An intelligent deduction method for small watershed rainstorm chain disaster
By employing dynamic graph neural networks and multi-physics constraints, the problems of topological change adaptability and physical consistency in small watershed rainstorm chain disasters were solved, achieving high-precision disaster chain prediction and early warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING YUFA WATER CONSERVANCY RES INST CO LTD
- Filing Date
- 2026-04-01
- Publication Date
- 2026-06-30
AI Technical Summary
Existing technologies are ill-suited to adapting to dynamic topological changes in cascading disasters such as landslides and flooding caused by torrential rains in small watersheds, and data-driven models lack physical consistency, leading to inaccurate prediction results.
By employing a dynamic graph neural network combined with multi-physics constraints, an initial dynamic graph is constructed through multi-source data access and feature building, and the graph topology is updated in real time. Feature fusion and deduction are performed by combining the Saint-Venant equation and the Moore-Coulomb criterion to achieve time-series prediction and early warning of disaster chains.
It achieves dynamic adaptation to the disaster chain process, improves the physical consistency of prediction results and the accuracy of early warning, and can maintain reasonable inference logic and high-precision prediction in extreme data-scarce scenarios.
Smart Images
Figure CN122309998A_ABST
Abstract
Description
Technical Field
[0001] This invention mainly relates to the technical field of small watershed disaster data extrapolation, specifically a method for intelligent extrapolation of small watershed rainstorm chain disasters. Background Technology
[0002] Chain disasters such as landslides and waterlogging triggered by torrential rains in small watersheds are characterized by their suddenness and complex spatiotemporal evolution. Existing extrapolation methods are mainly divided into physical models and data-driven models. Physical models (such as SWAT) rely on precise parameters and are computationally slow, making them difficult to handle sudden scenarios. Data-driven models (such as GNNs) have a fast response, but they usually use static graph structures, which cannot adapt to the topographic changes and confluence path reconstruction caused by landslides during disasters. Furthermore, pure data models lack physical consistency, and prediction results are prone to distortion in extreme data-scarce scenarios. Therefore, there is a need for an intelligent extrapolation method that can adapt to dynamic topological changes and has physical interpretability to improve data processing speed and prediction accuracy. Summary of the Invention
[0003] To address the shortcomings of current technologies, this invention combines existing technologies and, based on practical applications, provides an intelligent simulation method for small watershed rainstorm chain disasters.
[0004] The technical solution of the present invention is as follows: A method for intelligent simulation of chain disasters caused by rainstorms in small watersheds includes the following steps: S1. Multi-source data access and feature construction: Acquire relevant data of the target small watershed, perform anomaly processing and data completion on the data, and construct a standardized data matrix containing static geological features and dynamic hydrological features; S2. Constructing an initial dynamic graph driven by disaster chains: Divide the small watershed into several homogeneous hydrological response units as graph nodes, initialize the node features, calculate the edge weights between nodes based on the differences in hydrological status between nodes and the rules of potential impact of geological disasters, and construct an initial adjacency matrix. S3. Real-time updating of graph topology: During disaster simulation, the landslide risk status of nodes is monitored or predicted in real time. When it is determined that the landslide risk of a node is sufficient to change the watershed confluence path, the edge weights or connection status between the corresponding nodes are dynamically adjusted, and a dynamic adjacency matrix sequence that changes over time is generated. S4. Feature fusion and deduction under physical constraints: The updated dynamic graph data is input into a pre-set graph neural network model. Differentiable regularization terms based on physical equations are embedded in the model. The node features are updated and spatiotemporal evolution is deduced under physical constraints. S5. Chain Disaster Time Series Prediction and Early Warning: Decode and extrapolate the node characteristics, output the time series prediction results of landslide risk and waterlogging depth, and trigger graded early warnings based on preset disaster thresholds.
[0005] Furthermore, in step S1, the relevant data for the target small watershed includes at least rainfall data, topographic data, soil parameter data, hydrological monitoring data, and disaster chain data; For missing hydrological data, spatiotemporal interpolation using random forest was employed to complete the data; for missing soil data, matching and completion were performed based on lithology and slope similarity.
