] This closed-loop mechanism ensures a high degree of accuracy and consistency in the risk attributes of the generated scenario set, reflecting the idea of ​​proactive learning.

[0079] In some alternative implementations, the iterative process can employ an adaptive step-size strategy. This involves using a larger step size when the deviation is large, and gradually decreasing the step size as the deviation decreases, thereby accelerating convergence and preventing oscillations. Simultaneously, intermediate results from each iteration can be recorded to analyze the manifold evolution trajectory, providing a basis for further optimization of the model structure.

[0080] According to one aspect of this application, the convergence criterion may also be: employing a dual threshold joint determination mechanism. Specifically, the preset convergence conditions include:

[0081] Condition 1: In N consecutive iterations (e.g., 5 iterations), the rate of change of the average risk deviation of each risk level is less than the first threshold (e.g., 0.5%).

[0082] Condition 2: The rate of change of the standard deviation of the risk deviation distribution is less than the second threshold (e.g., 1%).

[0083] Only when both conditions one and two are met is it determined that the risk condition hydrological-risk joint manifold has evolved to a stable state, and the iteration process is terminated.

[0084] The formula for calculating risk deviation is: Δ _R =R _actual -R _target ;

[0085] Where, Δ _R R represents the risk deviation value. _actual This represents the actual risk index calculated in the reservoir group scheduling simulation model of the generated flood process; R_target This indicates the target risk level value set in the target hydrological-risk characteristic template.

[0086] The formula for calculating the convergence criterion is as follows:

[0087] Change _Rate (Mean _Δ_R )|<ε1 and|Change _Rate (Std _Δ_R )|<ε2;

[0088] Among them, Change _Rate The Mean function represents the rate of change and calculates the relative change between the current iteration and the previous iteration. _Δ_R Std represents the mean of the average risk deviation of the samples at each risk level. _Δ_R ε1 represents the standard deviation of the risk deviation distribution; ε2 represents the convergence threshold of the rate of change of the average deviation, for example, 0.5%; ε3 represents the convergence threshold of the rate of change of the standard deviation, for example, 1%.

[0089] Example 4 describes the process of constructing a risk-hierarchical scenario tree. It focuses on how to achieve dual topology preservation of hydrological features and risk indicators through an improved self-organizing map network (SOM), and how to construct a scenario tree structure with clearly defined risk hierarchy semantics.

[0090] Step 401: For each sample in the converged risk condition flood process sample set, extract multidimensional hydrological morphological features and corresponding actual risk indicators, and combine them into a standardized scenario feature risk vector set after standardization processing.

[0091] In this step, the complex flood process first needs to be transformed into a vector form that can be processed by a neural network. For each converged sample generated, its key hydrological features are extracted, including but not limited to: total flood volume V, peak flow Q. _max Time of flood peak T _peak Flood duration T _dur Maximum amplitude of rise (rate threshold) R _rise and the maximum amplitude R of the receding section _fall Simultaneously, obtain the actual risk indicators calculated for this sample in the simulation, such as the probability of exceeding the flood limit P. _over and destructive loss D _loss Hydrological characteristics are combined with risk indicators to form the original feature vector. To eliminate differences between different physical units (such as flow rate in cubic meters per second versus probability percentage), standardization is necessary. Common methods include Z-score standardization (zero mean, unit variance) or Min-Max normalization (mapping to the 0-1 interval). The resulting standardized scenario feature risk vector V is... _inputIt is a high-dimensional vector, the first part of which is V _hydro Representing hydrological morphology, the latter part V _risk This represents the risk level. The construction of combined inputs is the data foundation for realizing joint topology mapping.

[0092] Step 402: Construct a risk topology self-organizing mapping network containing mapping layer nodes, and define a hydrology-risk joint distance metric. The hydrology-risk joint distance metric is a weighted combination of the spatial distance of hydrological features and the spatial distance of risk indicators, which is used to characterize the similarity of scenarios in both hydrological morphology and risk level dimensions.

[0093] Traditional SOM networks typically only calculate Euclidean distance, which has limitations when handling flood scenarios: two floods with similar hydrological morphology may lead to drastically different risk consequences due to slight differences in peak values; conversely, two floods with the same risk level may have different hydrological morphologies. To address this contradiction, this embodiment defines a distance metric called the hydrological-risk joint distance metric. Specifically, for the input vector V... _input and the weight vector W of the mapping layer nodes _j Their joint distance D _j The calculation formula is:

[0094] D _j =α*D _hydro +(1-α)*D _risk ;

[0095] Among them, D _hydro It is the Euclidean or cosine distance between the hydrological part of the input vector and the hydrological part of the node weight vector; D _risk It is the distance between the risk part of the input vector and the risk part of the node weight vector; α is the balancing weight coefficient, which usually ranges from 0 to 1.

[0096] In some preferred embodiments, the weight coefficient α is not fixed but dynamically adjusted. For example, in the early stages of training, α is set to a larger value, such as 0.8, to focus on coarse clustering of hydrological morphology and to form a basic morphological topology in the network. As the number of training iterations increases, the value of α is gradually decreased, such as to 0.4, to increase the weight of risk distance and force the network to pay more attention to the consistency of risk during the fine-tuning phase. This dynamic strategy helps to form a clear risk gradient locally while maintaining the continuity of hydrological morphology.

[0097] Step 403: Input the normalized scene feature-risk vector set into the risk topology self-organizing mapping network to perform iterative training. In each iteration, calculate the distance between the input vector and the mapping layer nodes based on the hydrology-risk joint distance metric, and select the node with the smallest distance as the winning node.

[0098] Furthermore, a risk-sensitive neighborhood update rule is applied to optimize the weight vector distribution of the risk topology self-organizing mapping network: a risk-sensitive neighborhood update rule is configured, and a risk-sensitive amplification factor that changes positively with the risk level of the sample is introduced into the rule; when updating the weight vectors of the winning node and its neighboring nodes, the risk level contained in the current input vector is analyzed, and the weight vector update amplitude is dynamically adjusted through the risk-sensitive amplification factor, so that high-risk samples have a significant pulling effect on the topology of the risk topology self-organizing mapping network; through iterative updates, a trained risk topology self-organizing mapping network is formed, in which the weight vectors of the mapping layer nodes present a topology in the feature space that takes into account both hydrological morphological similarity and orderly distribution of risk levels.

