A flash flood simulation method based on physical perception residual learning and dynamic runoff compensation

Abstract

This invention discloses a flash flood simulation method based on physical perception residual learning and dynamic runoff compensation, belonging to the field of flash flood prevention engineering technology. The method includes: using a spatial indexing algorithm to match the locations of monitoring stations in the target watershed with the grid positions in the digital elevation model; constructing a high-precision computational domain using a station-specific DEM etching algorithm; combining the grid data in the high-precision computational domain with a soil saturation feedback factor to construct a dynamic runoff compensation mechanism; calculating the basic simulated flow rate using a two-dimensional diffusion wave equation; constructing a physical perception residual prediction model based on XGBoost, using the basic simulated flow rate and rainfall characteristics as input vectors, employing a combined optimization strategy of leave-one-station cross-validation and a tree-structured Parzen estimator, and using the difference between the measured flow rate and the basic simulated flow rate as the learning objective to predict nonlinear residuals; and introducing a recession inertia constraint mechanism and a water balance correction based on a geometric scaling factor, achieving high-precision, physically consistent flash flood simulation for watersheds with no or insufficient data.

Description

Technical Field

[0001] This invention belongs to the field of mountain flood control engineering technology, and relates to a mountain flood simulation method based on physical perception residual learning and dynamic runoff compensation. Background Technology

[0002] With the intensification of global climate change, extreme heavy rainfall events, especially typhoons, are exhibiting characteristics of "sudden onset, rapid confluence, and great destructive power" in small watersheds. my country has vast mountainous areas, and due to complex terrain and a lack of comprehensive hydrological monitoring infrastructure, flash floods have become the primary natural disaster threatening public safety and hindering regional sustainable development. Accurately simulating the evolution of flash floods is crucial for disaster prevention and mitigation. Existing methods for simulating flash flood evolution are as follows: The invention patent with publication number CN121167236A discloses a quantitative identification method for high-risk areas of flash floods based on ensemble learning. This method constructs a database of flash flood hazard influencing factors, consisting of "hazard-causing factors-hazard-bearing bodies," and builds a HECRAS two-dimensional hydrodynamic model to simulate flood elements under different rainfall scenarios. Combining multi-scale assessment units and GIS spatial analysis, it establishes a dynamic assessment method based on inundation elements and hazard level thresholds. It integrates scenario rainfall characteristics and hazard-bearing environmental factors as input variables, constructs a training dataset through index screening and discretization, and builds a disaster-bearing body-oriented flash flood hazard assessment model based on the Stacking algorithm framework.

[0003] The invention patent with publication number CN118643768A discloses a method and system for simulating flash flood processes based on the coupling of a statistical model and a physical model. The method obtains basic watershed data, establishes a linear regression model of runoff based on the basic data to predict the runoff of the simulated flash flood, establishes a fluid dynamics physical model to simulate the movement law of flash floods, couples the linear regression model of runoff and the fluid dynamics physical model to obtain a comprehensive flash flood simulation model, calibrates and verifies the parameters of the coupled model, and performs flash flood process simulation to obtain the final result of the flash flood process simulation.

[0004] However, traditional physical models are static parameterization schemes, such as distributed models based on standard SCS-CN and Green-Ampt methods. But in actual short-duration heavy rainfall, changes in soil water storage can trigger a nonlinear "threshold effect"—that is, when the soil reaches saturation, the runoff generation mechanism changes abruptly. Traditional physical models cannot dynamically capture this decline in runoff generation capacity dominated by the degree of initial wetting, leading to an overestimation of infiltration in the "initial loss" stage and an underestimation of runoff in the peak flood stage, resulting in systematic simulation bias.

[0005] Furthermore, for watersheds with widespread lack of or insufficient data, purely data-driven models struggle to capture the hydrological response patterns due to the absence of adequate historical flow records. More seriously, pure AI models lack physical constraints, and when faced with extreme weather events beyond the training set, they are prone to producing predictions that violate the law of conservation of mass or basic hydraulic principles, casting doubt on the reliability of their early warnings.

