Mountain torrent debris flow multi-scale intelligent coupling full-chain modeling method

Abstract

The invention provides a mountain torrent and debris flow multi-scale intelligent coupling full-chain modeling method, belongs to the technical field of mountain torrent and debris flow disaster prevention and control, and is used for solving the problems that single-scale modeling precision is insufficient, an optimization strategy is difficult to adapt to disaster dynamics, and error traceability is inaccurate in the prior art. According to the method, a coupling model is constructed according to macro, medium and micro scales, adaptive constraints are generated in combination with a knowledge base, disaster stages are dynamically recognized, optimization strategies are configured, model iteration is driven by targeted splitting errors, resources are adapted according to risk levels, generalization is enhanced through feedback calibration and knowledge base updating, and all modules cooperate to guarantee continuous processes. According to the method, the modeling precision and efficiency are improved, the disaster dynamic response capability is enhanced, and reliable technical support is provided for disaster prevention and control.

Description

Technical Field

[0001] This application relates to the field of flash flood and debris flow disaster prevention and control technology, and in particular to a multi-scale intelligent coupling full-chain modeling method for flash floods and debris flows. Background Technology

[0002] Flash floods and debris flows, as geological disasters characterized by their sudden onset and destructive power, involve multi-scale physical processes, including macroscopic watershed runoff, mesoscopic channel transport, and microscopic slope instability. Accurate modeling is crucial for disaster early warning and prevention. Currently, flash flood and debris flow modeling has evolved from traditional empirical formulas to numerical simulations, gradually incorporating intelligent algorithms to improve modeling efficiency.

[0003] Existing modeling methods mostly focus on simulating physical processes at a single scale or use fixed weights to achieve cross-scale coupling, lacking dynamic adaptation to multi-scale causal relationships; optimization strategies mostly use static parameter configurations, which cannot respond to the differences between the stable and abrupt stages of disaster evolution; error analysis is mostly an overall assessment, making it difficult to locate the specific error sources in intra-scale and cross-scale coupling; at the same time, the modeling process is disconnected from the disaster risk level, resulting in unreasonable allocation of computing resources, insufficient accuracy in high-risk scenarios or low efficiency in low-risk scenarios.

[0004] These shortcomings make it difficult for existing methods to form a full-chain modeling capability of "multi-scale coupling - dynamic optimization - precise source tracing - risk adaptation", which cannot meet the dual requirements of modeling accuracy and efficiency for the prevention and control of flash floods and debris flows. There is an urgent need for a full-chain modeling method that takes into account both multi-scale collaboration and intelligent optimization. Summary of the Invention

[0005] This application provides a multi-scale intelligent coupling full-chain modeling method for flash floods and debris flows, which can solve the technical problems of insufficient accuracy and low efficiency of existing methods through multi-scale collaboration, dynamic optimization and risk adaptation.

[0006] This application provides a multi-scale intelligent coupling full-chain modeling method for flash floods and debris flows. The method defines the coupling scales according to macro-level watersheds, meso-level channels, and micro-level slopes, constructing multi-physical process sub-models at each scale and cross-scale coupling models. Based on the output of the constructed multi-scale models, and combined with a multi-scenario-multi-scale causal knowledge base and real-time data, a multi-scale constrained feasible region is generated. This allows for the identification of disaster evolution stages and the configuration of multi-objective optimization strategies to solve for the Pareto optimal parameter set. The optimal parameter set is substituted into the model to obtain simulation results. Based on these results, simulation errors are decomposed and their contribution is calculated, driving bidirectional iteration between the surrogate and physical models. Using the iterated simulation results as a basis, disaster risk levels are quantified using safety thresholds, and the uncertainty quantification depth and computational resources are adapted according to the risk level. The uncertainty results output by risk quantification and constraint adaptability indicators are used to calibrate model parameters, dynamically update the knowledge base, and finally integrate the results of each step to output the full-chain result.

[0007] By adopting the above technical solutions, a multi-scale coupled full-chain modeling framework was constructed, which integrates causal constraints, stage identification, error tracing and risk assessment into the modeling process, achieving a balance between simulation accuracy and computational efficiency, and solving the limitations of single-scale modeling and static optimization.

[0008] Furthermore, the step of generating the multi-scale constrained feasible region includes: calculating the similarity between the new scene and the scene in the knowledge base; transferring the intra-scale and inter-scale causal matrices of the adapted scene; based on the multi-scale partitioning results, splitting the transfer matrix to obtain intra-scale and inter-scale sub-matrices; optimizing the sub-matrix strength by combining the real-time residuals of each scale; and generating scale-specific constraints and cross-scale weight constraints based on the optimized sub-matrices, scale weights, and cross-scale coupling strength, and superimposing them to form the total constrained feasible region.

[0009] By adopting the above technical solution, the knowledge base is used to realize the rapid migration and dynamic optimization of constraint parameters, so that the feasible domain of constraints can be adapted to the multi-scale features of new scenarios, thereby improving the rationality of optimization parameters and the accuracy of boundary constraints.

[0010] Furthermore, the step of identifying the disaster evolution stage and configuring a multi-objective optimization strategy includes: extracting the state change amount and change rate of the simulation results, calculating the mutation identification index; comparing the mutation index with a set threshold to determine whether the disaster evolution is in a stable period or a mutation period; based on the stage determination result, configuring a multi-objective weight with accuracy priority for the stable period, configuring a stability priority weight for the mutation period and introducing a mutation regularization term, and dynamically adjusting the number of iterations and the number of reference points in combination with strategy requirements.

[0011] By adopting the above technical solutions, dynamic identification and adaptive adjustment of optimization strategies for disaster evolution stages are achieved. The modeling accuracy is guaranteed during the stable period, and the stability of the results is enhanced during the abrupt change period, thereby improving the model's ability to respond to the dynamics of disaster evolution.

[0012] Furthermore, the step of splitting the simulation error and driving the bidirectional iteration of the surrogate-physical model includes: splitting the total simulation error based on the scale weights of the multi-scale model to obtain the intra-scale error and the cross-scale coupling error; calculating the contribution of the two types of errors to determine the dominant error type with the highest proportion; if the dominant error is an intra-scale error, focusing on the training of the corresponding scale surrogate model and the correction of the physical model parameters; if it is a dominant coupling error, optimizing the cross-scale coupling parameters, adjusting the iteration step size and correction magnitude based on the contribution.

[0013] By adopting the above technical solutions, the source of error can be accurately located and targeted optimization can be achieved, avoiding the waste of resources in the overall iteration and improving the efficiency and pertinence of model parameter correction.

[0014] Furthermore, the step of quantifying disaster risk levels and adapting resources includes comparing the iterative simulation results with safety thresholds to classify disaster risks into three levels: low, medium, and high. Based on the risk classification results, low risk only quantifies causal uncertainty, medium risk superimposes and optimizes uncertainty, and high risk covers uncertainty across the entire chain. Iteration parameters are configured in conjunction with risk levels, prioritizing computational accuracy in high-risk scenarios and prioritizing computational efficiency in low-risk scenarios.

[0015] By adopting the above technical solutions, a mechanism for matching risk levels with computing resources was established, ensuring the reliability of early warnings in high-risk scenarios and improving modeling efficiency in low-risk scenarios, thus achieving a rational allocation of resources.