[0099] During the competitive learning phase of network training, the joint distance D defined above is used. _j Find the winning node (BMU). Unlike traditional methods, the subsequent weight vector update process differs. This embodiment introduces a risk-sensitive neighborhood update rule to address the problem that high-risk samples are often a sparse minority and easily overwhelmed by low-risk majority samples. Specifically, the neighborhood function h is defined. _ci (t) includes a risk-sensitive amplification factor G(R). This factor is a monotonically increasing function of the sample risk level R, for example, G(R) = 1 + β*R, where β is the adjustment coefficient. When the input sample is high-risk, the value of G(R) is larger, leading to a significant increase in the update magnitude (learning rate) and influence range (neighborhood radius) of the weight vectors of the winning node and its neighboring nodes. The risk-sensitive amplification factor is calculated based on the risk level of the current input vector.

[0100] A high-risk sample has a stronger pulling effect on the network topology than a low-risk sample. Through the above mechanism, high-risk samples can occupy more node resources in the mapping layer, resulting in a higher resolution of the finally trained network in high-risk areas. This is particularly important for flood control; although extreme high-risk scenarios have a low probability of occurrence, their decision-making value is high and they must be fully represented in a scenario set, such as... Figure 5 As shown.

[0101] Step 404: Analyze the weight vectors of the mapping layer nodes of the trained risk topology self-organizing mapping network. Based on the risk dimension attribute in the weight vectors, divide the mapping layer nodes into multiple sets of non-overlapping risk level nodes to establish the first-level risk hierarchical structure of the scenario tree.

[0102] Once the SOM network training converges, each node in the mapping layer represents a typical flood scenario. Instead of directly clustering the nodes indiscriminately, the risk component in the weight vector of each node is first examined. Based on a preset risk level standard (consistent with the levels defined in Example 2), all nodes are divided into high-risk, medium-risk, and low-risk node sets. This step constitutes the first layer of the scenario tree's skeleton—the risk hierarchy. This structural design ensures that when users view the scenario tree, they first see the distribution of risks, reflecting a risk-oriented decision-making logic.

[0103] Step 405: Within each risk level node set, perform secondary clustering analysis based on the hydrological feature dimension attributes in the weight vector to identify several intra-layer feature clusters existing under that risk level, establishing the second-level hydrological morphology structure of the scenario tree. Based on the first-level risk stratification structure and the second-level hydrological morphology structure, construct a risk stratification scenario tree that sequentially represents risk levels and hydrological morphology from top to bottom, and statistically analyze the number of samples covered by each node and the average risk index, serving as the scenario basis for subsequent multi-stage risk budget optimization, such as... Figure 4 As shown.

[0104] After defining the risk levels, secondary clustering based on hydrological features is performed separately for the node set within each level, such as using the K-means algorithm or hierarchical clustering. For example, in the high-risk node set, there may be clusters of thin, high-risk nodes and clusters of large, high-risk nodes; while in the low-risk node set, there may be clusters of uniform, low-risk nodes. These morphological clusters are then used as the second-level nodes of the scene tree.

[0105] The final constructed risk-stratified scenario tree has a clear semantic structure: the root node branches downwards to different risk levels, and each risk level node further branches downwards to different hydrological morphologies. The system statistically analyzes the number of original samples, average risk value, and average hydrological characteristics covered by each leaf node (morphology cluster). This statistical information directly supports subsequent sampling and optimization.

[0106] In some alternative implementations, the scene tree can also include a third-level structure, further subdividing the morphological clusters based on the timing of the flood peaks. Furthermore, to reduce the number of scenes, only one sample closest to the cluster center can be extracted from each leaf node (morphological cluster) as a representative scene, and assigned a probability weight representing all samples in that cluster. The resulting representative scene set covers all risk levels and typical morphologies at each level, while significantly reducing the number of samples, thus improving the efficiency of subsequent optimization calculations.

[0107] According to one aspect of this application, the hydrological-risk joint distance metric can also be defined as: employing a hybrid distance formula based on dynamic weights. Specifically, for the input vector V _input and the weight vector W of the mapping layer nodes _j Their joint distance D _j It can be calculated as follows:

[0108] D _j =α*D _hydro +(1-α)*D _risk ;

[0109] Among them, D _j D represents the joint distance between the input vector and the j-th mapping layer node; _hydro D represents the distance (e.g., Euclidean distance) between the hydrological feature components of the input vector and the hydrological feature components of the node weight vector; _risk This represents the distance between the risk index components of the input vector and the risk index components of the node weight vector; α is the balancing weight coefficient, used to adjust the relative importance of hydrological features and risk indicators in clustering.

[0110] Furthermore, the value of α is not fixed, but decreases dynamically with the number of training iterations t', for example:

[0111] α(t')=α _start *(1-t' / T _max );

[0112] Where α(t') represents the weight coefficient at the t'th training iteration; α _start t' represents the initial weight coefficient, for example, 0.8; t' represents the current iteration number; T _max This indicates the preset maximum total number of iterations.

[0113] The above settings enable the network to primarily learn the topology of hydrological morphology in the early stages of training, while gradually focusing on the topology refinement based on risk levels in the later stages of training.

[0114] Optionally, the formula for calculating the risk-sensitive neighborhood function is:

[0115] h _ci (t')=η(t')*exp(-||r _c -r _i || 2 / (2*σ(t') 2 *(1+γ*R _sample )));

[0116] Among them, h _ci (t') represents the influence coefficient of the winning node c on the update of its neighboring node i in the t'th iteration; η(t') represents the learning rate; r_c and r _i Let t' and γ represent the position coordinates of the winning node c and its neighboring node i in the topological grid, respectively; σ(t') represents the neighborhood radius; γ represents the risk sensitivity coefficient; R _sample This represents the risk level (normalized value) of the current input sample. This term is used to amplify the range of influence of the neighborhood when the risk is high.

[0117] Example 5 describes the process of optimizing the execution discount parameters, focusing on how to allocate risk as a resource in multiple stages, and how to optimize the scheduling strategy while ensuring fairness between stages.

[0118] Step 501: Establish a risk control standard that includes total risk budget indicators and stage importance weights. Decompose the total risk budget indicators according to the scheduling stages to generate a multi-stage risk budget and quota set that includes the upper limit of risk quota for each stage.