[0006] Therefore, in light of the above situation, how to achieve dynamic correction of runoff generation errors and improve the accuracy of flood peak simulation has become an urgent problem to be solved. Summary of the Invention

[0007] The purpose of this invention is to provide a flash flood simulation method based on physical perception residual learning and dynamic runoff compensation. This method establishes a dynamic runoff compensation mechanism, constructs a physical perception residual prediction model based on XGBoost, and introduces a receding water inertia constraint mechanism and water balance correction based on geometric scaling factor to achieve high-precision and physically consistent flash flood simulation in watersheds with no or insufficient data.

[0008] A flash flood simulation method based on physical perception residual learning and dynamic runoff compensation includes the following steps: Step 1: Obtain the digital elevation model, land use data and meteorological monitoring data of the target watershed. Use the spatial indexing algorithm to match the monitoring station in the target watershed with the grid position in the digital elevation model. Use the station-specific DEM etching algorithm to force the elevation correction of the local terrain at the monitoring section and construct a high-precision calculation domain. Step 2: Combining land use data with raster data in the high-precision computational domain, a soil saturation feedback factor is introduced to construct a dynamic runoff compensation mechanism; and the basic simulated flow rate is calculated based on the two-dimensional diffusion wave equation. Step 3: Construct a physical sensing residual prediction model based on XGBoost. The calculated basic simulated flow rate is used as the physical sensing feature, and the rainfall features in the meteorological monitoring data are used as the input vector. The leave-one-station cross-validation and tree-structured Parzen estimator are used to perform a joint optimization strategy for physical parameters and model hyperparameters. The difference between the measured flow rate and the basic simulated flow rate is used as the learning target to predict the nonlinear residual. Step 4: Linearly superimpose the basic simulated flow rate with the predicted nonlinear residual, introduce the recession inertia constraint mechanism and the water balance correction based on the geometric scaling factor, and generate the final flash flood simulation process line.

[0009] Furthermore, the site-specific DEM etching algorithm described in step one is used to solve the "pseudo-flow path" problem caused by the deviation between the monitoring site and the actual river channel location under plain micro-topography. The specific execution logic is as follows: S1. Construct a kd-tree spatial index to map the vector coordinates of the monitoring stations to a two-dimensional raster space and lock the target raster. S2. Establish a local search window centered on the target grid, and set the search radius to 2 grid units; S3. Perform local minimum traversal within the local search window to identify the lowest point of the natural terrain; calculate the target riverbed elevation based on the preset etching depth parameters; S4. Execute the elevation update logic. If the original elevation of the current monitoring station grid is greater than the target riverbed elevation, then the updated elevation of the grid is assigned to the target riverbed elevation. If the monitoring station is located in a local depression, then a small disturbance value of 0.1 meters is applied to the target riverbed elevation as the updated elevation. In this way, the local funnel-shaped confluence structure at the monitoring section is reconstructed without changing the macro-topographic features of the watershed.

[0010] Furthermore, the dynamic runoff compensation mechanism described in step two is used to compensate for the deficiency of the traditional SCS-CN model in failing to reflect the nonlinear influence of previous soil water storage on the runoff threshold. The process is as follows: S1. Determine the number of baseline curves based on land use data. The specific land use types in the CLCD dataset are... The correspondence is shown in Table 1: Table 1. Land Use Type Determination in CLCD Dataset Parameter table Calculate the baseline value for maximum potential retention This is used as the upper limit of the grid's water storage capacity: ; S2. Set the grid soil water storage capacity at the initial moment of the simulation as follows: Based on the maximum potential retention benchmark value Calculate the nonlinear dynamic retention at the current moment. : ; Among them, the grid soil water storage at the previous moment ;ratio Represents the relative saturation of the soil at the current moment; It is a nonlinear attenuation coefficient used to control the degree of nonlinear response of dynamic retention as the relative saturation of the soil changes; the soil saturation feedback factor includes the relative saturation of the soil and the initial saturation of the soil. S3, The nonlinear dynamic retention at the current moment. Substitute into the time-varying SCS-CN equation to calculate surface runoff. : ; in, This represents the rainfall at the current moment. This is the initial loss amount, and ; This is the initial loss coefficient; S4. Calculate the actual infiltration rate into the soil based on the principle of conservation of mass. : ; in, This represents the rainfall at the current moment. The surface runoff at the current moment; S5. Utilize the actual infiltration volume. Update the current grid soil water storage. And use it as the input state for the next time step: ; S6, when Approaching At that time, since the relative saturation of the soil approaches 1, The nonlinear rapid decay and approach to 0 drive the surface runoff to smoothly transition from an infiltration-dominated type to a saturated storage type, thereby capturing the "initial loss-runoff" threshold mutation effect at the beginning of flash floods.

