Flood resilience evolution simulation method and system based on natural-social element interaction

Abstract

The application discloses a kind of based on natural-social element interaction's flood prevention resilience evolution simulation method and system, the method includes: obtaining the natural element space-time sequence and social element space-time sequence of target area, and it is aligned to the space-time grid node of uniformity, obtain heterogeneous node sequence;The time-varying interaction intensity between nodes and the interaction uncertainty of time-varying interaction intensity are calculated;Using pre-set rule compiler, the time-varying rule parameter between each pair of interaction nodes is compiled to generate;Time-varying rule parameter is input into pre-constructed resilience dynamics model, drives to update the resilience state value of each node at continuous time step, obtains flood prevention resilience evolution trajectory.The application solves the problem of data driving and mechanism model in the prior art by rule compiling mechanism, realizes the dynamic transformation of interaction characteristics from soft weight to hard rule, effectively improves the accuracy and interpretability of flood prevention resilience evaluation under complex time-varying scenario.

Description

Technical Field

[0001] This invention belongs to the field of urban flood disaster prevention and control, and in particular, it is a method and system for simulating the evolution of flood resilience based on the interaction of natural and social elements. Background Technology

[0002] With the intensification of global climate change and the rapid advancement of urbanization, urban areas face increasingly severe flood risks. Traditional flood control systems often rely on single hydrological and hydraulic models, neglecting the complex interactions and coupling effects between natural and social factors. Against this backdrop, urban flood resilience faces challenges, especially under multiple pressures (such as extreme weather events, population growth, and aging infrastructure), making it difficult to accurately identify the critical state of flood control systems, and the optimization of flood control strategies lacks dynamism and flexibility.

[0003] Existing technologies typically employ static models or traditional physical simulation methods, which, while capable of reflecting the hydrological processes of floods and the effectiveness of flood control measures to some extent, fail to fully consider the impact of social factors such as economic activities, social organization, and emergency response capabilities on flood control systems. Furthermore, traditional methods lack the ability to dynamically assess flood resilience and predict multiple scenarios, making it difficult to make accurate decisions in complex and uncertain environments.

[0004] In practical applications, existing technologies suffer from the following key problems: Natural flood processes and social response behaviors are modeled separately, neglecting the dynamic feedback of both and failing to accurately simulate the comprehensive response of flood control systems. Furthermore, flood resilience assessments often rely on static indicators, failing to consider the spatiotemporal evolution before, during, and after disasters, resulting in an inability to predict resilience critical points and lagging flood control strategies that are difficult to optimize. Moreover, existing flood scenario simulations are based solely on historical recurrence periods, ignoring the uncertainties of extreme events and social factors under climate change. Strategy optimization depends on expert experience and fails to effectively integrate multiple measures, leading to deviations between actual decision-making outcomes and expectations. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for simulating flood resilience evolution based on the interaction of natural and social elements, so as to solve the above-mentioned problems existing in the prior art.

[0006] The technical solution, based on a simulation method for flood resilience evolution based on the interaction of natural and social factors, includes:

[0007] Obtain the spatiotemporal sequences of natural and social elements in the target area and align them to a unified spatiotemporal grid node to obtain a heterogeneous node sequence.

[0008] Based on heterogeneous node sequences, the time-varying interaction strength between nodes and the interaction uncertainty of the time-varying interaction strength are calculated.

[0009] Using a pre-built rule compiler, time-varying rule parameters between each pair of interactive nodes are generated based on the time-varying interaction strength and interaction uncertainty.

[0010] By inputting time-varying rule parameters into a pre-built resilience dynamics model, the resilience state values ​​of each node at continuous time steps are updated, thus obtaining the flood control resilience evolution trajectory.

[0011] According to one aspect of this application, a flood resilience evolution simulation system based on the interaction of natural and social elements includes:

[0012] The data alignment module is used to acquire the spatiotemporal sequences of natural elements and social elements in the target area, and align them to a unified spatiotemporal grid node to obtain a heterogeneous node sequence.

[0013] The interactive computing module is used to calculate the time-varying interaction strength between nodes and the interaction uncertainty of the time-varying interaction strength based on the heterogeneous node sequence.

[0014] The rule compilation module is used to call the pre-built rule compiler to compile and generate time-varying rule parameters between each pair of interactive nodes based on the time-varying interaction strength and interaction uncertainty.

[0015] The evolution simulation module is used to input time-varying rule parameters into a pre-built resilience dynamics model, drive the update of the resilience state value of each node in continuous time steps, and obtain the flood control resilience evolution trajectory.

[0016] Beneficial effects: This invention solves the problem of the separation between data-driven and mechanism models in the prior art through a rule compilation mechanism, realizes the dynamic transformation of interactive features from soft weights to hard rules, and effectively improves the accuracy and interpretability of flood resilience assessment in complex time-varying scenarios. Attached Figure Description

[0017] Figure 1 A flowchart illustrating the steps of a flood resilience evolution simulation method based on the interaction of natural and social elements, provided in this application embodiment.

[0018] Figure 2 A flowchart illustrating the steps for calculating the time-varying interaction intensity between computing nodes as provided in this application embodiment.

[0019] Figure 3 A flowchart illustrating the steps for calculating the attention weight of each potential connection edge, provided in an embodiment of this application.

[0020] Figure 4 A flowchart illustrating the steps for calculating the interaction uncertainty of time-varying interaction intensity in an embodiment of this application. Detailed Implementation

[0021] 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.

[0022] It should be noted that the terms include and have, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0023] The study revealed significant shortcomings in existing technologies for handling natural-social dynamic feedback mechanisms and model interpretability. The main problems lie in the semantic gap between implicit features mined by deep learning and physically simulated entities, and the static coupling's neglect of the time-varying characteristics of disaster evolution. Specifically, firstly, there is a lack of effective mechanisms to transform data-driven nonlinear interaction features into interpretable and executable dynamic rules, preventing the weights of the black-box model from directly driving the mechanism-based agent evolution; secondly, existing coupled models struggle to dynamically quantify the interaction intensity and its uncertainty over time, leading to time lags or distortions when simulating complex adaptive behaviors such as hazard-response-feedback, making it difficult to maintain generalization ability and prediction accuracy under different flood scenarios.

[0024] like Figure 1 As shown, a simulation method for flood resilience evolution based on the interaction of natural and social factors is proposed, including the following steps:

[0025] Obtain the spatiotemporal sequences of natural elements and social elements in the target area, and align the spatiotemporal sequences of natural elements and social elements to a unified spatiotemporal grid node to obtain a heterogeneous node sequence.

[0026] In this embodiment, a spatiotemporal grid node refers to a regular grid unit that divides the research area into, for example, 50m×50m or 100m×100m. Each grid serves as a basic computational unit, i.e., an agent node. Optionally, the heterogeneous node sequence consists of four types of agents with different attributes, including environmental agents, social agents, facility agents, and management agents. Specifically, the state vector x of the environmental agent... env Includes attributes such as water level, flow velocity, and cumulative rainfall; the state vector x of the social intelligent agent. socIncludes attributes such as population density, migration intention, and risk perception; the state vector x of the facility agent. infra Includes attributes such as functional status (e.g., normal, damaged, failed), defense level, and current capacity; the state vector x of the management agent. gov It includes attributes such as emergency supplies reserves, response speed, and dispatch efficiency.

[0027] Specifically, acquiring the spatiotemporal sequence of natural elements involves the collection and fusion of multi-source data. For example, land cover types can be inverted using satellite remote sensing data; real-time rainfall intensity data can be obtained using meteorological radar or ground observation stations; and topographic elevation and slope information can be extracted using a digital elevation model (DEM). Preferably, for hydrological data, hydrodynamic models such as HEC-RAS or MIKE11 can be used to perform preliminary simulations of the precipitation-runoff process within the watershed, obtaining estimated values ​​of water level and flow velocity at different times for each grid node, which serve as the initial input for the environmental agent. Natural element data is spatially interpolated and temporally resampled, and then uniformly mapped onto a preset spatiotemporal grid. For example, for acquiring the spatiotemporal sequence of natural elements, a temporal resolution of 1 hour and a spatial resolution of 100m × 100m grid are used. The update frequency of meteorological radar rainfall data is 6 minutes, and after temporal aggregation, it is mapped to a 1-hour time step; the acquisition frequency of water level telemetry data is 5 minutes, and it is also temporally resampled.

[0028] Specifically, big data technology can be used to obtain the spatiotemporal sequence of social elements. For example, based on mobile signaling data provided by mobile communication operators, the population density distribution within each grid can be estimated in real time through dwell point analysis algorithms; based on point of interest (POI) data and basic geographic information data, the location and attributes of key infrastructure such as roads, bridges, hospitals, and schools can be identified; and based on nighttime light data or regional economic statistical yearbooks, the intensity of economic activity in each grid unit can be estimated. In addition, web crawling technology can be used to crawl disaster reports and requests for help on social media as a supplement to social perception data.