[0119] In this embodiment, risk is no longer a vague qualitative concept, but is quantified into specific budgetary indicators. Based on the basin flood control plan and the safety standards of downstream protected objects, the total risk budget R for the entire scheduling period is determined. _total For example, R _total This can be defined as a maximum permissible probability of exceeding the flood limit water level of 5%. Next, the total budget is broken down according to the functional positioning and risk tolerance of different scheduling phases. For example, the pre-storage phase without rain primarily serves power generation and water supply, with a lower risk tolerance, and the allocated quota R... _quota_1 The allocation is 0.5%; during the pre-release phase for water utilization, both water storage and reservoir emptying must be considered, and the allocation quota R... _quota_2 The allocation is 1.5%; during the flood control pre-discharge phase, the area directly faces the impact of floods, and the allocated quota R... _quota_3 The weight is 3.0%. At the same time, an importance weight W is set for each stage. _k For example, W _1 =0.2, W _2 =0.3, W _3 =0.5, to reflect the core position of the flood control stage in risk control.

[0120] Step 502: Extract representative scenarios from the risk stratification scenario tree, and combine them with the preset discount parameter candidate space to calculate the difference in stage risk indicators under the baseline scheduling rule and different discount parameter combinations, thereby generating a stage risk increment sequence describing the magnitude of risk change in each scheduling stage; based on the stage risk increment sequence and stage importance weight, calculate the contribution ratio of each scheduling stage to the total risk increment, and construct a risk responsibility allocation matrix set that reflects the risk burden distribution structure under different discount parameter combinations.

[0121] Select a representative set of scenarios S _repAnd a set of discount parameter combinations X to be optimized, for example, X includes the beneficial storage capacity discount coefficient λ. _1 and flood control storage capacity discount factor λ _2 For each scenario and each set of parameters X, the risk index R at each stage k is calculated through scheduling simulation. _k (X). Simultaneously, calculate the baseline risk R when no pre-leakage strategy (i.e., baseline rule) is adopted. _k (base). Stage risk increment Δ _R_k (X) is defined as Δ _R_k (X)=R _k (X)-R _k (base). This increment reflects the amount by which the risk increases or decreases in this phase due to the pre-discharge scheduling using the discount parameter X. The increments for all scenarios are statistically analyzed, such as by taking the average or the 95th percentile, to form a phase risk increment sequence, such as... Figure 6 As shown.

[0122] Step 503: Construct a discount parameter optimization model with the objective function of maximizing comprehensive benefit and the hard constraints of ensuring that the total risk budget index and the stage risk quota do not exceed the limit in the multi-stage risk budget and quota set. Introduce stage fairness constraints into the discount parameter optimization model, calculate the stage risk Gini coefficient or range ratio based on the risk responsibility sharing matrix set, eliminate solutions where the unevenness of risk responsibility distribution exceeds the preset fairness threshold, and select the optimal combination set of discount parameters. Substitute the optimal combination set of discount parameters into the reservoir group scheduling rules, perform multi-stage scheduling simulation on representative scenarios in the risk stratification scenario tree, and output a set of risk control scenarios that meet the risk budget constraints and stage fairness constraints.

[0123] Traditional optimization methods might choose an extreme discount parameter in pursuit of maximizing overall benefits, leading to excessive risk increments in a certain stage (such as the benefit-generating stage) and compromising the fairness of the scheduling rules. Therefore, this step defines a risk-sharing index C. _k C _k Equal to the weighted risk increment of this stage (Δ) _R_k *W _k The proportion of C to the sum of weighted increments across all stages. Based on C for each stage. _k The algorithm calculates fairness metrics, such as the Gini coefficient or the range ratio (maximum liability divided by minimum liability). A fairness threshold is set, such as a Gini value less than 0.4. During the optimization search, any discount parameter combination X that causes the fairness metric to exceed the threshold is directly eliminated, even if it has high economic benefits. This mechanism forces the algorithm to find a strategy that can evenly distribute risk across stages.

[0124] Step 504: Perform risk attribution analysis based on the risk control scenario set: Backtrack the features of scenarios with high residual risk in the risk control scenario set, and identify the sensitive scheduling stages and key hydrological morphology types that lead to risk exceeding limits by combining the risk responsibility sharing matrix set and the target hydrological-risk feature template; quantify the marginal contribution rate of different discount parameters to the risk increment of each scheduling stage, and construct a risk source identification dataset containing risk driving factors, sensitive stage identifiers and key morphological features to guide the targeted modification of flood control scheduling rules for reservoir groups.

[0125] After determining the optimal combination of discount parameters and generating a risk control scenario set, attribution analysis is performed on the few remaining high-risk scenarios—extreme cases where the risk remains high even under the optimal strategy. Using a risk responsibility allocation matrix, the system tracks which stage contributes the main risk increment. For example, analysis reveals that 80% of the total risk in a certain scenario originates from the flood pre-discharge stage, and this scenario belongs to a bimodal flood pattern cluster. This indicates that the current discount parameters have weak control over bimodal floods during the flood pre-discharge stage. Based on this, the system generates a risk source identification dataset, clearly pointing out that bimodal floods are a weak link in the current strategy, and the main risk stems from insufficient pre-discharge capacity during the flood pre-discharge stage. This conclusion can directly guide targeted modifications to scheduling rules, such as designing corresponding pre-discharge rules for bimodal floods.

[0126] In some alternative implementations, the optimization model can employ a multi-objective optimization architecture, simultaneously minimizing total risk and maximizing power generation benefits, resulting in a set of Pareto optimal solutions for decision-makers to choose from based on their prevailing risk preferences. Risk attribution analysis can also be combined with decision tree or random forest algorithms to automatically identify key hydrological characteristic thresholds that lead to high risk; for example, when the peak flow exceeds 15,000 cubic meters per second and the peak occurs more than 24 hours earlier, the risk increases significantly.

[0127] Example 6 describes the specific mathematical process of multi-stage risk budget allocation, risk responsibility sharing matrix construction, and optimization model solution based on fairness constraints. It focuses on how to transform abstract risk fairness into computable mathematical constraints to achieve scientific allocation of risk resources in multi-stage scheduling.

[0128] Step 601: Construct a multi-stage risk increment sequence and risk responsibility sharing matrix. In this embodiment, firstly, read the representative scenario set S from the risk stratification scenario tree, and the preset discount parameter candidate set X. For each discount parameter combination x in set X... _j For example, x _jIncluding predetermined discount coefficients for beneficial storage capacity and flood control storage capacity, these are substituted into the scheduling model to simulate scenario set S. For each scheduling stage k (k ranges from 1 to K, representing pre-storage without rainfall, pre-release for beneficial purposes, pre-release for flood control, etc.), the average risk index R of that stage is calculated. _k (x _j Simultaneously, the average risk indicator R under the baseline scheduling rules is obtained. _k (base).