[0011] Furthermore, the process of calculating the basic simulated flow rate based on the two-dimensional diffused wave equation described in step two is as follows: Based on the principle of mass conservation, a two-dimensional diffusion wave equation is constructed. The outflow flux, corrected by an adaptive flux limiter, is used to update the water depth state of each grid in the next time step. ; in, The current grid water depth (m) for the next time step; at the current time step, The current grid water depth (m) Rainfall rate (m / s) The actual infiltration rate (m / s) The outflow of the current grid into the upstream adjacent grid. The sum (m) 3 / s), The outflow from the current grid to its lower-level neighboring grids. The sum (m) 3 / s); Continuous flow calculation based on watershed grid topology: ; in, The basic physical simulation flow rate; Target location grid Outflow to lower neighboring cells at the current time step The sum (m) 3 / s).

[0012] Preferably, the outflow rate after correction by the adaptive flux limiter is calculated as follows: The steepest slope direction for each grid cell is calculated using the D8 algorithm, and the outflow between adjacent grid cells is calculated based on the Manning equation. ; in, The calculated outflow flux (m) between adjacent grid cells 3 / s); The benchmark Manning roughness coefficient; For calibration factor / calibration coefficient, h Water depth (meters) For grid width, For frictional gradient: ; in, The water level is the central grid level, which is the sum of the grid depth and the DEM (Digital Elevation Model). For the first Water level of each neighborhood grid; Numerical stability constraints are introduced, and an adaptive flux limiter is set during the explicit time integration process to limit the outflow flux within a single time step. No more than 50% of the current available water depth of the grid: ; in, The outflow rate between adjacent grid cells (m3 / s); For time step; h This represents the current available water depth for the grid. The grid area is used to prevent mass conservation errors or numerical oscillations in calculations on steep terrain.

[0013] Furthermore, the construction process of the physical perception residual prediction model described in step three is as follows: S1. Construct a physical-data hybrid feature vector: ； in, These represent instantaneous rainfall intensity, cumulative rainfall, basic physical simulation flow rate, and physical flow rate change rate, respectively; the instantaneous rainfall intensity and cumulative rainfall are derived from rainfall characteristics in meteorological monitoring data; S2. Construct a nonlinear mapping function based on XGBoost. And perform parameter optimization loop based on the tree-structured Parzen estimator. Using the constructed physical-data hybrid feature vector as input and the difference between the measured flow rate and the baseline simulated flow rate as the learning objective, a nonlinear mapping function based on the XGBoost algorithm is established. ; Define the hyperparameter search space, including physical parameters (CN adjustment factor, Manning coefficient, initial soil saturation) and XGBoost model parameters (tree depth, learning rate). Among them, the CN adjustment factor is a dynamic optimization variable used to correct watershed system bias. In each iteration, the probability model is continuously updated based on the evaluation results of historical parameters, and the next set of parameters that maximizes the expected improvement is selected accordingly. A leave-one-station cross-validation strategy is adopted, and the current parameters are substituted into the validation stations that have not participated in the training for evaluation, and the Kling-Gupta efficiency coefficient is calculated. Repeat the above iterations until the preset number of iterations is reached, and output the globally optimal parameter set to solve the problem of different parameters having the same effect when transferring parameters to areas without data. S3. Using the physical sensing residual prediction model, predict the nonlinear residual at the current moment. : ; in, This represents the physical-data hybrid feature vector at the current moment; The nonlinear mapping function; It is the difference between the basic physical simulation flow rate and the actual observed flow rate.

