A small watershed flood forecasting and warning method and system

Abstract

The application provides a small watershed flood forecasting and early warning method and system, and belongs to the technical field of mountain torrent disaster prevention. The method comprises the following steps: collecting multi-source data and preprocessing; constructing a physically guided neural network model, embedding the hydrodynamic physical equation as a constraint into the network training process; constructing a dynamic adaptability evaluation index system, calculating the initial values of the rainfall-runoff response consistency, model structure applicability and prediction uncertainty indicators; constructing a reinforcement learning agent, executing a cyclic correction mechanism with the initial values of the three indicators as the starting state to obtain the final converged values of the three indicators; configuring the parameters of the physically guided neural network model according to the converged indicator values, executing flood forecasting by adopting a hierarchical progressive strategy, and generating hierarchical early warning information in combination with dynamic hazard body information. The application realizes self-adaptive optimization and dynamic parameter adjustment of the prediction model, and solves the problems of poor model adaptability, parameter calibration dependence on historical data and single early warning indicator in the prior art.

Description

Technical Field

[0001] This invention relates to the field of flash flood disaster prevention technology, and in particular to a method and system for forecasting and warning floods in small watersheds. Background Technology

[0002] Small watershed floods are characterized by short confluence times, sudden onset, and high destructiveness, making them a major form of flash flood disaster. Because small watersheds typically lack comprehensive hydrological monitoring facilities, and their runoff generation and confluence mechanisms are significantly influenced by the heterogeneity of the underlying surface, the accuracy and lead time of flood forecasting have always been challenging issues for the industry.

[0003] Existing flood forecasting methods for small watersheds mainly include physical models based on distributed unit hydrographs, forecasting methods based on radar quantitative precipitation estimation, and data-driven models based on artificial intelligence. Distributed unit hydrograph-based flood forecasting methods rely on static underlying surface parameters and cannot respond to the dynamic impact of rainfall intensity changes on the velocity field, resulting in a significant decrease in forecast accuracy under heavy rainfall conditions. Radar quantitative precipitation estimation-based forecasting methods primarily address precipitation input issues and do not adequately consider the nonlinear characteristics of runoff generation and confluence processes. Artificial intelligence-based flood forecasting methods are highly dependent on historical data, making them difficult to apply effectively in small watersheds lacking long-term observational data, and their model interpretability is poor.

[0004] To address the aforementioned issues, methods coupling physical models and data-driven models have emerged in recent years. By embedding hydrophysical equations as constraints into the neural network training process, these methods balance forecast accuracy and model interpretability to some extent. However, existing coupling methods still suffer from the following technical drawbacks: complex model parameter calibration and low computational efficiency; fixed model structure, unable to be dynamically adjusted according to different rainfall intensities and underlying surface conditions; and single early warning indicators, failing to fully consider the combined impact of previous soil moisture content and dynamic disaster-bearing body information.

[0005] Therefore, there is an urgent need for a method for forecasting and early warning of floods in small watersheds. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for forecasting and warning floods in small watersheds. By constructing a dynamic adaptive evaluation index system and a reinforcement learning cyclic correction mechanism, the forecasting model can be dynamically optimized and its parameters can be adaptively adjusted. At the same time, it can generate refined warnings by combining dynamic disaster-bearing body information, so as to solve the technical problems of poor adaptability of single models, dependence on historical data for parameter calibration, and single warning indicators in the existing technology.

[0007] To achieve the above objectives, the present invention provides a method for forecasting and early warning of floods in small watersheds, comprising the following steps: Step S1: Collect multi-source data of the study area and perform preprocessing; Step S2: Construct a physical-guided neural network model. The physical-guided neural network model uses hydrodynamic physical equations as constraints to embed into the network training process. Step S3: Construct a dynamic adaptive evaluation index system and calculate the initial values ​​of the rainfall-runoff response consistency index, model structure applicability index, and forecast uncertainty index; Step S4: Construct a reinforcement learning agent. Starting with the three index values ​​calculated in step S3 as the initial state, execute a loop correction mechanism. The reinforcement learning agent outputs correction actions to iteratively correct the three indexes in turn until the convergence condition is met, and finally obtain the three converged index values. Step S5: Configure the parameters of the physical guided neural network model based on the three final converged index values, execute flood forecasting using a hierarchical progressive strategy, and generate graded early warning information based on the flood forecast results and dynamic disaster-bearing body information.