[0129] Therefore, the incremental risk at each stage is calculated as follows:

[0130] Δ _R_k (x _j )=R _k (x _j )-R _k (base);

[0131] Where, Δ _R_k (x _j ) indicates that when using the discount parameter combination x _j Under the scheduling strategy, the risk increment in the k-th scheduling stage; R _k (x _j R represents the risk indicator value in stage k under this strategy; _k (base) represents the risk indicator value for stage k under the baseline scheduling rule.

[0132] This incremental response strategy x _j Compared to the additional risk introduced by the baseline strategy in stage k.

[0133] Based on this, a stage importance weight w is introduced. _k To construct a risk responsibility sharing indicator. Specifically, for the discount parameter combination x _j Its risk responsibility sharing coefficient C in stage k _k (x _j The calculation formula is:

[0134] C _k (x _j )=(w _k *Δ _R_k (x _j )) / ∑ _i (w _i *Δ _R_i (x _j ));

[0135] Among them, C _k (x _j ) indicates the discount parameter combination x _j The proportion of risk responsibility borne by the k-th scheduling phase; w _kThe importance weight of the k-th scheduling stage is represented by ∑. _i This represents the summation calculation over all scheduling stages i; Δ _R_i (x _j ) represents the risk increment of the i-th scheduling phase.

[0136] This formula transforms the physical risk increment into a normalized responsibility ratio, forming a risk responsibility sharing matrix. Each row of the matrix corresponds to a combination of discount parameters, and each column corresponds to the responsibility proportion of a scheduling phase.

[0137] Step 602: Define the fairness constraint index for stage risk. To prevent excessive concentration of risk in a certain stage, this embodiment uses the Gini coefficient or range ratio as a fairness quantification index. In a preferred embodiment, the range ratio constraint is used. For any combination of discount parameters x... _j Calculate the risk liability coefficient C for each stage. _k (x _j The ratio of the maximum to the minimum value of ) Ratio(x) _j Set a fairness threshold θ _fair For example, θ _fair It equals 3.0. If Ratio(x) _j ) greater than θ _fair Then determine the combination x _j This strategy violates fairness constraints, resulting in some stages bearing more than three times the risk responsibility of others, and should be eliminated.

[0138] For example, the formula for calculating the stage risk fairness constraint index (range ratio) is as follows:

[0139] Ratio(x _j )=Max(C _k (x _j )) / Min(C _k (x _j ));

[0140] Where Ratio(x) _j ) represents the discount parameter combination x _j Risk liability range ratio; Max(C) _k (x _j ()) represents the maximum value of the risk liability coefficient at each stage under this combination; Min(C) _k (x _j )) represents the minimum risk liability coefficient for each stage under this combination.

[0141] Step 603: Construct and solve the discount parameter optimization model under multi-stage risk budget constraints. Establish a comprehensive benefit-generating model B(x) _j The optimization model is to maximize the objective function.

[0142] The model includes three types of hard constraints:

[0143] The first type is the total risk budget constraint, which is the sum of risk indicators for all stages ∑(R) _k (x _j The total risk budget R must be less than or equal to the total risk budget. _total ;

[0144] The second type is the stage risk quota constraint, that is, for any stage k, R _k (x _j The quota R for that stage must be less than or equal to the quota for that stage. _quota_k ;

[0145] The third category is the fairness constraint defined above, namely Ratio(x) _j ) must be less than or equal to θ _fair .

[0146] In some alternative implementations, the optimization problem can be formulated as a mixed-integer nonlinear programming problem, or solved using the multi-objective particle swarm optimization (MOPSO) algorithm. During the solution process, the algorithm first filters the feasible solution space based on three types of constraints, and then searches within the feasible solution space for the solution that maximizes the overall benefit B(x). _j The maximum optimal discount parameter combination x _opt The final output x _opt This approach not only improves the utilization rate of flood resources in the reservoir group, but also ensures that risks are fairly and controllably distributed across different scheduling stages, achieving a deep synergy between benefits and safety.

[0147] According to one aspect of this application, in the process of constructing risk-condition hydrological-risk joint manifolds and generating flood processes, the risk level classification and target hydrological-risk template construction are specifically as follows:

[0148] The risk indicators from the joint hydrological-risk sample set are obtained. Multiple risk indicators, such as the probability of exceeding the flood limit, the maximum extent of exceeding the flood limit, and tail loss, are jointly statistically analyzed to obtain the histogram distribution, quantile curves, and correlation matrix of the risk indicators, forming a statistical feature set for risk indicator classification. These statistical features become the basic input for subsequent risk level interval division.

[0149] Obtain a statistical feature set of risk indicators, and determine the number and boundaries of risk levels based on risk tolerance and risk distribution characteristics. For example, events can be divided into multiple intervals—low risk, low-to-medium risk, medium-to-high risk, and high risk—based on several quantiles in the risk distribution. Assign a unique level number and text label to each risk interval, forming a risk level classification rule set that clearly defines the range of risk indicators corresponding to each level and their management implications. This is not simply a matter of low, medium, and high labels, but rather a clearly defined risk condition input used in subsequent target templates and generation processes.

[0150] A joint hydrological-risk sample set and a risk level classification rule set are obtained, and the joint sample is divided into different risk level subsets. Within each risk level subset, cluster analysis is performed only on the hydrological characteristics to identify the main flood morphology types within that risk level, such as peak-short-duration floods and low-peak-long-duration floods. The cluster centers and their statistical characteristics are recorded as hydrological morphology clusters within the same risk level, achieving a structured description of the diversity of hydrological morphology under the same risk level. Unlike traditional methods that cluster based on the overall sample, this approach emphasizes morphological subdivision within the same risk level.

[0151] Obtain the hydrological morphology clusters and risk level classification rule set within each risk level. Construct a target hydrological-risk characteristic template for each combination of risk level and hydrological morphology cluster. The template includes the target total flood volume, target peak flow, target peak occurrence time, target peak duration, target rise steepness, target recession steepness, target risk level, and target risk index reference values ​​for that combination. Organize all combination templates into a target hydrological-risk characteristic template set. Each template explicitly binds a risk level and a typical hydrological morphology; these are then used directly as risk condition inputs during subsequent generation, rather than relying solely on hydrological morphology or a single risk dimension.