[0006] Further, in step S2, during node feature initialization, each node is assigned an 8-dimensional feature vector, including: average rainfall intensity, soil saturation, current water level, landslide susceptibility influence coefficient, waterlogging risk value, slope, cohesion c, and internal friction angle φ. The waterlogging risk value is a comprehensive quantitative value constructed based on the node's current water level, terrain depression level, runoff accumulation, and historical waterlogging disaster data, with a value range of [0,1]. The calculation method is: Waterlogging risk value = 0.4 × (current water level / node's historical highest waterlogging water level) + 0.3 × (runoff accumulation / maximum runoff accumulation in the basin) + 0.3 × node's historical waterlogging frequency. For nodes without historical waterlogging data, the average waterlogging risk value of neighboring nodes with the same slope and soil type is assigned.
[0007] Furthermore, in step S2, the rules for calculating the edge weights between nodes include: The study comprehensively considers the differences in soil saturation, water level, and landslide susceptibility among nodes. The landslide susceptibility influence coefficient is positively correlated with the landslide risk or landslide volume of the upstream node. Edge weights increase as the differences in soil saturation or water level decrease, representing enhanced water flow connectivity. The initial landslide susceptibility influence coefficient... =1.0, and the value is dynamically adjusted according to the following rules after the landslide occurs: (1) Landslide susceptibility influence coefficient The value range is [0,2]. The larger the landslide risk / volume at the upstream node, the better. The higher the value, the stronger the impact of the upstream landslide on the downstream node confluence path; (2) When the volume of the landslide body at the upstream node is used as the basis for quantification, =1.0+k×V / V0, where V is the actual landslide volume at the upstream node, V0 is the critical landslide volume in the basin (preset according to the scale of the river channel in the basin, with an example value of 500m³), and k is the influence coefficient correction factor (taken as 0.5, which can be adjusted according to the geological conditions of the basin). (3) When the landslide risk level of the upstream node is used as the quantitative basis, the landslide risk level is divided into 1-5 levels according to the probability of landslide occurrence (Level 1: probability < 20%, Level 2: 20% ≤ probability < 40%, Level 3: 40% ≤ probability < 60%, Level 4: 60% ≤ probability < 80%, Level 5: probability ≥ 80%), corresponding to The values are 1.0, 1.2, 1.4, 1.7, and 2.0, respectively. The nonlinear incremental value is adopted because when the landslide risk level reaches level 4-5 (high / extremely high risk), the influence of the landslide body on the confluence path increases nonlinearly. This value is more in line with the actual physical process of "landslide-confluence".
[0008] Furthermore, the edge weights between node i and node j The calculation formula is as follows: ; in, For soil saturation, For water level, α represents the landslide susceptibility influence coefficient, and α, β, and γ are weighting coefficients.
[0009] In further step S3, dynamically adjusting the edge weights or connection connectivity between corresponding nodes specifically includes: Set a landslide impact threshold. When the landslide risk parameter of a certain node exceeds the threshold, identify the confluence blocking effect or diversion effect of that node on downstream nodes. When a blocking effect is determined, the connection edge weight between the node and the downstream node is reset to disconnect. When a diversion effect is identified, the edge weights on the original confluence path are reduced, and new connections are established or the edge weights of the side paths are increased based on the new terrain trend.
[0010] Furthermore, in step S4, the differentiable regularization term based on the physical equation includes: Waterlogging evolution constraint term: constructed based on the Saint-Venant equation, used to constrain the mass and momentum conservation of water flow motion; Landslide triggering constraint: Constructed based on the Mohr-Coulomb criterion, used to constrain the mechanical determination of slope instability; The constraint terms are transformed into differentiable residual functions, which are then used as part of the loss function in the training or derivation optimization of the model.
[0011] Furthermore, in step S4, the continuity and momentum equation residuals of the waterlogging evolution constraint terms are calculated using the Preissmann difference scheme. ; The difference scheme expression for the continuity equation is as follows: ; The difference scheme for the momentum equation is as follows: ; In the formula, Let Q be the water level, A be the flow rate, n be the Manning coefficient, and R be the hydraulic radius. It is the acceleration due to gravity. For spatial step size, For time step; Let i be the water level value at the (i+1)th node at the j-th time step. Let be the water level value of the i-th node at the j-th time step; Let be the overcurrent value of the (i+1)th node at the j-th time step. Let be the overcurrent value of the i-th node at the j-th time step.
