A refined flood forecasting method with multiple runoff generation mechanisms in parallel

Abstract

This invention discloses a refined flood forecasting method based on multiple runoff generation mechanisms in parallel, comprising: collecting and organizing basic hydrological and meteorological data and remote sensing data of the study basin; constructing the Xin'anjiang model, calibrating and simulating the runoff processes of all flood event sub-basins, and building a flood process sample library; labeling the runoff generation mechanisms of the flood processes in the sample library sub-basins using the number of runoff curves and runoff coefficients; selecting characteristic factors affecting runoff generation mechanisms from the collected data and constructing a random forest runoff generation mechanism classification and prediction model; constructing a spatiotemporally differentiated hybrid hydrological model with multiple runoff generation mechanisms in parallel, and combining runoff generation strategies based on the runoff generation mechanism results predicted by the classification and prediction model for sub-basins; calibrating the model parameters based on the improved SCE-UA optimization algorithm, and finally simulating flood events. This invention improves the model's adaptability to differences in event scale and spatial underlying surface changes, achieving more stable and reliable flood forecasting performance.

Description

Technical Field

[0001] This invention belongs to the technical field of flood forecasting, specifically relating to a refined flood forecasting method with multiple runoff generation mechanisms operating in parallel. Background Technology

[0002] Hydrological forecasting is one of the core technologies for understanding and responding to flood disasters. Floods, as one of the most significant types of natural disasters globally, cause enormous economic losses and loss of life every year. In recent years, against the backdrop of global climate change, the frequency and intensity of extreme precipitation events have shown an upward trend, and the spatiotemporal evolution characteristics of flood processes have become more complex and variable. Floods triggered by different types of rainstorms in different regions exhibit significant non-stationarity in terms of occurrence time, scale, and impact range, which places higher demands on watershed flood risk assessment and forecasting. Against this backdrop, hydrological simulation, especially the characterization of runoff response processes, has become a crucial link in hydrological research and practical applications.

[0003] By constructing watershed hydrological models, processes such as rainfall, infiltration, soil moisture dynamics, and the interaction between surface runoff and groundwater can be quantitatively described. This allows for the reproduction of watershed runoff generation and concentration processes at continuous or event-based scales, enabling quantitative analysis and prediction of flood evolution and providing scientific support for flood control and disaster reduction scheduling, climate change impact assessment, and other related fields. With the continuous development of observational data, computing power, and modeling theories, hydrological models are constantly evolving in terms of physical process characterization, spatiotemporal resolution, and multi-source data fusion. However, the overall goal remains the same: to more realistically reflect watershed hydrological processes and improve the reliability and applicability of runoff simulation and forecasting.

[0004] To adapt to different climates and underlying surface conditions, researchers have developed various hybrid hydrological models and modeling strategies, all aiming to more closely reflect the actual runoff generation in watersheds. Common approaches include combining runoff generation mechanisms vertically, or spatially combining mechanisms such as saturation runoff and infiltration runoff by region / grid to more closely approximate the actual runoff generation process. However, these methods typically pre-set and maintain the runoff generation mechanism used in a particular region, making it difficult to reflect the mechanism transitions that occur under different rainfall structures and antecedent soil moisture conditions, thus failing to adequately characterize the spatiotemporal variability of runoff generation mechanisms. Some technical solutions introduce the "mechanism adaptation" approach: identifying runoff generation mechanisms based on rainfall and soil moisture information and switching calculation schemes accordingly to improve forecast flexibility. However, existing schemes only uniformly identify runoff at the entire watershed scale, ignoring the differences in rainfall distribution and antecedent soil moisture within the watershed during a single rainfall event, making it difficult to express the differences in dominant runoff generation mechanisms between sub-watersheds. Therefore, it is necessary to further develop methods for expressing runoff generation mechanisms that can be dynamically updated at the sub-watershed scale with rainfall processes and antecedent conditions, and that reflect spatiotemporal differentiation. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a refined flood forecasting method that integrates multiple runoff generation mechanisms, fully considering the decisive role of factors such as rainfall processes and soil conditions on runoff generation mechanisms, while identifying and characterizing the spatial differences in runoff generation mechanisms in different sub-basins during a single rainfall event.

[0006] Technical solution: The method described in this invention includes the following steps:

[0007] Collect and organize basic hydrological and meteorological data of the watershed, and collect and preprocess three types of remote sensing data: elevation DEM, ERA5-Land soil moisture content, and normalized vegetation index NDVI.

[0008] Based on the collected data, the Xin'anjiang model was constructed, and the runoff process of all flood event sub-basins was calibrated and simulated to build a flood process sample library.

[0009] For the flood process sample library, the runoff generation mechanism of the flood process in the sub-basin of the sample library is marked by the runoff curve number CN and the runoff coefficient C;

[0010] From the collected data, characteristic factors affecting runoff generation mechanisms were selected, including rainfall, short-term rainfall index, initial soil moisture content and normalized vegetation index, to construct a classification and prediction model for stochastic forest runoff generation mechanisms.

[0011] A spatiotemporally differentiated hybrid hydrological model with multiple runoff generation mechanisms in parallel was constructed, and runoff generation strategies were combined based on the runoff generation mechanism results of sub-basins predicted by the random forest runoff generation mechanism classification prediction model.

[0012] Using Nash efficiency coefficient, Kling-Gupta efficiency coefficient, flood peak error, and peak occurrence time error as evaluation indicators, the model parameters were calibrated based on the improved SCE-UA optimization algorithm, and finally, flood events were simulated and calculated.