[0008] This invention also provides a small watershed flood forecasting and early warning system for implementing the above-described small watershed flood forecasting and early warning method, comprising: The data acquisition and preprocessing module is used to collect multi-source data from the study area and perform preprocessing. A physics-guided neural network model building module is used to embed hydrodynamic physical equations as constraints into the network training process. The dynamic adaptive evaluation index construction module is used to calculate the initial values ​​of the rainfall-runoff response consistency index, model structure suitability index, and forecast uncertainty index. The reinforcement learning agent construction module is used to start with three initial values ​​as the initial state, execute a loop correction mechanism, output correction actions to iteratively correct the three indicators in turn until the convergence condition is met, and obtain the final converged values ​​of the three indicators. The flood forecasting and early warning generation module is used to configure the parameters of the physical-guided neural network model based on the three final converged index values, execute flood forecasting using a hierarchical progressive strategy, and generate graded early warning information based on the flood forecasting results combined with dynamic disaster-bearing body information.

[0009] Therefore, the present invention employs the above-mentioned method and system for forecasting and early warning of small watershed floods, and the beneficial technical effects are as follows: (1) This invention constructs a dynamic adaptive evaluation index system that includes rainfall-runoff response consistency index, model structure applicability index and forecast uncertainty index, and combines a reinforcement learning agent to execute a cyclic correction mechanism to achieve adaptive dynamic adjustment of model structure and parameters. This effectively solves the problem of poor adaptability of a single model under different hydrological conditions and can complete model optimization without relying on a large amount of historical data.

[0010] (2) The physical-guided neural network model constructed in this invention uses the water balance equation and Saint-Venant equations as constraints to embed into the network training process. While maintaining the fitting ability of the data-driven model, it ensures the physical consistency of the forecast results, solves the problem of poor interpretability of existing data-driven models, and improves the applicability of the model in areas with little historical data in small watersheds.

[0011] (3) The present invention uses the three final converged evaluation indicators for dynamic adjustment of the early warning threshold, and combines them with a two-dimensional hydrodynamic model to simulate the flood evolution process. Through spatial overlay analysis, it accurately assesses the information of the affected population, buildings and roads, etc., and realizes the refinement from "flood warning" to "risk warning", providing more targeted support for flood control decision-making. Attached Figure Description

[0012] Figure 1 This is a flowchart of a small watershed flood forecasting and early warning method according to the present invention; Figure 2 Here is a flowchart of the cyclical correction mechanism; Figure 3 A comparison diagram of flow process curves for different forecasting methods. Detailed Implementation

[0013] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0014] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0015] Example 1 A method for forecasting and warning floods in small watersheds includes the following steps: Step S1: Collect multi-source data of the study area and perform preprocessing.

[0016] This embodiment uses a small watershed in a mountainous area as the application object, with a watershed area of ​​156 km². 2 It belongs to a typical small watershed in a hilly area.

[0017] Step S11: Collect multi-source data for the study area: Rainfall data measured by rain gauge network Five automatic rain gauges were deployed within the watershed, with data collected every five minutes and transmitted in real time to the data center via a GPRS network.

[0018] Radar quantitative precipitation estimation data It connects to the meteorological department's X-band weather radar products, with a spatial resolution of 1km×1km and a temporal resolution of 5 minutes, covering the entire watershed area.

[0019] Digital Elevation Model Data : 12.5m resolution survey data was used to extract topographic parameters such as watershed boundaries, river networks, slope, and aspect.

[0020] Soil moisture content data Three soil moisture monitoring stations were set up in the watershed to collect the volumetric water content of three soil layers at depths of 10cm, 20cm, and 40cm, with a data collection frequency of 1 hour.

[0021] Actual flow data Radar level gauges and flow velocity meters are installed at the outlet section of the basin. The flow is induced by the water level-flow curve, and the data acquisition frequency is 5 minutes.

[0022] Population distribution data Based on the latest census data, and combined with the spatial processing of administrative village boundaries, a population density distribution map with a 100m×100m grid is formed.

[0023] Building distribution data Based on high-resolution satellite image interpretation, the location, outline, and number of floors of buildings within the watershed are extracted.

[0024] Road data Based on navigation map data, vector layers of national highways, provincial highways, county roads and rural roads within the basin are extracted.

[0025] Step S12: Perform spatiotemporal matching on the collected multi-source data. Unify all data to the CGCS2000 coordinate system, unify the time to Beijing time, and unify the spatial resolution to a 100m×100m grid.

[0026] Step S13: Perform outlier removal and missing value imputation on the matched data. Outliers are removed using the 3σ principle, and missing values ​​are imputed using inverse distance weighted interpolation.

[0027] Step S14, based on and Construct a function to fuse rainfall data, with the following expression: ; in, For the merged rainfall data, The weighting coefficient for the rain gauge station. This represents the radar weighting coefficient. In this embodiment, considering the moderate density of the rain gauge network within the watershed and the complete spatial coverage of radar data, we take... =0.4, =0.6, which ensures that the fusion results take into account both the accuracy of rainfall stations and the spatial representativeness of radar.