[0152] According to one aspect of this application, the risk condition morphology transformation model and physical constraint construction during the construction of the risk-condition hydrological-risk joint manifold and the generation of flood processes are specifically as follows:

[0153] A multi-stage flood event sample library is acquired. For each historical flood event, the inflow time series is standardized in terms of both time length and time step. Flood processes of different durations are mapped to a reference time axis of uniform length through linear interpolation or time scale transformation. The normalized inflow series are then compiled into a normalized baseline inflow process set, which serves as the baseline input for the morphological transformation model. This step ensures that all subsequent morphological transformations are performed in a uniform time domain, avoiding unnecessary complexity caused by time scale differences.

[0154] The normalized baseline inflow process set and reservoir group structural parameter set are obtained. A risk condition morphological transformation operator is defined to map the baseline inflow process, target hydrological characteristics, and target risk level together into the transformed inflow process. The morphological transformation operator adopts a parameterized structure, specifically a formula, which has been shown in the previous embodiments and will not be detailed here.

[0155] Subsequently, the target risk level is explicitly used as the input to the set of shape parameters for calculation, so that the shape transformation operator responds not only to hydrological targets, but also to risk targets.

[0156] Obtain the structural parameter set and normalized baseline inflow process set of the reservoir group, and transform the reservoir capacity-water level relationship, outflow capacity constraints, and cascade connection relationships of the reservoir group into physical constraints for the morphological transformation model. Specifically, this includes:

[0157] The total flood volume conservation constraint is applied to the flow rate Q. _new The integral of the (morphological transformation operator) on the reference time axis requires that the total difference before and after the transformation be within a set error range;

[0158] The flow rate non-negativity constraint requires that the inbound flow rate at any time step after the transformation is not less than zero.

[0159] A smoothing constraint adapted to downstream water level changes is implemented, which restricts the rate of change of flow rate in adjacent time steps to a limit to avoid sudden changes in water level.

[0160] The above constraints are expressed in the form of equations or inequalities and organized into a set of physical constraints for risk condition morphological transformation.

[0161] The structure of the morphological transformation operator and the set of physical constraints for risk-condition morphological transformation are obtained. Combined with the target hydrological features and target risk level to be used subsequently, the inputs, outputs, parameters, and constraints of the morphological transformation are encapsulated into a unified data structure, forming the parameter set describing the risk-condition morphological transformation model. This dataset explicitly specifies that the model inputs are the normalized baseline inflow process, the target hydrological features, and the target risk level; the model's internal parameters are the amplitude parameter a, time parameters b and c, baseline parameter d, and the shape parameter set θ; and the model output is the transformed inflow process. The risk level is used as an explicit input dimension of the model, and the model description stipulates that all subsequent optimization processes must approximate the target risk level as closely as possible while simultaneously satisfying both hydrological features and physical constraints.

[0162] According to one aspect of this application, the risk condition morphology transformation parameter mapping learning is performed during the construction of the risk-condition hydrological-risk joint manifold and the generation of flood processes, specifically as follows:

[0163] We acquire a joint hydrological-risk sample set, a target hydrological-risk feature template set, and a parameter set describing the risk condition morphology transformation model. For each historical flood event, we assign one or more target templates, ensuring that the templates cover typical situations in terms of hydrological features and risk levels. For each historical event-target template combination, we record the baseline data entry process, target hydrological features, and target risk level, forming a training sample set for risk condition morphology transformation, which is used for subsequent parameter optimization and mapping learning.

[0164] A training sample set and a parameter set describing the risk condition morphological transformation model are obtained. A constrained optimization problem is constructed for each training sample pair. The goal is to make the hydrological characteristics of the transformed inflow process as close as possible to the target hydrological characteristics, and to ensure that the risk indicators obtained from simulations under the current or baseline scheduling rules are as close as possible to the target risk level, while strictly satisfying physical constraints. A constrained numerical optimization algorithm is used to solve for the morphological transformation parameter vector for each sample pair, obtaining a set of optimal parameters that satisfy both hydrological and risk conditions, forming a conditionally optimal morphological transformation parameter set.

[0165] This process involves obtaining the conditionally optimal morphological transformation parameter set and the target hydrological features and target risk levels from the training samples. A supervised learning relationship is established between each target hydrological feature + target risk level and its corresponding optimal morphological transformation parameter vector. A suitable function approximation model, such as a multilayer perceptron or radial basis function network, is selected. Using the target feature-risk vector as input and the morphological transformation parameter vector as output, a risk-conditional morphological transformation parameter mapping model is trained. This model differs structurally from traditional fixed-ratio models because it explicitly includes the risk level dimension in the input and learns the parameter distribution under different risk levels.

[0166] A risk condition morphological transformation parameter mapping model and an independent validation sample set were obtained. The inflow process generated by the parameters output by the mapping model was compared with the baseline results obtained through direct optimization, and the model was evaluated in terms of hydrological feature error and risk index error. By adding regularization terms or constraining the model parameter range, the excessive extrapolation of the mapping model outside the training samples was suppressed, ensuring that the mapping parameters still meet physical constraints and risk rationality in the actual generation process. The validated and adjusted mapping model serves as the final version of the risk condition morphological transformation parameter mapping model and is used for subsequent risk condition sample generation.

[0167] According to one aspect of this application, in the process of constructing risk-conditional hydrological-risk joint manifolds and generating flood processes, the generation of risk-conditional flood process samples specifically includes:

[0168] A multi-stage flood event sample library and a target hydrological-risk feature template set are obtained. For each target hydrological-risk feature template, a set of historical entry processes that are closest to its hydrological features are searched in the multi-stage flood event sample library as the starting benchmark process for morphological transformation. The correspondence between the template and the benchmark process is recorded to form a risk condition generation benchmark process matching set.

[0169] The process involves acquiring a risk condition generation baseline matching set and a risk condition morphological transformation parameter mapping model. The hydrological characteristics and risk level of each target hydrological-risk feature template are input into the mapping model to obtain the corresponding morphological transformation parameter vector. The generated parameters are then associated with their corresponding baseline data entry process to form a risk condition generation parameter set, which is used in the specific generation and data entry process.

[0170] Obtain the parameter set for risk conditions and the parameter set for the risk condition morphological transformation model. For each template-baseline process-parameter combination, input the morphological transformation parameters into the risk condition morphological transformation model to calculate the transformed inflow process, resulting in a candidate inflow process that satisfies the target hydrological characteristics and target risk level. Organize all generated candidate inflow processes according to template number and baseline event number to form the initial set of risk condition flood process samples.