[0013] Furthermore, the study involves collecting and organizing basic hydrological and meteorological data for the watershed, and preprocessing remote sensing data; including:

[0014] Collect and process elevation DEM data of the study watershed, and divide the watershed into several sub-watersheds based on the elevation DEM data;

[0015] Collect and organize basic hydrological and meteorological data of the basin, including daily and hourly data from rain gauges and hydrological stations in the basin, extract historical flood events, and organize the total rainfall, average rainfall intensity, maximum 1-hour average rainfall intensity, maximum 3-hour average rainfall intensity, rainfall in the 7 days before the rain, and the percentage of maximum 6-hour rainfall in each sub-basin for all flood events.

[0016] Collect and process ERA5-Land soil moisture content data concurrent with flood events, and compile the pre-rainfall shallow soil moisture content and deep soil moisture content variation coefficients for each sub-basin of each flood event.

[0017] Collect and process Normalized Difference Vegetation Index (NDVI) data contemporaneous with flood events, and compile NDVI data for each sub-basin of each flood event.

[0018] Furthermore, based on the collected data, a Xin'anjiang model was constructed to calibrate and simulate the runoff processes of all flood event sub-basins, and a flood process sample library was built, including:

[0019] Based on the collection and collation of basic hydrological and meteorological data of the watershed, a daily-scale model of the Xin'an River was constructed to output the initial soil moisture content and initial free water volume for each flood event, and the basic parameters were calibrated to ensure water balance.

[0020] Based on the given basic parameters, initial soil moisture content and initial free water, a Xin'anjiang hourly-scale model was constructed to calibrate and simulate each flood event one by one.

[0021] Based on the simulation of each flood event, the local runoff process of each sub-basin in each flood event is output, and a flood process sample library is constructed.

[0022] Furthermore, for the flood process sample library, through runoff curve data... and runoff coefficient Runoff generation mechanisms were labeled for flood events in sub-basins within the sample database; including:

[0023] Calculate the runoff coefficient for all flood events in the sample database. and number of runoff curves A dual-indicator joint judgment system was established, based on the runoff coefficient. Preliminary assessment: number of runoff curves Correction;

[0024] When runoff coefficient When the runoff coefficient is less than the threshold C1, the flood event is initially judged to be an over-infiltration runoff generation mechanism; when the runoff coefficient is less than the threshold C1, the flood event is judged to be an over-infiltration runoff generation mechanism. When the runoff coefficient is greater than the threshold C2, the flood event is initially judged to be a full-storage runoff generation mechanism; when the runoff coefficient is greater than the threshold C2, the flood event is judged to be a full-storage runoff generation mechanism. When the value is greater than threshold C1 and less than threshold C2, the flood process is initially judged to be a mixed runoff generation mechanism.

[0025] Based on the initial assessment, the runoff coefficient was... The number of runoff curves is less than the threshold C1. Flood events with runoff below the threshold CN1 are reclassified as mixed runoff generation mechanisms; the runoff coefficient is adjusted accordingly. The number of runoff curves is greater than the threshold C2. Flood events exceeding the threshold CN2 are reclassified as mixed runoff generation mechanisms.

[0026] Furthermore, runoff coefficient and runoff coefficient The calculation formula is as follows:

[0027] ;

[0028] ;

[0029] ;

[0030] in, For runoff depth; Rainfall; Potential maximum retention capacity represents the maximum possible retention capacity of a watershed before surface runoff is generated.

[0031] Furthermore, a classification and prediction model for random forest runoff generation mechanisms is constructed, including:

[0032] All flood events were divided into two parts according to time sequence: a training set and a test set, which correspond to the calibration period and validation period for subsequent model calibration.

[0033] Several potential impact mechanisms of runoff generation were statistically analyzed as input features. The input features included: total rainfall in each sub-basin of all flood events, maximum 1-hour average rainfall intensity, maximum 3-hour average rainfall intensity, percentage of maximum 6-hour rainfall, average rainfall intensity, rainfall in the 7 days before the rain, remote sensing shallow soil moisture content before the rain, coefficient of variation of remote sensing deep soil moisture content, initial soil moisture content provided by the daily-scale Xin'anjiang model, and normalized vegetation index.

[0034] A random forest runoff generation mechanism classification and prediction model is constructed based on input features. The model adopts the SMOTEENN sampling strategy to solve the class imbalance problem, and a preset physical constraint is added to handle samples at the model boundary for extreme climate conditions.

[0035] The constructed random forest classification prediction model was used to predict the runoff generation mechanism of flood events in the test set for sub-basins, and the results were compared with the judgment results based on the runoff coefficient C and the number of runoff curves CN to ensure accuracy.

[0036] Furthermore, a spatiotemporally differentiated hybrid hydrological model with multiple runoff generation mechanisms in parallel is constructed. Runoff generation strategies are combined based on the sub-basin runoff generation mechanism prediction results obtained from the random forest runoff generation mechanism classification prediction model; including:

[0037] Based on the Xin'anjiang model, a spatiotemporal differentiated mixed hydrological model with multiple runoff generation mechanisms in parallel was constructed. The model includes: evapotranspiration module, runoff generation module, water source division module and runoff confluence module. The runoff generation module adopts three runoff generation strategies: infiltration runoff, storage runoff, and mixed runoff.