[0014] Preferably, the optimized formula for calculating the Kling-Gupta efficiency coefficient is as follows: ; Where KGE is the Kling-Gupta efficiency coefficient; The Pearson correlation coefficient between simulated and measured flow rates is used to characterize the consistency of trends; The ratio of the mean values ​​represents the deviation in water balance. The coefficient of variation is the ratio, which characterizes the deviation in flow fluctuation. The average of the simulated flow rate; This represents the average of the measured flow rates. The standard deviation of the basic simulated flow; This represents the standard deviation of the measured flow rate.

[0015] By optimizing the Kling-Gupta efficiency coefficient to maximize the KGE value, the errors among trend, peak flow, and process variability are automatically balanced.

[0016] Furthermore, the receding inertia constraint mechanism described in step four ensures that the receding process follows an exponential decay law, eliminating non-physical fluctuation noise that may be generated in the high-frequency output of the physical perception residual prediction model, and ensuring that the generated flash flood process line is smooth and physically reliable. The execution process is as follows: The current physical simulation flow rate is linearly superimposed with the nonlinear residual to obtain the original predicted flow rate at the current moment: ; The sequence of the original predicted flow rate is subjected to monotonicity detection. When the original predicted flow rate at the current moment is less than the final output flow rate at the previous moment, it is determined that the flow rate is entering the receding stage. The exponential decay smoothing filter is activated during the receding phase, so that the final output flow rate at the current moment is determined by a weighted average of the final output flow rate at the previous moment and the original predicted flow rate at the current moment. ; in, This represents the final output flow rate at the current moment. This represents the final output flow rate from the previous moment; This represents the original predicted flow rate at the current moment; The value is the inertia coefficient of the receding water, ranging from 0.8 to 0.99.

[0017] Furthermore, the water balance correction based on the geometric scaling factor described in step four eliminates systematic water volume deviations by geometrically scaling the basic simulated flow rate based on the ratio of the grid discretized area to the actual watershed area. The execution process is as follows: To address the discretization bias in catchment area caused by DEM raster resolution limitations, a linear geometric scaling factor is constructed: ; in, The physical model represents the catchment area. This represents the actual catchment area. Before the physical model outputs the basic simulated flow rate, a baseline correction is performed to compensate for the systematic water volume deviation caused by grid discretization: ; in, The corrected basic physical simulation flow rate; Based on physical simulation of flow; This is the linear geometric scaling factor.

[0018] Preferably, the physical parameters are jointly optimized based on the tree-structured Parzen estimator, and the contribution of each parameter to the simulation accuracy is quantified using SHAP interpretive analysis technology; the sensitivity differences between runoff generation parameters (CN adjustment factor, initial soil saturation) and runoff confluence parameters (Manning coefficient) are clearly distinguished, thereby guiding the model's parameter dynamic adjustment strategy under different soil relative saturation.

[0019] The beneficial effects of this invention are: This invention provides a flash flood simulation method based on physical perception residual learning and dynamic runoff compensation. This method compensates for the shortcomings of static parameters in traditional physical models by establishing a dynamic runoff compensation mechanism. It solves the problem of lack of physical interpretability of pure data-driven models by constructing a physical perception residual prediction model based on XGBoost. Furthermore, it introduces a receding inertia constraint mechanism and water balance correction based on geometric scaling factor to ensure that the model output conforms to hydraulic common sense. Finally, it achieves high-precision and physically consistent flash flood simulation in watersheds with no or insufficient data. Attached Figure Description

[0020] Figure 1 A flowchart illustrating the overall technical process of a flash flood simulation method based on physical perception residual learning and dynamic runoff compensation, provided in this embodiment of the invention. Figure 2 This is a spatial distribution map of the maximum inundation depth of a watershed generated based on a site-specific DEM etching algorithm, provided in an embodiment of the present invention. Figure 3 A schematic diagram of physical parameter sensitivity analysis and nonlinear physical threshold response based on the SHAP method provided in this embodiment of the invention; Figure 4 A comparison diagram of flood process simulations between the traditional physical model and the physical perception residual prediction model of the present invention under a typical typhoon and rainstorm event provided in an embodiment of the present invention; Figure 5 Box plots of model simulation error distribution and scatter plots of peak capture capability under different flow levels are provided for embodiments of the present invention. Detailed Implementation