[0012] The landslide triggering constraint term is constructed by calculating the residuals of shear stress and shear strength, where shear stress... The calculation expression is as follows: ; shear strength The calculation expression is as follows: ; residual The expression is as follows: ; In the formula, θ is the natural unit weight of the soil, h is the soil layer thickness, and θ is the slope angle. For normal stress, It is the internal friction angle. It is cohesive force.
[0013] Furthermore, the total loss function of the graph neural network model is expressed as: ; in, The data-driven prediction error between the model output and the actual observed values; The physical residuals obtained from the Saint-Venant equations; The physical residuals obtained from the Mohr-Coulomb criterion; and The coefficients are used to balance physical constraints and data fitting weights.
[0014] Furthermore, in step S5, the tiered early warning adopts a progressive triggering mechanism: When the predicted landslide risk exceeds the first-level threshold, a landslide warning is triggered. Based on the updated topology after the landslide, a second-level simulation is performed. If the predicted waterlogging depth or water level rise of downstream nodes exceeds the second-level threshold, a waterlogging association warning is triggered, and warning information containing the disaster chain transmission path is output.
[0015] The beneficial effects of this invention are: Strong dynamic adaptability: It breaks through the limitations of traditional static GNNs, transforms the evolution of disaster chains into graph topology updates, and can cut off or reconnect confluence paths in real time according to the occurrence of landslides, so as to truly restore the disaster chain process.
[0016] High physical consistency: Differentiable physical regularization is introduced to achieve deep integration of data and physics, ensuring that the prediction results conform to the basic principles of fluid mechanics and soil mechanics, and maintaining reasonable inference logic even in areas without monitoring data.
[0017] High accuracy of early warning: The progressive early warning mechanism clarifies the causal chain of disasters and effectively avoids the one-sidedness of early warning for single disasters. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating the overall process of this invention.
[0019] Figure 2 This is a schematic diagram of the real-time update mechanism of the graph topology in this invention, where -- represents a normal merging connection (weight 1.0), ---- represents a weak connection (weight 0.2), ... represents a semi-blocking connection (weight 0.5). To block the connection (weight 0).
[0020] Figure 3 This is a schematic diagram of the feature fusion and deduction model structure under physical constraints of the present invention.
[0021] Figure 4 This is a schematic diagram of the progressive early warning triggering logic of the present invention.
[0022] Figure 5 This is a schematic diagram of the time-series prediction results and visualization report of chain disasters in this invention. Detailed Implementation
[0023] The present invention will be further described in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.
[0024] This embodiment provides an intelligent prediction method for cascading disasters caused by torrential rain in small watersheds. Specifically, it is an intelligent prediction method for cascading disasters caused by torrential rain, landslides, and urban flooding that integrates dynamic graph neural networks and multi-physics constraints. By constructing a dynamic graph construction mechanism driven by both hydrological conditions and disaster impacts, the evolution of disaster chains is transformed into graph topology updates. Utilizing differentiable programming frameworks such as JAX, the Saint-Venant equation and the Mohr-Coulomb criterion are embedded into the neural network loss function, achieving a deep fusion of data and physics.
[0025] The intelligent simulation method for rainstorm chain disasters in this embodiment, through... Figure 1 The standardized steps shown are implemented. Figure 1 The document details the complete logical steps from multi-source data access (S1), initial dynamic graph construction (S2), real-time graph topology update (S3), physical constraint feature deduction (S4), to chain-related disaster early warning (S5), presenting the data flow transmission relationship between each step. Each step deeply integrates the core innovations of dynamic graphs and multiple physical constraints, forming a closed-loop disaster chain deduction system of "data-topology-constraint-early warning," as detailed below.