[0028] Step S2: Construct a physical guidance neural network model.

[0029] Step S21: Construct an input layer to receive fused rainfall data. and digital elevation model data The input time window is set to the past 6 hours, the time step is 10 minutes, and the rainfall sequence is 36 time steps in total.

[0030] Step S22: Construct a physical constraint embedding layer. The water balance equation and Saint-Venant equations are used as physical constraints in the network training process. A soft constraint approach is used to construct the loss function, expressed as: ; in, For the total loss function, For data-driven loss terms, For physical constraint loss terms, The physical constraint weighting coefficients are initially set in this embodiment. =0.1; The expression is: ; in, For the sample size, The sample number. For the first Forecast flow for each sample The first data collected in step S11 The measured flow rate of each sample; The expression is: ; in, Because of the water depth, For flow rate, For time, For spatial coordinates, For lateral inflow, It is the acceleration due to gravity. The riverbed slope, Frictional slope, riverbed slope Digital elevation model data acquired in step S11 Extracted.

[0031] Step S23: Construct the attention mechanism layer, including a spatial attention sublayer and a temporal attention sublayer. The spatial attention sublayer sets up 8 spatial attention heads to capture the contribution weights of different sub-basins within the watershed to the flood at the outlet section; the temporal attention sublayer sets up 6 temporal attention heads to capture the impact weights of rainfall in different historical periods on the current flow.

[0032] Step S24: Construct the output layer to output the flood forecast flow. The output is the flow rate process for the next 3 hours, with a time step of 10 minutes.

[0033] Step S3: Construct a dynamic adaptive evaluation index system and calculate the initial values ​​of the rainfall-runoff response consistency index, model structure applicability index, and forecast uncertainty index.

[0034] Step S31: Calculate the rainfall-runoff response consistency index using the mutual information method. The expression is: ; in, The fused rainfall sequence output in step S14 Compared with the measured flow sequence collected in step S11 mutual information content This represents the total number of states in the rainfall sequence. The total number of states in the flow sequence. This is the sequence number of the rainfall status. This is the flow status sequence number. Rainfall With traffic status The joint probability distribution function, Rainfall The marginal probability distribution function, Traffic status The marginal probability distribution function.

[0035] In this embodiment, take =10, =10, dividing rainfall and flow rate into 10 state intervals respectively. The initial values ​​are calculated. =0.62, indicating a strong correlation between rainfall and runoff.

[0036] Step S32: Calculate the model structure suitability index using the cosine similarity method. The expression is: ; in, The applicable feature vectors for the physically guided neural network model constructed in step S2. The feature vector of the underlying surface of the watershed. Digital elevation model data acquired in step S11 and soil moisture content data The construction includes features such as average slope, river network density, soil saturation, and topographic index. This embodiment calculates the initial... =0.58, indicating that the model structure has a moderate degree of matching with the characteristics of the watershed underlying surface.

[0037] Step S33: Calculate the forecast uncertainty index using the Monte Carlo dropout method. The expression is: ; in, The flood forecast flow rate output in step S24 At half the width of the 90% confidence interval, The flood forecast flow rate output in step S24 The mean value. In this embodiment, the initial value is obtained through 50 dropout forward propagations. =0.23, indicating that the uncertainty of the model forecast is at a moderate level.

[0038] Step S4: Construct a reinforcement learning agent. Starting with the three initial values ​​of the indicators calculated in step S3, execute a loop correction mechanism. The reinforcement learning agent outputs correction actions to iteratively correct the three indicators in turn until the convergence condition is met, and finally obtain the three converged indicator values.

[0039] Step S41: Construct the state space, which is defined by the rainfall-runoff response consistency index output in step S31. The model structure suitability index output in step S32 And the forecast uncertainty index output in step S33 The composition, the state vector is represented as superscript This represents the matrix transpose. The initial state in this embodiment... .

[0040] Step S42: Construct the action space, which includes the first correction coefficient. Second correction factor and the third correction factor The action vector is represented as ,in , , .

[0041] Step S43: Train the reinforcement learning agent using the deep deterministic policy gradient algorithm, which includes an Actor network and a Critic network. Both the Actor network and the Critic network use a 3-layer fully connected structure with 128 neurons per layer. The training run is 500 rounds with a learning rate of 0.001.

[0042] Step S44: Construct the reward function, the expression of which is: ; in, As a reward value, This is the amount of improvement in the Nash efficiency coefficient of the model before and after correction. This represents the current iteration step. Action vector L2 norm, , , As a balancing factor, in this embodiment, the values ​​are set to 0.6, 0.3, and 0.1 respectively, prioritizing the improvement of forecast accuracy while taking into account convergence speed and smoothness of action.

