Intelligent real-time basin flood warning method and system

Abstract

The application relates to the technical field of hydrology and water conservancy, and particularly discloses an intelligent real-time basin flood early warning method and system, which comprises the following steps: acquiring basin data, constructing a fusion prediction model, calculating multi-level early warning thresholds, generating early warnings, and performing online evaluation. The method constructs real-time residuals of a physical hydrological model, adopts a residual prediction model based on machine learning to learn the complex mode of the residuals changing with real-time rainfall intensity and basin state, provides an interpretable, robust prediction baseline and high-precision error correction through a hybrid architecture of forward coupling and feedback coupling, effectively improves the overall robustness of flood prediction, dynamically inverts the fusion prediction model, identifies the flood risk differences of different soil conditions, greatly improves the matching degree of the early warning signal and the actual risk, calculates the comprehensive risk value in real time on the basis of a unified grid division system, and updates and refreshes the risk zoning map in real time with the updating of the rainfall prediction, thereby effectively improving the practicability of the early warning information and the pertinence of the emergency response.

Description

Technical Field

[0001] This invention relates to the field of hydrology and water conservancy technology, specifically to an intelligent real-time basin flood early warning method and system. Background Technology

[0002] Floods are among the most frequent and devastating natural disasters globally. Flash floods, triggered by short-duration heavy rainfall, are particularly prominent due to their suddenness, rapid spread, and immense destructive power, making them a key focus and challenge in current flood prevention and disaster reduction efforts. Traditional physical hydrological models, once calibrated offline, have fixed parameters and structures. However, in real-time forecasting, these models suffer from systematic biases due to rainfall input errors and model simplification. While purely data-driven AI forecasting models can fit complex relationships, they lack physical interpretability and extrapolation robustness. Furthermore, traditional methods typically use fixed values ​​based on historical statistics or typical design scenarios as water level warning thresholds, failing to reflect the actual wet and dry conditions of the watershed before rainfall. Warning information is often limited to cross-sectional water levels, lacking spatial accuracy that is dynamically integrated with real-time rainfall distribution, making it difficult to support precise point-to-point emergency responses. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent real-time watershed flood early warning method and system. Traditional physical hydrological models, after offline calibration, have fixed parameters and structures, leading to systematic biases in real-time forecasting due to factors such as rainfall input errors and model simplification. While purely data-driven AI forecasting models can fit complex relationships, they lack physical interpretability and extrapolation robustness. This solution constructs real-time residuals of the physical hydrological model and employs a machine learning-based residual prediction model to specifically learn the complex patterns of residual changes with real-time rainfall intensity and watershed conditions. Through a hybrid architecture of forward and feedback coupling, it provides an interpretable and robust forecast baseline while offering high-precision error correction, effectively improving the overall accuracy of flood forecasting. Robustness: Traditional methods typically use fixed values ​​based on historical statistics or typical design scenarios as water level warning thresholds, failing to reflect the actual wet and dry conditions of the watershed before rainfall. Furthermore, warning information is often limited to cross-sectional water levels, lacking spatial accuracy that is dynamically integrated with real-time rainfall distribution, making it difficult to support precise point-to-point emergency responses. This solution proposes a warning mechanism based on dynamic critical rainfall inversion. It utilizes a fusion forecast model for dynamic inversion, identifying differences in flood risk under different soil conditions, significantly improving the matching degree between warning signals and actual risks. On a unified grid division system, the comprehensive risk value of each grid cell is calculated in real time, and the risk zoning map is updated in real time with rainfall forecast updates, expanding the warning from lines to areas, effectively improving the practicality of warning information and the targeting of emergency responses.

[0004] The technical solution adopted in this invention is as follows: This invention provides an intelligent real-time watershed flood early warning method, which includes the following steps:

[0005] Step S1: Watershed data acquisition. Collect spatial and time series data of the target watershed and preprocess them to form a historical flood event dataset for inundation prediction.

[0006] Step S2: Constructing a fusion forecast model. Based on spatial data and time series data, a physical hydrological model of the target watershed is constructed. A residual prediction model is constructed to correct the simulation data of the physical hydrological model in real time. The physical hydrological model and the residual prediction model are combined into a fusion forecast model through a two-way coupling mechanism to output the final forecast flow sequence.

[0007] Step S3: Calculate multi-level early warning thresholds, perform dynamic critical rainfall inversion based on the fusion forecast model, conduct real-time risk assessment of the target watershed, and generate a dynamic flood risk zoning map;

[0008] Step S4: Early warning generation, converting the final forecast flow sequence and dynamic flood risk zoning map into early warning instructions and generating early warning cases;

[0009] Step S5: Online evaluation. Package the data of each complete flood event into standardized cases, perform automated post-evaluation on each early warning case, calculate the forecast accuracy of the fusion forecast model, and fine-tune the weights of the fusion forecast model online using the gradient descent method.