[0021] Figure 1This invention provides an overall technical flowchart for a flash flood simulation method based on physical perception residual learning and dynamic runoff compensation. This method addresses the common challenges of weak hydrological monitoring infrastructure and a lack of comprehensive historical data in small watersheds. Instead of relying on massive amounts of historical data for training, it mines valuable chart data from publicly available academic literature and uses image-text dual correction technology to reconstruct high-frequency hydrological and rainfall sequences, achieving engineering-level reproduction of extreme flash flood events at a specific period.

[0022] The implementation steps of a flash flood simulation method based on physical perception residual learning and dynamic runoff compensation in this embodiment are as follows: (1) Reconstruction of high-frequency hydrological sequences of extreme events based on multimodal data mining.

[0023] like Figure 1 As shown, this embodiment selects the A region watershed as the demonstration and verification object. Targeting typical short-duration flash flood events occurring during typhoon impacts, a high spatiotemporal resolution hydrological and rainfall dataset is constructed using a dual technical approach of image-text fusion and physical etching. It should be noted that this embodiment employs literature mining to reconstruct historical disasters in the absence of measured data. In actual real-time flash flood early warning applications, the rainfall and water level data described in steps I to III below can be directly input through real-time acquisition interfaces from rain gauges, water level gauges, or radar remote sensing, without requiring image-text reconstruction; the remaining processing logic of the model remains unchanged.

[0024] The specific restructuring process is as follows: I. Raster Image Digitization Extraction: Obtain the "rainfall-water level process line" illustration from the automatic early warning stations in area A of the watershed. Using the WebPlotDigitizer digitization tool, define a coordinate system (X-axis from 08:00 on 2024-10-31 to 03:00 on 2024-11-02, Y-axis for water level / rainfall), and automatically extract the original time series coordinate points using the "X-axis step interpolation algorithm".

[0025] II. Text Anchor Point Correction Mechanism: To address extraction errors caused by image resolution limitations, a text anchor point correction logic is established. Discrete key "anchor point" values ​​are extracted from the textual descriptions in the literature (e.g., "initial warning threshold 168.8m", "immediate evacuation water level 169.13m"). Using these absolutely accurate textual records as a benchmark (Ground Truth), a linear correction equation is constructed to perform amplitude correction on the digitized water level sequence, eliminating image pixel deviations.

[0026] III. 5-minute high-frequency resampling: Spline interpolation is performed on the corrected non-uniform time series to reconstruct a standard 5-minute equally spaced time series (a total of 516 time periods), forming the synchronous water and rainfall data dataset required for model input.

[0027] IV. Site-Specific Terrain Etching: Based on the geographical location of the automatic early warning station, a spatial index is established in the DEM using the kd-tree algorithm. A site-specific DEM etching algorithm is employed, centered on the detection station, in... The lowest point is searched internally, and a forced topographic correction is performed based on the riverbed elevation inferred from the reconstructed water level data to ensure that upstream runoff accurately converges at the virtual monitoring section. Figure 2 (a) and Figure 2 As shown in (b) of the diagram.

[0028] (2) Dynamic runoff compensation and baseflow calculation at a 5-minute scale.

[0029] A dynamic runoff compensation mechanism is established based on reconstructed 5-minute high-frequency rainfall data. A soil saturation feedback factor is introduced, and the maximum potential retention capacity is updated in real time using an exponential decay equation within each 5-minute calculation step. The system accurately captures the sudden change in the "initial loss-runoff" threshold at the moment of the surge in rainfall intensity at noon on November 1st. Subsequently, high-frequency confluence calculations are performed using the two-dimensional diffusion wave equation to generate a basic simulated flow sequence that includes fundamental physical trends.

[0030] (3) Learning the characteristics of physical perception residuals.