[0171] An initial set of samples for flood processes under risk conditions and a set of structural parameters for the reservoir group are obtained. Each generated inflow process is input into the reservoir group simulation model to calculate the corresponding water level time series and actual risk indicators. It is then checked whether the preset tolerance range is met in terms of total flood volume, peak value, water level change, and risk indicators. For samples that do not meet physical constraints or deviate significantly from the target risk, corrections are made by slightly adjusting the morphological transformation parameters or switching to a more suitable baseline process. Finally, the validated samples are retained to form a sample set of flood processes under risk conditions and a corresponding set of actual risk indicators for the risk conditions.

[0172] According to one aspect of this application, in the process of constructing the risk topology self-organizing map and the risk hierarchical scenario tree, the training of the risk topology self-organizing map network specifically involves:

[0173] Obtain the scene feature-risk vector set, and standardize both the hydrological feature part and the risk indicator part separately, for example, by using zero-mean unit variance standardization or interval scaling to eliminate dimensional differences. Concatenate the standardized hydrological feature vector and risk indicator vector to form a unified standardized scene feature-risk vector set, which serves as the input to the risk topology self-organizing mapping network.

[0174] Obtain a standardized scene feature-risk vector set, define a joint distance metric for competitive learning, and combine the hydrological feature distance and risk indicator distance in a weighted manner. The combination process is specifically described by the formula, which has been elaborated in the previous embodiments and will not be detailed here.

[0175] This joint metric enables the self-organizing map network to simultaneously focus on hydrological morphological similarity and risk level similarity during training, thus forming a risk topology joint distance parameter set.

[0176] To obtain the risk component from the normalized scene feature-risk vector set, a risk-sensitive neighborhood function and learning rate function are designed for training the self-organizing map network. Specifically, based on the traditional neighborhood function, an amplification factor that varies with the sample risk level is introduced, so that the winning node corresponding to high-risk samples and its neighboring nodes have a larger weight vector update amplitude, while the update amplitude corresponding to low-risk samples is relatively smaller. The risk-sensitive neighborhood function, risk-sensitive learning rate function, and their hyperparameters are organized into a risk-sensitive training parameter set.

[0177] A standardized set of scene features—a risk vector set, a risk-topological joint distance parameter set, and a risk-sensitive training parameter set—is obtained. The self-organizing map network is then trained within a predetermined number of iterations. In each iteration, for each scene sample, the distance between it and all mapping layer nodes is calculated based on the joint distance metric, and the node with the smallest distance is selected as the winning node. Based on the risk level of the sample, the corresponding learning rate and neighborhood width are obtained from the risk-sensitive training parameter set, and the weight vectors of the winning node and its neighboring nodes are updated. After multiple rounds of training, the network weight vectors in the feature space form a structure that maintains the hydrological topology while being orderly distributed according to risk levels, ultimately yielding the risk-topological self-organizing map network parameter set.

[0178] According to one aspect of this application, in the process of risk topology self-organizing mapping and risk hierarchical scenario tree construction, the construction of the risk hierarchical scenario tree structure is specifically as follows:

[0179] Obtain the parameter set and risk level classification rule set of the risk topology self-organizing mapping network. Based on the risk component in the weight vector of each mapping layer node, divide the nodes into different risk level sets, such as low-risk, medium-risk, and high-risk node sets. This forms a risk node grouping set of the mapping layer organized by risk level, laying the foundation for constructing a higher-level tree structure with risk semantics.

[0180] Obtain the hydrological feature weight vector from the risk node grouping set and network parameters of the mapping layer. Perform hydrological feature clustering on the node set under each risk level, dividing nodes with similar hydrological features into several sub-clusters. Each sub-cluster represents a typical hydrological morphology, which may include multiple different water forms under the same risk level. Organize all hydrological feature sub-clusters within the risk layer into a hydrological morphology sub-cluster set within the risk layer.

[0181] Obtain the risk node grouping set and the hydrological morphology subclusters within the risk layer of the mapping layer. Treat each risk level as a first-level node in the scene tree, and each hydrological morphology subcluster under that risk level as a child node of the corresponding first-level node, constructing a two-level structure from risk level to hydrological morphology. If further subdivision is needed, secondary nodes can be constructed within the hydrological morphology subclusters to represent fine-grained feature differences. By recording the parent-child node relationships and the mapping layer node numbers covered by the nodes, a risk-layered scene tree structure with risk level meaning is formed.

[0182] The risk-stratified scenario tree structure and scenario feature-risk vector set are obtained. Each scenario sample is mapped to the corresponding leaf node or intermediate node in the scenario tree according to the mapping layer determined during training. The number of scenarios covered by each node, the average hydrological feature vector, the average risk index, and the cumulative risk weight are counted to form a scenario tree node statistical set. This statistical set is used for subsequent representative scenario selection and also provides quantifiable data support for illustrating typical hydrological morphologies and their risk distribution under each risk level.

[0183] According to one aspect of this application, an alternative scheme for a method of identifying risk sources and generating scene sets for flood resource utilization in reservoir groups is provided, such as... Figure 7 As shown, it includes the following steps:

[0184] Meteorological, hydrological, and engineering data of the multi-reservoir system in the study area were acquired; historical rainfall and forecast data were processed in a unified manner to construct measured and forecast samples of average daily rainfall; the marginal distribution model of meteorological forecast errors was optimized, and ensemble forecast sequences were generated based on the Monte Carlo method to provide an input basis for subsequent error-driven modeling.

[0185] Key uncertainties and risk sources in the utilization of flood resources are identified, including meteorological forecast errors, hydrological forecast errors, and uncertainties in the spatial combination of flood volume. Based on meteorological and hydrological data, error margin distributions for rainless periods and receding water sections are constructed. A Vine Copula function is constructed to establish the spatial joint distribution of total reservoir flood volume. A set of random scenarios with reasonable structure and physical consistency is generated. Typical scenarios are extracted by dimensionality reduction through self-organizing mapping method to improve scenario representativeness and computational efficiency.

[0186] Based on the optimized error distribution and typical historical flood processes, a multi-factor superposition strategy is adopted to simulate the propagation and coupling of forecast errors in time and space. Using typical historical flood processes as a reference baseline, a multi-dimensional dynamic scaling ratio is constructed to realize the spatial reconstruction of the total flood volume scenario and generate a flood process sample set coupled with meteorological and hydrological multi-source uncertainties.

[0187] With reservoir water level regulation safety as the core, risk indicators reflecting the system's operational safety are defined, including the risk rate of a single reservoir exceeding the flood limit and the system's risk rate of exceeding the limit during multi-reservoir coordinated scheduling, and a quantitative expression method for risk is established.