[0038] Based on the prediction results of the random forest model on the runoff generation mechanism of the flood event sub-basin in the test set, the corresponding sub-basin runoff generation strategy combination is selected to simulate the flood event, that is, different runoff generation strategies are considered for different sub-basin units; and three hydrological models that only consider a single runoff generation strategy are constructed at the same time for comparative analysis.

[0039] Furthermore, the calculation formula for the super-permeability flow generation strategy is as follows:

[0040] ;

[0041] ;

[0042] ;

[0043] in, This refers to the infiltration capacity of the watershed. To stabilize the infiltration rate; The average tensile water storage capacity of the basin; This refers to the water storage capacity of soil tension at a single point. Permeability coefficient, reflecting the impact of soil water shortage on infiltration; Real-time infiltration rate; For clean rain; The infiltration rate distribution curve index; Surface runoff;

[0044] The formula for calculating the full flow generation strategy is:

[0045] ;

[0046] ;

[0047] ;

[0048] in, For tension water storage capacity less than or equal to Partially permeable area; The permeable area of ​​the entire watershed; This represents the maximum tension water storage capacity at a single point. The index of the tensile water storage capacity curve; Free water storage capacity is less than or equal to The area; The area of ​​runoff generation; This refers to the free water storage capacity at a single point in the watershed. This represents the maximum free water storage capacity. The free water storage capacity curve index; Total output flow; This represents the maximum value of the average initial tension water storage capacity of the basin.

[0049] The mixed runoff generation strategy includes surface runoff and subsurface runoff. The calculation of surface runoff is the same as that of the super-permeable runoff generation strategy.

[0050] The groundwater runoff calculation considers the tensional water storage capacity curve under the full runoff strategy. The formula for calculating groundwater runoff is as follows:

[0051] ;

[0052] in, This refers to the flow generated below ground level.

[0053] Furthermore, using Nash efficiency coefficient, Kling-Gupta efficiency coefficient, peak flow error, and peak occurrence time error as evaluation indicators, the model parameters were calibrated based on the improved SCE-UA optimization algorithm, and finally, flood events were simulated and calculated; including:

[0054] Selecting the Nash efficiency coefficient Kling-Gupta efficiency coefficient Flood peak error and peak occurrence time error To evaluate the indicators, a comprehensive objective function is constructed. ;

[0055] Based on the improved SCE-UA optimization algorithm and rate-periodic flood events, a comprehensive objective function is proposed. The parameters of the three single runoff generation mechanism hydrological models and the spatiotemporally differentiated mixed hydrological model were calibrated using the minimization principle. Based on the calibrated models and the prediction of runoff generation mechanisms in sub-basins for each flood event during the validation period, simulation calculations were performed on the flood events during the validation period.

[0056] The improved SCE-UA optimization algorithm introduces a multiple restart mechanism and adopts three differentiated initial sampling strategies: completely random sampling, sampling near the boundary, and improved random sampling. During the evolution process, an adaptive reflection coefficient and a perturbation mechanism are introduced to dynamically adjust the search step size to improve the ability to escape local optima.

[0057] Furthermore, the Nash efficiency coefficient Kling-Gupta efficiency coefficient Flood peak error Peak occurrence time error and the comprehensive objective function The calculation formula is:

[0058] ;

[0059] ;

[0060] ;

[0061] ;

[0062] ;

[0063] in, For the first Simulated traffic flow over a given time period; For the first Observed flow rate during the specified time period; This represents the average value of the observed flow rate; This represents the total number of time periods; The correlation coefficient between simulated flow and observed flow; The ratio of standard deviations. To simulate the standard deviation of the flow rate, The standard deviation of the observed flow rate; This is the ratio of the means. To simulate the average flow rate, This represents the average of the observed flow rates; To simulate peak flood levels; To observe the peak flood level; , These represent the simulated and observed times of flood peak occurrence, respectively. This represents the total number of flood periods; , , , These are the weighting coefficients for each indicator.

[0064] Beneficial effects: Compared with the prior art, the significant technical effects of this invention are as follows: (1) By selecting several characteristic factors that may affect runoff generation mechanisms, a random forest runoff generation mechanism classification and prediction model is constructed. This fully utilizes the ability of random forest to effectively mine the nonlinear response characteristics of hydrological processes and the complex interaction between multiple environmental factors. At the same time, physical constraints are introduced to further correct the samples under boundary conditions, thereby achieving accurate prediction of runoff generation mechanisms in sub-basins. (2) Based on the basic framework of the Xin'anjiang model, a spatiotemporal differentiated hybrid hydrological model with multiple runoff generation mechanisms in parallel is constructed. Combined with the constructed random forest runoff generation mechanism classification and prediction model, it can comprehensively consider the spatiotemporal differences of meteorological conditions and underlying surface conditions in the basin, so as to achieve a refined expression of the spatial runoff situation in the basin. Overall, this makes the hydrological model closer to the actual runoff situation in the basin, thereby improving the simulation accuracy. Attached Figure Description

[0065] Figure 1 This is a flowchart of the method of the present invention;

[0066] Figure 2 This is a detailed flowchart of the method of the present invention;

[0067] Figure 3 A map showing the distribution of sub-basins within the Tunxi River Basin;

[0068] Figure 4 Confusion matrix diagram of the random forest runoff generation mechanism classification prediction model on the test set;

[0069] Figure 5 NSE box plot for 44 regular flood events;

[0070] Figure 6 NSE box plots for 16 validation flood events;

[0071] Figure 7 KGE box plot for 44 periodic flood events;

[0072] Figure 8 KGE box plots for 16 flood events during the validation period;

[0073] Figure 9 The following are flood process diagrams for four different flood events simulated by four model strategies: (a) for event 20080527, (b) for event 20120808, (c) for event 20190515, and (d) for event 20200701. Detailed Implementation

[0074] The present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0075] This embodiment takes the Tunxi River Basin as an example and simulates a flood event according to this method.