[0043] Nash efficiency coefficient The expression is: ; in, This represents the average measured flow rate.

[0044] The cyclical correction mechanism specifically includes: Step S45: Initial state value calculated in step S3 Input into the reinforcement learning agent trained in step S43.

[0045] Step S46: The reinforcement learning agent outputs a correction action. Perform cyclical corrections sequentially: First correction: using the first correction factor Correct the model structure suitability index output in step S32 and with the revised Adjusting the rainfall-runoff response consistency index in step S31 Update the calculation parameters. The expression is: ; in, For the updated rainfall-runoff response consistency index, The previous rainfall-runoff response consistency index, As an indicator of the applicability of the current model structure, This is used as a reference model structure applicability index. This embodiment uses... =0.70. In the first iteration, the agent outputs... =1.12, calculated as follows =0.62×(1+1.12×(0.58−0.70))=0.62×0.8656=0.54. The significance of this correction is that when the model structure suitability index is low, it indicates that the current model structure does not match the characteristics of the watershed underlying surface. It is necessary to recalculate the rainfall-runoff response relationship by adjusting parameters such as the time lag window, so that the model can more accurately capture the runoff generation and concentration characteristics of the current watershed.

[0046] Second correction: using the second correction factor The updated As a priori for uncertainty quantification, the forecast uncertainty index in step S33 is corrected. The confidence interval boundaries are updated. The expression is: ; in, For the updated forecast uncertainty index, This is the forecast uncertainty index before the update. In the first iteration, the agent output... =0.95, calculated as follows =0.23×(1+0.95×(1−0.54))=0.23×1.437=0.33. The significance of this correction is that when the consistency between rainfall and runoff response decreases, the model's simulation accuracy of the current hydrological process declines. The uncertainty interval should be expanded to reflect the increased forecast risk and avoid decision-making errors.

[0047] Third correction: using the third correction factor The updated As feedback, the model structure suitability index in step S32 is corrected in reverse. ,renew The expression is: ; in, For the updated model structure applicability index, This is an indicator of the suitability of the model structure before the update. In the first iteration, the agent output... =1.05, calculated as follows =0.58×(1−1.05×0.33)=0.58×0.6535=0.38. The significance of this correction is that when the forecast uncertainty increases significantly, even if the model structure applicability index is still acceptable, it indicates that the current model structure is difficult to accurately simulate the actual process, and it is necessary to appropriately lower the applicability threshold to trigger model structure adjustment.

[0048] Step S47: Update the status As the state input for the next iteration, repeat steps S46 to S47 until the convergence condition is met.

[0049] The convergence condition is: ; or number of iterations Reaching the preset maximum number of iterations ,in, For the first The consistency index of rainfall-runoff response in the next iteration For the first The model structure suitability index for the next iteration. For the first The forecast uncertainty index for the next iteration. The first convergence threshold is... The second convergence threshold, This is the third convergence threshold.

[0050] After five rounds of iteration, the model's structural suitability was significantly improved to 0.72, the consistency of rainfall-runoff response was improved to 0.67, the forecast uncertainty was reduced to 0.18, and the overall adaptability of the model was optimized.

[0051] Step S5: Configure the parameters of the physical guided neural network model based on the three final converged index values, execute flood forecasting using a hierarchical progressive strategy, and generate graded early warning information based on the flood forecast results and dynamic disaster-bearing body information.

[0052] The parameter configuration of the physics-guided neural network model based on the three final convergence metrics specifically includes: Step S51: Calculate the final converged rainfall-runoff response consistency index output in step S47. Model structure applicability index and forecast uncertainty indicators Parameter configuration vector .

[0053] Step S52: Adjust the physical constraint weight coefficients of the physical-guided neural network model in step S22 according to the parameter configuration vector. The expression is: ; in, The initial physical constraint weight coefficient is set to 0.1. The adjustment coefficient for the rainfall-runoff response consistency index is set to 0.5. The adjustment coefficient for the forecast uncertainty index is set to 0.5. Due to the high consistency between rainfall and runoff responses, the weights of physical constraints are appropriately increased to make the model more strictly follow physical laws.

[0054] Step S53: Adjust the output weights and spatial attention weights of the attention mechanism layer in the physically guided neural network model in step S23 according to the parameter configuration vector. And time attention weight The expression is: ; ; in, and For learnable parameter matrix, Let be the dimension of the spatial attention weight vector. For dimension number, Let the dimension of the time attention weight vector be . For dimension number, For learnable parameter matrix The One portion, For learnable parameter matrix The Each component.

[0055] The tiered and progressive strategy for implementing flood forecasting specifically includes: Step S54: Perform short-term forecasting. Extrapolate the rainfall for the next 3 hours using the radar quantitative precipitation estimation data from the most recent 6 hours, and combine it with the adjusted physical-guided neural network model to perform rolling forecasts, thereby obtaining the short-term forecast flow. Forecast lead time is 0-3 hours, with a time resolution of 10 minutes.