[0016] Furthermore, the step of driving the bidirectional iteration of the agent-physical model also includes: dividing the agent model into multi-scale sub-modules according to the multi-scale division, configuring the training weights of each module based on the error contribution; inputting the trained module into a validation set composed of historical monitoring data, and calculating the validation error; if the validation error exceeds the convergence threshold, returning to retraining; if convergence is achieved, updating the agent-physical model with the module parameters, and outputting the optimized simulation results.

[0017] By adopting the above technical solution, the closed-loop verification of the surrogate-physics model iteration is strengthened, ensuring the training accuracy of each scale module and improving the reliability of the simulation results after iteration.

[0018] Furthermore, the steps of configuring the multi-objective optimization strategy include: integrating the simulation error, the mutation identification index, and the disaster risk level to construct an optimized state space; designing a dynamic reward function that adapts to the stage and risk; adjusting the distribution of reference points based on the state space and reward function, combined with historical Pareto solution sets; and selecting the optimal solution according to the stage determination results during the stable period using robustness indicators, and selecting the Pareto optimal solution using a composite index of robustness and mutation fitness during the mutation period.

[0019] By adopting the above technical solutions, the optimization strategy is deeply integrated with the disaster stage and risk level, and the selected optimal solution is more in line with the actual modeling needs, thus improving the applicability of the optimization results.

[0020] Furthermore, the steps of the feedback calibration model include: calculating the fit index between the optimized weights and the optimized submatrix, setting a threshold to determine the degree of matching; based on the fit evaluation results, analyzing the correlation between the causal uncertainty and the fit index, and determining the direction of constraint adjustment; calibrating the strength of the causal matrix, updating the constraint elasticity coefficient, and generating a new constraint feasible region by combining the optimized matrix, according to the correlation analysis conclusions and risk level.

[0021] By adopting the above technical solution, a dual feedback calibration mechanism for constraints and optimization results was established, which dynamically corrects constraint parameters and improves the adaptability of the constraint feasible region to the modeling process.

[0022] Furthermore, the step of dynamically updating the knowledge base includes: extracting multi-scale features of the new scene, the mutation features, and the risk features to construct a comprehensive feature vector; calculating the similarity between the vector and the scenes in the knowledge base, selecting a preset number of suitable scenes with the highest similarity, and determining the update strategy in conjunction with the constraint adaptability index; if the adaptability is high, adding scene data; otherwise, merging and updating the optimal scene; associating and storing features and constraint information, and providing initial constraint parameters for the new scene based on the updated knowledge base.

[0023] By adopting the above technical solutions, the knowledge base can be dynamically expanded and optimized, providing accurate initial parameters for modeling new scenarios and improving the scenario transfer capability of the method.

[0024] Furthermore, it also includes multi-module collaborative steps: real-time data flow between modules is achieved according to the set data interaction frequency and feedback calibration cycle; the constraints output by the causal constraint module serve as the boundary input of the multi-objective optimization module; the output of the stage identification and risk assessment module guides the switching of optimization strategies; the results of the error tracing module drive the iteration of the surrogate-physical model; the conclusions of the feedback calibration module reversely correct the constraints and optimization parameters; the causal knowledge base provides initial adaptation data for each module; and the linkage of each module forms a closed-loop collaboration.

[0025] By adopting the above technical solutions, a closed-loop collaborative mechanism for each module was constructed, enabling real-time interaction between data and strategies and improving the integrity and efficiency of the entire chain modeling.

[0026] In summary, this application has at least the following beneficial effects:

[0027] A multi-scale, full-chain intelligent modeling method for flash floods and debris flows is provided, achieving a dual improvement in accuracy and efficiency; the adaptability of modeling parameters is enhanced through multi-scale constraint optimization.

[0028] By combining phased and risk-based dynamic optimization, disaster response capabilities can be improved.

[0029] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0030] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0031] Figure 1 shows a schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented.

[0032] Figure 2 A flowchart of a multi-scale intelligent coupling full-chain modeling method for flash floods and debris flows is shown in an embodiment of this application. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0035] This application provides a multi-scale intelligent coupling full-chain modeling method for flash floods and debris flows. Through multi-scale collaboration, dynamic optimization and risk adaptation, it improves modeling accuracy and efficiency, enhances parameter adaptability, strengthens the dynamic response capability of disaster evolution, and provides reliable support for disaster prevention and control.

[0036] Figure 1 shows a schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented.

[0037] Referring to Figure 1, the operating environment includes a multi-module collaborative hardware system that supports the "multi-scale intelligent coupling full-chain modeling method for flash floods and debris flows". This system takes data flow as the main line and integrates five core links: data acquisition, computing and processing, storage and interaction, communication transmission and deployment and operation. Each link is interconnected through a stable network to form a closed loop, providing continuous and reliable hardware support for full-chain modeling.

[0038] The core component of this operating environment is a cluster of multi-scale data acquisition devices, which includes satellite remote sensing equipment and UAV monitoring systems for macro-basins, rainfall and mud depth sensors for meso-level gullies, and soil moisture and slope sensors for micro-level slopes. Its function is to collect multi-scale geographic and environmental data across the entire region. All acquisition devices are connected to edge computing nodes via wired or wireless links to achieve real-time data upload.

[0039] Directly connected to the edge computing nodes are high-performance computing support devices, specifically distributed server clusters equipped with multi-node CPU and GPU acceleration modules. Their core functions are to run intelligent algorithms such as agent model training and multi-objective optimization, process real-time data transmitted from edge nodes and historical data called by the storage system, and feed the computing results back to the edge nodes for real-time decision-making and synchronize them to the storage system for archiving.

[0040] The storage system adopts a combined architecture of enterprise-grade distributed storage devices and cloud backup nodes. Local distributed storage is responsible for caching real-time computing data and frequently accessed causal knowledge bases, while cloud nodes are used for long-term storage of historical monitoring data and simulation results. The system is connected to computing support devices through a high-speed data bus to ensure the efficiency of data reading and writing and dynamic updates of the knowledge base.

[0041] A low-latency communication network runs through all hardware modules, consisting of a 5G communication module and an industrial Ethernet. 5G is responsible for long-distance data transmission between field monitoring equipment and edge nodes, while the industrial Ethernet supports high-speed interaction between core devices such as computing and storage, ensuring that the data interaction latency of each module is controlled at the millisecond level, and meeting the real-time requirements of closed-loop collaboration across the entire chain.

[0042] All hardware components work together in an orderly manner through the link of "acquisition device - edge node - computing cluster - storage system". The edge node receives the acquired data and performs preliminary preprocessing. The computing cluster executes the core modeling algorithm based on the preprocessed data. The storage system provides data support for the algorithm operation. The communication network ensures seamless data flow in each link, and together they build a complete operating environment that is compatible with multi-scale coupling and intelligent optimization.

[0043] This application discloses a multi-scale intelligent coupling full-chain modeling method for flash floods and debris flows.

[0044] Figure 2A flowchart of a multi-scale intelligent coupling full-chain modeling method for flash floods and debris flows is shown in an embodiment of this application.

[0045] Reference Figure 2 The method specifically includes the following steps:

[0046] S1: Define the coupling scale according to macro watershed, meso channel, and micro slope, and construct multi-physical process sub-models and cross-scale coupling models at each scale.