[0029] Furthermore, to achieve alignment, the system establishes a unified spatiotemporal coordinate system. Spatially, projection transformation is used to unify all geographic data to the same projected coordinate system, and the data is reorganized according to the grid index. Temporally, a unified time step is set, such as 1 hour or 10 minutes, and linear interpolation or aggregation is performed on data with different sampling frequencies so that all elements have corresponding values ​​on the same time segment.

[0030] In some preferred embodiments, to construct a more realistic physical-social dual-layer coupled network, the connection edges between nodes are defined to include not only physical connections, such as road connectivity and water flow direction, but also social connections, such as information propagation paths and organizational affiliations. This dual-layer network structure lays the topological foundation for subsequent simulations of the interaction between flood spread and information diffusion. Furthermore, in the dual-layer coupled network, when the management agent (social layer node) makes a flood diversion decision based on the current disaster assessment, this decision is not merely an information signal but also directly modifies the attribute parameters of the physical layer environmental agent. Specifically, this interaction rule is defined as follows: once the decision variable of the management node is set to 1 (activated), the corresponding hydrological model boundary conditions change; for example, the maximum flow capacity parameter of a river cross-section is forcibly modified to the value after flood diversion. Through cross-layer parameter dynamic binding, the reverse intervention effect of human social behavior on natural hydrological processes is accurately simulated.

[0031] Preferably, the spatiotemporal sequence of natural elements includes, but is not limited to: watershed precipitation intensity, river water level and flow velocity, topographic elevation data and soil permeability; the spatiotemporal sequence of social elements includes, but is not limited to: regional population density, infrastructure vulnerability level, intensity of economic activity and distribution of emergency resources.

[0032] Based on heterogeneous node sequences, the time-varying interaction strength between nodes and the interaction uncertainty of the time-varying interaction strength are calculated.

[0033] Specifically, time-varying interaction relationships are established between natural and social elements, as well as between nature and nature, and between society and society, and quantified as interaction strength, reliability, and uncertainty that change over time. Time-varying interaction strength can quantitatively describe the impact of a state change at a specific time step t on the resilience-related state of node B. Considering that the interaction structure may differ in different stages of a flood process (warning period, overflow period, and receding period), interaction strength can be estimated separately for each time step, or a rolling value can be calculated based on a time window. Interaction uncertainty can describe the confidence level, stability, and reliability of the interaction strength estimate. Uncertainty from different sources, such as measurement error, model error, and scenario error, can be decomposed to form a decomposed uncertainty index.

[0034] Using a pre-built rule compiler, time-varying rule parameters between each pair of interactive nodes are generated based on the time-varying interaction strength and interaction uncertainty.

[0035] Alternatively, it can be said that by calling a pre-configured rule compiler, based on the time-varying interaction strength and interaction uncertainty, and through gating functions and parameter mapping, time-varying rule parameters between each pair of interaction nodes are compiled and generated.

[0036] In this embodiment, the abstract interaction strength and uncertainty are translated into formalized time-varying rule parameters that can be directly invoked by the resilience dynamics model. The rule compiler is a software module that transforms the implicit weights output by the data-driven model into executable rule parameters. This module receives time-varying interaction strength and interaction uncertainty as input, and through gating functions and parameter mapping mechanisms, outputs time-varying rule parameters that can be directly used to drive the dynamics model, thus realizing the transformation of soft weights into hard rules.

[0037] By inputting time-varying rule parameters into a pre-built resilience dynamics model, the resilience state values ​​of each node at continuous time steps are updated, thus obtaining the flood control resilience evolution trajectory.

[0038] In this embodiment, the toughness state value R i (t) is not an abstract scalar, but a comprehensive score calculated based on a multidimensional flood control resilience evaluation index system. The multidimensional flood control resilience evaluation index system serves as the node's own supporting term E in the dynamic model. i The calculation basis of (t) includes three dimensions: resilience, adaptability, and recovery capability. Specifically, the resilience index E resist It can be quantified through flood control engineering standards and early warning response capabilities. For example, for a facility intelligence agent, its resilience can be calculated using the following formula:

[0039] E resist = w1 * (H defense / H water ) + w2 * A warning ;

[0040] Where H defense H is the design defense water level for dikes or facilities. water A represents the current actual flood level. warning The value ranges from 0 to 1, representing the coverage or response efficiency of the early warning system. w1 and w2 are the corresponding weighting coefficients. When the water level exceeds the defense standard, the resilience index will decrease. Adaptability index E adapt This reflects the social system's ability to adapt during disasters. For example, it can be quantified through social learning capacity and technological innovation capacity. Its calculation formula can involve the frequency of flood prevention plan updates (times / year) based on historical disaster data and the public's flood prevention knowledge dissemination rate (percentage). In the model, it is reflected in the efficiency of social agents in adjusting their behavioral patterns, such as changing evacuation routes, when facing risks. Resilience index E recover The focus is on assessing the speed of post-disaster reconstruction and recovery. For example, for the economic system, this can be quantified by the projected timeframe for the affected area's GDP to recover to pre-disaster levels. A possible formula is:

[0041] E recover = 1 / (Tgdp + c);

[0042] Where T gdp The number of months required for recovery is estimated, where c is a constant. For lifeline projects (water supply, power supply), this can be represented by the reciprocal of the Mean Time to Repair (MTTR). After normalization, the quantified indicators serve as inputs to the kinetic equations, determining the slope of the resilience curve's recovery after the disaster. This embodiment completes the transformation from raw multi-source data to standardized heterogeneous node sequences and fundamental resilience indicators, clarifying the specific physical meaning and quantification methods of the system inputs, providing solid data support for subsequent processing of interaction relationships and driving dynamic evolution.

[0043] Optionally, in the flood resilience evaluation index system, the determination of the weights of each level of index can make the final resilience state value R... i (t) More accurate. A combined subjective and objective weighting method is preferred for calculating the weights of each indicator, and a time decay factor is introduced to adapt to the characteristics of different stages of a disaster. Specifically, for each indicator, such as flood control engineering standards and the completeness of the early warning system, the objective weight is calculated using the entropy weight method. The system collects normalized data sequences of the indicator in similar historical flood events and calculates its information entropy. The smaller the information entropy of an indicator, the greater its degree of variation and the more information it provides, thus assigning a larger objective weight. Simultaneously, the analytic hierarchy process (AHP) is used to determine the subjective weights. Experts in flood control are invited to compare and score the relative importance of each indicator pairwise, constructing a judgment matrix and calculating the eigenvector corresponding to its largest eigenvalue. After consistency testing, the subjective weights are obtained. The final comprehensive weight is the weighted sum of the objective and subjective weights. Furthermore, considering the timeliness of disaster evolution, a time decay factor can be introduced. In the pre-disaster prevention stage, the weight of resilience indicators is emphasized; in the disaster response stage, the weight of adaptability and emergency response indicators is dynamically increased; and in the post-disaster recovery stage, the weight of recovery capacity indicators is emphasized. The system adjusts the weight distribution ratio of each dimension in real time according to the current simulation time progress to ensure that the evaluation results match the actual needs.

[0044] This embodiment is not only applicable to urban flooding scenarios, but can also be extended to the following scenarios: watershed-scale flood resilience assessment, in which case the grid resolution can be adjusted to the 1km×1km level; storm surge disaster simulation in coastal cities, which requires expanding the input elements to include marine meteorological data such as tide level and wind speed; and flash flood and debris flow disaster early warning, which requires introducing geological elements such as slope and vegetation cover. The rule template library needs to be adjusted accordingly for different scenarios, but the overall technical framework remains unchanged.

[0045] like Figure 2 As shown, in one possible implementation, calculating the time-varying interaction strength between nodes over time includes:

[0046] Construct a heterogeneous element interaction graph containing natural nodes, social nodes, and facility nodes, and define the potential connection edges between nodes.

[0047] In other words, based on the heterogeneous node sequence, a heterogeneous element interaction graph is constructed, which includes natural nodes, social nodes, and facility nodes represented by the infrastructure vulnerability level in the spatiotemporal sequence of social elements. Potential connection edges between nodes are defined to obtain a graph structure with edge connection relationships.

[0048] In this embodiment, the heterogeneous element interaction graph is typically represented as G=(V, E), where V is the set of nodes and E is the set of edges. To be processed by a graph neural network, an adjacency matrix A needs to be defined. Unlike traditional binary adjacency matrices, the adjacency matrix in this embodiment is constructed as a weighted matrix considering both spatial distance and functional dependency. Specifically, for spatial adjacency relationships, if the Euclidean distance between two grid nodes is less than a preset threshold, such as 500 meters, a potential connection edge is considered to exist; for functional dependency relationships, if there is an upstream / downstream relationship in water flow or a road connection, a connection edge is established regardless of the physical distance.

[0049] Furthermore, to differentiate between different types of interactions, such as flood impacting facilities versus facilities supporting population, a heterogeneous graph approach is adopted, assigning a specific type label k to each type of edge. For example, type k=1 represents the influence of natural nodes pointing to facility nodes, while type k=2 represents the feedback of social nodes pointing to management nodes. This allows the model to learn different parameter weights for different interaction patterns, avoiding information confusion caused by modeling a single relationship.