[0031] A physical sensing residual prediction model based on XGBoost is constructed. The basic simulated flow generated in step (2) is used as a physical sensing feature, and the instantaneous rainfall intensity and cumulative rainfall from the 5-minute water and rainfall data set are combined to construct the input vector. The model uses the difference between the reconstructed measured water level and the basic simulated flow (i.e., structural bias) as the learning objective. Figure 3 The SHAP interpretability analysis shows that physical parameters (CN adjustment factor, initial soil saturation) have been confirmed as the most critical dominant factors affecting the accuracy of residual correction (e.g., Figure 3 (a) and Figure 3 In (b) of the model, the values ​​of each parameter show a consistent feedback pattern to the model output that conforms to the hydrological mechanism (e.g., Figure 3 (c) in the model, and the model successfully captured the nonlinear threshold response characteristics of the runoff generation parameter when the soil tends to saturate. Figure 3 (d)). This evidence strongly demonstrates that the physical perception residual prediction model is not simply a data fit, but rather effectively utilizes the physical laws implicit in the reconstructed data.

[0032] (4) Drainage inertia constraint mechanism and water balance correction based on geometric scaling factor.

[0033] The basic simulated flow rate is linearly superimposed with the nonlinear residual, and a recession inertia constraint mechanism is introduced. When the flow rate is detected to be in the recession phase (after 14:00 on November 1st), the current flow rate is constrained by the state of the previous 5 minutes, ensuring it strictly follows an exponential decay law. For example... Figure 4 As shown, the physical sensing residual prediction model significantly corrects the peak flood phase lag and amplitude underestimation phenomena existing in the physical model at various stations in the A region watershed (e.g. Figure 4 In (a) of the model, the simulation fit remained extremely high at the validation station, verifying the spatial generalization of the method (e.g., Figure 4 (b) In this context, the receding water curves at various stations in the B region watershed exhibit a smooth characteristic that strictly conforms to hydraulic laws, effectively eliminating non-physical oscillation noise (such as...) commonly found in pure data-driven models. Figure 4 (c) He Figure 4 (d) in the middle.

[0034] (5) Parameter sensitivity decoupling and optimization.

[0035] A tree-structured Parzen estimator is used to jointly optimize the physical parameters. For example... Figure 5 As shown, the absolute error distribution of the physical sensing residual prediction model is more convergent under the three different flow levels of low, medium, and high, and the median error is significantly lower than that of the traditional physical model (e.g., Figure 5 In (a) of the above, and in the verification of flood peak capture capability, the prediction points of the physical perception residual prediction model are closely distributed around the 1:1 diagonal, overcoming the systematic underestimation of extreme water levels by the physical model (e.g. Figure 5 (b)). This demonstrates that even based on single-field short sequence data, this method still exhibits excellent stability and robustness.

[0036] (6) Refined verification based on extreme typhoon events.

[0037] To verify the model's accuracy under extreme conditions, a massive flash flood triggered by a typhoon during a certain period was selected for a 5-minute accuracy test. The traditional physical model was used as the control group, and the physical perception residual prediction model was used as the experimental group. The focus was on examining the model's performance at the peak of the flood.

[0038] Results analysis: At noon on November 1, the A region experienced short-duration heavy rainfall.

[0039] Because the control group could not detect that the soil was already saturated in the early stage, the runoff calculation was too low. The simulated peak water level was only 168.42m, which was 0.72m lower than the actual value, and the peak water level was delayed by 45 minutes. This means that the early warning would be seriously ineffective.

[0040] By learning high-frequency residuals, the experimental group accurately identified the abrupt change in runoff generation after soil saturation, simulating a peak flood level of 169.11m with an error of only 3 centimeters and a peak flood time deviation of only 5 minutes, achieving an NSE as high as 0.94. This strongly demonstrates that even under extreme conditions with only single-event data, the proposed image-text fusion data reconstruction technology and physical perception learning algorithm can still achieve a precise "digital twin" of flash flood processes.