[0047] This step specifically includes the following methods: defining coupled-scale boundaries based on macro-watershed, meso-channel, and micro-slope, with the definition centered on topographic features and the extent of disaster impact, using a characteristic length quantification standard—the macro-watershed is bounded by the watershed, and the characteristic length... It encompasses complete catchment units; the mesoscopic channel is defined by its mouth and head, with a characteristic length. Focusing on the main channels of debris flow transport; the microscopic slope body is defined by the slope fracture surface, with characteristic length... This study targets key areas of slope instability. The scale parameters were all derived from high-precision DEM data (resolution...). Extract and calibrate the boundary position by combining the results of field geological surveys.

[0048] Based on geological and hydrological characteristics at various scales, corresponding multi-physical process sub-models are constructed. The core mathematical principles and formulas of each sub-model are as follows: The macro-basin sub-model takes the runoff generation and confluence process as its core, and uses the improved SCS runoff generation formula to calculate the watershed runoff. The formula is as follows: Where R is the runoff (unit: mm), P is the rainfall during the period (unit: mm, from real-time monitoring data of distributed meteorological stations within the basin), and S is the maximum retention capacity of the basin (unit: mm, determined by soil type and vegetation cover, obtained from the SCS curve number table); the runoff process is calculated using the Muskingan method, with the formula: ,in They are respectively Flow rate at outlet section at any time (unit: ), , They are respectively Net rainfall input to the basin at any time (unit: ), The confluence coefficient is determined by the watershed length, slope, and roughness.

[0049] The mesoscopic channel sub-model focuses on debris flow transport processes, using Bingham's fluid momentum equation to describe the motion state, the formula being: ,in Debris flow density (unit: (Measured by on-site sampling), Q is the flow rate (unit: (output from the macro-basin sub-model), v is the flow velocity (unit: x is the channel length (in meters), and A is the cross-sectional area of ​​the flow path (in meters). (obtained from topographic survey of the ditch). The slope of the ditch (dimensionless, extracted from DEM data). Bingham yield stress (unit: (Indoor rheological test determination) Wet period (unit: ), Let gravitational acceleration be (take) ), It is the Manning roughness coefficient (dimensionless, determined based on the material composition of the channel wall). Hydraulic radius (unit: (Calculated).

[0050] The microscopic slope sub-model focuses on slope stability and uses a simplified Bishop method to calculate the safety factor, as shown in the formula: ,in The slope safety factor (dimensionless) For stability) The weight of the i-th soil strip (unit: kN, calculated from the soil strip volume and natural unit weight) is given. The dip angle of the slip surface of the i-th soil strip (unit: (as determined by geological survey) The pore water pressure of the i-th soil strip is expressed as kPa, which is monitored in real time by a pore water pressure gauge. The width of the i-th soil strip (in meters) The internal friction angle of the i-th soil strip (unit: (Determined by direct shear test) The cohesion of the i-th soil strip (unit: kPa, determined by direct shear test).

[0051] A cross-scale coupled model is constructed by combining the laws of matter and energy transfer across different scales to achieve organic integration of multi-scale models. The coupling relationship is realized through "flux transfer + parameter feedback": the yield output of the macro-basin sub-model is used as the inlet flow boundary condition of the meso-channel sub-model, and the channel interception coefficient is introduced during the transfer process. ( Determined by the ratio of the catchment area of ​​the ditch to the watershed area. ,Right now ,in, For the inlet flow rate of the channel, The catchment area of ​​the gully; the flow velocity and impact force output from the mesoscopic gully sub-model serve as the external load inputs for the microscopic slope sub-model, with the load transfer coefficient... ( Determined by the distance between the ditch and the slope. ,Right now ,in The impact force of debris flow on the slope. The contact area between the slope and the gully is represented by the slope sub-model. At the same time, the landslide volume output by the micro slope sub-model is fed back to the meso gully sub-model to correct the debris flow density parameters, forming a closed-loop coupling of "macro-meso-micro".

[0052] In the foregoing content, the channel interception coefficient The quantization formula is ,in Direct catchment area of ​​the ditch (unit: (Extracted from DEM data using ArcGIS hydrological analysis tools) Total area of ​​the macro-basin (unit: ), Vegetation interception correction factor (range of values) (0.95 for bare land and 0.85 for woodland, determined by the remote sensing vegetation index NDVI). Load transfer factor. The quantization formula is ,in The horizontal distance between the slope and the ditch (unit: m, GPS measured). The attenuation coefficient (take) (The coefficients are obtained by fitting historical landslide-debris flow correlation data), ensuring that the coefficients dynamically decrease with distance, which conforms to the energy transfer law.

[0053] S2: Based on the output of the constructed multi-scale model, combined with the multi-scenario-multi-scale causal knowledge base and real-time data, a multi-scale constrained feasible region is generated, which then identifies the disaster evolution stage and configures a multi-objective optimization strategy to solve the Pareto optimal parameter set.

[0054] The specific methods in this step include: calculating the similarity between the new scene and scenes in the knowledge base, and using the cosine similarity algorithm to achieve scene matching. The similarity calculation formula is as follows: ,in The new scene's comprehensive feature vector consists of macroscopic watershed runoff (S1 macroscopic sub-model output), mesoscopic channel flow velocity (S1 mesoscopic sub-model output), microscopic slope safety factor (S1 microscopic sub-model output), and real-time rainfall (monitored by a field weather station, sampling frequency 1 Hz). Standardized composition; For the feature vectors of historical scenes in the knowledge base, the structure and Consistent; The similarity value ranges from [value range missing]. ,when It is determined to be an appropriate scenario.

[0055] The intra-scale and inter-scale causal matrices of the transfer adaptation scene, the causal matrix is ​​based on Indicates that n is the total number of key influencing factors at multiple scales (including macroscopic watershed area, mesoscopic channel slope, microscopic soil weight, etc., with a total of 12 core factors selected), and matrix elements. This represents the causal influence strength of the i-th factor on the j-th factor, with a value range of... The larger the value, the stronger the influence. Based on the multi-scale partitioning results, the total matrix is... Decomposed into intra-scale submatrices (Macro) (Middle level) (Microscopic) and cross-scale submatrices (Macro-Meso) (Mesoscale-Microscale), the splitting rule is to classify factors according to their scale, and the cross-scale matrix corresponds to the interaction terms of factors at different scales.

[0056] Combine the strength of the submatrix by optimizing the real-time residuals at each scale. The real-time residual is defined as the deviation between the model output value and the measured value at each scale, and the formula is as follows: ( (corresponding to macro, meso, and micro levels respectively), where Output for k-scale models (e.g., macroscopic export flow). These are measured values ​​at the k-scale (from monitoring equipment at the corresponding scale). Optimization uses gradient descent to update matrix elements, with the update formula being: ,in Let t be the matrix element values ​​after t iterations. The learning rate is set to 0.01, calibrated using historical optimization data. The partial derivatives of the residuals with respect to the matrix elements reflect the degree to which changes in the elements affect the residuals.