[0026] S1: Multi-source data access and feature construction (1) Data access: Multi-source data access is achieved in seconds through OGC SensorThings APIv1.1 (Open Geospatial Information Consortium (OGC) Internet of Things Sensor Service Framework version 1.1). Specific data types include: ① Rainfall data (integrating hourly observation data from regional meteorological stations with X-band radar inversion data to generate rainfall grid data with a spatial resolution of 1km and a temporal resolution of 15min); ② Topographic data (using 30m resolution topographic data (DEM) to extract topographic factors such as slope, aspect, and runoff accumulation); ③ Soil data (based on the results of the Second National Soil Survey, including cohesion). =15-50kPa, internal friction angle =15°-35°, saturated water content S=0.45cm³ / cm³ and other core parameters); ④ Hydrological data (access to water level and flow monitoring data at the watershed outlet station, sampling frequency 5min); ⑤ Disaster chain data (including landslide susceptibility zones (levels 1-5) and historical disaster chain time records).
[0027] (2) Anomaly handling and completion: The “3σ criterion + watershed hydrological threshold” was used to remove rainfall anomalies; for missing water level data, random forest spatiotemporal interpolation was used to complete the data (RMSE (root mean square error) was controlled within 0.08m); for missing geological parameters, matching and completion were performed based on lithology and slope similarity.
[0028] (3) Standardization and Correlation Construction: Normalize all feature data to the [0,1] interval and construct a standardized data matrix containing static geological features and dynamic hydrological features as input to S2.
[0029] S2: Constructing the initial dynamic graph driven by the disaster chain (1) Node division: Based on topographic data with a precision of 30 meters, the small watershed was divided into 100 homogeneous hydrological response units using the ArcGIS (Geographic Information System software) hydrological analysis module (filling depressions, flow direction, flow accumulation), with each unit serving as a graph node.
[0030] (2) Feature initialization: Each node is assigned an 8-dimensional feature vector, including: average rainfall intensity, soil saturation, current water level, landslide susceptibility influence coefficient, waterlogging risk value, slope, cohesion c, and internal friction angle. .
[0031] (3) Initial edge weight calculation: The edge weight between node i and node j is calculated based on the "dual-driven" rule of "hydrological state-disaster impact". In this embodiment, the calculation formula is as follows: ; in, For soil saturation, For water level, The landslide susceptibility impact coefficient is defined, and all input parameters have been normalized to the [0,1] interval through step S1. Weighting coefficients are set to α=0.4, β=0.3, and γ=0.3, with a sum of 1, ensuring the initial edge weights are accurate. Naturally located within the normalized interval [0,1], the calculation results form the initial adjacency matrix. In the initial state, the baseline weight of the edge of the normal confluence path without landslide influence is 1.0, representing complete connectivity.
[0032] S3: Real-time updates of graph topology This step employs a unique and innovative method that can be used to simulate the dynamic changes in the confluence path caused by landslides. Figure 2 The diagram vividly illustrates the evolution of the adjacency matrix over time. It includes the tree-like confluence structure at the beginning of rainfall (0h), as well as the local network or broken structure formed by landslides during the peak of the rainstorm (e.g., 3h). The specific manifestations of the "blocking effect" (disconnection) and "rerouting effect" (establishment of weak connections) caused by landslides are specifically marked, as follows.
[0033] (1) Landslide risk monitoring: During the simulation, the landslide risk of each node is calculated in real time. The landslide volume threshold is set at 500m³ (an exemplary threshold, which can be adjusted according to the geological conditions of the watershed).
[0034] (2) Topology reconstruction logic: (2.1) Blocking effect: When a landslide is detected at node k and the estimated landslide volume is >1000m³ (exemplary threshold, which can be adjusted according to the scale of the river channel in the basin), it is determined that it strongly blocks the river channel, and the weight w of the connection edge between this node and the downstream node is forcibly set to 0. (Blocking the connection).
[0035] (2.2) Diversion effect: When the monitored landslide volume is between 500-1000m³, it is determined that it has a semi-blocking effect on the original confluence path. The weight of the corresponding edge of the original confluence path is directly assigned to 0.5, which represents that the original main channel retains 50% of the confluence connectivity. At the same time, based on the topographic data (DEM), potential confluence paths on the side are searched and new weak connections on the side are established. The weight of the newly added weak connection edge is directly assigned to 0.2, which represents that the side overflow path only has 20% of the confluence connectivity.
[0036] (3) Dynamic update: The above judgment is performed every 30 minutes to generate a dynamic adjacency matrix sequence that changes over time (e.g., Figure 2 As shown, the tree-like structure at 0h evolves into a localized network / broken-circuit structure at 3h, truly recreating the physical process of "landslide blocking the river - water level rising".