[0056] Step S55: Perform short-term forecasting. Use numerical weather prediction data (future 24 hours, time resolution 1 hour) combined with an adjusted physical-guided neural network model to obtain the short-term forecast flow. Forecast lead time is 3-24 hours, with a time resolution of 1 hour.

[0057] Step S56: Perform a medium-term forecast by combining multi-model ensemble weather forecast data (3 days ahead, 3-hour time resolution) with an adjusted physical-guided neural network model to obtain the medium-term forecast flow. The forecast lead time is 1-3 days, and the time resolution is 3 hours.

[0058] Step S57: The three-layer forecast results obtained in steps S54 to S56 are fused using the three final converged index values ​​output in step S47 as weights. The fusion expression is as follows: ; in, This refers to the merged forecast flow.

[0059] The generation of tiered early warning information specifically includes: Step S58: Dynamically adjust the warning threshold for each cross-section based on the three final converged index values ​​output in step S47. Warning water level threshold. The expression is: ; in, The baseline warning water level is determined based on historical flood frequency analysis. =0.15 is the adjustment factor. This is the water level offset (taken as 0.5m).

[0060] Step S59: Simulate flood evolution using a two-dimensional hydrodynamic model, using the fused forecast flow rate output in step S57. As an upstream boundary condition, the digital elevation model data collected in step S11 As the terrain input, the hourly inundation range was obtained through simulation. Combined with the population distribution data collected in step S11 Building distribution data Road data Spatial overlay analysis is performed to calculate the number and extent of affected disaster-bearing bodies. The expression is as follows: ; ; in, For the number of people affected, and For the number of rows and columns of the population distribution grid, and For grid row and column numbers, The coordinates of the grid center are, Population density within the grid, For the length of the affected roads, The total number of road segments. For the first section of road, The length of the road segment. This is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

[0061] Step S510: Using the fuzzy comprehensive evaluation method combined with the three finally converged index values ​​output in step S47, determine the warning level. The expression is: ; in, The weighting coefficient for the rainfall-runoff response consistency index is set to 0.4. The weighting coefficient for the model structure suitability index is set to 0.3. The weighting coefficient for the forecast uncertainty index is set to 0.3.

[0062] Step S511: Generate early warning information including early warning location, early warning level, expected flood peak arrival time, expected inundation range calculated in step S59, number of affected people, number of affected buildings, length of affected roads, and model confidence level output in step S47, and publish it to relevant townships and administrative villages in the basin through SMS, broadcast, APP push, etc.

[0063] The invention will be further illustrated by simulation experiments below.

[0064] A typical heavy rainfall and flood event in the target watershed was selected as the experimental subject. This rainfall event lasted 12 hours, with a cumulative rainfall of 128 mm and a measured peak flow of 156 m³ / h. 3 / s. Comparative experiments on flood forecasting were conducted using the method of this invention, the traditional distributed unit line method, and a single LSTM neural network model.

[0065] Experimental data: Using the data acquisition scheme described in Example 1, data from 6 hours prior to the flood event was selected as input to forecast the flow rate over the next 3 hours. Training data consisted of 15 historical floods from the past 5 years in the basin, with 10 used for training and 5 for validation.

[0066] Evaluation metrics: Nash efficiency coefficient (NSE), relative peak error (EQP), peak time error (ETP), and forecast uncertainty index (U). fore ).

[0067] The experimental results are shown in Table 1.

[0068] Table 1. Test Results

[0069] The experimental results show that the method of this invention significantly outperforms traditional methods in terms of NSE, relative error of flood peak, and peak occurrence time error, while also significantly reducing forecast uncertainty. This is because the present invention achieves dynamic optimization of model structure and parameters through a dynamic adaptive evaluation index system and a reinforcement learning cyclic correction mechanism, enabling the model to better adapt to runoff generation and confluence characteristics under different hydrological conditions. Simultaneously, the embedding of physical constraints ensures the physical consistency of forecast results, thereby effectively improving forecast accuracy and reliability. Furthermore, regarding the timeliness of early warning, the present invention adopts a hierarchical progressive strategy, issuing warnings 3 hours before the predicted flood peak, approximately 1 hour earlier than traditional methods, providing a more ample time window for flood control decision-making.

[0070] To further demonstrate the forecasting effect of the method of the present invention, flow process lines of the three methods are plotted for comparison. Figure 3 As shown.