[0050] The heterogeneous element interaction graph is input into a pre-built spatiotemporal graph neural network model. The spatial features of nodes are captured by graph convolutional layers, and the temporal dependencies of nodes are captured by temporal convolutional layers.

[0051] In other words, the heterogeneous element interaction graph is input into a pre-set spatiotemporal graph neural network model. The spatial features of the nodes are captured by the graph convolutional layer, and the temporal dependencies of the nodes are captured by the temporal convolutional layer, thus obtaining the spatiotemporal feature representation of the nodes.

[0052] In this embodiment, the Spatiotemporal Graph Neural Network (ST-GCN) model adopts an architecture consisting of alternating stacks of spatial graph convolutional layers and temporal convolutional layers. Specifically, for spatial feature extraction, graph convolution operators based on Chebyshev polynomial approximation or spatial domain aggregation operators can be used. For temporal feature extraction, one-dimensional dilated convolution sliding along the time axis is preferred to expand the temporal receptive field without significantly increasing computational cost, thereby capturing long-term flood evolution trends.

[0053] Based on the spatial features and temporal dependencies of nodes, the attention mechanism of the spatiotemporal graph neural network model is used to calculate the attention weight of each potential connection edge, and the normalized attention weight is used as the time-varying interaction intensity.

[0054] Specifically, to achieve dynamic quantization of interaction intensity, a graph attention mechanism (GAT) is embedded in the graph convolutional layer. For any edge of type k connecting source node i and target node j, the node features can be mapped to a high-dimensional feature space through a linear transformation.

[0055] like Figure 3 As shown, in a preferred implementation, the attention weight for each potential connection edge is calculated, including:

[0056] For each potential connection edge, an unnormalized attention score is calculated using a learnable linear transformation matrix, based on the source node features, target node features, and edge type features connected by each potential connection edge.

[0057] In other words, for each potential connection edge, based on the source node features, target node features, and edge type features of its connection in the spatiotemporal feature representation, an unnormalized attention score is calculated using a learnable linear transformation matrix.

[0058] In this embodiment, the unnormalized attention score s ijk The formula for calculating (t) can be:

[0059] s ijk (t) = LeakyReLU( a k T [W] k h i (t) || W k h j (t) ||u k e ij (t)]);

[0060] Where h i (t) and h j (t) represent the feature vectors of source node i and target node j at time t, respectively; W k Let be a learnable linear transformation matrix for the corresponding edge type k, used to map the feature dimension to the attention space; || denotes the vector concatenation operation; a k Let be the learnable attention vector corresponding to edge type k; LeakyReLU is a non-linear activation function used to introduce non-linear characteristics and prevent gradient vanishing. T For transpose; u k The mapping weights for edge features; e ij(t) represents the static or dynamic features of edge ij, such as distance and connectivity. By evaluating the degree of matching between the features of two nodes under a specific relationship type, the potential influence of the source node on the target node can be preliminarily determined.

[0061] The attention scores of all potential connection edges connected to the same target node are normalized using Softmax to obtain the normalization coefficients.

[0062] Specifically, to make the influence of different neighboring nodes comparable, the scores need to be normalized. The calculation formula is as follows:

[0063] α ijk (t) = exp(s ijk (t)) / Σ m∈Ni exp(s imk (t));

[0064] Where N i α represents the set of all neighboring nodes of node i; ijk (t) is the normalization coefficient. This ensures that for any target node, the sum of the weights of all its input edges is 1, thus forming an effective probability distribution or weight distribution.

[0065] The normalized coefficients are used as attention weights to quantify the time-varying interaction intensity of the source node to the target node.

[0066] In this embodiment, the calculated α ijk (t) is defined as the time-varying interaction intensity W ij (t). For example, if the calculation results show that the interaction strength between the flood node and the dam node is 0.9 at a certain moment, it physically means that the state of the flood at that moment will dominate the state change of the dam, such as a surge in water level that may lead to dam failure.

[0067] For example, W ij (t)=clip((1 / H) *∑ h=1 H α ijk，h (t), 0, 1);

[0068] Among them W ij (t) represents the time-varying interaction strength between node i and node j in the final output; H represents the total number of attention heads; α ijk，h (t) is the normalization coefficient calculated by the h-th attention head; clip(…, 0, 1) is the truncation function to ensure that the output value is between 0 and 1.

[0069] In a detailed numerical example, suppose node i (the flooded area) has two neighboring nodes: node j (dam A) and node k (dam B). After linear transformation and activation function, the unnormalized attention score s is calculated. ij =2.5, s ik =1.2. Perform Softmax normalization: α ij = exp(2.5) / (exp(2.5)+exp(1.2)) = 12.18 / (12.18+3.32) = 0.786, α ik = 3.32 / 15.50 = 0.214. This indicates that the impact intensity of the flood node on dam A (0.786) is higher than that on dam B (0.214), possibly because dam A is closer to the center of the flood or the water flow direction is towards dam A.

[0070] like Figure 4 As shown, in a further implementation, the interaction uncertainty of the time-varying interaction strength is calculated, including:

[0071] A Monte Carlo random deactivation strategy is applied to the spatiotemporal graph neural network model to perform a predetermined number of forward inferences, resulting in a predetermined number of attention weight samples for each potential connection edge.

[0072] In other words, by enabling the Monte Carlo random deactivation strategy for the spatiotemporal graph neural network model, the random deactivation layer is kept on during the inference phase, and multiple forward inferences are performed to obtain multiple attention weight samples for each potential connection edge.

[0073] Calculate the variance of a predetermined number of attention weight samples and use the variance as the interaction uncertainty of potential connection edges.

[0074] Specifically, in flood control decision-making, simply knowing the interaction strength is insufficient; it's also necessary to know the degree of certainty in this judgment. This embodiment employs Monte Carlo Dropout (MC Dropout) as a Bayesian approximation method. For example, during the model inference (prediction) phase, the Dropout layer (randomly deactivated layer) used during training is not disabled but remains active, for example, with a deactivation rate set to 0.2, and the same input data is processed B times, for example, 50 times, repeating the forward propagation. Due to the randomness of Dropout, the attention weight W obtained in each forward propagation... ij (t) will all be slightly different, thus forming the sample set {W} ij (1) (t), W ij (2) (t), ..., W ij (B) (t)}. By calculating the variance Var(W) of this set. ij(t)) can then be used to obtain the interactive uncertainty U. ij (t). If the variance is large, it indicates that the model's judgment of this interaction relationship is very unstable, possibly because there is a lack of similar scenarios in the training data. The subsequent rule compiler will suppress the influence of this interaction to avoid the risk of false alarms. Optionally, a multi-model ensemble strategy can also be used for a predetermined number of forward inferences. This embodiment is suitable for urban flood control scenarios with sufficient data and highly complex interaction patterns.

[0075] For example, the average interaction strength obtained from multiple inferences is: μ ij (t) = (1 / B)* ∑ b=1 B W ij b (t);

[0076] The variance of the interaction strength is: σ ij (t) = (1 / (B-1)) * ∑ b=1 B (W ij b (t) - μ ij (t)) 2 ;

[0077] The final output interaction uncertainty is: U ij (t) = clip(γ*σ ij (t), 0, 1);

[0078] Where B represents the number of Monte Carlo samplings or the number of ensemble models; W ij b(t) is the interaction strength sample value obtained from the b-th inference; γ is the adjustment coefficient.

[0079] In another possible implementation, the time-varying interaction strength between nodes can also be calculated as follows:

[0080] Define structural variables for each node in the heterogeneous node sequence and construct a time-varying structural equation model with time delay terms to describe the causal dependencies between structural variables.

[0081] In this embodiment, the structural variable z i (t) represents key proxy variables selected from the node features. For example, for environmental nodes, water level is selected as a structural variable; for facility nodes, functional loss rate is selected as a structural variable. The specific form of the constructed time-varying structural equation model is as follows:

[0082] z j (t+Δt) =Σ i∈Nj A ij (t) z i(t) + b j T u j (t) +ε j (t);

[0083] Where z j (t) represents the structural variable value of the target node j at the next time step; z i (t) represents the structural variable value of source node i at the current time; Δt is the time step; N j Let A be the set of associated nodes of node j; ij (t) represents the time-varying structure coefficients of source node i with respect to target node j at time t; b j T u is the transpose of the exogenous input coefficient vector; j (t) represents exogenous input variables, such as rainfall-driven variables; ε j (t) represents the random error term.

[0084] Within a sliding window moving along the time axis, the time-varying structure equation model is solved to obtain the time-varying structure coefficients between nodes.

[0085] In other words, a sliding window with a preset window length is set, and within the sliding window that moves along the time axis, the time-varying structure equation model is solved to obtain the time-varying structure coefficients between nodes.

[0086] Specifically, to capture the dynamic evolution of interactions over time—for example, a dam that can block floods when intact but undergoes abrupt changes in its relationships after breach—a sliding window technique can be used. The window length is set to L, for example, 24 hours, and the window moves forward one step along the time axis each time. Within each window [tL, t], assuming the structural coefficient A... ij It is relatively stable, and regression solutions are performed using a subset of the observed data within this window.