[0076] like Figure 1 and Figure 2 As shown, the refined flood forecasting method based on multiple parallel runoff generation mechanisms of the present invention includes the following steps:

[0077] S1. Collect and organize basic hydrological and meteorological data of the research basin; collect and preprocess three types of remote sensing data: elevation DEM, ERA5-Land soil moisture content, and normalized difference vegetation index (NDVI); including:

[0078] S1.1 Collect and process the elevation DEM data of the study watershed, and divide the watershed into several sub-watersheds based on the elevation DEM data;

[0079] S1.2 Collect and organize basic hydrological and meteorological data of the watershed, including daily and hourly data of rain gauges and hydrological stations in the watershed, extract historical flood events, and organize the total rainfall, average rainfall intensity, maximum 1-hour average rainfall intensity, maximum 3-hour average rainfall intensity, rainfall in the 7 days before the rain, and the percentage of maximum 6-hour rainfall in each sub-watershed for all flood events.

[0080] S1.3 Collect and process ERA5-Land soil moisture content data concurrent with flood events, and compile the pre-rainfall shallow soil moisture content and deep soil moisture content variation coefficients for each sub-basin of each flood event.

[0081] S1.4 Collect and process Normalized Difference Vegetation Index (NDVI) data contemporaneous with flood events, and compile NDVI data for each sub-basin of each flood event.

[0082] The Tunxi River Basin is located in southern Anhui Province, with a drainage area of ​​2687 km². 2 This is a typical small to medium-sized river basin. The basin has a subtropical humid monsoon climate, with abundant but highly concentrated rainfall; the average annual rainfall exceeds 1500 mm. Continuous torrential rains during the plum rain season and heavy rainfall during typhoon season are the main flood-causing factors. Furthermore, the basin's topography is predominantly low mountains and hills, with significant topographic relief. This combination of "mountainous terrain + torrential rain" forms the natural basis for the basin's extremely high flood risk. The dominant runoff generation mechanisms in each tributary basin also exhibit significant spatiotemporal heterogeneity due to rainfall heterogeneity, topography, and pre-existing soil moisture storage.

[0083] The rainfall and flood data in this embodiment were selected from daily measured rainfall, evaporation, and flow data of the Tunxi River Basin from 2007 to 2022, as well as hourly measured rainfall, evaporation, and flow data extracted during the flood season (March to October). The data included 12 rain gauge stations and 1 hydrological station. Sixty flood events were selected (Table 1), and the first 44 flood events were selected in chronological order as the calibration period, and the last 16 flood events as the validation period.

[0084] This embodiment also collected Normalized Difference Vegetation Index (NDVI) data and ERA5-Land hourly soil moisture content data concurrent with the flood event.

[0085] Table 1 Historical Flood Events in the Tunxi River Basin

[0086]

[0087] S2. Based on the data collected in step S1, construct the Xin'anjiang model, perform calibration simulations on the runoff processes of all flood event sub-basins, and build a flood process sample library; specifically, this includes the following steps:

[0088] S2.1 Based on the data from steps S1.1 and S1.2, construct a daily-scale model of the Xin'an River, output the initial soil moisture content and initial free water for each flood event, and calibrate the basic parameters to ensure water balance;

[0089] S2.2. Based on the basic parameters given in step S2.1, the initial soil moisture content and the initial free water content, the Xin'anjiang hourly-scale model is constructed to calibrate and simulate each flood event one by one;

[0090] S2.3 Based on the simulation of each flood event in step S2.2, output the local runoff process of each sub-basin in each flood event and construct a flood process sample library.

[0091] Figure 3 The results of the sub-basin division of the Tunxi River Basin are presented. The Tunxi River Basin is divided into 11 sub-basins. By constructing the Xin'anjiang model at the hourly scale, each flood event is calibrated and simulated one by one, and the local runoff processes of the 11 sub-basins in all flood events are output, thus obtaining 660 flood events and completing the construction of the flood process sample library.

[0092] S3. For the flood process sample library in step S2, analyze the runoff curve data... and runoff coefficient The runoff generation mechanism is labeled for flood processes in sub-basins within the sample database; this specifically includes the following steps:

[0093] S3.1 Calculate the runoff coefficient for all flood events in the sample database. and number of runoff curves A dual-indicator joint judgment system was established, based on the runoff coefficient. Preliminary assessment: number of runoff curves Correction;

[0094] runoff coefficient and runoff coefficient The calculation formula is as follows:

[0095] (1)

[0096] (2)

[0097] (3)

[0098] in, The runoff depth is expressed in mm. Rainfall amount, in mm; Potential maximum retention capacity represents the maximum possible retention capacity of a watershed before surface runoff is generated, expressed in mm.