[0057] Based on the optimized submatrix, scale weights, and cross-scale coupling strength, scale-specific constraints and cross-scale weight constraints are generated, respectively. Scale weights The analytic hierarchy process (AHP) is used to determine the condition by constructing a judgment matrix and calculating eigenvectors, thus satisfying the condition. (Typical value:) , (Adjustable depending on the scenario); cross-scale coupling strength The calculation is performed using mutual information, and the formula is as follows: ,in for The set of factors of scale, For joint probability density, The marginal probability density is obtained from historical data. Scale-specific constraints are... ( Let k be the scale parameter vector. (The safety threshold is at scale k, derived from disaster prevention and control standards), and the cross-scale constraint is... ( The cross-scale interaction threshold (determined by physical experiments) is superimposed with two types of constraints to form the total constrained feasible region.

[0058] Extract the state changes and rates of change from the simulation results, calculate the mutation identification index, and construct the index using cusp catastrophe theory. The formula is as follows: ,in This refers to the state change between two adjacent simulation times (such as the change in flow velocity in the meso-level channel). The simulation time step is set to 10 min to match the monitoring data frequency. This represents the average state over that period; a larger T value indicates a more drastic change in state. The mutation index is compared to a set threshold. use The criteria are determined, that is ,in The mean of the T values ​​during historical stable periods. The standard deviation of historical T-values ​​is calculated from over 100 sets of stable period data in the knowledge base; when If the time frame is right, it is considered a mutation period; otherwise, it is considered a stable period.

[0059] Based on the stage assessment results, during the stable period, multi-objective weights are configured with priority given to accuracy, while during the abrupt change period, weights are configured with priority given to stability, and a change regularization term is introduced. The multi-objective optimization objectives include minimizing simulation error. Minimize parameter fluctuations The stationary option repositioning vector is The mutation period is The formula for the mutation regularization term is: ,in Let be the regularization coefficient (taken as 0.05), which is incorporated into the optimization objective to obtain the overall objective during the mutation period. The number of iterations and reference points are dynamically adjusted and optimized based on strategy requirements, with a focus on the number of iterations during stable periods. Number of reference points The mutation period is designed to reduce response time. The parameter values ​​are determined based on historical optimization efficiency and accuracy statistics.

[0060] By integrating the simulation error, the mutation identification index, and the disaster risk level, an optimized state space is constructed. The state space uses the model's core parameters as variables (e.g., curve numbering of the macroscopic SCS model, yield stress of the mesoscopic Bingham fluid, cohesion of the microscopic Bishop method, etc., totaling 8 core parameters), and the variable boundaries are defined by the aforementioned total constraint feasible region. A dynamic reward function for the adaptation stage and risk is designed, with the formula Reward. ,in To allow for the maximum simulation error (taken as 0.15, from industry standards), This is the largest mutation indicator in history. Quantification of risk level (low) middle high ), Stable period mutation period The sum of the coefficients is 1 to ensure weight balance.

[0061] Based on the state space and reward function, and combined with adjusting the reference point distribution using the historical Pareto solution set, the NSGA-III algorithm is used to solve the multi-objective optimization problem. The core steps of the algorithm are: initializing the population (size 200, randomly generating parameter combinations that satisfy the constraints), fast non-dominated sorting, calculating crowding distance, selecting offspring based on reference points, and iterating until the termination condition is met (reaching the set number of iterations). The historical Pareto solution set comes from the optimization results of similar scenarios in the knowledge base. The reference point adjustment rule is to use the mean of the parameters of the historical best solution as the center of the new reference point to ensure that the search direction focuses on the efficient region. The reference point adjustment adopts a "historical best + current state" fusion model, and the formula is as follows: ,in The center vector of the historical Pareto solution set (calculated from the mean of the parameters of 50 optimal solutions in the knowledge base). ), This is the initial optimization parameter vector for the current scenario (derived from the causal knowledge base matching results). Weight (stable period) Increased to 0.6, mutation period (Raised to 0.7). This adjustment logic reflects the technological evolution from "fixed experience reference to dynamic scene adaptation reference", solving the problem of the disconnect between the traditional NSGA-III reference point and the real-time scene.

[0062] Based on the stage determination results, the optimal solution is selected during the stable period using the robustness index. The robustness index formula is as follows: ,in For parameter perturbation (take) ), Rb represents the standard deviation of the objective function under perturbation; a larger Rb value indicates stronger robustness. During the mutation period, a composite index of robustness and mutation fitness is used for screening. The solution with the largest C value is selected as the Pareto optimal parameter set, and the Pareto optimal parameter set is obtained by solving the solution.

[0063] S3: Substitute the optimal parameter set into the model to obtain the simulation results, split the simulation error based on the results and calculate the error contribution, and drive the bidirectional iteration of the surrogate-physical model.

[0064] The specific steps in this process include: The optimal parameter set is the Pareto optimal solution obtained in S2, containing eight core parameters such as the macroscopic watershed SCS curve number, the mesoscopic channel Bingham yield stress, and the microscopic slope cohesion. These parameters are substituted into the multi-scale coupled model constructed in S1, outputting simulated values ​​of key physical quantities at each scale (e.g., macroscopic outlet flow rate, mesoscopic debris flow velocity, and microscopic slope displacement), forming a complete simulation result set. Based on these results, the total simulation error is calculated and quantified using the root mean square error (RMSE), as shown in the formula: ,in This is the simulated value (model output) of the i-th physical quantity. The corresponding measured values ​​(from synchronous monitoring data of macro-level hydrological stations, meso-level current meters, and micro-level displacement meters), N is the total number of physical quantities (taking 15 core monitoring indicators), and the smaller the RMSE value, the higher the simulation accuracy.

[0065] Based on the scale weights of the multi-scale model, the total simulation error is decomposed to obtain the intra-scale error and cross-scale coupling error. Scale weights The results determined by the analytic hierarchy process in S2 are used. The total error is decomposed using a weighted sum of squares, as shown in the formula: ,in The in-scale error at the k-th scale is calculated as follows: ( The number of physical quantities at the k-th scale, such as macroscopic... The cross-scale coupling error is obtained by subtracting the sum of squared weighted errors within each scale from the squared total error, and then taking the square root. This reflects the error caused by the interaction between scales.

[0066] The contribution of the two types of errors is calculated, and the contribution is quantified using the error contribution rate, as shown in the formula: (Intra-scale error contribution rate) (Cross-scale error contribution rate), where This represents the percentage of error contribution within the k-th scale. The contribution ratio of cross-scale error, satisfying Identify the dominant error type with the highest proportion, i.e., when a certain... or Greater than When the error is less than 1, it is determined to be the current dominant error; if all are less than 1, it is determined to be the dominant error. If the two types of errors with the largest proportion are selected as the joint dominant error, then the two types of errors with the largest proportion are selected as the joint dominant error.