[0087] In a preferred implementation, the time-varying structure coefficients are obtained, including:

[0088] Based on the time-varying structural equation model, an optimization objective function is constructed, which includes a prediction error term for structural variables and an L1 regularization term for time-varying structural coefficients.

[0089] Minimize the objective function to obtain a sparse solution for the time-varying structural coefficients.

[0090] In this embodiment, to filter out the truly effective critical paths from numerous potential connections, the idea of ​​minimum absolute shrinkage and Lasso regression is introduced. The optimization objective function J(A) is defined as:

[0091] J(A) =Σk=t-L t || z(k) - A z(k-τ)|| 2 +λ||A||1;

[0092] The first term is the prediction error term, used to ensure the accuracy of the model's fit to the data, where τ is the lag time and z(k) is the value of the structural variable. The second term is the L1 regularization term, where ||A||1 represents the sum of the absolute values ​​of all elements in the structural coefficient matrix; λ is the regularization parameter. Due to the geometric properties of L1 regularization, it tends to compress unimportant coefficients directly to 0, thus obtaining a sparse solution. Physically, this corresponds to eliminating spurious or weak interactions while retaining the main causal chains.

[0093] In another possible embodiment, the minimization of the optimization objective function can also be expressed as:

[0094] min A(t) ∑ k=t-L t ||z(k +Δt) - A(t)z(k) - Bu(k) ||2 2 +λ* ||A(t) ||1 + ξ *||A(t) - A(t-Δt)|| F 2 ;

[0095] Where A(t) are time-varying structure coefficients; min represents minimizing the objective function; L is the sliding window length; || ||2 2 ξ is the squared L2 norm, i.e., the prediction error term; λ is the L1 regularization coefficient; ||A(t)||1 is the L1 norm of the structure coefficient matrix, i.e., the induced sparsity; ξ is the smoothing constraint coefficient; ||A(t) - A(t-Δt)|| F 2 is the squared Frobenius norm, used to smooth the change of constraint coefficients over time; Bu(k) is the linear effect term of exogenous inputs on structural variables.

[0096] The absolute values ​​of the time-varying structural coefficients are normalized, and the normalized results are used as the time-varying interaction intensity.

[0097] Specifically, the solution obtained A ij (t) may contain negative values, indicating a suppressive effect, and its dimensions are not uniform. To ensure compatibility with the rule compiler interface, its absolute value is taken and normalized, i.e., W. ij (t)= |A ij (t) | / Σ m |A mj (t)|.

[0098] In a further implementation, the interaction uncertainty of the time-varying interaction strength can also be calculated as follows:

[0099] Bootstrap resampling is performed on the data within the sliding window to generate a predetermined number of data subsamples.

[0100] Solve the time-varying structure equation model for each data subsample to obtain the distribution set of time-varying structure coefficients;

[0101] Calculate the variance of the distribution set and map the variance to interactive uncertainty.

[0102] In this embodiment, to quantify the uncertainty of statistical inference, a bootstrap resampling method can be used. Specifically, in each sliding window's dataset, N sets of samples (e.g., N=100) are randomly drawn with replacement, forming multiple data subsets. The Lasso algorithm is run independently for each subset to obtain the coefficients A. ij The empirical distribution of (t). The variance (or confidence interval width) of this distribution reflects the statistical significance and stability of the causal relationship, and is mapped to the interaction uncertainty U. ij (t) can provide a basis for confidence in subsequent decisions.

[0103] In a preferred embodiment, a hybrid system can also be constructed, employing structural equation modeling (SEM) for critical nodes (such as dams) to achieve high interpretability, while using graph neural networks (GNN) for general large-area nodes (such as urban grids) to achieve high accuracy. The W outputs from both methods... ij (t) and U ij (t) Unify the input of subsequent rule compilers.

[0104] This embodiment uses a time-varying structural equation model (TV-SEM), which is suitable for environments with small sample data or scenarios with extremely high requirements for model interpretability, such as accident investigations that require clear attribution.

[0105] According to one aspect of this application, a pre-built rule compiler is used to compile and generate time-varying rule parameters between each pair of interactive nodes based on the time-varying interaction strength and interaction uncertainty, including calculating the gating coefficients, specifically:

[0106] A gating function is constructed, which is configured to be positively correlated with the time-varying interaction intensity and negatively correlated with the interaction uncertainty.

[0107] In this embodiment, the gate function g ij (t) acts as a soft switch, determining whether the previously calculated interactions should be adopted and incorporated into subsequent dynamical evolution. For example, the mathematical form of the gating function can be:

[0108] g ij (t) = σ(k1 * W ij (t) - k2 * U ij (t) - k3);

[0109] Where g ij (t) represents the gating coefficient between node i and node j; σ() is the Sigmoid activation function, used to restrict the output to between 0 and 1; k1 and k2 are positive weight coefficients, which control the time-varying interaction strength W respectively. ij The positive contribution of (t) and the interactive uncertainty U ij The negative penalty of (t); k3 is a bias term used to set the activation threshold. g is only effective when the time-varying interaction strength between the two nodes is sufficiently large, and the model is sufficiently confident in its judgment of this interaction, i.e., the uncertainty is low. ij The value of (t) will only approach 1 if the interaction is weak or the model's judgment is extremely unstable (i.e., large variance), thus allowing the rule to take effect; otherwise, if the interaction is weak or the model's judgment is extremely unstable (i.e., large variance), g ij (t) will approach 0, thus masking the noise interference. Optionally, the weighting coefficient k1 typically ranges from 5 to 20, with a larger value indicating a stronger tendency to retain high-intensity interactions; the weighting coefficient k2 typically takes a value close to k1 to balance the trade-off between intensity and confidence; the bias term k3 typically ranges from 3 to 8 and is used to set the activation threshold. In practical applications, parameters can be tuned through grid search on the validation set. Parameter sensitivity analysis shows that when the k2 / k1 ratio increases, the system tends to be conservative, and more interactions are suppressed; when k3 increases, the activation threshold increases, and only the strongest interactions are retained.

[0110] By substituting the time-varying interaction intensity and interaction uncertainty into the gate function, the gate coefficient of each pair of interaction nodes is calculated, so as to suppress high-uncertainty interaction while retaining high-intensity interaction.

[0111] For example, suppose the scenario is: node A (flood-inundated area) affects node B (a dam). Set the model parameters k1=10, k2=10, k3=5. Case 1 (high intensity, low uncertainty): the time-varying interaction intensity W output by the GNN model. AB =0.9, interactive uncertainty U AB =0.1. Substituting into the formula, we get: x = k1 * W ij (t) - k2 * U ij (t) - k3 = 10*0.9 - 10*0.1 - 5 = 3. Therefore, the gating coefficient g is... AB=σ(3) ≈ 0.95. The results show that the interaction is strongly preserved, and the threat of flood to the levee is confirmed. Case 2 (high intensity, high uncertainty): GNN model output W AB =0.9, but due to severe data gaps in this region, the uncertainty soared to U. AB =0.8. Substituting into the formula, we get: x = k1 * W ij (t) - k2 * U ij (t) -k3 = 10*0.9 - 10*0.8 - 5 = -4. Therefore, the gating coefficient g... AB =σ(-4) ≈ 0.018. The results show that although the interaction appears strong, the compiler automatically suppresses it due to low confidence, preventing the system from issuing erroneous dam-break alarms due to misjudgment. Furthermore, time-varying rule parameters are obtained by modulating the parameters of the pre-stored rule template based on the gating coefficients.

[0112] In a further embodiment, the compilation and generation of time-varying rule parameters also includes:

[0113] Retrieve the pre-stored rule template and the corresponding basic parameter vector.

[0114] In this embodiment, the rule template R k These are predefined parameterized physical or social rules. For example, for an interaction of type k where a flood causes damage to facilities, the rule template could be defined as the damage probability P. dam Functional form:

[0115] p dam，i (t) =σ(θ0 +θ1 * (h i (t) - h0) +θ2 * v i (t) );

[0116] Where p dam，i (t) represents the real-time failure probability of facility i; θ0, θ1, and θ2 are the compiled rule parameter components, i.e., the basic parameter vector, which are usually set by historical experience or specifications and stored in the database as default values; h i (t) represents the current water level; h0 represents the defense elevation; v i (t) represents the current flow rate. Optionally, the rule template can also define a flow distribution coefficient q. m Functional form:

[0117] q m (t) = exp( q m0 +∑ j g mj (t) * κ mj (t) ) / ∑r exp( q r0 + ∑ j g rj (t) * κ rj (t));

[0118] Where q m (t) represents the flow distribution ratio in the flood diversion area or river branch m; q m0 Assign parameters initially; g mj (t) represents the gating influence coefficient of management decision node j on branch m; κ mj (t) represents the decision adjustment strength; the denominator is the normalized summation over all branches r.