[0099] S3.2, when the runoff coefficient When the runoff coefficient is less than the threshold C1 (in this embodiment, the threshold C1 is 0.3), the flood process is initially judged to be an over-infiltration runoff generation mechanism; when the runoff coefficient is less than the threshold C1, the flood process is .... When the runoff coefficient is greater than the threshold C2 (in this embodiment, the threshold C2 is 0.4), the flood process is initially judged to be a full-storage runoff generation mechanism; when the runoff coefficient is greater than the threshold C2, the flood process is judged to be a full-storage runoff generation mechanism. When the runoff coefficient is greater than threshold C1 and less than threshold C2 (in this embodiment, when the runoff coefficient is greater than threshold C1 and less than threshold C2), the runoff coefficient is less than threshold C2 and less than threshold C1. When the value is greater than 0.3 and less than 0.4, the flood event is initially judged to be a mixed runoff generation mechanism;

[0100] S3.3, Based on the initial judgment in S3.2, the runoff coefficient is... The number of runoff curves is less than the threshold C1. Flood events with runoff coefficients less than the threshold CN1 (in this embodiment, threshold CN1 is set to 60) are reclassified as mixed runoff generation mechanisms; the runoff coefficient is... The number of runoff curves is greater than the threshold C2. Flood processes exceeding the threshold CN2 (in this embodiment, the threshold CN2 is set to 80) are reclassified as mixed runoff generation mechanisms.

[0101] S4. From the data in steps S1 and S2, select 10 characteristic factors that affect the runoff generation mechanism, including total rainfall, maximum 1-hour average rainfall intensity, maximum 3-hour average rainfall intensity, maximum 6-hour rainfall percentage, average rainfall intensity, rainfall in the 7 days before the rain, remote sensing shallow soil moisture content before the rain, remote sensing deep soil moisture content variation coefficient, initial soil moisture content provided by the daily-scale Xin'anjiang model, and normalized vegetation index (NDVI), and construct a classification and prediction model for the runoff generation mechanism of stochastic forests.

[0102] Specifically, the following steps are included:

[0103] S4.1 Divide all flood events into two parts in chronological order: a training set and a test set. These correspond to the calibration period and validation period for subsequent model calibration, with a ratio of approximately 7:3.

[0104] S4.2 Statistically analyze the characteristics of several potential impact mechanisms on runoff generation as input features. Input features include: total rainfall in each sub-basin of all flood events, maximum 1-hour average rainfall intensity, maximum 3-hour average rainfall intensity, percentage of maximum 6-hour rainfall, average rainfall intensity, rainfall in the 7 days before the rain, remote sensing shallow soil moisture content before the rain, coefficient of variation of remote sensing deep soil moisture content, initial soil moisture content provided by the daily-scale Xin'anjiang model, and normalized vegetation index (NDVI).

[0105] S4.3. Based on the input features collected in step S4.2, a random forest runoff generation mechanism classification and prediction model is constructed. The model uses the SMOTEENN sampling strategy to solve the class imbalance problem. At the same time, appropriate physical constraints are added to handle samples at the model boundary (for example, setting a threshold to determine flood events with total rainfall greater than the threshold as runoff generation mechanisms. In this embodiment, the threshold is 255 mm, which is the 90th percentile of the rainfall corresponding to all flood events in the sample library) to ensure applicability to extreme climate conditions.

[0106] SMOTE (Synthetic Minority Oversampling) expands the minority class sample size by interpolating and synthesizing new samples among minority class samples, while ENN (Nearest Neighbor Editing) removes misclassified majority class samples. SMOTEENN first performs SMOTE oversampling on the minority class and then applies ENN to the entire data for noise reduction, which increases the proportion of minority class samples and removes noise. Appropriate physical constraint rules are introduced, such as for extreme events where the total rainfall exceeds a set threshold (255 mm in this example, which is the 90th percentile of the rainfall corresponding to all flood events in the sample library), which can be directly regarded as being dominated by runoff generation. These rules are based on experience and basic hydrological principles and help to correct the model output.

[0107] S4.4. Using the random forest classification prediction model constructed in step S4.3, predict the runoff generation mechanism of flood events in the test set for sub-basins, and compare it with the runoff coefficient-based prediction in step S3. and number of runoff curves The judgment results are compared to ensure that the accuracy rate reaches a high level.

[0108] Table 2 shows the performance of the random forest runoff mechanism classification and prediction model. It can be seen that the random forest runoff mechanism classification and prediction model has high accuracy and generalization ability.

[0109] Table 2. Simulation Evaluation of Random Forest Runoff Mechanism Classification Prediction Model

[0110]

[0111] In this embodiment, the training set corresponds to 484 flood events from 44 flood events during the rate-setting period, and the test set corresponds to 176 flood events from 16 flood events during the validation period.

[0112] Figure 4 The confusion matrix of the random forest runoff generation mechanism classification prediction model is shown on the test set. Overall, it can be seen that the recall rate of the model for all three categories is at a high level.

[0113] S5. Construct a spatiotemporally differentiated hybrid hydrological model with multiple runoff generation mechanisms in parallel, and combine runoff generation strategies based on the sub-basin runoff generation mechanism prediction results predicted by the random forest runoff generation mechanism classification prediction model in step S4.

[0114] Specifically, the following steps are included:

[0115] S5.1. Based on the Xin'anjiang model, a spatiotemporal differentiated mixed hydrological model with multiple runoff generation mechanisms in parallel is constructed. The model is mainly divided into four modules: evapotranspiration module, runoff generation module, water source division module, and runoff confluence module. Among them, the runoff generation module adopts three runoff generation strategies: infiltration runoff generation, storage runoff generation, and mixed runoff generation.

[0116] The main calculation formulas for the over-permeability flow generation strategy are as follows:

[0117] (4)

[0118] (5)

[0119] (6)

[0120] in, This refers to the infiltration capacity of the watershed. To stabilize the infiltration rate; The average tensile water storage capacity of the basin; This refers to the water storage capacity of soil tension at a single point. Permeability coefficient, reflecting the impact of soil water shortage on infiltration; Real-time infiltration rate; For clean rain; The infiltration rate distribution curve index; It refers to surface runoff.