[0071] like Figure 3 As shown, the flow process curve predicted by the method of this invention has the highest degree of agreement with the measured flow process curve, with a peak flow error of only 8.6% and a peak occurrence time error of only 0.5 hours. In contrast, the predicted flow process curves of traditional distributed unit hydrograph methods and LSTM neural network methods deviate significantly from the measured flow process curves, especially near the peak flow time. This graph visually verifies the superiority of the method of this invention in fitting flood processes and provides visual support for the numerical comparison in Table 1 above.

[0072] To further verify the effectiveness of the cyclic correction mechanism of this invention, the following ablation test was conducted: Experimental Design: Two comparative experiments were set up. Group A was the complete method of this invention (including dynamic adaptive evaluation indicators and reinforcement learning cyclic correction mechanisms), and Group B was the ablation control group (only retaining the physical-guided neural network model and hierarchical progressive prediction, without including the cyclic correction mechanisms of steps S3 and S4, i.e., the model structure and parameters were fixed). Both groups used the same input data, training samples, and prediction time periods.

[0073] Experimental data: Five flood events with different rainfall intensities (light rain, moderate rain, heavy rain, rainstorm, and torrential rain) in the target watershed were selected for verification. Two sets of methods were used to forecast each flood event.

[0074] Evaluation metrics: Nash efficiency coefficient (NSE), relative peak error (EQP), peak time error (ETP), forecast uncertainty index (U). fore .

[0075] The experimental results are shown in Table 2.

[0076] Table 2 Test Results

[0077] Overall, Group A showed better forecast accuracy than Group B across all rainfall intensities, indicating that the cyclic correction mechanism significantly improved model performance.

[0078] As rainfall intensity increased, the accuracy of Group B forecasts decreased significantly, with the NSE dropping from 0.78 to 0.58, the relative error of the flood peak increasing from 10.5% to 28.6%, the peak occurrence time error increasing from 0.8 h to 2.1 h, and the forecast uncertainty increasing from 0.18 to 0.41. This indicates that the fixed model structure is difficult to adapt to changes in runoff generation and confluence characteristics under different hydrological conditions.

[0079] Group A maintained relatively stable forecast accuracy under different rainfall intensities, with the NSE consistently above 0.84, the relative error of the flood peak controlled within 12%, and the peak occurrence time error controlled within 1 hour. This is because the cyclic correction mechanism can dynamically adjust the model parameters and structure according to the current hydrological conditions, ensuring that the model always maintains good adaptability.

[0080] Under heavy rainfall conditions, Group A showed a 0.26-point improvement in NSE compared to Group B, a 17.3 percentage point reduction in the relative error of the flood peak, a 1.4-hour reduction in the peak occurrence time error, and a 0.20-point reduction in forecast uncertainty. These results demonstrate that the cyclic correction mechanism of this invention exhibits the most significant advantages under high-intensity rainfall conditions, effectively addressing the problem of a sharp decline in forecast accuracy under heavy rainfall conditions in existing technologies.

[0081] Taking the number of rainstorm events as an example, the changes in the three indicators of Group A after 5 rounds of iterative correction are shown in Table 3.

[0082] Table 3 Change Process Table

[0083] The iterative process shows that after five rounds of cyclical correction, the model structure suitability index improved from 0.52 to 0.71, the rainfall-runoff response consistency index improved from 0.58 to 0.67, the forecast uncertainty index decreased from 0.31 to 0.18, and the model NSE improved from 0.65 to 0.87. This result demonstrates that the cyclical correction mechanism can effectively drive the three evaluation indices to mutually verify and converge, achieving continuous optimization of model performance.

[0084] Example 2 A small watershed flood forecasting and early warning system includes: The data acquisition and preprocessing module is used to collect multi-source data from the study area and perform preprocessing. A physics-guided neural network model building module is used to embed hydrodynamic physical equations as constraints into the network training process. The dynamic adaptive evaluation index construction module is used to calculate the initial values ​​of the rainfall-runoff response consistency index, model structure suitability index, and forecast uncertainty index. The reinforcement learning agent construction module is used to start with three initial values ​​as the initial state, execute a loop correction mechanism, output correction actions to iteratively correct the three indicators in turn until the convergence condition is met, and obtain the final converged values ​​of the three indicators. The flood forecasting and early warning generation module is used to configure the parameters of the physical-guided neural network model based on the three final converged index values, execute flood forecasting using a hierarchical progressive strategy, and generate graded early warning information based on the flood forecasting results combined with dynamic disaster-bearing body information.