[0010] Further, in step S2, the construction of the fusion forecast model specifically includes the following steps:

[0011] Step S21: Physical hydrological model construction. Construct a physical model suitable for the target watershed, divide the target watershed into multiple hydrological response units, and simulate the physical mechanism of the entire process of rainfall, runoff generation, confluence, evolution to flood evolution in the river channel and floodplain by combining time series data.

[0012] Step S22: Define the simulation residuals of the physical hydrological model at any time, using the following formula:

[0013] ;

[0014] In the formula, Indicates time, Indicating physical hydrological models in Simulated residuals at time points, express The actual observed flow rate at the outlet section of the basin at any given time. express The flow rate simulated by the physical hydrological model at any given time;

[0015] Step S23: Construct a residual prediction model based on the random forest algorithm, learn the residuals generated by the physical hydrological model in real time, extract the feature vector at the current moment, including current and historical surface rainfall, soil moisture saturation, and flow change rate, input them into the residual prediction model, train the residual prediction model using historical flood event dataset, and output the residual prediction values ​​at each moment within the future early warning forecast period.

[0016] Step S24: Fusion forecast generation. The physical hydrological model and the residual prediction model are combined to form a fusion forecast model. The time step of the forecast period is set. In each forecast period, the flow simulated by the physical hydrological model at the current time is added to the residual prediction value to obtain the corrected forecast flow. The final forecast flow sequence is output. The formula used is as follows:

[0017] ;

[0018] In the formula, Indicates the sequence number of a future time period. express Forecast flow after time correction express The flow rate simulated by the physical hydrological model at any given time. Indicates the early warning period. express The residual value predicted at each time step.

[0019] Furthermore, in step S3, the calculation of the multi-level early warning threshold specifically includes the following steps:

[0020] Step S31: Water level setting, setting early warning control sections for the target watershed, and determining multi-level characteristic water levels;

[0021] Step S32: Calculate the cross-sectional flow rate. Using the Manning formula, establish the relationship between water level and flow rate based on spatial data, and calculate the cross-sectional flow rate thresholds corresponding to multiple characteristic water levels. The formula used is as follows:

[0022] ;

[0023] ;

[0024] In the formula, Indicates the cross-sectional flow threshold. Indicates the cross-sectional area of ​​the water passage. This represents the Manning roughness coefficient. Indicates the hydraulic radius. Indicates wet period, Indicates hydraulic energy slope;

[0025] Step S33: Dynamic critical rainfall inversion. Set different anterior soil moisture conditions and rainfall duration, drive the fusion forecast model to perform automatic iterative inversion until the simulated cross-sectional flow accurately matches the cross-sectional flow threshold of the multi-level characteristic water level, and obtain the corresponding watershed average rainfall.

[0026] Step S34: Query table generation. A dynamic critical rainfall query table is constructed and stored in the form of a three-dimensional array, with dimensions being warning level, rainfall duration, and previous soil moisture, as follows:

[0027] ;

[0028] In the formula, This indicates the critical rainfall level. Indicates the warning level. Indicates the duration of rainfall. Indicates the initial soil moisture. Represents a relational mapping function;

[0029] Step S35: Establish a risk assessment indicator system. Establish a risk assessment indicator system, set the weight of each indicator, calculate the comprehensive risk value of each grid cell in the target watershed, and generate a dynamic flood risk zoning map of the target watershed.

[0030] Furthermore, in step S4, the generation of the warning specifically includes the following steps:

[0031] Step S41: Determine the early warning status, monitor the final forecast flow sequence and its peak flow and peak time within the early warning period, simultaneously monitor the quantitative precipitation forecast values ​​for different durations in the future, compare them in real time with the cross-sectional flow thresholds corresponding to the current soil moisture conditions, perform spatial overlay analysis in conjunction with the dynamic flood risk zoning map of the target watershed, delineate the affected area, and determine the evacuation range, recommended evacuation routes and refuge sites.

[0032] Step S42: Warning Trigger. Based on the dynamic critical rainfall query table, a graded warning trigger is performed, which is converted into a warning instruction and a warning case is generated. The warning case includes: warning level, affected area, expected peak arrival time, expected highest water level and flow rate, recommended evacuation routes and refuge sites.

[0033] The present invention provides an intelligent real-time watershed flood early warning system, comprising a watershed data acquisition module, a fusion forecast model construction module, a multi-level early warning threshold calculation module, an early warning generation module, and an online evaluation module;

[0034] The watershed data acquisition module collects spatial and time series data of the target watershed and preprocesses them to form a historical flood event dataset. The spatial data, time series data, and historical flood event dataset are then sent to the fusion forecast model construction module.

[0035] The fusion forecast model construction module constructs a physical hydrological model of the target watershed based on spatial data and time series data, constructs a residual prediction model to correct the simulation data of the physical hydrological model in real time, and combines the physical hydrological model and the residual prediction model into a fusion forecast model through a two-way coupling mechanism, outputs the final forecast flow sequence, sends the fusion forecast model to the multi-level early warning threshold calculation module, and sends the final forecast flow sequence to the early warning generation module.