[0037] S4: Feature Fusion and Deduction under Physical Constraints Figure 3 This demonstrates the internal processing flow of a graph neural network (GNN) based on the JAX framework for dynamic graph data input. The diagram highlights how the physically constrained attention layer aggregates neighborhood features, and how the Saint-Venant equation (flooding constraint) and the Moore-Coulomb criterion (landslide constraint) are transformed into differentiable residuals and embedded into the total loss function. The calculation path is as follows.
[0038] (1) Physically Constrained Attention Layer: The dynamic graph generated by S3 is input into the GNN model built based on the JAX framework. The model uses a multi-head attention mechanism to aggregate neighborhood features. The attention weights are adjusted using a two-layer mechanism of "static prior first-round assignment + dynamic temporal iterative update": ① During the first calculation, based on the static geological prior information such as the landslide susceptibility influence coefficient initialized by the nodes in step S2 and the landslide susceptibility zoning level pre-divided in the target watershed, the features of nodes with high landslide susceptibility of level 4-5 are given an amplified attention weight of 1.2-1.5 times, providing a clear basis for the assignment for the first calculation; ② During subsequent temporal iterative calculations, based on the landslide risk level results output by the model in the previous time step, nodes judged as having high landslide risk of level 4-5 are given an attention weight amplification of 1.2-1.5 times, so as to achieve continuous focus on the features of high-risk nodes and improve the inference accuracy.
[0039] (2) Differentiable physical regularization computation: Waterlogging constraint (Saint-Venant equation): Calculation of continuity and momentum equation residuals using the Preissmann difference scheme. The constrained water flow evolution conforms to the conservation of mass and momentum.
[0040] The difference scheme for the continuity equation is as follows: ; The difference scheme for the momentum equation is as follows: ; In the formula, Let Q be the water level, A be the flow rate, n be the Manning coefficient, and R be the hydraulic radius. It is the acceleration due to gravity. For spatial step size, For time step.
[0041] Landslide constraint (Mohr-Coulomb criterion): Calculate the shear stress using the following formula: ; In the formula, θ is the natural unit weight of the soil, h is the soil layer thickness, and θ is the slope angle. The formula for calculating shear strength is as follows: ; In the formula, σ is the normal stress, c is the cohesion, and the residual formula is as follows: ; The stability of slopes is determined by constructing a residual constraint.
[0042] Loss function optimization: Define the total loss function ,in, To predict losses (cross-entropy loss for landslide grade and MSE loss for waterlogging depth), the model output is minimized to approximate historical observations while satisfying fluid mechanics and soil mechanics constraints, thus addressing the problem of prediction distortion in areas without monitoring data.
[0043] S5: Sequence Prediction and Early Warning of Chain-Related Disasters
[0044] Figure 4 The diagram illustrates the judgment logic for the transmission from a "Level 1 Warning (Landslide)" to a "Level 2 Warning (Water Flooding)". The diagram clarifies how the system automatically performs topology reconstruction and secondary deduction after the landslide probability or risk level triggers the Level 1 threshold, thereby determining whether to trigger a closed-loop process for downstream water flooding-related warnings, as detailed below.