[0085] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0086] Therefore, this invention adopts the above-mentioned method and system for small watershed flood forecasting and early warning. Through the synergistic effect of dynamic adaptive evaluation index system and reinforcement learning cyclic correction mechanism, it realizes adaptive optimization and dynamic adjustment of forecast model parameters. Combined with physical-guided neural network model, it ensures the physical consistency of forecast results. Furthermore, based on refined disaster-bearing body assessment, it improves the pertinence and practicality of early warning. It effectively solves the technical problems of poor model adaptability, parameter calibration relying on historical data, and single early warning index in the prior art, and provides reliable technical support for the prevention of flash floods in small watersheds.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for forecasting and early warning of floods in small watersheds, characterized in that, Includes the following steps: Step S1: Collect multi-source data of the study area and perform preprocessing; Step S2: Construct a physical-guided neural network model. The physical-guided neural network model uses hydrodynamic physical equations as constraints to embed into the network training process. Step S3: Construct a dynamic adaptive evaluation index system and calculate the initial values ​​of the rainfall-runoff response consistency index, model structure applicability index, and forecast uncertainty index; Step S4: Construct a reinforcement learning agent. Starting with the three index values ​​calculated in step S3 as the initial state, execute a loop correction mechanism. The reinforcement learning agent outputs correction actions to iteratively correct the three indexes in turn until the convergence condition is met, and finally obtain the three converged index values. Step S5: Configure the parameters of the physical guided neural network model based on the three final converged index values, execute flood forecasting using a hierarchical progressive strategy, and generate graded early warning information based on the flood forecast results and dynamic disaster-bearing body information.

2. The method for small watershed flood forecasting and early warning according to claim 1, characterized in that, Step S1 specifically includes: Step S11: Collect measured rainfall data from rain gauge networks. Radar quantitative precipitation estimation data Digital elevation model data Soil moisture content data Actual flow data Population distribution data Building distribution data Road data ; Step S12: Perform spatiotemporal matching on the collected multi-source data and unify them to the same coordinate system and time reference. Step S13: Perform outlier removal and missing value imputation on the matched data; Step S14, based on and Construct a function to fuse rainfall data, with the following expression: ； in, For the merged rainfall data, The weighting coefficient for the rain gauge station. This represents the radar weighting coefficient.

3. The method for small watershed flood forecasting and early warning according to claim 2, characterized in that, The physical-guided neural network model in step S2 specifically includes: Step S21: Construct an input layer to receive fused rainfall data. and digital elevation model data ; Step S22: Construct a physical constraint embedding layer. The water balance equation and Saint-Venant equations are used as physical constraints in the network training process. A soft constraint approach is used to construct the loss function, expressed as: ； in, For the total loss function, For data-driven loss terms, For physical constraint loss terms, These are the physical constraint weighting coefficients; Step S23: Construct the attention mechanism layer, including a spatial attention sublayer and a temporal attention sublayer; Step S24: Construct the output layer to output the flood forecast flow. .

4. The method for small watershed flood forecasting and early warning according to claim 3, characterized in that, Step S3 specifically includes: Step S31: Calculate the rainfall-runoff response consistency index using the mutual information method. The expression is: ； in, For the merged rainfall sequence Compared with the measured flow sequence mutual information content This represents the total number of states in the rainfall sequence. The total number of states in the flow sequence. This is the sequence number of the rainfall status. This is the flow status sequence number. Rainfall With traffic status The joint probability distribution function, Rainfall The marginal probability distribution function, Traffic status The marginal probability distribution function; Step S32: Calculate the model structure suitability index using the cosine similarity method. The expression is: ； in, The applicable feature vectors for the physically guided neural network model constructed in step S2. The feature vector of the underlying surface of the watershed; Step S33: Calculate the forecast uncertainty index using the Monte Carlo dropout method. The expression is: ； in, Forecasting flow rates for floods At half the width of the 90% confidence interval, Forecasting flow rates for floods The mean.

5. The method for small watershed flood forecasting and early warning according to claim 4, characterized in that, The reinforcement learning agent in step S4 specifically includes: Step S41: Construct the state space, which consists of the rainfall-runoff response consistency index. Model structure applicability index and forecast uncertainty indicators The composition, the state vector is represented as superscript Indicates matrix transpose; Step S42: Construct the action space, which includes the first correction coefficient. Second correction factor and the third correction factor The action vector is represented as ; Step S43: Train the reinforcement learning agent using the deep deterministic policy gradient algorithm, which includes the Actor network and the Critic network. Step S44: Construct the reward function, the expression of which is: ； in, As a reward value, This is the amount of improvement in the Nash efficiency coefficient of the model before and after correction. This represents the current iteration step. Action vector L2 norm, , , This is the balance coefficient; Nash efficiency coefficient The expression is: ； in, This represents the average measured flow rate.