[0119] For example, the pre-stored rule template library includes, but is not limited to, the following types: Facility damage rules: used to describe the probability of flood damage to infrastructure, with inputs being water level difference and flow velocity, and output being the probability of damage; Population evacuation rules: used to describe residents' evacuation behavior, with inputs being risk perception and early warning information, and output being the probability of evacuation; Resource allocation rules: used to describe the allocation efficiency of emergency supplies, with inputs being the amount of supplies in reserve and transportation distance, and output being the arrival time; Cascade failure rules: used to describe the propagation of cascading failures between facilities, with inputs being the status and connection strength of upstream facilities, and output being the degree of impact on downstream facilities.

[0120] The feature interaction terms are obtained and constructed based on the node features of the interactive nodes; the feature interaction terms and gating coefficients are weighted and combined using a mapping matrix, and the weighted combination result is superimposed on the basic parameter vector to obtain the time-varying rule parameters.

[0121] Alternatively, based on the node features of interactive nodes, feature interaction terms can be constructed through the difference, ratio, or vector concatenation operations of node features.

[0122] In this embodiment, the static template is transformed into a dynamic instance. The specific calculation formula is as follows:

[0123] θ k，ij (t) = θ k，base + g ij (t) * B k * φ(x i x j );

[0124] Where θ k，ij (t) represents the time-varying rule parameters for edge ij generated during compilation; θ k，base Let g be the pre-defined basic parameter vector for this type of rule; ij (t) is the gating coefficient; φ(x) i x j) represents the feature interaction term, which can be constructed as the difference, ratio, or concatenation vector of node features; B k x is a learnable feature mapping matrix; i x j These represent the real-time state characteristics of nodes i and j, respectively. The system can dynamically fine-tune the rule parameters θ based on the real-time flood situation, facility status, and the reliability of the interaction. For example, in the case of severe rainstorms, the probability of damage P can be temporarily increased by parameter superposition. dam The θ1 value in the formula simulates the nonlinear physical characteristics that make facilities more susceptible to damage under extreme pressure.

[0125] In one embodiment of this application, time-varying rule parameters are input into a pre-built toughness dynamics model to drive the updating of the toughness state values ​​of each node at consecutive time steps, including:

[0126] A discrete dynamic equation is constructed, which includes a node’s self-support term, a disaster load term, and a neighborhood interaction term weighted by time-varying rule parameters.

[0127] In other words, a discrete dynamic equation is constructed, which includes the node's self-support term calculated based on the node's flood control capacity index, the disaster load term determined based on the current flood state, and the neighborhood interaction term weighted by time-varying rule parameters.

[0128] In this embodiment, the resilience state evolution of each agent node follows the following discrete dynamic equation:

[0129] R i (t+Δt) = R i (t) +Δt * [α i * E i (t) -β i * D i (t) +Σ j∈Ni g ij (t) * F k (R j R i ;θ k，ij ) ];

[0130] Where R i (t) represents the resilience state value of node i at time t; Δt is the time step; E i (t) represents its own supporting components, such as flood control standards and material reserves; D i (t) represents the disaster load, such as inundation depth and rainfall intensity; α i and β i These are the influence coefficients for the self-supporting item and the disaster load item, respectively; Σ j∈Ni g ij(t) * F k (R j R i ;θ k，ij ) represents the neighborhood interaction term, N i Let F be the set of neighboring nodes of node i. k For the time-varying rule parameter θ k，ij The interaction function of control, such as cascading failure based on the probability of damage, and this term is subject to a gating coefficient g. ij Global weighting of (t).

[0131] In another alternative embodiment, the discrete dynamic equations may also be:

[0132] R i (t+Δt)= clip( R i (t) +Δt * (α i * E i (t) -β i * D i (t) +Σ j∈Ni g ij (t) *w k *I j (t)), 0, 1);

[0133] Where R i (t+Δt) represents the resilience state value at the next moment; the Σ term represents the sum of neighborhood interaction effects; w k For interaction weights; I j (t) represents the state input of neighbor node j; clip(…, 0, 1) is the truncation function.

[0134] Based on the toughness state value at the current time step, the state increment is calculated using discrete dynamic equations; the state increment is then superimposed on the toughness state value at the current time step to obtain the toughness state value at the next time step.

[0135] In this embodiment, the resilience trajectory R of the global nodes over time is generated through iterative calculation. i (t) can accurately reflect the entire process of flood control system from resistance to absorption to recovery. In particular, it can capture the resilience drop point caused by interaction (such as downstream flooding caused by dike breach, and slow rescue caused by information interruption), i.e. the critical point, and provide intuitive basis for emergency decision-making.

[0136] For example, suppose the toughness state value R of a dam node i at time t=0 is... i (0) = 0.8, self-supporting term E i (0) = 0.7, Disaster load term D i (0) = 0.3, coefficient α i=0.5, β i =0.8, time step Δt = 1 hour. Let the resilience state value R of its unique neighbor node j be 0.8. j (0) = 0.6, gating coefficient g ij (0) = 0.9, interaction weight ω k =0.2. Calculate the state increment: ΔR = 1 × (0.5 × 0.7 - 0.8 × 0.3 + 0.9 × 0.2 × 0.6) = 1 × (0.35 - 0.24 + 0.108) = 0.218. Then the toughness state value R at the next moment is... i (1) = clip(0.8 + 0.218, 0, 1) = 1.0 (truncated to the upper limit). This result indicates that under the current working conditions, the resilience of the dam nodes tends to improve, mainly due to their strong self-supporting capacity and the positive influence of neighborhood interactions.

[0137] In some optional embodiments, a toughness failure threshold is also set for determination, specifically:

[0138] R crit，i (t) =ρ0 + ρ1 * (1 - E resist，i (t)) + ρ2 * Risk i (t);

[0139] Among them, R crit，i (t) represents the dynamic failure threshold of node i at time t; ρ0 is the basic threshold; ρ1 is the correction coefficient for insufficient resilience to the threshold; E resist，i (t) represents the node's current resilience index; ρ2 is the external risk correction coefficient; Risk i (t) represents the environmental risk level. When R i (t) <R crit，i When (t), the node is determined to be invalid.

[0140] In one possible embodiment, to address the problem of traditional models failing to generalize effectively when faced with unprecedented extreme flood events, a cross-scenario invariant calibration mechanism can be introduced. This ensures that the model learns robust physical / social mechanisms, rather than statistical biases specific to a particular dataset. In other words, the model parameters of the resilience dynamics model and the rule compiler are obtained through pre-training using cross-scenario invariant calibration, which includes:

[0141] Obtain a dataset of a predetermined number of historical flood events.

[0142] In this embodiment, the training data is no longer a single mixed data pool, but is explicitly divided into different sets of environments or events E = {e1, e2, ..., e...} MFor example, e1 corresponds to the Meiyu-type flood data of a certain river basin in 2010, and e2 corresponds to the typhoon-type flood data of another city in 2015. These events show significant differences in rainfall type, topography, and social structure.

[0143] Construct a loss function that includes a prediction error term for a single event and an invariance constraint term for gradients across events.

[0144] For example, the objective function is designed as follows:

[0145] L total = Σ e∈E L task (e) + λ'* L inv ;

[0146] Where L task (e) represents conventional task losses, such as the mean square error (MSE) between the predicted and actual water levels; L inv For invariant constraints, specifically, the gradient penalty form in Invariant Risk Minimization (IRM) can be used, i.e., L inv = Σ e∈E || ▽ w L task (e) || 2 , representing the search for a set of optimal parameters w (including the aforementioned GNN weights, mapping matrix B, etc.) such that these parameters are locally optimal in all different flood events e, i.e., their gradients are close to 0, ensuring that the functional mechanism of the parameters remains unchanged regardless of environmental variations. Where ▽ w Let w be the gradient with respect to the parameter w.

[0147] In some possible embodiments, the objective function can also be:

[0148] L(Θ)=Σ e∈E L e (Θ)+ λ'* Σ e∈E || ▽ Θ L e (Θ)- ▽ Θ L avg (Θ)||2 2 ;

[0149] Where L(Θ) is the total training loss; Θ is the model parameter set; E is the set of historical flood events; L e (Θ) represents the model's task prediction error on event e, such as mean squared error; λ' represents the invariance penalty weight; ▽ Θ L e (Θ) represents the gradient of the loss for event e with respect to the parameters; ▽ Θ L avg(Θ) is the gradient of the average loss across all events; || ||2 2 The L2 norm squared of the gradient is used to force the gradient direction to remain consistent under different environments, i.e., to find an invariant mechanism.

[0150] With the goal of minimizing the loss function, the model parameters are iteratively optimized to ensure that the model parameters maintain a consistent distribution under different flood scenarios.

[0151] In this embodiment, the total loss mentioned above can be minimized using gradient-based optimization algorithms (such as Adam or SGD). The trained model parameters can remove background noise specific to the event while retaining the universal causal mechanism. For example, whether in a plains city or a mountainous city, the model can correctly compile the invariant physical rule that high-velocity floods impacting earthen levees lead to a high probability of breach.