[0041] In summary, this embodiment achieves this through... Figure 1 Data reconstruction and terrain construction Figure 2 Spatial etching Figure 3 Mechanistic analysis and Figure 4 and Figure 5 Through model optimization, a complete high-frequency flash flood process simulation system has been formed, providing a brand-new technical approach for solving the problem of historical disaster review and early warning threshold determination in areas without data.

[0042] The embodiments described above are merely one application of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation are within the protection scope of the present invention.

Claims

1. A flash flood simulation method based on physical perception residual learning and dynamic runoff compensation, characterized in that, Includes the following steps: Step 1: Obtain the digital elevation model, land use data and meteorological monitoring data of the target watershed. Use the spatial indexing algorithm to match the monitoring station in the target watershed with the grid position in the digital elevation model. Use the station-specific DEM etching algorithm to force the elevation correction of the local terrain at the monitoring section and construct a high-precision calculation domain. Step 2: Combining land use data with raster data in the high-precision computational domain, a soil saturation feedback factor is introduced to construct a dynamic runoff compensation mechanism; and the basic simulated flow rate is calculated based on the two-dimensional diffusion wave equation. Step 3: Construct a physical sensing residual prediction model based on XGBoost. The calculated basic simulated flow rate is used as the physical sensing feature, and the rainfall features in the meteorological monitoring data are used as the input vector. The leave-one-station cross-validation and tree-structured Parzen estimator are used to perform a joint optimization strategy for physical parameters and model hyperparameters. The difference between the measured flow rate and the basic simulated flow rate is used as the learning target to predict the nonlinear residual. Step 4: Linearly superimpose the basic simulated flow rate with the nonlinear residual, introduce the receding inertia constraint mechanism and the water balance correction based on the geometric scaling factor, and generate the final mountain flood simulation process line.

2. The flash flood simulation method of claim 1, wherein, The execution logic of the site-specific DEM etching algorithm described in step one is as follows: S1. Construct a kd-tree spatial index to map the vector coordinates of the monitoring stations to a two-dimensional raster space and lock the target raster. S2. Establish a local search window centered on the target grid, and set the search radius to 2 grid units; S3. Perform local minimum traversal within the local search window to identify the lowest point of the natural terrain; calculate the target riverbed elevation based on the preset etching depth parameters; S4. Execute the elevation update logic. If the original elevation of the current monitoring station grid is greater than the target riverbed elevation, then the updated elevation of the grid is assigned to the target riverbed elevation. If the monitoring station is located in a local depression, then a small disturbance value of 0.1 meters is applied to the target riverbed elevation as the updated elevation, thereby reconstructing the local funnel-shaped confluence structure at the monitoring section.

3. The flash flood simulation method of claim 1, wherein, The process of the dynamic abortion compensation mechanism described in step two is as follows: S1. Determine the baseline curve number based on land use data , calculate the maximum potential retention baseline value , use it as the upper limit of the grid water storage capacity: ； S2, set the grid soil water storage at the initial time point as , according to the maximum potential retention amount reference value , calculate the nonlinear dynamic retention amount at the current time point : ； Among them, the grid soil water storage at the previous moment ;ratio Represents the relative saturation of the soil at the current moment; It is a nonlinear attenuation coefficient used to control the degree of nonlinear response of dynamic retention as the relative saturation of the soil changes; the soil saturation feedback factor includes the relative saturation of the soil and the initial saturation of the soil. S3, nonlinear dynamic retention amount of the current time Substitute the time-varying SCS-CN equation to calculate the surface yield : ； wherein, is the rainfall amount at the current time; is the initial damage amount, and ; is the initial damage coefficient; S4. Calculate the actual amount of water that enters the soil according to the principle of conservation of mass : ； in, This represents the rainfall at the current moment. This represents the surface runoff at the current moment; S5, using the actual infiltration amount updating the current time step grid soil water storage and using it as the input state for the next time step. ； S6, when Approaching At that time, since the relative saturation of the soil approaches 1, The nonlinear rapid decay and approach to 0 drive the surface runoff to smoothly transition from an infiltration-dominated type to a saturated storage type, thereby capturing the "initial loss-runoff" threshold mutation effect at the beginning of flash floods.