[0036] The multi-level early warning threshold calculation module performs dynamic critical rainfall inversion based on the fusion forecast model, conducts real-time risk assessment of the target watershed, generates a dynamic flood risk zoning map, and sends it to the early warning generation module.

[0037] The early warning generation module converts the final forecast flow sequence and dynamic flood risk zoning map into early warning instructions, generates early warning cases, and sends the early warning cases to the online evaluation module.

[0038] The online assessment module packages the complete flood event process data into standardized cases, performs automated post-assessment of each early warning case, calculates the forecast accuracy of the fusion forecast model, and uses the gradient descent method to fine-tune the weights of the fusion forecast model online.

[0039] The beneficial effects achieved by the present invention using the above solution are as follows:

[0040] (1) Traditional physical hydrological models have fixed parameters and structure after offline calibration. In real-time forecasting, there are systematic deviations due to factors such as rainfall input error and model structure simplification. Although pure data-driven AI forecasting models can fit complex relationships, they lack physical interpretability and extrapolation robustness. This scheme constructs the real-time residuals of physical hydrological models and adopts a machine learning-based residual prediction model to specifically learn the complex patterns of residual changes with real-time rainfall intensity and watershed status. Through a hybrid architecture of forward coupling and feedback coupling, it provides interpretable and robust forecasting baselines while providing high-precision error correction, effectively improving the overall robustness of flood forecasting.

[0041] (2) In view of the fact that traditional methods usually use fixed values ​​based on historical statistics or typical design scenarios as water level warning thresholds, which cannot reflect the actual dry and wet conditions of the watershed before rainfall, and the warning information is mostly limited to the cross-sectional water level, lacking spatial accuracy combined with the real-time rainfall area dynamics, making it difficult to support point-to-point precise emergency response, this scheme proposes a warning mechanism based on dynamic critical rainfall inversion. It uses a fusion forecast model for dynamic inversion to identify the differences in flood risk under different soil conditions, which greatly improves the matching degree between the warning signal and the actual risk. On a unified grid division system, the comprehensive risk value of each grid unit is calculated in real time, and the risk zoning map is updated in real time with the update of rainfall forecast, expanding the warning from lines to areas, effectively improving the practicality of warning information and the pertinence of emergency response. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating an intelligent real-time watershed flood early warning method proposed in this invention.

[0043] Figure 2 This is a schematic diagram of an intelligent real-time watershed flood early warning system proposed in this invention.

[0044] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0045] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0046] Example 1, see Figure 1 The present invention provides an intelligent real-time watershed flood early warning method, which includes the following steps:

[0047] Step S1: Watershed data acquisition, collect spatial and time series data of the target watershed and preprocess them to form a historical flood event dataset;

[0048] Step S2: Constructing a fusion forecast model. Based on spatial data and time series data, a physical hydrological model of the target watershed is constructed. A residual prediction model is constructed to correct the simulation data of the physical hydrological model in real time. The physical hydrological model and the residual prediction model are combined into a fusion forecast model through a two-way coupling mechanism to output the final forecast flow sequence.

[0049] Step S3: Calculate multi-level early warning thresholds, perform dynamic critical rainfall inversion based on the fusion forecast model, conduct real-time risk assessment of the target watershed, and generate a dynamic flood risk zoning map;

[0050] Step S4: Early warning generation, converting the final forecast flow sequence and dynamic flood risk zoning map into early warning instructions and generating early warning cases;

[0051] Step S5: Online evaluation. Package the data of each complete flood event into standardized cases, perform automated post-evaluation on each early warning case, calculate the forecast accuracy of the fusion forecast model, and fine-tune the weights of the fusion forecast model online using the gradient descent method.

[0052] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the acquisition of watershed data specifically includes the following steps:

[0053] Step S11: Spatial data acquisition, acquiring DEM (Digital Elevation Model), LULC (Land Use and Land Cover) data, soil type distribution data, river cross-section data, and underlying surface data of the target watershed;

[0054] Step S12: Time series data acquisition, accessing the real-time multi-source rainfall monitoring network data composed of ground rain gauges, weather radar, and meteorological satellite inversion, acquiring real-time water level monitoring data, including water level, flow rate, and rainfall data, and simultaneously accessing short-term heavy rainfall forecast data;

[0055] Step S13: Data preprocessing, performing noise reduction, spatiotemporal alignment, missing value completion, and format standardization on spatial and time series data to organize them into a historical flood event dataset.

[0056] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S2, the fusion forecast model is constructed, which specifically includes the following steps:

[0057] Step S21: Physical hydrological model construction. Using the iWater.FMS watershed flood simulation system as the core engine, a physical hydrological model suitable for the target watershed is constructed. A one-dimensional river channel and a two-dimensional floodplain hydrodynamic model are coupled. Based on spatial data, the target watershed is divided into separate hydrological response units and their grid units. Combined with time series data, the physical mechanism of the entire process from rainfall, runoff generation, confluence, evolution to flood evolution in the river channel and floodplain is simulated.