[0152] According to one aspect of this application, it also includes the automatic generation and optimization of flood control strategies based on the results of flood resilience evolution simulation, including: after obtaining high-precision resilience evolution simulation capabilities, executing a strategy optimization process. Specifically, diverse future scenarios are generated by random sampling using the Monte Carlo method, including rainfall combinations with different return periods (20 years, 50 years, 100 years) and sudden engineering accidents, such as pump station power outages. For each scenario, the system generates candidate combinations of flood control measures, such as a 2-hour advance warning + activation of the A-zone flood storage area + dispatching of B-team rescue forces. A non-dominated sorting genetic algorithm II (NSGA-II) is used for multi-objective optimization. The total cost of implementing measures and the global resilience improvement value are set as two mutually exclusive optimization objectives, where the total cost of implementing measures and the global resilience improvement value are the difference between the integral areas of the resilience curves with and without measures. The simulation model is used as the fitness evaluation function of the genetic algorithm to quickly simulate and evaluate each generation of the population (strategy combination). After multiple generations of evolution, a Pareto optimal solution set is output. Decision-makers can choose from this solution set based on their actual budget and risk appetite. For example, they can select a strategy with moderate costs that avoids critical node resilience failures as the final implementation plan. In addition, the Sobol global sensitivity analysis method is introduced to calculate the contribution index of each input factor, such as early warning delay time and material allocation, to the final resilience value, identifying the weak links and bottlenecks in the flood control system and providing scientific guidance for projects to address these weaknesses during peacetime.

[0153] Optionally, after generating and selecting the Pareto optimal combination of flood control strategies, a dynamic strategy knowledge base is also constructed to enable long-term reuse and intelligent retrieval of the strategies. Specifically, the system utilizes knowledge graph technology to store verified optimal strategies, such as the activation of flood diversion area A + material allocation plan B, as entity nodes in the knowledge graph database. This knowledge graph not only stores the strategies themselves but also the context in which the strategies take effect and the associated element interaction rules, where the element interaction rules can be generated by a rule compiler. For example, an applicable relationship edge is established between the rainstorm scenario entity and the dike heightening strategy entity, with an attached success rate attribute. When the system encounters similar flood scenarios in future simulations, it can quickly retrieve the historically optimal strategy from the knowledge graph using a graph matching algorithm, serving as the initial population for the genetic algorithm, thereby significantly accelerating the convergence speed of strategy optimization. Through a closed-loop learning mechanism based on the knowledge graph, continuous evolution of flood control intelligence is achieved.

[0154] According to another aspect of this application, a flood resilience evolution simulation system based on the interaction of natural and social elements is also provided, including a data alignment module, an interactive calculation module, a rule compilation module, and an evolution simulation module:

[0155] The data alignment module is used to acquire the spatiotemporal sequences of natural elements and social elements in the target area, and align them to a unified spatiotemporal grid node to obtain a heterogeneous node sequence.

[0156] The interactive computing module is used to calculate the time-varying interaction strength between nodes and the interaction uncertainty of the time-varying interaction strength based on the heterogeneous node sequence.

[0157] The rule compilation module is used to call the pre-built rule compiler to compile and generate time-varying rule parameters between each pair of interactive nodes based on the time-varying interaction strength and interaction uncertainty.

[0158] The evolution simulation module is used to input time-varying rule parameters into a pre-built resilience dynamics model, drive the update of the resilience state value of each node in continuous time steps, and obtain the flood control resilience evolution trajectory.

[0159] In other words, the data alignment module is used to access the meteorological bureau's rainfall data interface, the water resources bureau's hydrological telemetry terminal, and the operator's population heat map database. It performs spatiotemporal grid mapping and interpolation operations in memory, outputting a standardized heterogeneous node sequence. The interactive computing module is typically deployed on a computing server equipped with a high-performance graphics processing unit (GPU, such as NVIDIA A100), running an offline-trained ST-GCN or SEM model to calculate the interaction strength and uncertainty between millions of node pairs across the entire domain in real time with high concurrency. The rule compilation module, as a lightweight middleware, is embedded within the simulation engine, converting the weight matrix returned by the GPU into a CPU-executable rule parameter table in real time. The evolution simulation module, based on a multi-threaded parallel computing architecture (such as OpenMP or MPI), uses numerical integration to solve discrete dynamic equations, quickly extrapolating the resilience evolution trend over the next 24 to 72 hours, and displaying the results to the command center user in the form of a dynamic heat map through the WebGIS front end.

[0160] Furthermore, in terms of hardware deployment, the system can adopt a distributed cluster architecture. The data layer uses Hadoop or Spark for cleaning and storing massive amounts of multi-source data; the computing layer uses CPU+GPU heterogeneous computing nodes, respectively responsible for logic-intensive dynamic evolution and computation-intensive neural network inference; the application layer is based on a Spring Boot microservice architecture, providing RESTful API interfaces to support multi-terminal access from mobile and PC. This ensures the system's efficiency, stability, and scalability when handling large-scale urban flood simulations.

[0161] In some possible implementations, a fault-tolerance mechanism is also included to address potential anomalies. Specifically, when a node's input features have missing values, the system attempts to fill them using linear interpolation along the time dimension. If the consecutive missing time exceeds a preset threshold, such as 6 hours, a weighted average of the spatial neighborhood is used as a substitute, and the node's data quality is marked as impaired, automatically increasing its interaction uncertainty in subsequent calculations. When the neural network inference outputs a non-numerical (NaN) or infinite (Inf) value, the system automatically reverts to the interaction strength value of the previous valid time step and triggers an alarm log. If anomalies occur for three consecutive time steps, the system switches to a rule-based simplified model for degraded operation. When communication with an external data source is interrupted, the system automatically uses the most recently valid data cached locally and gradually reduces the trust weight of that data source in an exponential decay manner.

[0162] In a detailed embodiment, it is assumed that the watershed area of ​​a medium-sized city is approximately 500 square kilometers, divided into 5000 100m × 100m grid nodes. In July 2025, the watershed experienced a once-in-a-century rainstorm, with a 24-hour cumulative rainfall of 280mm. The system receives real-time rainfall data from meteorological radar, water level data from 12 hydrological stations, and population heat data from mobile phone signals. After spatiotemporal interpolation, a heterogeneous node sequence from T=0 to T=72 hours is generated. A pre-trained ST-GCN model is used to calculate the time-varying interaction strength and uncertainty of approximately 250,000 potential interaction edges. Among them, the interaction strength between the main river channel and the dikes on both banks reaches a peak of 0.92 at T=8 hours. For the above high-intensity interaction, the gating coefficient is calculated to be 0.89. The rule compiler combines this value with the current water level margin of the dikes (design flood level - actual water level) to generate time-varying failure probability parameters. Dynamic model simulations show that at T=12 hours, the resilience values ​​of three nodes in the main urban area will fall below the failure threshold of 0.3; at T=18 hours, without intervention, it is estimated that 15 nodes will fail. The system generates an early warning report, recommending the activation of flood diversion zone A at T=6 hours, which could control the number of failed nodes to within 5.

[0163] According to one aspect of this application, the flood resilience evolution simulation method based on the interaction of natural and social elements can also be used for:

[0164] Step S1: Construct a multi-scale natural element model. Use deep learning fusion algorithms to process multi-source data such as satellite remote sensing, UAV aerial survey, and ground sensors to establish a four-dimensional spatiotemporal distribution model of natural elements including hydrology, topography, meteorology, and ecology.

[0165] Specifically, constructing multi-scale natural element models includes: analyzing factors such as precipitation, soil permeability, and topographic slope in different regions within the watershed; establishing a hydrological response model of the watershed using hydrodynamic models (such as HEC-RAS and MIKE11); simulating precipitation processes, runoff processes, and hydrodynamic characteristics; and accurately simulating the distribution and changes of water flow within the watershed, including key hydrological variables such as velocity, water level, and discharge. Through simulations of different flood scenarios, the hydrological response of the watershed is analyzed to obtain real-time flood risk assessments and behavioral responses of flood control systems, such as water retention, dike overflow, and channel water conveyance capacity, reflecting actual changes in the natural environment and their impact on flood resilience. Using remote sensing technology and Geographic Information System (GIS) data, a digital elevation model (DEM) of the watershed or region is constructed to accurately describe topographic features, including parameters such as elevation, slope, and aspect of each topographic unit within the watershed. Through topographic analysis, potential flood accumulation areas and vulnerable regions are further identified, providing support for flood control planning.

[0166] Based on historical meteorological data, climate change trends, and regional characteristics, spatiotemporal distribution models of meteorological elements such as rainfall, temperature, and evaporation are established to accurately describe the changing patterns of meteorological elements at different time scales (e.g., daily, monthly, yearly) and spatial scales (e.g., different locations within a region or watershed). Through meteorological data analysis, the frequency of seasonal climate change and extreme climate events (e.g., rainstorms, droughts) and their potential future trends are captured. Using remote sensing imagery and land use data, spatial distribution models of different land use types (e.g., urban, agricultural, forest) are constructed within the watershed or region to describe the hydrological response characteristics under different land use types, including the impact of ground cover type on precipitation infiltration, evaporation, and the hydrological cycle. Simultaneously, considering the dynamic processes of land use change, such as urban expansion and agricultural land change, their impact on flood resilience is assessed.