[0045] Temporal Decoding and Output: The model output layer decodes node features to generate temporal prediction results for the next 6 hours, including: the probability of landslide occurrence, landslide risk level (1-5), waterlogging depth, and water level change curves for each node. The model output layer decodes the inferred high-dimensional features of the nodes through a fully connected layer and a temporal convolutional network (TCN), mapping the node feature vectors output by the graph neural network to temporal prediction values. Based on a time step (15 minutes in the example), it generates temporal prediction results for the next 6 hours. The specific generation process for each result is as follows: ① Probability of landslide occurrence: Input the relevant features of nodal landslides (landslide susceptibility influence coefficient, shear stress, shear strength, soil saturation) into the sigmoid activation function, and output a probability value of [0,1], that is, probability of landslide occurrence = σ(ω1×landslide susceptibility influence coefficient + ω2×(τ / )+ω3×soil saturation+b), where σ is the sigmoid activation function, ω1, ω2, and ω3 are feature weights (exemplary values are 0.4, 0.35, and 0.25 respectively), and b is the bias term (value -0.1). ② Landslide risk levels (1-5): Based on the probability of landslide occurrence, the risk levels are quantified and classified as follows: probability < 20% is Level 1 (extremely low risk), 20% ≤ probability < 40% is Level 2 (low risk), 40% ≤ probability < 60% is Level 3 (medium risk), 60% ≤ probability < 80% is Level 4 (high risk), and probability ≥ 80% is Level 5 (extremely high risk). ③Inland waterlogging depth: Input the node's hydrological features (current water level, runoff, soil saturation, slope) into the linear regression layer, and combine the water flow evolution law derived from the Saint-Venant equation to output the actual inland waterlogging depth value of the node. The calculation method is: Inland waterlogging depth = h0 + Δh × (Q / Q_max), where h0 is the initial water accumulation depth of the node, Δh is the water level rise value (calculated by the residual constraint of the Saint-Venant equation), Q is the flow rate, and Q_max is the maximum flow rate of the node; ④ Water level change curve: With time as the horizontal axis and node water level as the vertical axis, interpolation fitting is performed based on the water level projection values at each time step (15min) to generate a continuous water level change curve. The fitting method adopts cubic spline interpolation to ensure the smoothness and physical consistency of the curve.
[0046] (2) Progressive early warning triggering: Construct a two-level early warning logic and dynamically trigger based on real-time prediction results.
[0047] ① Level 1 Warning (Landslide Warning): When the probability of a landslide occurring in a certain unit is ≥60% or the risk level is ≥3, a Level 1 warning is immediately triggered. The warning information includes "Unit Number - Expected Occurrence Time - Risk Level - Impact Range (Radius 500m)".
[0048] ② Level II Warning (Waterlogging-Related Warning): The system automatically updates the topology of the map based on the landslide assumption (execute S3) and performs a secondary simulation. If the predicted waterlogging depth of the downstream node is ≥0.5m or the water level rise is ≥0.3m within 30 minutes, a Level II related warning is triggered. The warning information clearly marks the chain relationship of "landslide blocking the confluence → sudden rise in water level → aggravated waterlogging".
[0049] (3) Early warning report generation and push: Automatically generate standardized early warning reports, including text descriptions (chain relationship, risk level) and visualization charts (disaster chain time-series evolution curve, risk heat map); push to the basin disaster prevention and mitigation command platform via HTTP / 2 protocol.
[0050] Figure 5 The standardized early warning report interface output by the model is shown, which includes a landslide risk heat map of the core early warning nodes (showing the spatial distribution of risk levels 1-5), a curve of waterlogging depth change in the next 6 hours (showing the comparison between predicted and measured values), and a text description example containing the disaster chain transmission path (landslide blockage and confluence → sudden rise in water level).
Claims
1. A method for intelligent simulation of chain disasters caused by torrential rains in small watersheds, characterized in that, Includes the following steps: S1. Multi-source data access and feature construction: Acquire relevant data of the target small watershed, perform anomaly processing and data completion on the data, and construct a standardized data matrix containing static geological features and dynamic hydrological features; S2. Constructing an initial dynamic graph driven by disaster chains: Divide the small watershed into several homogeneous hydrological response units as graph nodes, initialize the node features, calculate the edge weights between nodes based on the differences in hydrological status between nodes and the rules of potential impact of geological disasters, and construct an initial adjacency matrix. S3. Real-time updating of graph topology: During disaster simulation, the landslide risk status of nodes is monitored or predicted in real time. When it is determined that the landslide risk of a node is sufficient to change the watershed confluence path, the edge weights or connection status between the corresponding nodes are dynamically adjusted, and a dynamic adjacency matrix sequence that changes over time is generated. S4. Feature fusion and deduction under physical constraints: The updated dynamic graph data is input into a pre-set graph neural network model. Differentiable regularization terms based on physical equations are embedded in the model. The node features are updated and spatiotemporal evolution is deduced under physical constraints. S5. Chain Disaster Time Series Prediction and Early Warning: Decode and extrapolate the node characteristics, output the time series prediction results of landslide risk and waterlogging depth, and trigger graded early warnings based on preset disaster thresholds.