[0058] Step S22: Define the physical hydrological model at any time The simulated residuals are calculated using the following formula:

[0059] ;

[0060] In the formula, Indicates time, Indicating physical hydrological models in Simulated residuals at time points, express The actual observed flow rate at the outlet section of the target basin at any given time. express The flow rate simulated by the physical hydrological model at any given time;

[0061] Step S23: Construct a residual prediction model based on the random forest algorithm, learn the residuals generated by the physical hydrological model in real time, extract the feature vector at the current moment, including current and historical surface rainfall, soil moisture saturation, and flow change rate, input them into the residual prediction model, train the residual prediction model using historical flood event dataset, and output the residual prediction values ​​at each moment within the future early warning forecast period.

[0062] Step S24: Fusion forecast generation. The physical hydrological model and the residual prediction model are combined to form a fusion forecast model. The time step of the forecast period is set. In each forecast period, the flow simulated by the physical hydrological model at the current time is added to the residual prediction value to obtain the corrected forecast flow. The final forecast flow sequence is output. The formula used is as follows:

[0063] ;

[0064] In the formula, Indicates the sequence number of a future time period. express Forecast flow after time correction express The flow rate simulated by the physical hydrological model at any given time. Indicates the early warning period. express The residual value predicted at each time step.

[0065] Example 4, see Figure 1 This embodiment is based on the above embodiment. In step S21, the construction of the physical hydrological model specifically includes the following steps:

[0066] Step S211: Meteorological unit construction. The Kriging spatial interpolation method is used to interpolate the discrete real-time water level monitoring data in the time series data into the average rainfall of the target watershed. The Penman-Montes formula is used to estimate the potential evapotranspiration of the target watershed.

[0067] Step S212: Runoff generation unit construction. A distributed runoff generation model based on the improved Xin'anjiang model is constructed. Combining soil type distribution data from the spatial data, the total runoff depth of each hydrological response unit is calculated, including the following steps:

[0068] Step S2121: Calculation of the average water storage capacity of the target watershed, using the following formula:

[0069] ;

[0070] In the formula, This represents the average tensile water storage capacity of the target watershed. This represents the maximum single-point water storage capacity within the target watershed. An index representing the distribution curve of water storage capacity;

[0071] Step S2122: Calculation of maximum point water storage capacity, using the following formula:

[0072] ;

[0073] In the formula, This represents the water storage capacity at a point on the water storage capacity distribution curve. Indicates time, express Average soil water storage in the target watershed at any given time;

[0074] Step S2123: Calculation of total runoff depth at each time point, using the following formula:

[0075] ;

[0076] In the formula, Indicates time period, Indicates in The total runoff depth generated at the outlet section of the target basin at any given time. express Effective net rainfall during the period;

[0077] Step S213: Convergence cell construction. The DEM-based kinematic wave method is used to simulate the runoff generation process on the slope and in the river network. The convergence time from each grid cell to the target watershed outlet is calculated using the following formula:

[0078] ;

[0079] ;

[0080] In the formula, Indicates the index of the raster cell. Indicates the total number of grid cells. Indicates the first The distance from one grid cell to the next grid cell along the flow direction. Indicates the first The water flow velocity of each grid cell Indicates the first Manning roughness coefficient corresponding to each grid cell Indicates the first The hydraulic radius of each grid cell Indicates the first The surface slope of each grid cell;

[0081] Step S214: Constructing the river evolution unit, inputting the runoff generation process lines in the slope and river network simulated by the confluence unit, and solving the flow process lines of the target warning section based on the river cross-section data in the spatial data using the dynamic wave form of the one-dimensional Saint-Venant equations.

[0082] Step S215: Initialize parameters. The meteorological unit, runoff generation unit, confluence unit, and channel evolution unit together constitute the physical hydrological model. The parameters of the physical hydrological model are calibrated using historical flood event datasets to obtain a benchmark parameter set applicable to the target watershed.

[0083] By performing the aforementioned operations, this solution addresses the issue that traditional physical hydrological models, after offline calibration, have fixed parameters and structures, leading to systematic biases in real-time forecasts due to factors such as rainfall input errors and model structure simplification. While purely data-driven AI forecasting models can fit complex relationships, they lack physical interpretability and extrapolation robustness. This solution constructs real-time residuals of the physical hydrological model and employs a machine learning-based residual prediction model to specifically learn the complex patterns of residual changes with real-time rainfall intensity and watershed status. Through a hybrid architecture of forward and feedback coupling, it provides interpretable and robust forecast baselines while offering high-precision error correction, effectively improving the overall robustness of flood forecasts.

[0084] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S3, the multi-level early warning threshold calculation specifically includes the following steps:

[0085] Step S31: Water level setting. Based on flood control engineering standards, set early warning control sections for the target basin and determine the three levels of characteristic water levels: early warning, danger, and extremely high danger.