[0121] The main calculation formulas for the full flow generation strategy are as follows:

[0122] (7)

[0123] (8)

[0124] (9)

[0125] in, For tension water storage capacity less than or equal to Partially permeable area; The permeable area of ​​the entire watershed; This represents the maximum tension water storage capacity at a single point. The index of the tensile water storage capacity curve; Free water storage capacity is less than or equal to The area; The area of ​​runoff generation; This refers to the free water storage capacity at a single point in the watershed. This represents the maximum free water storage capacity. The free water storage capacity curve index; Total output flow; This represents the maximum value of the average initial tension water storage capacity of the basin.

[0126] The main calculation formulas for the hybrid runoff generation strategy are as follows:

[0127] The calculation of surface runoff is the same as that of over-permeability runoff. The calculation of groundwater runoff takes into account the tension water storage capacity curve of equation (7). The formula for calculating groundwater runoff is as follows:

[0128] (10)

[0129] in, This refers to the flow generated below ground level.

[0130] S5.2 Based on the prediction results of the random forest model on the runoff generation mechanism of the test set flood event sub-basin in step S4, select the corresponding sub-basin runoff generation strategy combination to simulate the flood event, that is, different runoff generation strategies are considered for different sub-basin units; and at the same time, three hydrological models that only consider a single runoff generation strategy are constructed for comparative analysis.

[0131] S6, using the Nash efficiency coefficient Kling-Gupta efficiency coefficient Flood peak error and peak occurrence time error To evaluate the performance, the model parameters were calibrated based on the improved SCE-UA optimization algorithm, and finally, a flood event simulation was performed.

[0132] Specifically, the following steps are included:

[0133] S6.1 Selecting the Nash efficiency coefficient Kling-Gupta efficiency coefficient Flood peak error and peak occurrence time error To evaluate the indicators, a comprehensive objective function is constructed. The calculation formula is as follows:

[0134] (11)

[0135] (12)

[0136] (13)

[0137] (14)

[0138] (15)

[0139] in, For the first Simulated traffic flow over a given time period; For the first Observed flow rate during the specified time period; This represents the average value of the observed flow rate; This represents the total number of time periods; The correlation coefficient between simulated flow and observed flow; The ratio of standard deviations. To simulate the standard deviation of the flow rate, The standard deviation of the observed flow rate; This is the ratio of the means. To simulate the average flow rate, This represents the average of the observed flow rates; To simulate peak flood levels; To observe the peak flood level; , These represent the simulated and observed times of flood peak occurrence, respectively. This represents the total number of flood periods; , , , These are the weighting coefficients for each indicator.

[0140] S6.2. Based on the improved SCE-UA optimization algorithm and periodic flood events, a comprehensive objective function is proposed. The parameters of the three single runoff generation mechanism hydrological models and the spatiotemporally differentiated mixed hydrological model were calibrated using the minimization principle. Based on the calibrated models and the prediction of runoff generation mechanisms for each flood event sub-basin in step S4 during the validation period, the flood events during the validation period were simulated.

[0141] To enhance the robustness of parameter calibration, a multiple restart mechanism is introduced for this optimization algorithm. Three differentiated initial sampling strategies are adopted: completely random sampling, sampling near the boundary, and improved random sampling, in order to reduce the sensitivity to initial values ​​and enhance the global search capability. In addition, an adaptive reflection coefficient and perturbation mechanism are introduced during the evolution process to dynamically adjust the search step size to improve the ability to escape local optima.

[0142] In this embodiment, model M1 was constructed using the super-permeability runoff generation strategy, model M2 was constructed using the full-storage runoff generation strategy, and model M3 was constructed using the mixed runoff generation strategy. The spatiotemporal differentiated mixed hydrological model strategy is referred to as model M4. The four model strategies were used to simulate and evaluate flood events during the rate-setting and validation periods. The average performance of the four model strategies is shown in Table 3. It can be seen that the overall performance of model M4 is significantly better than the other models.

[0143] Table 3. Average evaluation metrics for four model strategies in flood event simulation.

[0144]

[0145] Figure 5 , Figure 6 , Figure 7 , Figure 8 Box plots of NSE and KGE for the four model strategies across all flood events are presented. Overall, model M4 demonstrates more robust performance in simulating flood events, and it also performs well in simulating events where the other three models perform poorly.

[0146] Figure 9 Tables (a) to (d) show the performance of four model strategies in four flood events of different magnitudes. Table 4 shows the sub-basin runoff generation mechanism combinations of the M4 model strategy in these four flood events (0 represents infiltration excess, 1 represents storage fullness, and 2 represents mixing). The events 20080527 and 20120808 are rate-period events, while 20190515 and 20200701 are validation-period events. The rainfall events of these four flood events are typical of the basin, including continuous rainfall during the plum rain season and sudden summer downpours. The 20200701 flood event was a once-in-fifty-year flood, causing 127 reservoirs in downstream Huangshan City to exceed their flood limits, with 114 of them overflowing. Overall, model M4 shows the best simulation performance, followed by model M3. This is especially true for the two flood peaks of approximately 2000m during the validation period. 3 / s and 5000m 3 For flood events with a magnitude of / s, model M3, which considers a mixed runoff generation mechanism, performs exceptionally well in simulating large flood events. Its simulation of the peak value of the 20200701 flood event is superior to that of M4, initially highlighting the advantages of combined runoff generation mechanisms. However, overall, model M4, which considers the adaptation of runoff generation mechanisms to sub-basins, yields better results. This strategy is adaptable to floods of various magnitudes and can accurately simulate multiple peak values ​​of the flood process.