[0167] Step S2: Establish a dynamic social element model. Based on big data mining and social network analysis technologies, construct a real-time dynamic model of population flow, infrastructure status, economic activity intensity, and social organization networks.

[0168] Specifically, establishing dynamic social element models includes: using spatiotemporal clustering algorithms to build a real-time dynamic distribution model of population density based on mobile signaling data and social media data, capturing population flow and aggregation in different time periods and regions in real time; constructing spatial layout models of infrastructure such as roads, buildings, and water conservancy facilities, and establishing distribution models of infrastructure by combining spatial analysis techniques with the spatiotemporal characteristics of infrastructure to assess its role in flood control emergency response; using spatial econometric models to establish spatial distribution and temporal variation models of economic activities based on nighttime light remote sensing data, POI data, and economic statistics, analyzing the spatiotemporal evolution of economic activities within the region, and assessing its impact on flood resilience and disaster recovery; and using complex network analysis methods to construct social relationship network and information dissemination network models based on social media data and communication data, simulating social interaction behaviors such as information flow, public opinion dissemination, and disaster response through topological analysis of social networks, revealing the impact of social organization and information transmission on flood control decision-making.

[0169] Step S3: Design a natural-social element interaction mechanism based on graph neural networks, establish an element node graph and a relationship edge graph, and learn complex nonlinear interaction relationships through graph convolutional neural networks to achieve intelligent modeling of the influence between elements.

[0170] In this embodiment, the design of a natural-social interaction mechanism based on graph neural networks includes: the impact mechanism of nature on society: the damage mechanism of floods on population, infrastructure, and economic activities, specifically including the direct impact of natural factors such as flood level, flow velocity, and duration on social systems (such as population distribution, infrastructure damage, and economic activity interruption). These damages are quantified using a post-disaster assessment model, providing a basis for optimizing flood control strategies. The feedback mechanism of society on nature: the impact mechanism of flood control projects and land use change on hydrological processes, including the regulating effect of flood flow patterns on projects such as flood dams and river channel modifications, and the impact of land use changes (such as urbanization and agricultural land conversion) on the hydrological cycle. The feedback effect of social behavior on hydrological processes is simulated using a system dynamics model. Natural and social elements are abstracted as graph nodes, and the influence relationships between elements are abstracted as graph edges, constructing a heterogeneous element interaction graph. Specifically, this includes node modeling of natural factors (such as precipitation, topography, and climate) and social factors (such as population, infrastructure, and policies), using graph theory methods to describe the interactions between nodes, forming a dynamic interaction network.

[0171] This study employs Graph Attention Networks (GAT) and Graph Convolutional Networks (GCN) to learn node features and edge weights, capturing complex nonlinear relationships between elements. Through learning node features and edge weights, it automatically identifies the importance of different elements and effectively models their mutual influence, providing intelligent support for decision-making. A Spatiotemporal Graph Convolutional Network (ST-GCN) is used to handle the temporal dependencies of element interactions, establishing a dynamic interaction mechanism model. By introducing a time dimension, it models the dynamic characteristics of the interaction relationships between elements as they change over time, capturing the complex patterns of the combined effects of temporal and spatial dependencies. Graph embedding technology maps high-dimensional interaction relationships to a low-dimensional space, achieving a quantitative expression of interaction intensity. This maps complex interaction relationships into low-dimensional vectors, facilitating subsequent analysis and application, and supporting efficient multi-dimensional relationship mining and optimization.

[0172] Step S4: Construct a multi-dimensional flood resilience evaluation index system, conduct quantitative evaluation from three dimensions: resistance capacity, adaptability and recovery capacity, and determine the weight of each index.

[0173] Specifically, the construction of a multi-dimensional flood resilience evaluation index system includes: constructing quantitative indicators of resilience, including flood control engineering standards, the completeness of the early warning system, and emergency response capabilities. Flood control engineering standards are quantified by assessing the compliance rate (%) of dike design flood control standards, the compliance rate (%) of drainage network design return period standards, and the reservoir regulation capacity index; the completeness of the early warning system is measured by the proportion of flood warning coverage areas (%), the delay time (minutes) of early warning information transmission, and the accuracy rate (%) of early warnings; emergency response capabilities are quantified by the adequacy rate of emergency material reserves (%) and the average time (hours) for rescue teams to reach high-risk areas. The construction of quantitative indicators of adaptability includes social learning capacity, institutional adjustment capacity, and technological innovation capacity. Social learning capacity is quantified by assessing the frequency of flood control plan updates based on historical disaster data (times / year) and the public's flood control knowledge dissemination rate (%); institutional adjustment capacity is assessed by the cross-departmental collaborative response efficiency index and the timeliness score of flood control policy and regulation revisions; and technological innovation capacity is measured by the coverage rate (%) of intelligent monitoring equipment and the response speed (seconds) of the real-time data analysis platform.

[0174] Quantitative indicators of recovery capacity were constructed, including the speed of post-disaster reconstruction, economic recovery capacity, and social recovery capacity. The speed of post-disaster reconstruction was assessed using the average recovery time (days) for lifeline infrastructure (water / electricity / transportation) and the housing reconstruction completion rate (% / month). Economic recovery capacity was quantified by the time (months) it took for the affected area's GDP to recover to pre-disaster levels and the rate of business resumption (% / week). Social recovery capacity was measured by the coverage rate of psychological assistance services (%) and the community function recovery index (based on the reopening rate of public service facilities). Dynamic optimization of indicator weights employed a combined weighting method, including the entropy weight method to calculate objective weights and the AHP (Analytic Hierarchy Process) to determine subjective weights, ensuring the rationality of weights in the assessment process. A time decay factor was also introduced to dynamically adjust the indicator weights at different stages—pre-disaster, during-disaster, and post-disaster—optimizing the resilience assessment based on the actual conditions at each stage.

[0175] Step S5: Establish a resilience evolution simulation model. Based on multi-agent system and complex network theory, construct a dynamic model of the spatiotemporal evolution of flood control resilience.

[0176] In this embodiment, the resilience evolution simulation model is constructed using the following method: Environmental agents are generated based on a spatiotemporal distribution model of natural elements, including hydrological agents (carrying attributes such as water level, flow velocity, and flow rate), meteorological agents (carrying attributes such as rainfall intensity, temperature, and evaporation), and topographic agents (carrying attributes such as elevation, slope, and catchment area). Social agents are generated based on a dynamic model of social elements, including population agents (carrying attributes such as density, migration direction, and risk perception), infrastructure agents (carrying attributes such as location, flood control level, and damage status), economic organization agents (carrying attributes such as output value, shutdown rate, and recovery cost), and management agency agents (carrying attributes such as emergency resources, decision-making rules, and coordination efficiency). The graph neural network interaction mechanism (the node influence weights output by the GNN) is transformed into rules for interaction between agents. Specifically, these include: a nature-to-society rule, where the damage probability calculation of the infrastructure agent is triggered when the flood agent's water level exceeds a threshold; a society-to-nature rule, where the flow allocation parameters of the hydrological agent are dynamically modified when the management agency agent initiates flood diversion decisions; and an internal social rule, where the evacuation path selection of the population agent is constrained by the connectivity of the infrastructure agent.

[0177] Based on an infrastructure spatial layout model and a social relationship network, a two-layer coupled network is constructed. The physical layer network uses roads, pipelines, and dams as edges, with key facilities as nodes; the social layer network uses information propagation paths and organizational collaboration relationships as edges, with communities / institutions as nodes. Inter-layer coupling mechanisms are defined, such as physical layer node failure triggering a social layer node response delay. A spatiotemporal dynamics model is used to describe the agent's behavior. The agent's behavior is described by combining differential equations and state machines: dR i (t) / dt=f(E i (t), ∑ j∈Nj W ij ·I j (t)); where R i (t) represents the resilience state value of agent i at time t, E i (t) represents the quantitative result of the indicator (resistance / adaptability / recovery ability), W ij I represents the interaction weights output by the graph neural network. j (t) represents the state input of the neighboring agent j. The spatiotemporal evolution of global resilience is driven by agent state transitions. Specifically, this involves dynamically updating the resilience state value of each agent based on its state changes before, during, and after the disaster, and generating a heatmap sequence of global resilience changes through a spatiotemporal evolution model. This heatmap sequence can intuitively display the spatial distribution and evolution trend of flood resilience states in various regions and systems at different time points, providing real-time and accurate prediction and assessment basis for disaster prevention and mitigation decisions.

[0178] Step S6: Generate diverse scenarios with different combinations of flood scenarios and flood control measures based on Monte Carlo methods and genetic algorithms, and identify key risk factors and optimal flood control strategies using sensitivity analysis and uncertainty analysis.