2. The intelligent simulation method for small watershed rainstorm chain disasters according to claim 1, characterized in that, In step S1, the relevant data for the target small watershed includes at least rainfall data, topographic data, soil parameter data, hydrological monitoring data, and disaster chain data; For missing hydrological data, spatiotemporal interpolation using random forest was employed to complete the data; for missing soil data, matching and completion were performed based on lithology and slope similarity.
3. The intelligent simulation method for small watershed rainstorm chain disasters according to claim 1, characterized in that, In step S2, during node feature initialization, each node is assigned an 8-dimensional feature vector, including: average rainfall intensity, soil saturation, current water level, landslide susceptibility influence coefficient, waterlogging risk value, slope, cohesion c, and internal friction angle φ.
4. The intelligent simulation method for small watershed rainstorm chain disasters according to claim 1, characterized in that, In step S2, the rules for calculating the edge weights between nodes include: The study comprehensively considers the differences in soil saturation, water level, and landslide susceptibility among nodes. The landslide susceptibility coefficient is positively correlated with the landslide risk or landslide volume of the upstream node. The edge weight increases as the differences in soil saturation or water level decrease, representing the enhancement of water flow connectivity.
5. The intelligent simulation method for small watershed rainstorm chain disasters according to claim 4, characterized in that, Edge weight between node i and node j The calculation formula is as follows: ； in, For soil saturation, For water level, α represents the landslide susceptibility influence coefficient, and α, β, and γ are weighting coefficients.
6. The intelligent simulation method for small watershed rainstorm chain disasters according to claim 1, characterized in that, In step S3, dynamically adjusting the edge weights or connectivity states between corresponding nodes specifically includes: Set a landslide impact threshold. When the landslide risk parameter of a certain node exceeds the threshold, identify the confluence blocking effect or diversion effect of that node on downstream nodes. When a blocking effect is determined, the connection edge weight between the node and the downstream node is reset to disconnect. When a diversion effect is identified, the edge weights on the original confluence path are reduced, and new connections are established or the edge weights of the side paths are increased based on the new terrain trend.
7. The intelligent simulation method for small watershed rainstorm chain disasters according to claim 1, characterized in that, In step S4, the differentiable regularization terms based on the physical equations include: Waterlogging evolution constraint term: constructed based on the Saint-Venant equation, used to constrain the mass and momentum conservation of water flow motion; Landslide triggering constraint: Constructed based on the Mohr-Coulomb criterion, used to constrain the mechanical determination of slope instability; The constraint terms are transformed into differentiable residual functions, which are then used as part of the loss function in the training or derivation optimization of the model.
8. The intelligent simulation method for small watershed rainstorm chain disasters according to claim 7, characterized in that, In step S4, the continuity and momentum equation residuals of the waterlogging evolution constraint terms are calculated using the Preissmann difference scheme. ; The difference scheme expression for the continuity equation is as follows: ； The difference scheme for the momentum equation is as follows: ； In the formula, Let Q be the water level, A be the flow rate, n be the Manning coefficient, and R be the hydraulic radius. It is the acceleration due to gravity. For spatial step size, For time step; The landslide triggering constraint term is constructed by calculating the residuals of shear stress and shear strength, where shear stress... The calculation expression is as follows: ； shear strength The calculation expression is as follows: ； residual The expression is as follows: ； In the formula, θ is the natural unit weight of the soil, h is the soil layer thickness, and θ is the slope angle. For normal stress, It is the internal friction angle. It is cohesive force.
9. The intelligent simulation method for small watershed rainstorm chain disasters according to claim 8, characterized in that, The total loss function of the graph neural network model is expressed as: ； in, The data-driven prediction error between the model output and the actual observed values; The physical residuals obtained from the Saint-Venant equations; The physical residuals obtained from the Mohr-Coulomb criterion; and The coefficients are used to balance physical constraints and data fitting weights.
10. The intelligent simulation method for small watershed rainstorm chain disasters according to claim 1, characterized in that, In step S5, the tiered early warning adopts a progressive triggering mechanism: When the predicted landslide risk exceeds the first-level threshold, a landslide warning is triggered. Based on the updated topology after the landslide, a second-level simulation is performed. If the predicted waterlogging depth or water level rise of downstream nodes exceeds the second-level threshold, a waterlogging association warning is triggered, and warning information containing the disaster chain transmission path is output.