[0147] Table 4. Combinations of subbasin runoff generation mechanisms for the M4 model strategy in four flood events.

[0148]

[0149] In summary, the refined flood forecasting method based on multiple runoff generation mechanisms provided by this invention compensates for the discrepancy between traditional models and actual runoff generation behavior. It can fully consider the spatiotemporal differences in runoff generation mechanisms in small and medium-sized watersheds, improve the simulation accuracy of flood events, and provide a more physically consistent and generalizable technical approach for flood forecasting in small and medium-sized watersheds under complex conditions. This can provide strong support for flood control and disaster reduction decision-making.

Claims

1. A refined flood forecasting method with multiple parallel runoff generation mechanisms, characterized in that, Includes the following steps: Collect and organize basic hydrological and meteorological data of the watershed, and collect and preprocess three types of remote sensing data: elevation DEM, ERA5-Land soil moisture content, and normalized vegetation index NDVI. Based on the collected data, the Xin'anjiang model was constructed, and the runoff process of all flood event sub-basins was calibrated and simulated to build a flood process sample library. For the flood process sample library, the runoff generation mechanism of the flood process in the sub-basin of the sample library is marked by the runoff curve number CN and the runoff coefficient C; From the collected data, characteristic factors influencing runoff generation mechanisms were selected, including rainfall, short-term rainfall indices, initial soil moisture content, and normalized difference vegetation index (NDI), to construct a classification and prediction model for stochastic forest runoff generation mechanisms; including: All flood events were divided into two parts according to time sequence: a training set and a test set, which correspond to the calibration period and validation period for subsequent model calibration. Several potential impact mechanisms of runoff generation were statistically analyzed as input features. The input features included: total rainfall in each sub-basin of all flood events, maximum 1-hour average rainfall intensity, maximum 3-hour average rainfall intensity, percentage of maximum 6-hour rainfall, average rainfall intensity, rainfall in the 7 days before the rain, remote sensing shallow soil moisture content before the rain, coefficient of variation of remote sensing deep soil moisture content, initial soil moisture content provided by the daily-scale Xin'anjiang model, and normalized vegetation index. A random forest runoff generation mechanism classification and prediction model is constructed based on input features. The model adopts the SMOTEENN sampling strategy to solve the class imbalance problem, and a preset physical constraint is added to handle samples at the model boundary for extreme climate conditions. The constructed random forest classification prediction model was used to predict the runoff generation mechanism of flood events in the test set for sub-basins, and the results were compared with the judgment results based on the runoff coefficient C and the number of runoff curves CN to ensure accuracy. A spatiotemporally differentiated hybrid hydrological model with multiple runoff generation mechanisms in parallel was constructed, and runoff generation strategies were combined based on the runoff generation mechanism results of sub-basins predicted by the random forest runoff generation mechanism classification prediction model. Using Nash efficiency coefficient, Kling-Gupta efficiency coefficient, flood peak error, and peak occurrence time error as evaluation indicators, the model parameters were calibrated based on the improved SCE-UA optimization algorithm, and finally, flood events were simulated and calculated.

2. The method according to claim 1, characterized in that, Collect and organize basic hydrological and meteorological data for the research basin, and preprocess remote sensing data; including: Collect and process elevation DEM data of the study watershed, and divide the watershed into several sub-watersheds based on the elevation DEM data; Collect and organize basic hydrological and meteorological data of the basin, including daily and hourly data from rain gauges and hydrological stations in the basin, extract historical flood events, and organize the total rainfall, average rainfall intensity, maximum 1-hour average rainfall intensity, maximum 3-hour average rainfall intensity, rainfall in the 7 days before the rain, and the percentage of maximum 6-hour rainfall in each sub-basin for all flood events. Collect and process ERA5-Land soil moisture content data concurrent with flood events, and compile the pre-rainfall shallow soil moisture content and deep soil moisture content variation coefficients for each sub-basin of each flood event. Collect and process Normalized Difference Vegetation Index (NDVI) data contemporaneous with flood events, and compile NDVI data for each sub-basin of each flood event.

3. The method according to claim 1, characterized in that, Based on the collected data, a Xin'anjiang model was constructed, and calibration simulations were performed on the runoff processes of all flood event sub-basins to build a flood process sample library, including: Based on the collection and collation of basic hydrological and meteorological data of the watershed, a daily-scale model of the Xin'an River was constructed to output the initial soil moisture content and initial free water volume for each flood event, and the basic parameters were calibrated to ensure water balance. Based on the given basic parameters, initial soil moisture content and initial free water, a Xin'anjiang hourly-scale model was constructed to calibrate and simulate each flood event one by one. Based on the simulation of each flood event, the local runoff process of each sub-basin in each flood event is output, and a flood process sample library is constructed.