[0067] If the dominant error is an in-scale error, the focus is on training the corresponding scale surrogate model and correcting the physical model parameters. The surrogate model is constructed using the Fourier Neural Operator (FNO), which excels at handling high-dimensional spatiotemporal data. The input is a vector of influencing factors at that scale (e.g., microscale inputs such as soil moisture, slope, and pore water pressure), and the output is the simulated value at that scale. The training loss function is... ,in The surrogate model outputs values, and the Adam optimizer is used to minimize the loss and learning rate. (Calibrated by model convergence speed). Physical model parameter correction uses gradient descent; the correction formula is as follows: ,in For physical parameters (such as the cohesion c of the microscopic slope) at iteration number t. Let be the partial derivative of the scale error with respect to this parameter. The higher the contribution rate to the error, the greater the parameter correction, ensuring that optimization focuses on the core error sources. The input dimensions of FNO are determined by the number of core influencing factors at each scale. The macro-level module has 4 input dimensions (rainfall, watershed area, vegetation cover, soil type coding), the meso-level module has 5 (gully slope, roughness, mud depth, density, flow velocity), and the micro-level module has 3 (soil moisture, slope angle, pore water pressure). Dimension selection is achieved through mutual information entropy. ( As a factor, (For output) Determined. Number of Fourier modes. The selection formula is ,in The frequency features of the input data (extracted by Fast Fourier Transform (FFT), macroscopic module) Therefore Micro-modules Therefore This ensures that the model strikes a balance between accuracy and computational efficiency, reflecting a refined modeling evolution path of "factor selection - frequency adaptation".

[0068] If the dominant coupling error is to be optimized, the core optimization object is the cross-scale coupling strength defined in S2. Optimization was achieved using a genetic algorithm, with a population size of 50, 30 iterations, and a fitness function of [function name missing]. (That is, maximizing fitness is equivalent to minimizing coupling error). The selection operator uses roulette wheel selection, the crossover operator uses single-point crossover (crossover probability 0.8), and the mutation operator uses Gaussian mutation (mutation probability 0.05). The final output is the one with the highest fitness. As optimized parameters, the interaction term calculation logic of the cross-scale coupling model is updated. In both cases, the iteration step size and correction magnitude are adjusted based on the contribution, specifically the iteration step size... ,in The contribution rate of the corresponding error (e.g., when the dominant error is an in-scale error). This ensures that errors with high contribution rates are optimized in a more targeted manner.

[0069] Based on the aforementioned multi-scale division, the surrogate model is split into three sub-modules: macroscopic, mesoscopic, and microscopic. Each module corresponds to the simulation of physical processes at its respective scale, and data interaction between modules is achieved through cross-scale coupling parameters. The training weights of each module are configured based on the aforementioned error contribution, and the weight calculation formula is as follows: ,satisfy —If the macroscopic scale error contribution rate Then the training weights of the macro module During training, this module has a higher weight in the loss function, prioritizing the reduction of core errors. The training dataset uses a hybrid set of "historical monitoring data + numerical simulation data". The historical data comes from monitoring records of typical flash floods and debris flows in the past 10 years (200 sets of samples in total), while the numerical simulation data is generated by the physical model under different parameter combinations (500 sets of samples in total), ensuring that the data covers multiple scenarios.

[0070] The trained module is then fed into a validation set consisting of historical monitoring data. The validation set is randomly selected from the historical data (which constitutes a portion of the total historical data). (A total of 40 samples) The validation error was calculated using the mean absolute percentage error (MAPE), and the formula is: ,in Number of validation set samples. Convergence threshold. use The criteria are determined, that is ,in The mean of the validation error during historical training (approximately) ), The standard deviation of historical verification error (approximately) ),final Pick If the verification error Return to retraining, this time adjusting the learning rate to... To avoid shocks; if Once the model converges, the optimal parameters of the module (the weights and biases corresponding to the minimum loss function) are used to update the coupling interface between the surrogate and physical models. In other words, the output of the surrogate model is used as the basis for correcting the boundary conditions of the physical model, realizing a two-way iteration of "surrogate model approximating physical laws and physical model providing real constraints". Finally, the optimized simulation results are output, and the RMSE of the results is reduced by more than 30% compared with the previous iteration (based on historical optimization effect statistics).

[0071] S4: Based on the iterative simulation results, combine the safety threshold to quantify the disaster risk level, and adapt the uncertainty quantification depth and computing resources according to the risk level.

[0072] The specific methods in this step include: the iterative simulation results are the optimized data output by S3, and the core indicators cover the maximum flow rate at the macro-basin outlet section. (unit: Peak velocity of debris flow in mid-level gullies (unit: ), Maximum cumulative displacement of micro slope (Unit: m), all three factors together reflect the degree of disaster damage. The safety threshold needs to be determined by combining the current "Technical Standards for Mountain Flood Disaster Prevention Engineering" with historical disaster data of the study area, denoted as... (Macro safety threshold, such as) (Determined by the basin flood control standards) (Meso-level safety threshold, such as) (Refer to the mudslide kinematic safety specifications) (Microscopic safety threshold, such as 0.5 m, based on measured data of critical displacement for slope stability). The threshold needs to be revised annually based on geological survey results.

[0073] The iterative simulation results are compared with a safety threshold, and the risk level is quantified using a risk index method, the formula of which is: ,in Scale weights (using the results of the S2 analytic hierarchy process) This is a risk index, with a range of values. .based on Disaster risk is classified into three levels: low risk. This indicates that all simulated indicators are below the safety threshold. The probability of disasters occurring is lower than Medium risk This indicates that some indicators are close to or slightly exceed the threshold, and the probability of disaster occurrence is high. High risk ( This indicates that the core indicator has significantly exceeded the threshold, and the probability of disaster occurrence exceeds [a certain threshold]. Risk classification boundary values ​​are derived from historical disaster cases. Statistical value calibration (e.g., selecting 10 disasters that have already occurred) The minimum value of 1.2 is used as the high-risk threshold.

[0074] Based on the risk classification results, low risk only quantifies causal uncertainty. Causal uncertainty stems from the randomness of the correlation between multi-scale factors, and is quantified using conditional entropy, with the formula as follows: ,in This is a set of influencing factors (such as rainfall and soil moisture). For simulation result indicators (such as) ), For joint probability, The conditional probability is obtained from more than 1,000 sets of data in a multi-scenario, multi-scale causal knowledge base. Only this indicator is calculated for low-risk scenarios, and the quantification depth is controlled within 10 core factors.

[0075] The medium-risk factor is superimposed with optimization uncertainty, which is caused by fluctuations in the Pareto optimal parameter set. This uncertainty is quantified using a parameter sensitivity coefficient, the formula of which is: ,in To optimize the number of parameters (8 core parameters determined by S2), The sensitivity coefficient of the i-th parameter ( For simulation results, (The parameter values ​​are calculated using the controlled variable method). The standard deviation of the parameters in the Pareto solution set needs to be calculated simultaneously for medium-risk scenarios. and And analyze the coupling relationship between the two.

[0076] In medium-risk scenarios, and The coupling relationship is quantized using covariance, and the formula is as follows: ,in These are the means of the two types of uncertainty, respectively. For the expectation operator, the combined uncertainty after coupling is: .when (That is, when the causal uncertainty increases, the optimization uncertainty also increases), so the optimization weights need to be adjusted. promote This formula embodies the evolution of mathematical principles from "independent quantization to coupled correlation quantization," and solves the shortcomings of traditional methods that ignore the interactive effects of uncertainties.

[0077] High risk covers uncertainty across the entire chain. This uncertainty is a superposition of uncertainties related to causality, optimization, and model structure, as shown in the formula: ,in To account for model structural uncertainties, the relative deviations output by different models (such as the physical model of S1 and the surrogate model of S3) are calculated. High-risk scenarios require output The quantification depth covers all 12 influencing factors and 8 optimization parameters, including the proportion of each component.