[0179] Specifically, the scenario simulation analysis includes setting a baseline scenario, an extreme flood scenario, and a scenario for enhanced flood control measures. The baseline scenario is a normal flood scenario without any disaster intervention. The extreme flood scenario considers adverse conditions such as extreme rainfall and floods exceeding standards. The scenario for enhanced flood control measures assumes that existing flood control projects and management measures have been strengthened, such as raising the standards of dikes and improving the drainage system. Flood scenarios are generated based on meteorological models. The Monte Carlo method is used to randomly sample and generate combinations of rainfall intensity and duration with different return periods (20 / 50 / 100 years), as well as the spatial distribution probability of extreme climate events such as typhoon rainstorms and snowmelt floods. The spatiotemporal distribution of flood inundation range, water depth, and flow velocity under the above scenarios is simulated using a hydrological response model. The combination of flood control measures is generated using a genetic algorithm (with the fitness function being a comprehensive resilience evaluation value). This includes combinations of engineering measures, such as dike heightening, activation of flood diversion zones, and reservoir scheduling schemes, as well as combinations of non-engineering measures, such as advance warning periods, evacuation route planning, and resource allocation strategies.

[0180] The model inputs scenarios into a resilience evolution simulation model and outputs a comparison of resilience trajectories, showcasing spatiotemporal evolution heatmaps of resilience under baseline, extreme flood, and enhanced measures scenarios. Threshold breakthrough analysis statistically analyzes the frequency and spatial distribution of resilience threshold breakthroughs in different scenarios, and quantifies the rate of decline in resilience and the rate of increase in recovery capacity based on dynamic equations. Sensitivity analysis uses the Sobol index method to calculate the contribution of natural and social factors to the resilience evaluation value, identifying key risk factors with a sensitivity greater than 0.3, such as the location of levee breaches and bottlenecks in evacuation routes. Uncertainty analysis quantifies the impact of the randomness of information propagation delays on recovery capacity based on a social network model, and assesses the confidence interval for resilience assessment caused by missing data using a Bayesian network. Strategy effectiveness evaluation defines the strategy benefit ratio (η = disaster loss reduction / measure cost), and calculates the investment payback period in conjunction with economic recovery capacity indicators. Pareto optimal solution screening uses a genetic algorithm to screen non-dominated solutions in a two-dimensional space of measure cost and resilience improvement, outputting the optimal strategy combination that meets a preset threshold and improves resilience by more than 20%. A dynamic strategy library is built, and the validated strategies are written into the knowledge graph and associated with the interaction rules of the elements to support real-time decision retrieval.

[0181] In summary, the flood resilience evolution simulation method based on the interaction of natural and social elements includes: acquiring the spatiotemporal sequences of natural and social elements and aligning them into heterogeneous node sequences; using spatiotemporal graph neural networks or time-varying structural equation models to calculate the time-varying interaction intensity and interaction uncertainty between nodes; using a pre-built rule compiler to construct an uncertainty-based gating function to compile the time-varying interaction intensity into executable time-varying rule parameters; and inputting these parameters into the resilience dynamics model to drive the update of the resilience state values ​​of each node and generate evolution trajectories.

[0182] This invention introduces a rule-based compiler mechanism, which, through parameter instantiation mapping, transforms the implicit soft attention weights mined by the neural network into explicit hard physical parameters in the dynamic equations in real time, such as facility vulnerability coefficients. This establishes a logical link from feature mining to behavior-driven processes, enabling the direct and dynamic driving of physical mechanisms by implicit data features. An uncertainty-based gated filtering mechanism is employed, quantifying interaction variance in real time through ensemble learning or bootstrap methods. A gate function positively correlated with strength and negatively correlated with uncertainty is constructed, automatically eliminating false positives and retaining key interactions while suppressing low-confidence noise interference, ensuring the accuracy of the flood control system's response in complex environments. Cross-scenario invariant calibration training is used, introducing a loss function with gradient consistency constraints to force the model to learn robust causal mechanisms common to different flood events, rather than statistical biases specific to certain data, effectively improving the model's prediction accuracy under unknown extreme scenarios.

[0183] 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 simulating the evolution of flood resilience based on the interaction of natural-social elements, characterized by, The method comprises the following steps: acquiring a natural element space-time sequence and a social element space-time sequence of a target area, and aligning them to a unified space-time grid node to obtain a heterogeneous node sequence; based on the heterogeneous node sequence, calculating a time-varying interaction intensity between nodes and an interaction uncertainty of the time-varying interaction intensity; using a preset rule compiler, compiling time-varying rule parameters between each pair of interaction nodes based on the time-varying interaction intensity and the interaction uncertainty; inputting the time-varying rule parameters into a pre-constructed resilience dynamics model to drive the update of the resilience state value of each node at a continuous time step, and obtaining a flood prevention resilience evolution trajectory; compiling the time-varying rule parameters, including calculating a gating coefficient, specifically: constructing a gating function configured to be positively correlated with the time-varying interaction intensity and negatively correlated with the interaction uncertainty; substituting the time-varying interaction intensity and the interaction uncertainty into the gating function to obtain the gating coefficient of each pair of interaction nodes; compiling the time-varying rule parameters also includes: acquiring a pre-stored rule template and its corresponding basic parameter vector; acquiring and constructing a feature interaction term based on the node characteristics of the interaction nodes; using a mapping matrix to weight and combine the feature interaction term and the gating coefficient, and superimposing the weighted combination result to the basic parameter vector to obtain the time-varying rule parameters.

2. The method of claim 1, wherein, calculating the time-varying interaction intensity between nodes, including: constructing a heterogeneous element interaction graph containing natural nodes, social nodes and facility nodes, and defining potential connection edges between nodes; inputting the heterogeneous element interaction graph into a preset spatio-temporal graph neural network model, capturing node spatial features through a graph convolution layer, and capturing node time dependence through a time convolution layer; based on the node spatial features and the node time dependence, using the attention mechanism of the spatio-temporal graph neural network model to calculate the attention weight of each potential connection edge, and taking the normalized attention weight as the time-varying interaction intensity.

3. The method of claim 2, wherein, calculating the attention weight of each potential connection edge, including: for each potential connection edge, based on the source node features, target node features and edge type features connected thereto, using a learnable linear transformation matrix to calculate an unnormalized attention score; performing Softmax normalization processing on the attention scores of all potential connection edges connected to the same target node to obtain a normalization coefficient; taking the normalization coefficient as the attention weight to quantify the time-varying interaction intensity of the source node to the target node.

4. The method of claim 2, wherein, calculating the interaction uncertainty of the time-varying interaction intensity, including: enabling a Monte Carlo random inactivation strategy for the spatio-temporal graph neural network model, performing a predetermined number of forward inferences to obtain a predetermined number of attention weight samples for each potential connection edge; calculating the variance of the predetermined number of attention weight samples, and taking the variance as the interaction uncertainty of the potential connection edge.

5. The method of claim 1, wherein, calculating the time-varying interaction intensity between nodes can also be: defining a structure variable for each node in the heterogeneous node sequence, and constructing a time-varying structural equation model containing a time lag term to describe the causal dependence relationship between the structure variables; solving the time-varying structural equation model in a sliding window moving along the time axis to obtain time-varying structure coefficients between nodes; The absolute value of the time-varying structural coefficient is normalized, and the normalized result is taken as the time-varying interaction intensity.

6. The method of claim 5, wherein, The time-varying structural coefficient is obtained, including: Based on the time-varying structural equation model, an optimization objective function is constructed, which includes the prediction error term of the structural variable and the L1 regularization term for the time-varying structural coefficient; The optimization objective function is minimized to obtain the sparse solution of the time-varying structural coefficient.

7. The method of claim 5, wherein, The interaction uncertainty of the time-varying interaction intensity can also be calculated, including: Bootstrap self-sampling of data in the sliding window is performed to generate a predetermined number of data sub-samples; The time-varying structural equation model is solved for each data sub-sample to obtain a distribution set of the time-varying structural coefficient; The variance of the distribution set is calculated, and the variance is mapped to the interaction uncertainty.

8. A flood resilience evolution simulation system based on natural-social element interaction, characterized in that, It includes: The data alignment module is used to obtain the natural element spatiotemporal sequence and the social element spatiotemporal sequence of the target area, and align them to the unified spatiotemporal grid node to obtain the heterogeneous node sequence; The interaction calculation module is used to calculate the time-varying interaction intensity and the interaction uncertainty of the time-varying interaction intensity between nodes based on the heterogeneous node sequence; The rule compilation module is used to call a preconfigured rule compiler to compile time-varying rule parameters between each pair of interaction nodes based on the time-varying interaction intensity and the interaction uncertainty; wherein the time-varying rule parameters are compiled, including calculating the gating coefficient, specifically: constructing a gating function, which is configured to be positively correlated with the time-varying interaction intensity and negatively correlated with the interaction uncertainty; The time-varying interaction intensity and the interaction uncertainty are substituted into the gating function to calculate the gating coefficient of each pair of interaction nodes; The time-varying rule parameters are also compiled, including: obtaining a pre-stored rule template and its corresponding basic parameter vector; obtaining and constructing a feature interaction item based on the node features of the interaction nodes; using a mapping matrix to weight and combine the feature interaction item and the gating coefficient, and adding the weighted combination result to the basic parameter vector to obtain the time-varying rule parameters; The evolution simulation module is used to input the time-varying rule parameters into the pre-constructed resilience dynamics model to drive the update of the resilience state value of each node at consecutive time steps, and obtain the flood prevention resilience evolution trajectory.

Images