4. The flash flood simulation method of claim 1, wherein, The flow rate calculated based on the two-dimensional diffused wave equation in step two is a fundamental physical simulation. The process is as follows: Based on the principle of mass conservation, a two-dimensional diffusion wave equation is constructed. The outflow flux, corrected by an adaptive flux limiter, is used to update the water depth state of each grid in the next time step. ； in, This represents the current grid depth at the next time step; at the current time step, Given the current grid water depth, For rainfall rate, This represents the actual infiltration rate. The outflow of the current grid into the upstream adjacent grid. The sum, The outflow from the current grid to its lower-level neighboring grids. The sum; Continuous flow calculation based on watershed grid topology: ； wherein, is the base physical analog flux; is the target position grid is the outflow flux at the current time step out of the current grid to the surrounding lower adjacent grids is the sum.

5. The flash flood simulation method of claim 1, wherein, The construction process of the physical perception residual prediction model described in step three is as follows: S1. Construct a physical-data hybrid feature vector: ； wherein, respectively represent the instantaneous rainfall intensity, the cumulative rainfall, the basic physically simulated flow rate and the physically simulated flow rate change rate; the instantaneous rainfall intensity and the cumulative rainfall are derived from the rainfall characteristics in the meteorological monitoring data; S2. Construct a nonlinear mapping function based on XGBoost. And perform parameter optimization loops based on the tree-structured Parzen estimator: With the constructed physical-data hybrid feature vector as input and the difference between the measured flow and the basic simulation flow as learning target, a nonlinear mapping function based on XGBoost algorithm is established ; Define the hyperparameter search space, including physical parameters and XGBoost model parameters; in each iteration, continuously update the probability model based on the evaluation results of historical parameters, and select the next set of parameters that maximizes the expected improvement; adopt the leave-one-out cross-validation strategy, substitute the current parameters into the validation sites that have not participated in the training for evaluation, and calculate the Kling-Gupta efficiency coefficient; repeat the above iterations until the preset number of times is reached, and output the globally optimal parameter set; S3, predicting the nonlinear residual error at the current time point by using the physical perception residual error prediction model : ； wherein is the physical-data hybrid feature vector at the current time instant; is the non-linear mapping function; is the difference between the base physical analog flow and the actual observed flow.

6. The flash flood simulation method of claim 5, wherein, The optimized formula for calculating the Kling-Gupta efficiency coefficient is as follows: ； in, KGE The Kling-Gupta efficiency coefficient; The Pearson correlation coefficient between simulated and measured flow rates is used to characterize the consistency of trends; The ratio of the mean values ​​represents the deviation in water balance. The coefficient of variation is the ratio, which characterizes the deviation in flow fluctuation. The average of the simulated flow rate; This represents the average of the measured flow rates. The standard deviation of the basic simulated flow; This represents the standard deviation of the measured flow rate.

7. The flash flood simulation method of claim 1, wherein, The execution process of the water receding inertia constraint mechanism described in step four is as follows: The current physical simulation flow rate is linearly superimposed with the nonlinear residual to calculate the current original predicted flow rate. When the current original predicted flow rate is detected to be less than the final output flow rate of the previous time, it is determined that the water level is receding. The exponential decay smoothing filter is activated during the receding phase, so that the final output flow rate at the current moment is determined by a weighted average of the final output flow rate at the previous moment and the original predicted flow rate at the current moment. ； in, This represents the final output flow rate at the current moment. This represents the final output flow rate from the previous moment; This represents the original predicted flow rate at the current moment; The value is the inertia coefficient of the receding water, ranging from 0.8 to 0.

99.

8. The flash flood simulation method of claim 1, wherein, The execution process of water balance correction based on geometric scaling factor described in step four is as follows: To address the discretization bias in catchment area caused by DEM raster resolution limitations, a linear geometric scaling factor is constructed: ； wherein, is the physical model catchment area; is the actual catchment area; Before the physical model outputs the basic simulated flow rate, a baseline correction is performed to compensate for the systematic water volume deviation caused by grid discretization: ； wherein, is the corrected base physical analog flow rate; is the base physical analog flow rate; is the linear geometric scaling factor.

Images