[0086] Step S32: Calculate cross-sectional flow rate. Using the Manning formula, establish the relationship between water level and flow rate based on river cross-sectional data, and calculate the cross-sectional flow rate threshold corresponding to the three characteristic water levels. The formula used is as follows:

[0087] ;

[0088] ;

[0089] In the formula, Indicates the cross-sectional flow threshold. Indicates the cross-sectional area of ​​the water passage. This represents the Manning roughness coefficient. Indicates the hydraulic radius. Indicates wet period, Indicates hydraulic energy slope;

[0090] Step S33: Dynamic critical rainfall inversion. Different anterior soil moisture conditions are set by simulating the average soil water storage in the runoff generation unit. The rainfall duration is set, and the batch calculation function of the iWater.FMS watershed flood simulation system is called to drive the fusion forecast model to perform automatic iterative inversion until the simulated cross-sectional flow accurately matches the cross-sectional flow threshold of the three-level characteristic water level. The corresponding watershed average rainfall is then obtained through inversion.

[0091] Step S34: Query table generation. A dynamic critical rainfall query table is constructed and stored in the form of a three-dimensional array, with dimensions including warning level, rainfall duration, and previous soil moisture. Real-time interpolation queries are supported, and the format is as follows:

[0092] ;

[0093] In the formula, This indicates the critical rainfall level. Indicates the warning level. Indicates the duration of rainfall. Indicates the initial soil moisture. Represents a relational mapping function;

[0094] Step S35: Establish a risk assessment index system. This involves establishing a risk assessment index system that includes the hazard of disaster-causing factors and the vulnerability of disaster-bearing bodies. Weights are set for the hazard indicators of disaster-causing factors and the vulnerability indicators of disaster-bearing bodies. The comprehensive risk value of each grid cell in the target watershed is calculated, and a dynamic flood risk zoning map of the target watershed is generated. The formulas used are as follows:

[0095] ;

[0096] In the formula, This represents the overall risk value. This represents a set of hazard indicators for disaster-causing factors. This represents a set of vulnerability indicators for disaster-bearing entities. Index of hazard indicators for disaster-causing factors Index representing the vulnerability index of disaster-bearing bodies. , These represent the weights of the hazard index of the disaster-causing factor and the vulnerability index of the disaster-bearing body, respectively. , These represent the standardized scores of the hazard index of the disaster-causing factor and the vulnerability index of the disaster-bearing body, respectively.

[0097] By performing the aforementioned operations, this solution addresses the problems of traditional methods that typically use fixed values ​​based on historical statistics or typical design scenarios as water level warning thresholds, failing to reflect the actual wet and dry conditions of the watershed before rainfall, and having warning information mostly limited to cross-sectional water levels, lacking spatial accuracy in conjunction with real-time rainfall dynamics, and struggling to support precise point-to-point emergency responses. This solution proposes a warning mechanism based on dynamic critical rainfall inversion. It utilizes a fusion forecast model for dynamic inversion, identifying differences in flood risk under different soil conditions, significantly improving the matching degree between warning signals and actual risks. On a unified grid division system, the comprehensive risk value of each grid cell is calculated in real time, and the risk zoning map is updated in real time with rainfall forecast updates, expanding the warning from lines to areas, effectively improving the practicality of warning information and the targeting of emergency responses.

[0098] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S4, the warning is generated, which specifically includes the following steps:

[0099] Step S41: Determine the early warning status, monitor the final forecast flow sequence and its peak flow and peak time within the early warning period, simultaneously monitor the quantitative precipitation forecast values ​​for different durations in the future, compare them in real time with the cross-sectional flow thresholds corresponding to the current soil moisture conditions, perform spatial overlay analysis in conjunction with the dynamic flood risk zoning map of the target watershed, delineate the affected area, and determine the evacuation range, recommended evacuation routes and refuge sites.

[0100] Step S42: Warning Trigger. Based on the dynamic critical rainfall query table, a graded warning trigger is performed, which is converted into a warning instruction and a warning case is generated. The warning case includes: warning level, affected area, expected peak arrival time, expected highest water level and flow rate, recommended evacuation routes and refuge sites.

[0101] Example 7, see Figure 1 This embodiment is based on the above embodiment. In step S5, an online evaluation is performed, using NSE (Nash efficiency coefficient), relative error of runoff depth, relative error of peak flow, and peak occurrence time to evaluate the forecast accuracy of the fusion forecast model. The NSE calculation formula is as follows:

[0102] ;

[0103] In the formula, This represents the Nash efficiency coefficient. Indicates time, This indicates the total number of time steps within the current evaluation period. and They represent Real-time observation and fusion forecast of flow, This represents the average measured flow rate during this flood event.

[0104] Example 8, see Figure 2 Based on the above embodiments, this embodiment provides an intelligent real-time watershed flood early warning system, including a watershed data acquisition module, a fusion forecast model construction module, a multi-level early warning threshold calculation module, an early warning generation module, and an online evaluation module;

[0105] The watershed data acquisition module collects spatial and time series data of the target watershed and preprocesses them to form a historical flood event dataset. The spatial data, time series data, and historical flood event dataset are then sent to the fusion forecast model construction module.