[0078] Iteration parameters are configured in conjunction with risk levels, following the principle of "risk matching resources." A linear correlation formula is used to determine the core calculation parameters: the number of iterations. Calculate the number of nodes ,in As the baseline iteration number, The baseline number of computing nodes is (single node configuration: 8-core CPU + 16GB GPU). In high-risk scenarios, priority is given to ensuring computational accuracy. hour, With a maximum value of 200 and a maximum value of 8 for K, computation time is allowed to be extended to within 2 hours; in low-risk scenarios, computational efficiency is prioritized. hour, Drop to 30 Keep the value at 2, keep the calculation time within 30 minutes, and use linear interpolation of parameters for medium-risk scenarios according to the formula to achieve a balance between accuracy and efficiency.

[0079] S5: Utilize the uncertainty results and constraint fit indicators from risk quantification output to provide feedback for calibrating model parameters, dynamically update the knowledge base, and finally integrate the results of each step to output the full-chain outcome.

[0080] The specific method of this step includes: the uncertainty result of the risk quantification output is the causal uncertainty calculated by S4. Optimize uncertainty and uncertainty across the entire chain (Select corresponding indicators according to risk level), the optimized weights are the weight vector of S2 multi-objective optimization. (Stability period) mutation period The optimized submatrix is ​​the S² optimized intra- / inter-scale causal submatrix. Corresponding to macro, meso, and micro levels, t represents the current iteration number. The Pearson correlation coefficient is used to calculate the fit between the two, and the formula is: , where n is the number of dimensions that match the weights and the matrix (taken as 2, corresponding to two optimization objectives). The mean of the weight vector. submatrix The row mean, Let A be the overall mean of the submatrix, and its value range is... The larger the absolute value, the stronger the fit. Let a threshold be used to determine the degree of matching. Based on historical adaptation data statistics, the following approach was adopted. Criterion Calculation: ,in This is the average of 100 historical fitness metrics (approximately 0.6). The standard deviation is approximately 0.1. ;when It is judged as a good match at that time. If the timing indicates a mismatch, constraint adjustment needs to be initiated. Based on the fit assessment results, the causal uncertainty is analyzed. The correlation with the fit index A is quantified using mutual information, and the formula is as follows: ,in , A and The marginal probability, For joint probability, by Historical output data statistics were obtained; correlation degree At that time, the judgment Strongly correlated with A, constraint adjustments should prioritize reducing it. .

[0081] After determining the direction of constraint adjustment, based on the correlation analysis results and the risk level RI (S4 output), the strength of the causal matrix is ​​calibrated, and the matrix elements are updated using gradient descent with risk weights, as shown in the formula. ,in To calibrate the learning rate (when well matched) When the matching deviation occurs The introduction of RI (Individualized Resilience Index) is to define the partial derivative of the absolute value of the adaptability index with respect to the matrix elements. This makes high-risk scenarios (…) Calibration range increased That's all. Update the constraint elasticity coefficients. The elasticity coefficient reflects the constraint's ability to adapt to parameter changes, and the formula is: ,in As the reference elastic coefficient, The larger the value, the stronger the flexibility, avoiding overly strict constraints that could lead to an unsolvable optimization problem; combined with the optimized matrix... With the new elastic coefficient Regenerate the total constraint feasible region, and the new constraint boundary is (Scale-specific constraints) (Cross-scale constraints) ensure that constraints adapt to changes in parameters.

[0082] Extracting multi-scale features of the new scene, the mutation features, and the risk features, the multi-scale features include the scale features of S1 (macro-basin area). Mid-level gully slope Microscopic slope inclination angle ) and model output features ( Mutation index with mutation characteristic of S3 Risk index with risk characteristics of S4 Risk level labels; after standardizing these features, a comprehensive feature vector is constructed. The standardized formula is ( (These are the original feature values; extreme values ​​are derived from historical data in the knowledge base). The similarity between this vector and scenes in the knowledge base is calculated using the cosine similarity formula of S². ( (For historical scene vectors in the knowledge base), select the preset number of matching scenes with the highest similarity (set as Top-3 to balance computational load and adaptability).

[0083] Based on the constraint adaptability index A, the decision update strategy is as follows: if the Top-1 scenario... and If the adaptability is high, the new scenario will be considered suitable. Optimized , and Wait for the data to be added to the knowledge base; if or Then, the new scene data will be merged and updated with the Top-1 scene data. The fusion will use a weighted average method, and the feature values ​​after fusion will be... (Top-1 scene feature values), causal matrix ( (This is a Top-1 scene matrix). It associates and stores feature and constraint information, with a storage structure of "comprehensive feature vector". - Causal matrix - Optimize weights "Risk Level - Uncertainty Indicator" provides initial constraint parameters for new scenarios based on the updated knowledge base; these initial parameters constitute the new scenario. Most similar historical scene and This eliminates the need to start modeling new scenes from scratch.

[0084] The integrated data comprises the output data from each step of multi-scale modeling, constraint generation, optimization iteration, risk quantification, and calibration update. The integrated data includes: the multi-scale sub-model structure and core parameters of S1 (such as macroscopic SCS curve numbering, mesoscopic Bingham yield stress, and microscopic cohesion); the constraint feasible region boundary and Pareto optimal parameter set of S2; the surrogate model training weights and optimized simulation results of S3; and the risk index RI, risk level labels, and corresponding uncertainty indicators of S4. ; Causal matrix of S5 after calibration Along with the updated knowledge base index, this data is organized logically according to "scale-process-results" to form a complete modeling output that includes data tables, visualization charts (such as risk level heatmaps and error contribution pie charts), and parameter description documents. The output formats support CSV (data), PNG (charts), and PDF (documents), meeting the data analysis and reporting needs for disaster prevention and control decision-making.

[0085] S6: Real-time data flow between modules is achieved according to the set data interaction frequency and feedback calibration cycle. The constraints output by the causal constraint module serve as the boundary input of the multi-objective optimization module. The output of the stage identification and risk assessment module guides the switching of optimization strategies. The results of the error tracing module drive the iteration of the proxy-physical model. The conclusions of the feedback calibration module correct the constraints and optimization parameters in reverse. The causal knowledge base provides initial adaptation data for each module. The linkage of each module forms a closed-loop collaboration.

[0086] This step specifically includes: establishing data transmission links between modules according to a set data interaction frequency and feedback calibration cycle. The formula is determined based on the update cycle of the core monitoring data. ,in The sampling period of the monitoring equipment in S1 (such as the rainfall sensor) ,but (times / hour) to ensure data flow and monitoring rhythm are synchronized; feedback calibration cycle Combined with the model iteration efficiency setting, The number of iterations for S3 (e.g.) ), If the time consumed in a single iteration is then... This avoids resource waste caused by frequent calibration. Data transmission adopts a hybrid link of "edge node - industrial Ethernet - cloud cluster", and field modules (such as micro slope monitoring) transmit via 5G protocol (latency...). The core computing module interacts via industrial Ethernet (latency). The transmitted data is encapsulated in JSON format, containing a data identifier, timestamp, value, and checksum. The checksum formula is Check. ( (Data bytes, Check is the checksum) to ensure the integrity and accuracy of data transmission.