[0188] Based on engineering data and pre-storage and pre-release strategies, the scheduling operation results under different risk scenario sets are calculated, and the risk levels corresponding to different scheduling strategies are identified. By adjusting the combination of discount coefficients in the system, scheduling sensitivity analysis is carried out to reveal the response mechanism between the benefit and flood control objectives, and finally, strategy selection suggestions for actual operation are formed.

[0189] According to one aspect of this application, in the process of generating a sample set of flood processes with coupled meteorological-hydrological multi-source uncertainties, the coupling of multi-stage risk source scenario sets is specifically as follows:

[0190] Calculate the actual excess water storage volume of the Xingli pre-release section:

[0191] ;

[0192] In the formula, For the first The first reservoir The actual excess water storage capacity of the day To maximize the utilization of water flow in the reservoir, Because the period of no rain forecast is long, This represents the first [unit / item] calculated from the forecast rainfall. The cumulative inflow of water into the pre-release section of each reservoir, λ _1 The discount coefficient for pre-release of beneficial storage capacity (λ) _1 ≤1);

[0193] Calculate the planned excess water storage capacity of the Xingli pre-release section considering dual uncertainties:

[0194] ;

[0195] In the formula, To consider the dual uncertainty of the first The Xingli pre-release section of the reservoir exceeded its storage capacity. This is a sample of inflow water volume with dual uncertainties;

[0196] Calculate the actual excess water storage in the flood control pre-discharge section:

[0197] ；

[0198] In the formula, For the first The actual excess water storage capacity of the flood control pre-discharge section of each reservoir. For the first The downstream safe flow of each reservoir, To predict the incoming water process for the next flood, λ represents the duration of flood forecasts. _2 The flood control pre-discharge discount factor (flood control reservoir capacity discount factor) (λ) _2 ≤1), λ _2 The smaller the value, the greater the safety margin reserved for downstream applications;

[0199] Calculate the planned excess storage capacity of the flood control pre-release section considering the uncertainty of the spatial composition of flood volume:

[0200] ;

[0201] In the formula, For the first The planned flood discharge sections of the reservoirs exceed their storage capacity;

[0202] Calculate the water level of the Xingli pre-release section:

[0203] ;

[0204] in, For the relationship between water level and reservoir capacity, For the first Random sample of water levels in the pre-release section of a reservoir. This refers to the reservoir capacity corresponding to the flood control limit water level. For the first The actual water storage capacity of the pre-release section of a reservoir, when d=d * (d is the day index, d=1,2,…,d) * ), This refers to the cumulative inflow of water into the entire Xingli pre-discharge section. Become the first The final water level of the Xingli pre-release section of the reservoir :

[0205] ;

[0206] Calculate the final water level of the flood control pre-discharge section:

[0207] ;

[0208] In the formula, For the first The final water level of the flood control pre-discharge section of the reservoir.

[0209] In this embodiment, this step calculates the actual and planned excess water storage volume for the two pre-release stages respectively, and combines the water level-storage capacity relationship to transform the uncertainty of inflow into water level status indicators, thereby realizing the stage-by-stage extrapolation of reservoir operation results under typical scenarios and providing quantitative basis for scheduling risk identification and comparison.

[0210] This application establishes an inverse mapping from the target risk level to morphological transformation parameters and uses the risk deviations generated by actual simulations as feedback to correct the generation rules, thus realizing a mechanism for proactively generating corresponding flood processes based on the target risk level. This approach ensures that the generated scenario set can accurately cover the expected risk level, and in particular, effectively fills in the high-risk tail region that is difficult to capture by traditional methods, thereby improving the completeness of risk identification.

[0211] By introducing a joint hydrological-risk distance metric and a risk-sensitive neighborhood update rule, the forced clustering algorithm considers both hydrological morphology and risk attributes, and assigns stronger topological influence to high-risk samples. This process results in a scene tree with a clear risk-layered-morphological clustering structure, preventing high-risk scenes from being submerged by mean distribution and ensuring that the reduced scene set still retains the risk skeleton structure of the original data.

[0212] By decomposing the total risk into phased quotas and introducing fairness indicators such as the Gini coefficient to constrain the allocation of risk responsibility, the algorithm is forced to find a scheduling strategy that evenly distributes the risk increment across the pre-accumulation and pre-release phases. This not only avoids the excessive accumulation of scheduling risk in a single phase but also makes the generated risk control scenario set and scheduling rules more in line with the safety baseline of actual operation, thus possessing stronger feasibility.

[0213] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for identifying risk sources and generating scenario sets for flood resource utilization in reservoir groups, characterized in that, include: Based on basic hydrological operation data, a risk-condition hydrological-risk joint manifold is constructed. By using the preset target risk level as the input condition for morphological transformation, the risk-condition hydrological-risk joint manifold is iteratively corrected using the pre-calculated simulation risk deviation, thereby generating a risk-condition flood process sample set. Input the sample set of flood process under risk conditions into the risk topology self-organizing mapping network, perform topology mapping based on hydrology-risk joint distance metric and risk-sensitive neighborhood update rule, and construct a risk hierarchical scenario tree; Based on the risk-layered scenario tree, discount parameters are optimized under the constraints of multi-stage risk budgets, and scheduling strategy combinations that meet the stage risk fairness requirements are selected. A risk control scenario set is generated and risk attribution analysis is performed to identify risk sources. A sample set of flood processes under risk conditions is input into a risk topology self-organizing mapping network. Topology mapping is performed based on the hydrological-risk joint distance metric and risk-sensitive neighborhood update rules, including: For each sample in the convergent risk condition flood process sample set, multidimensional hydrological morphological features and corresponding actual risk indicators are extracted, and then combined into a standardized scenario feature-risk vector set after standardization processing. Construct a risk topology self-organizing mapping network containing mapping layer nodes, and define a hydrology-risk joint distance metric as a weighted combination of the spatial distance of hydrological features and the spatial distance of risk indicators; The standardized scene feature-risk vector set is input into the risk topology self-organizing map network for iterative training. In each iteration, the distance between the input vector and the mapping layer node is calculated based on the hydrology-risk joint distance metric, and the node with the smallest distance is selected as the winning node. Construct a risk-layered scenario tree, including: The weight vectors of the mapping layer nodes of the trained risk topology self-organizing mapping network are analyzed. Based on the risk dimension attribute in the weight vectors, the mapping layer nodes are divided into multiple sets of non-overlapping risk level nodes, and the first-level risk hierarchical structure of the scenario tree is established. Within each risk level node set, a secondary clustering analysis is performed based on the hydrological feature dimension attributes in the weight vector to identify the intra-layer feature clusters existing under that risk level and establish the second-level hydrological morphological structure of the scene tree. Based on the first-level risk stratification structure and the second-level hydrological morphology structure, a risk stratification scenario tree is constructed from top to bottom, which sequentially represents the risk level and hydrological morphology. The number of samples covered by each node and the average risk index are counted as the scenario basis for subsequent multi-stage risk budget optimization. Discount parameters refer to the coefficients used to adjust the flood control limit water level or pre-discharge flow rate in the pre-storage and pre-release scheduling rules.