[0106] The fusion forecast model construction module constructs a physical hydrological model of the target watershed based on spatial data and time series data, constructs a residual prediction model to correct the simulation data of the physical hydrological model in real time, and combines the physical hydrological model and the residual prediction model into a fusion forecast model through a two-way coupling mechanism, outputs the final forecast flow sequence, sends the fusion forecast model to the multi-level early warning threshold calculation module, and sends the final forecast flow sequence to the early warning generation module.

[0107] The multi-level early warning threshold calculation module performs dynamic critical rainfall inversion based on the fusion forecast model, conducts real-time risk assessment of the target watershed, generates a dynamic flood risk zoning map, and sends it to the early warning generation module.

[0108] The early warning generation module converts the final forecast flow sequence and dynamic flood risk zoning map into early warning instructions, generates early warning cases, and sends the early warning cases to the online evaluation module.

[0109] The online assessment module packages the complete flood event process data into standardized cases, performs automated post-assessment of each early warning case, calculates the forecast accuracy of the fusion forecast model, and uses the gradient descent method to fine-tune the weights of the fusion forecast model online.

[0110] Example 9, see Figure 1 This embodiment, based on the above embodiment, details the specific application of the fusion forecasting model in a typical urban flash flood early warning scenario. The target watershed is a hilly urban watershed with an area of ​​150 square kilometers, including the main urban area, the urban-rural fringe, and the upstream mountainous area. The watershed has one hydrological station and eight automatic rain gauges, and access to regional meteorological radar data. The disaster-bearing entities within the watershed include 12 administrative villages, three key industrial parks, two schools, and one hospital. Each area has designated flood control personnel and emergency plans. The specific steps include:

[0111] Step A1: Model deployment and initialization, including the following steps:

[0112] Step A11: Deployment of physical hydrological model. Based on the iWater.FMS watershed flood simulation system, construct a physical hydrological model of the target watershed. Set the model grid resolution to 30m×30m and divide the target watershed into 167,000 computing units. Based on historical flood event datasets, calibrate runoff parameters and Manning roughness coefficients to obtain a baseline parameter set.

[0113] Step A12: Deployment of residual prediction model. Based on the Scikit-learn machine learning library in Python, construct a residual prediction model based on the random forest algorithm. Input the data of 23 historical flood events in the past 10 years of the target watershed. Set the early warning prediction period to 6 hours, with a prediction time every 10 minutes, for a total of 36 prediction times. Train the model to obtain the residual prediction model for the future early warning prediction period.

[0114] Step A2: Execute the real-time business process at each hour on the hour, including the following steps:

[0115] Step A21: Data assimilation and status update. Collect measured rainfall data from 8 rain gauge stations in the basin over the past 6 hours, generate the basin surface rainfall process using Kriging spatial interpolation, and read the actual observed flow rate at the current outlet section of the hydrological station. The measured rainfall data is input into the physical hydrological model, which automatically runs to the current time, updates soil moisture and water level to the latest observation status, and completes data assimilation.

[0116] Step A22: Processing future rainfall inputs, obtaining gridded QPF (Quantitative Precipitation Forecast) data for the next 6 hours with a spatial resolution of 0.1°×0.1° and a temporal resolution of 1 hour, downscaling the QPF data to a 30-meter grid in the model, and generating a watershed surface rainfall forecast sequence with a step size of 10 minutes for the next 6 hours;

[0117] Step A23: Fusion forecast generation, including the following steps:

[0118] Step A231: Physical hydrological model forecast. The assimilated data is used as the initial condition. The future watershed isal rainfall forecast sequence is input into the physical hydrological model, and the preliminary flow forecast sequence of the watershed outlet for the next 6 hours is output.

[0119] Step A232: Real-time construction of feature vectors, extracting the feature vectors at the current time. This includes the current isal rainfall and the rainfall in the previous 3 hours, the current soil moisture saturation in the model, and the current simulated flow rate. Its rate of change, QPF data for the next hour, and simulated residuals for the first three time points. , , ;

[0120] Step A233: Residual prediction, using the feature vector Input into the residual prediction model, output the residual prediction values ​​for the next 36 time points { };

[0121] Step A234: Forecast Fusion. The preliminary flow forecast sequence from the physical hydrological model is added to the residual prediction values ​​from the residual prediction model time-by-time to obtain the final forecast flow sequence. ;

[0122] Step A24: Forecast output, including:

[0123] Core output: A graph and table showing the 6-hour watershed outlet flow process.

[0124] Intermediate products: Animation of spatial evolution of soil moisture in the watershed over the next 6 hours, and a breakdown map of runoff components;

[0125] Forecast uncertainty assessment: Based on the prediction differences of each decision tree in the random forest algorithm, calculate the 90% confidence interval of the flow forecast at each foreseeable time.

[0126] Step A3: Warning Level Determination. Based on the final forecast flow sequence and dynamic critical rainfall lookup table output by the fusion forecast model, a dynamic flood risk zoning map of the target watershed is generated, including the following steps:

[0127] Step A31: Determine the risk level threshold based on the comprehensive risk value of each grid cell, specifically as follows:

[0128] A blue alert has been issued due to the low risk level.