[0087] Ensure that the constraints generated by the causal constraint module are transmitted to the multi-objective optimization module in real time as boundary inputs. The output of the causal constraint module is the total constraint feasible region parameters optimized by S² (scale-specific constraints). Cross-scale constraints and elasticity coefficient After being transmitted to the multi-objective optimization module, it is directly used as the boundary conditions of the NSGA-III algorithm, i.e., the parameters in the algorithm. Must meet and The constraint update trigger mechanism is "transfer immediately after calibration". After S5 completes the causal matrix calibration, it triggers an interrupt signal to notify the optimization module to load the new constraints.

[0088] The phase identification and risk assessment modules synchronize the phase determination results and risk levels to the optimization module to guide strategy switching. The phase identification module outputs the mutation index T and phase label (stable period / mutation period) of S2, and the risk assessment module outputs the risk index RI and level label of S4. Strategy switching uses threshold triggering logic, when... ( (The mutation threshold determined by S2) or At that time, the optimization module immediately changed the weight vector From the stable period Switch to mutation period And enable the mutation regularization term for S2. ;when and At that point, switch back to precision-priority weighting and adjust the number of iterations. (Press S4) Switch response time This ensures that strategies are matched with disaster conditions in real time.

[0089] The error contribution data of the error tracing module drives the agent-physical model to initiate targeted iterations, and the error contribution rate of S3 is output by the error tracing module. (Within scale) and (Cross-scale), the iteration trigger threshold is set to When a certain Immediately upon receiving the instruction, a training command is sent to the surrogate model module at the corresponding scale, carrying the error contribution at that scale. With core error source parameters (such as microscale) At that time, carrying cohesion error partial derivative ); the proxy model module is based on Configure training weights and according to the loss function of S3 Training is initiated, and once completed, the optimized parameters are fed back to the physical model module, triggering parameter correction and forming a linkage chain of "error localization - targeted training - parameter update".

[0090] The conclusions of the feedback calibration module are used to correct the parameters of the causal constraint module and the optimization module. The output of the feedback calibration module is the calibrated causal matrix S5. Constraint elasticity coefficient and the adaptability index A; the modification to the causal constraint module is to directly replace the original matrix. for And update the constraint boundary calculation logic, the new boundary formula is: ( (where S5 is the fitness threshold), when hour, The constraint flexibility is increased by 1.2 times; the optimization module is modified by adjusting the reference point distribution, updating the historical Pareto solution set center of S2 to the mean of "calibrated parameters + high-adaptability scene parameters", and adjusting the correction period and feedback calibration period. Consistency ensures coordinated calibration of constraints and optimizations.

[0091] The causal knowledge base provides adaptive initial parameters for the initialization of each module. The knowledge base invocation mechanism is "triggered upon startup of a new scene". When there is a new modeling task, the multi-scale modeling module first extracts the comprehensive feature vector of the new scene. (S5 definition), using the cosine similarity formula of S5 Match the top 3 historical scenarios; generate the causal matrix of the scenario with the highest matching degree. Scale weight Optimize initial parameters The initialization data is pushed to the causal constraint, multi-objective optimization, and physical model modules respectively. The adaptability of the initialization parameters is verified by the adaptability index A of S5. If enabled, it will be activated directly; otherwise, a quick calibration will be triggered. (Iterate 10 times) to ensure the validity of the initial parameters.

[0092] The maximum permissible delay threshold for module collaboration is categorized by function: causal constraints → constraint transmission delay for optimization modules. (Affecting the real-time performance of optimization boundaries), stage identification → strategy switching delay of the optimization module. (Affecting disaster response speed), error tracing → iteration triggering delay of the surrogate model (Affects error correction efficiency). The delay threshold is calibrated using the formula relating "disaster evolution rate - response time": When the latency exceeds the threshold, a temporary local cache parameter activation mechanism is triggered. The cache parameter is the weighted average of the past 5 valid data points. ( Decays over time, This ensures modeling continuity even when collaboration is interrupted. It achieves data linkage and closed-loop collaboration between modules. The effectiveness of closed-loop collaboration is verified by the data consistency index between modules, expressed as Cons. Where K is the number of data items in the module interaction (taking 10 core data items, such as constraint boundaries). Optimize weight input and error wait), To output module data, Receive data for the input module; Cons For effective collaboration, when Cons At the same time, data retransmission and link diagnosis are initiated to ensure that there are no data breaks in the closed loop of "causal constraint → optimization → iteration → risk assessment → calibration → constraint", ultimately achieving real-time, accurate and collaborative full-chain modeling.

[0093] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0094] This application employs a progressive combination of technical means—"multi-scale modeling as the foundation—constraint optimization and orientation—error iteration and refinement—risk adaptation and effectiveness control—feedback update and generalization—collaborative mechanism to ensure smooth operation"—to form a complete logical chain from data input to output. The technical roles and effects of each component are derived as follows:

[0095] First, S1 defines the boundaries and constructs sub-models according to macro-, meso-, and micro-scales. It accurately describes the physical processes at each level through scale-specific mathematical models such as the SCS runoff formula and Bingham fluid equation. Combined with flux transfer coefficients, it achieves cross-scale coupling. This approach solves the problem of "macro-scale inaccuracy and micro-scale redundancy" in single-scale modeling from a spatial dimension. It quantifies the correlation between watershed runoff, channel transport, and slope instability, providing a realistic physical basis for subsequent full-chain calculations.

[0096] Based on the model output of S1, S2 matches the scene transfer causal matrix of the knowledge base using cosine similarity, and generates a constrained feasible region adapted to the new scene through gradient descent optimization. Simultaneously, it quantifies the mutation index using cusp catastrophe theory, dynamically adjusts the multi-objective optimization weights and regularization terms, and finally uses the NSGA-III algorithm to solve for the Pareto optimal parameter set. The scene adaptability of the constrained feasible region and the stage responsiveness of the optimization strategy complement each other, avoiding the blindness of parameter search while ensuring both accuracy during the stationary period and stability during the mutation period, making the optimized parameter set both scientific and practical.

[0097] S3 inherits the optimal parameters from S2, using a weighted sum of squares to separate intra-scale and cross-scale errors. It identifies the dominant error source by the error contribution rate—if it's an intra-scale error, it focuses on training the corresponding FNO surrogate model and correcting physical parameters; if it's a cross-scale error, it optimizes the coupling strength. This "targeted source tracing—precise iteration" logic eliminates the resource waste of traditional overall iteration. The introduction of the surrogate model further reduces the computational cost of repeatedly correcting the physical model, directly achieving a simultaneous improvement in simulation accuracy and efficiency.

[0098] S4 compares the iteration results of S3 with the safety threshold, quantifies the level using the risk index RI, adapts the uncertainty calculation depth according to the principle of "simplified quantification for low risk and full quantification for high risk," and then allocates computing resources using a linear correlation formula. The linkage between risk level and resource allocation ensures the reliability of uncertainty quantification across the entire chain in high-risk scenarios, while reducing resource redundancy through an efficiency-first strategy in low-risk scenarios, achieving refined management of "risk-adapted resources."