4. The method according to claim 1, characterized in that, For the flood process sample library, through runoff curve data and runoff coefficient Runoff generation mechanisms were labeled for flood events in sub-basins within the sample database; including: Calculate the runoff coefficient for all flood events in the sample database. and number of runoff curves A dual-indicator joint judgment system was established, based on the runoff coefficient. Preliminary assessment: number of runoff curves Correction; When runoff coefficient When the runoff coefficient is less than the threshold C1, the flood event is initially judged to be an over-infiltration runoff generation mechanism; when the runoff coefficient is less than the threshold C1, the flood event is judged to be an over-infiltration runoff generation mechanism. When the runoff coefficient is greater than the threshold C2, the flood event is initially judged to be a flood-generating mechanism; when the runoff coefficient is greater than the threshold C2, the flood event is judged to be a flood-generating mechanism. When the value is greater than threshold C1 and less than threshold C2, the flood process is initially judged to be a mixed runoff generation mechanism. Based on the initial assessment, the runoff coefficient was... The number of runoff curves is less than the threshold C1. Flood events with a runoff coefficient below the threshold CN1 are reclassified as mixed runoff generation mechanisms; the runoff coefficient is adjusted accordingly. The number of runoff curves is greater than the threshold C2. Flood events exceeding the threshold CN2 are reclassified as mixed runoff generation mechanisms.

5. The method according to claim 4, characterized in that, runoff coefficient and runoff coefficient The calculation formula is as follows: ； ； ； in, For runoff depth; Rainfall; Potential maximum retention capacity represents the maximum possible retention capacity of a watershed before surface runoff is generated.

6. The method according to claim 1, characterized in that, A spatiotemporally differentiated hybrid hydrological model with multiple runoff generation mechanisms in parallel was constructed, and runoff generation strategies were combined based on the runoff generation mechanism results of sub-basins predicted by the random forest runoff generation mechanism classification prediction model. include: Based on the Xin'anjiang model, a spatiotemporal differentiated mixed hydrological model with multiple runoff generation mechanisms in parallel was constructed. The model includes: evapotranspiration module, runoff generation module, water source division module and runoff confluence module. The runoff generation module adopts three runoff generation strategies: infiltration runoff, storage runoff, and mixed runoff. Based on the prediction results of the random forest model on the runoff generation mechanism of the flood event sub-basin in the test set, the corresponding sub-basin runoff generation strategy combination is selected to simulate the flood event, that is, different runoff generation strategies are considered for different sub-basin units; and three hydrological models that only consider a single runoff generation strategy are constructed at the same time for comparative analysis.

7. The method according to claim 6, characterized in that, The calculation formula for the over-permeability flow generation strategy is as follows: ； ； ； in, This refers to the infiltration capacity of the watershed. To stabilize the infiltration rate; The average tensile water storage capacity of the basin; This refers to the water storage capacity of soil tension at a single point. Permeability coefficient; Real-time infiltration rate; For clean rain; The infiltration rate distribution curve index; Surface runoff; The formula for calculating the full flow generation strategy is: ； ； ； in, For tension water storage capacity less than or equal to Partially permeable area; The permeable area of ​​the entire watershed; This represents the maximum tension water storage capacity at a single point. The index of the tensile water storage capacity curve; Free water storage capacity is less than or equal to The area; The area of ​​runoff generation; This refers to the free water storage capacity at a single point in the watershed. This represents the maximum free water storage capacity. The free water storage capacity curve index; Total output flow; This represents the maximum value of the average initial tension water storage capacity of the basin. The mixed runoff generation strategy includes surface runoff and subsurface runoff. The calculation of surface runoff is the same as that of the super-permeable runoff generation strategy. The groundwater runoff calculation considers the tensional water storage capacity curve under the full runoff strategy. The formula for calculating groundwater runoff is as follows: ； in, This refers to the flow generated below ground level.

8. The method according to claim 1, characterized in that, Using Nash efficiency coefficient, Kling-Gupta efficiency coefficient, peak flow error, and peak occurrence time error as evaluation indicators, the model parameters are calibrated based on the improved SCE-UA optimization algorithm, and finally, flood events are simulated and calculated; including: Selecting the Nash efficiency coefficient Kling-Gupta efficiency coefficient Flood peak error and peak occurrence time error To evaluate the indicators, a comprehensive objective function is constructed. ; Based on the improved SCE-UA optimization algorithm and rate-periodic flood events, a comprehensive objective function is proposed. The parameters of the three single runoff generation mechanism hydrological models and the spatiotemporally differentiated mixed hydrological model were calibrated using the minimization principle. Based on the calibrated models and the prediction of runoff generation mechanisms in sub-basins for each flood event during the validation period, simulation calculations were performed on the flood events during the validation period. The improved SCE-UA optimization algorithm introduces a multiple restart mechanism and adopts three differentiated initial sampling strategies: completely random sampling, sampling near the boundary, and improved random sampling. During the evolution process, an adaptive reflection coefficient and a perturbation mechanism are introduced to dynamically adjust the search step size to improve the ability to escape local optima.

9. The method according to claim 1, characterized in that, Nash efficiency coefficient Kling-Gupta efficiency coefficient Flood peak error Peak occurrence time error and the comprehensive objective function The calculation formula is: ； ； ； ； ； in, For the first Simulated traffic flow over a given time period; For the first Observed flow rate during the specified time period; This represents the average value of the observed flow rate; This represents the total number of time periods; The correlation coefficient between simulated flow and observed flow; The ratio of standard deviations. To simulate the standard deviation of the flow rate, The standard deviation of the observed flow rate; This is the ratio of the means. To simulate the average flow rate, This represents the average of the observed flow rates; To simulate peak flood levels; To observe the peak flood level; , These represent the simulated and observed times of flood peak occurrence, respectively. This represents the total number of flood periods; , , , These are the weighting coefficients for each indicator.

Images