[0129] A yellow alert has been issued for medium-risk areas.

[0130] An orange alert has been issued due to the high risk level.

[0131] A red alert has been issued due to extremely high risk.

[0132] Step A32: Statistically determine the overall warning level of the watershed. Based on the comprehensive risk value of all grid units, if there is a continuous area with an area greater than 0.1 square kilometers that reaches a certain risk threshold, the overall warning level of the watershed is determined to be that level, and a warning is issued for the entire watershed.

[0133] Step A33: Identify the core impact area, extract all grid cells with a risk level higher than the yellow alert, form a vector surface of the warning impact area, perform clustering and boundary smoothing on the vector surface to obtain independent high-risk patches, and each high-risk patch is a specific impact area;

[0134] Step A4: Targeted Early Warning Generation. After determining the overall early warning level and high-risk patches in the watershed, targeted early warnings are generated, including the following steps:

[0135] Step A41: Match the affected area with the disaster-bearing body. Perform spatial overlay analysis on each high-risk patch and the disaster-bearing body to generate a table of association between the affected area and the disaster-bearing body.

[0136] Step A42: Generate graded early warning information. Based on the dynamic flood risk zoning map and the association table between affected areas and disaster-bearing bodies, construct early warning instructions and push evacuation ranges, suggested evacuation routes, and refuge sites to the flood control personnel in each affected area.

[0137] Step A43: Multi-channel priority push, establish a four-level push channel to confirm the sending of warning commands, specifically as follows:

[0138] Level 1 Channel: Instant messaging, pushed via a dedicated flood control APP, with vibration and strong alerts, requiring the person in charge of flood control to click to confirm within 15 minutes. If not confirmed, the push will be repeated every 5 minutes.

[0139] Secondary channel: voice call, automatically initiating voice calls to the flood control personnel responsible for high-risk patches, and using TTS technology to read out the early warning instructions;

[0140] Level 3 channel: SMS platform, which sends standard format SMS messages to flood control personnel, providing all flood control personnel with the dynamic flood risk zoning map of the target watershed for viewing;

[0141] Level 4 Channel: Emergency Broadcast, which automatically triggers the emergency broadcast system in the affected area to continuously broadcast warning instructions, especially reminding vulnerable groups such as the elderly, the infirm, and the disabled;

[0142] Step A44: Real-time monitoring of response status, including the following steps:

[0143] Step A441: Confirm feedback collection, display the confirmation status of each flood control person in real time, count the confirmation rate of each area, and automatically remind superiors when the rate is below 80%;

[0144] Step A442: Tracking the evacuation progress. The person in charge of flood control reports the start time of the evacuation and the number of people evacuated through the flood control APP. The system automatically calculates the percentage of evacuation progress, generates a evacuation heat map, and displays the progress of each area.

[0145] Step A443: Reporting abnormal situations. Set up options for rapid reporting of road closures, personnel stranded, and material shortages. The system will automatically recommend emergency resource dispatching plans.

[0146] Step A45: Dynamic adjustment and downgrading of early warnings, including the following steps:

[0147] Step A451: Receive the latest QPF data every 30 minutes for rolling updates, rerun the fusion forecast model, and if the forecast rainfall decreases, the system automatically suggests downgrading the warning; if the forecast rainfall increases, upgrade the warning and expand the affected area.

[0148] Step A452: Perform feedback calibration based on real-time water level monitoring data. If the actual rainfall is significantly weaker than the forecast, initiate a rapid assessment. After manual confirmation, appropriately lower the warning level.

[0149] Step A453: Warning cancellation mechanism. When the rainfall process has basically ended and the river water level continues to drop, the warning will be cancelled and the cancellation information will be pushed to the flood control personnel in each area in a tiered manner.

[0150] Step A5: Online learning and model update. After each flood event, the rainfall input, model status, fusion forecast results, and measured flow are packaged into standardized cases and added to the historical flood event dataset. The NSE, peak flow relative error, and peak time error are calculated. The parameters in the residual prediction model involving the peak period are fine-tuned to complete the incremental training of the model.