[0099] S5 uses the Pearson coefficient to evaluate the fit between the optimization weights and the causal matrix, and combines mutual information analysis to determine the correlation between uncertainty and fit, calibrating the matrix and updating the constraints. Simultaneously, it extracts new scene features to construct vectors, and updates the knowledge base through similarity matching and weighted fusion. This feedback mechanism allows constraint parameters to be dynamically adjusted according to the scene, and the continuous expansion of the knowledge base provides accurate initial parameters for subsequent new scenes, forming a virtuous cycle of "modeling—calibration—reuse," strengthening the method's scene generalization ability.

[0100] Ultimately, S6 constructs a closed-loop data flow for each module through a hybrid "5G + Industrial Ethernet" link and tiered latency thresholds: boundary conditions drive optimization in the causal constraint module, stage and risk results guide strategy switching, error data triggers iteration, calibration conclusions correct parameters, and the knowledge base ensures initialization efficiency. Real-time linkage between modules eliminates data breakpoints, ensuring the continuity of the entire process from scale division to output, enabling the modeling process to dynamically respond to changes in monitoring data and disaster status.

[0101] The aforementioned technical methods are interconnected, forming a complete logical chain from physical foundation, parameter optimization, error correction, risk management to collaborative mechanisms: "data input - pattern quantification - parameter optimization - accuracy improvement - results output". Ultimately, this achieves the core effects of "physical accuracy, scientific parameters, timely response, and efficient resource utilization" in flash flood and debris flow modeling, providing reliable technical support for disaster prevention and control.

[0102] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A multi-scale intelligent coupled full-chain modeling method for flash floods and debris flows, characterized in that, include: The coupling scales are defined by macro-watershed, meso-channel, and micro-slope, and multi-physical process sub-models and cross-scale coupling models are constructed at each scale. Based on the output of the constructed multi-scale model, combined with a multi-scenario-multi-scale causal knowledge base and real-time data, a multi-scale constrained feasible region is generated, which then identifies the disaster evolution stage and configures a multi-objective optimization strategy to solve the Pareto optimal parameter set. Substitute the optimal parameter set into the model to obtain the simulation results. Based on the results, decompose the simulation error and calculate the error contribution, driving the bidirectional iteration of the agent-physical model. Based on the iterative simulation results, the disaster risk level is quantified by combining the safety threshold, and the uncertainty quantification depth and computing resources are adapted according to the risk level. By utilizing the uncertainty results and constraint fit indicators from risk quantification, the model parameters are calibrated, the knowledge base is dynamically updated, and finally, the results of each step are integrated to output the full-chain results.

2. The method according to claim 1, characterized in that, The step of generating the multi-scale constrained feasible region includes: Calculate the similarity between the new scene and the scenes in the knowledge base, and transfer the intra-scale and inter-scale causal matrices of the adapted scene. Based on the multi-scale partitioning results, the migration matrix is ​​split to obtain intra-scale and inter-scale submatrices, and the strength of the submatrices is optimized by combining the real-time residuals of each scale. Based on the optimized submatrix, scale weights, and cross-scale coupling strength, scale-specific constraints and cross-scale weight constraints are generated respectively, and superimposed to form the total constraint feasible region.

3. The method according to claim 2, characterized in that, The steps of identifying the disaster evolution stages and configuring multi-objective optimization strategies include, Extract the state change amount and rate of change from the simulation results, and calculate the mutation identification index; By comparing the mutation index with a set threshold, it can be determined whether the disaster evolution is in a stable period or a period of sudden change. Based on the stage judgment results, during the stable period, multi-objective weights with priority given to accuracy are configured, and during the mutation period, weights with priority given to stability are configured and mutation regularization terms are introduced. The number of iterations and the number of reference points are dynamically adjusted in combination with strategy requirements.

4. The method according to claim 1, characterized in that, The steps of splitting simulation errors and driving bidirectional iteration of the surrogate-physics model include: Based on the scale weights of the multi-scale model, the total simulation error is split to obtain the intra-scale error and cross-scale coupling error. Calculate the contribution of the two types of errors and determine the dominant error type with the highest proportion; If the dominant error is an intra-scale error, focus on training the corresponding scale surrogate model and correcting the physical model parameters; if it is a dominant coupling error, optimize the cross-scale coupling parameters, adjusting the iteration step size and correction magnitude based on the contribution.

5. The method according to claim 3, characterized in that, The steps of quantifying disaster risk levels and allocating resources include: The iterative simulation results are compared with safety thresholds to classify disaster risks into three levels: low, medium, and high. Based on the risk classification results, low risk only quantifies causal uncertainty, medium risk includes optimization uncertainty, and high risk covers uncertainty across the entire chain. The iteration parameters are configured in accordance with the risk level, prioritizing computational accuracy in high-risk scenarios and computational efficiency in low-risk scenarios.

6. The method according to claim 1 or 4, characterized in that, The steps of driving the agent-physical model bidirectional iteration also include... According to the multi-scale division, the surrogate model is split into multi-scale sub-modules, and the training weights of each module are configured based on the error contribution. The trained module is input into a validation set consisting of historical monitoring data, and the validation error is calculated. If the verification error exceeds the convergence threshold, retraining is required. If convergence is achieved, update the agent-physics model with the parameters of this module and output the optimized simulation results.

7. The method according to claim 5, characterized in that, The steps for configuring the multi-objective optimization strategy include: By integrating the simulation error, the mutation identification index, and the disaster risk level, an optimized state space is constructed, and a dynamic reward function that adapts to the stage and risk is designed. Based on the state space and reward function, the distribution of reference points is adjusted by combining the historical Pareto solution set. Based on the stage determination results, the optimal solution is selected according to the robustness index during the stable period, and the Pareto optimal solution is selected using a composite index of robustness and mutation fitness during the mutation period.

8. The method according to claim 5, characterized in that, The steps of the feedback calibration model include: Calculate the fit index between the optimized weights and the optimized submatrix, and set a threshold to judge the degree of matching; Based on the results of the adaptability assessment, the correlation between the causal uncertainty and the adaptability index is analyzed to determine the direction of constraint adjustment; Based on the correlation analysis results and risk level, the strength of the causal matrix is ​​calibrated, the constraint elasticity coefficient is updated, and a new constraint feasible region is generated by combining the optimized matrix.

9. The method according to claim 5, characterized in that, The steps for dynamically updating the knowledge base include: Extract multi-scale features, mutation features, and risk features of the new scene to construct a comprehensive feature vector; Calculate the similarity between the vector and the scenes in the knowledge base, select the preset number of suitable scenes with the highest similarity, and determine the update strategy in combination with the constraint adaptability index; If the adaptability is high, new scene data is added; otherwise, the optimal scene is merged and updated. The associated storage features and constraint information are used to provide initial constraint parameters for new scenes based on the updated knowledge base.

10. The method according to any one of claims 1 to 9, characterized in that, It also includes a multi-module collaborative process: real-time data flow between modules is achieved according to the set data interaction frequency and feedback calibration cycle; the constraints output by the causal constraint module serve as the boundary input of the multi-objective optimization module; the output of the stage identification and risk assessment module guides the switching of optimization strategies; the results of the error tracing module drive the iteration of the surrogate-physical model; the conclusions of the feedback calibration module reversely correct the constraints and optimization parameters; the causal knowledge base provides initial adaptation data for each module; and the linkage of each module forms a closed-loop collaboration.

Images