2. The method according to claim 1, characterized in that, Based on basic hydrological operational data, a risk-condition hydrological-risk joint manifold is constructed, and a sample set of risk-condition flood processes is generated, including: Historical inflow sequences that match the preset target hydrological-risk characteristic template are retrieved from basic hydrological operation data and normalized over time, serving as the normalization baseline for the inflow process. The pre-built risk condition morphology transformation parameter mapping model is invoked to map the target hydrological features and target risk levels contained in the target hydrological-risk feature template into morphology transformation parameters; The morphological transformation parameters and the normalized benchmark are input into the risk condition morphological transformation model, which drives the parameterized transformation operator containing physical constraints to perform morphological reconstruction, generating a risk condition flood process sample set that combines the target hydrological statistical characteristics and the target risk level structure.

3. The method according to claim 2, characterized in that, This further includes constructing a target hydrological-risk characteristic template: Perform multidimensional risk index statistics on historical flood events in basic hydrological operation data, analyze risk distribution characteristics, and divide non-overlapping risk level intervals; Within each risk level interval, hydrological feature clustering analysis is performed independently to identify typical hydrological morphological clusters existing under that risk level. The risk level intervals are then structurally bound to the typical hydrological morphological clusters to construct a target hydrological-risk feature template that includes the target risk level and the target typical hydrological features.

4. The method according to claim 3, characterized in that, Further steps include constructing a risk feedback-driven manifold correction closed loop to perform physical simulation verification and adaptive correction on a sample set of flood processes under risk conditions. Input the sample set of flood process under risk conditions into the reservoir group scheduling simulation model, analyze the actual risk indicators under the preset scheduling rules, measure the difference between them and the target risk level contained in the target hydrological-risk characteristic template, and generate a risk deviation statistical set. Based on the risk deviation statistics set, an adaptive correction operation is performed on the data parameters that constitute the risk condition hydrological-risk joint manifold to form an updated risk condition hydrological-risk joint manifold; Based on the updated risk-condition hydrological-risk joint manifold, an iterative generation mechanism is triggered, and the risk-condition morphological transformation parameter mapping model is called again to perform parameter mapping and morphological transformation until the risk deviation statistics set meets the preset convergence condition, and the converged risk-condition flood process sample set is output.

5. The method according to claim 1, characterized in that, This further includes applying risk-sensitive neighborhood update rules to optimize the weight vector distribution of risky topology self-organizing mapping networks: Configure risk-sensitive neighborhood update rules and introduce a risk-sensitive amplification factor into the rules that changes positively with the sample risk level; When updating the weight vectors of the winning node and its neighboring nodes, the risk level contained in the current input vector is analyzed, and the weight vector update amplitude is dynamically adjusted through a risk-sensitive amplification factor so that high-risk samples have a pulling effect on the topology of the risk topology self-organizing mapping network. Through iterative updates, a trained risk topology self-organizing mapping network is formed, in which the weight vectors of the mapping layer nodes present a topological structure in the feature space that takes into account both hydrological morphological similarity and orderly distribution of risk levels.

6. The method according to claim 1, characterized in that, Based on a risk-layered scenario tree, before optimizing discount parameters under the constraints of multi-stage risk budgeting, the process includes constructing a risk budgeting system and calculating risk liability data: Establish risk control standards that include total risk budget indicators and stage importance weights, decompose the total risk budget indicators according to the scheduling stages, and generate a multi-stage risk budget and quota set that includes the risk quota upper limit for each stage; Representative scenarios are extracted from the risk-layered scenario tree. Combined with the preset discount parameter candidate space, the stage risk index difference under the baseline scheduling rule and different discount parameter combinations is calculated to generate a stage risk increment sequence describing the risk change magnitude of each scheduling stage. Based on the phase risk increment sequence and phase importance weight, the contribution ratio of each scheduling phase to the total risk increment is calculated, and a risk responsibility allocation matrix set reflecting the risk burden distribution structure under different discount parameter combinations is constructed.

7. The method according to claim 6, characterized in that, Based on a risk-layered scenario tree, discount parameters are optimized under the constraints of multi-stage risk budgets to select scheduling strategy combinations that meet the stage-specific risk fairness requirements, generating a risk control scenario set, including: A discount parameter optimization model is constructed, with the objective function being to maximize the overall benefit, and the hard constraints being that the total risk budget indicator of multi-stage risk budget and quota concentration does not exceed the limit and the stage risk quota does not exceed the limit. In the discount parameter optimization model, a stage fairness constraint is introduced. The stage risk Gini coefficient or range ratio is calculated based on the risk responsibility sharing matrix set. Solutions with risk responsibility distribution imbalance exceeding the preset fairness threshold are eliminated, and the optimal combination set of discount parameters is obtained. Substitute the optimal combination of discount parameters into the reservoir group scheduling rules, perform multi-stage scheduling simulation on representative scenarios in the risk stratification scenario tree, and output a set of risk control scenarios that satisfy risk budget constraints and stage fairness constraints.

8. The method according to claim 2, characterized in that, The construction and training process of the risk condition morphology transformation model and the risk condition morphology transformation parameter mapping model includes: Define a risk condition morphological transformation operator that includes amplitude scaling parameters, time scale parameters, time translation parameters, and shape correction parameters. Construct a set of physical constraints with the conservation of total flood volume after transformation, non-negative inflow, and smooth water level connection as hard boundaries, and encapsulate them to form a risk condition morphological transformation model. A training sample set is constructed based on historical flood events. For each sample, a parameter optimization problem is constructed with the objective of minimizing the deviation of hydrological characteristics and risk indicators and the constraint condition of the physical constraint set. The conditionally optimal morphological transformation parameters are obtained through numerical solution. Using target hydrological features and target risk level as input features, and conditionally optimal morphological transformation parameters as output labels, a supervised learning algorithm is used to train a neural network to establish a nonlinear mapping relationship from the hydrological-risk joint space to the morphological parameter space, thereby generating a risk conditional morphological transformation parameter mapping model.

Images