[0151] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0152] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0153] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent real-time watershed flood early warning method, characterized in that: The method includes the following steps: Step S1: Watershed data acquisition, collect spatial and time series data of the target watershed and preprocess them to form a historical flood event dataset; Step S2: Constructing the fusion forecast model. Based on spatial and time-series data, a physical hydrological model of the target watershed is constructed. A residual prediction model is built to correct the simulation data of the physical hydrological model in real time. The physical hydrological model and the residual prediction model are combined into a fusion forecast model through a two-way coupling mechanism to output the final forecast flow sequence. This includes the following steps: Step S21: Physical hydrological model construction. Construct a physical hydrological model suitable for the target watershed, coupled with a one-dimensional river channel and a two-dimensional floodplain hydrodynamic model. Based on spatial data, divide the target watershed into multiple hydrological response units and grid units. Combine time series data to simulate the physical mechanism of the entire process from rainfall, runoff generation, confluence, evolution to flood evolution in the river channel and floodplain. Step S22: Define the simulation residuals of the physical hydrological model at any time, using the following formula: ； In the formula, Indicates time, Indicating physical hydrological models in Simulated residuals at time points express The actual observed flow rate at the outlet section of the target basin at any given time. express The flow rate simulated by the physical hydrological model at any given time; Step S23: Construct a residual prediction model based on the random forest algorithm, extract the feature vector at the current moment, input it into the residual prediction model, train the residual prediction model using the historical flood event dataset, and output the residual prediction values ​​at each moment within the future early warning prediction period. Step S24: Fusion forecast generation. The physical hydrological model and the residual prediction model are combined to form a fusion forecast model. The time step of the forecast period is set. In each forecast period, the flow simulated by the physical hydrological model at the current time is added to the residual prediction value to obtain the corrected forecast flow. The final forecast flow sequence is output. The formula used is as follows: ； In the formula, Indicates the sequence number of a future time period. express Forecast flow after time correction express The flow rate simulated by the physical hydrological model at any given time. Indicates the early warning period. express The residual value predicted at each time step; Step S3: Calculate multi-level early warning thresholds, perform dynamic critical rainfall inversion based on the fusion forecast model, conduct real-time risk assessment of the target watershed, and generate a dynamic flood risk zoning map, including the following steps: Step S31: Water level setting, setting early warning control sections for the target watershed, and determining multi-level characteristic water levels; Step S32: Calculate the cross-sectional flow rate. Use the Manning formula to establish the relationship between water level and flow rate based on spatial data, and calculate the cross-sectional flow rate threshold corresponding to the multi-level characteristic water levels. Step S33: Dynamic critical rainfall inversion. Set different anterior soil moisture conditions and rainfall duration, drive the fusion forecast model to perform automatic iterative inversion until the simulated cross-sectional flow accurately matches the cross-sectional flow threshold of the multi-level characteristic water level, and obtain the corresponding watershed average rainfall. Step S34: Query table generation, constructing and storing the dynamic critical rainfall query table in the form of a three-dimensional array; Step S35: Establish a risk assessment indicator system. Establish a risk assessment indicator system, set the weight of each indicator, calculate the comprehensive risk value of each grid cell in the target watershed, and generate a dynamic flood risk zoning map of the target watershed. Step S4: Early Warning Generation. The final forecast flow sequence and dynamic flood risk zoning map are converted into early warning instructions to generate early warning cases, including the following steps: Step S41: Determine the early warning status, monitor the final forecast flow sequence and its peak flow and peak time within the early warning period, simultaneously monitor the quantitative precipitation forecast values ​​for different durations in the future, compare them in real time with the cross-sectional flow thresholds corresponding to the current soil moisture conditions, perform spatial overlay analysis in conjunction with the dynamic flood risk zoning map of the target watershed, delineate the affected area, and determine the evacuation range, recommended evacuation routes and refuge sites. Step S42: Warning triggering: Based on the dynamic critical rainfall lookup table, graded warning triggering is performed, which is converted into warning instructions and warning cases are generated; Step S5: Online evaluation. Package the data of each complete flood event into standardized cases, perform automated post-evaluation on each early warning case, calculate the forecast accuracy of the fusion forecast model, and fine-tune the weights of the fusion forecast model online.

2. An intelligent real-time watershed flood early warning system, used to implement the intelligent real-time watershed flood early warning method as described in claim 1, characterized in that: It includes a watershed data acquisition module, a fusion forecast model construction module, a multi-level early warning threshold calculation module, an early warning generation module, and an online evaluation module; The watershed data acquisition module collects spatial and time series data of the target watershed and preprocesses them, organizes them into historical flood event data, and sends the spatial data, time series data and historical flood event data to the fusion forecast model construction module. The fusion forecast model construction module constructs a physical hydrological model of the target watershed based on spatial data and time series data, constructs a residual prediction model to correct the simulation data of the physical hydrological model in real time, and combines the physical hydrological model and the residual prediction model into a fusion forecast model through a two-way coupling mechanism, outputs the final forecast flow sequence, sends the fusion forecast model to the multi-level early warning threshold calculation module, and sends the final forecast flow sequence to the early warning generation module. The multi-level early warning threshold calculation module performs dynamic critical rainfall inversion based on the fusion forecast model, conducts real-time risk assessment of the target watershed, generates a dynamic flood risk zoning map, and sends it to the early warning generation module. The early warning generation module converts the final forecast flow sequence and dynamic flood risk zoning map into early warning instructions, generates early warning cases, and sends the early warning cases to the online evaluation module. The online assessment module packages the complete flood event process data into standardized cases, performs automated post-assessment of each early warning case, calculates the forecast accuracy of the fusion forecast model, and uses the gradient descent method to fine-tune the weights of the fusion forecast model online.

Images