6. The method for small watershed flood forecasting and early warning according to claim 5, characterized in that, The cyclic correction mechanism in step S4 specifically includes: Step S45: Initial state value calculated in step S3 Input is fed into the trained reinforcement learning agent; Step S46: The reinforcement learning agent outputs a correction action. Perform cyclical corrections sequentially: First correction: using the first correction factor Correct the model structure suitability index output in step S32 and with the revised Adjusting the rainfall-runoff response consistency index in step S31 The calculation parameters are updated. The expression is: ； in, For the updated rainfall-runoff response consistency index, The previous rainfall-runoff response consistency index, As an indicator of the applicability of the current model structure, As a reference model structure applicability index; Second correction: using the second correction factor The updated As a priori for uncertainty quantification, the forecast uncertainty index in step S33 is corrected. The confidence interval boundaries are updated. The expression is: ； in, For the updated forecast uncertainty index, This refers to the uncertainty indicators of the forecast before the update; Third correction: using the third correction factor The updated As feedback, the model structure suitability index in step S32 is corrected in reverse. ,renew The expression is: ； in, For the updated model structure applicability index, The applicability index of the model structure before the update; Step S47: Update the status As the state input for the next iteration, repeat steps S46 to S47 until the convergence condition is met.

7. The method for small watershed flood forecasting and early warning according to claim 6, characterized in that, Step S5 involves configuring the parameters of the physics-guided neural network model based on the three final convergence metrics. This specifically includes: Step S51: Calculate the final converged rainfall-runoff response consistency index output in step S47. Model structure applicability index and forecast uncertainty indicators Parameter configuration vector ; Step S52: Adjust the physical constraint weight coefficients of the physical-guided neural network model in step S22 according to the parameter configuration vector. The expression is: ； in, These are the initial physical constraint weighting coefficients. The adjustment coefficient for the consistency index of rainfall-runoff response. This is an adjustment coefficient for the forecast uncertainty index; Step S53: Adjust the output weights and spatial attention weights of the attention mechanism layer in the physically guided neural network model in step S23 according to the parameter configuration vector. And time attention weight The expression is: ； ； in, and For learnable parameter matrix, Let be the dimension of the spatial attention weight vector. For dimension number, Let the dimension of the time attention weight vector be . For dimension number, For learnable parameter matrix The One portion, For learnable parameter matrix The Each component.

8. The method for forecasting and early warning of small watershed floods according to claim 7, characterized in that, Step S5, the hierarchical and progressive strategy for flood forecasting, specifically includes: Step S54: Perform short-term forecasting using the radar quantitative precipitation estimation data collected in step S11. Extrapolation, combined with the physical-guided neural network model constructed in step S2, is used for rolling forecasting to obtain short-term forecast flows. Forecast lead time is 0-3 hours; Step S55: Perform short-term forecasting using numerical weather prediction data. By combining the physical-guided neural network model constructed in step S2, short-term forecast flows can be obtained. Forecast lead time is 3-24 hours; Step S56: Execute medium-range forecasts using multi-model ensemble weather forecast data. By combining the physical-guided neural network model constructed in step S2, the medium-term forecast flow can be obtained. The forecast period is 1-3 days. Step S57: The three-layer forecast results obtained in steps S54 to S56 are fused using the three final converged index values ​​output in step S47 as weights. The fusion expression is as follows: ； in, This refers to the merged forecast flow.

9. The method for small watershed flood forecasting and early warning according to claim 1, characterized in that, Step S5, which generates tiered early warning information, specifically includes: Step S58: Dynamically adjust the warning threshold for each cross-section based on the three final converged index values ​​output in step S47. Warning water level threshold. The expression is: ； in, As the benchmark warning water level, To adjust the coefficient, This represents the water level offset. Step S59: Simulate flood evolution using a two-dimensional hydrodynamic model, using the fused forecast flow rate output in step S57. As an upstream boundary condition, the collected digital elevation model data As the terrain input, the hourly inundation range was obtained through simulation. Combined with the collected population distribution data Building distribution data Road data Spatial overlay analysis is performed to calculate the number and extent of affected disaster-bearing bodies. The expression is as follows: ； ； in, For the number of people affected, and For the number of rows and columns of the population distribution grid, and For grid row and column numbers, The coordinates of the grid center are Population density within the grid, For the length of the affected roads, The total number of road segments. For the first section of road, The length of the road segment This is an indicator function that takes the value 1 when the condition is true and 0 otherwise. Step S510: Using the fuzzy comprehensive evaluation method combined with the three finally converged index values ​​output in step S47, determine the warning level. The expression is: ； in, The weighting coefficients for the rainfall-runoff response consistency index. These are the weighting coefficients for the model structure suitability index. The weighting coefficients for forecast uncertainty indicators; Step S511: Generate early warning information including the warning location, warning level, expected peak arrival time, expected inundation range calculated in step S59, number of affected people, number of affected buildings, length of affected roads, and model confidence level output in step S47.

10. A small watershed flood forecasting and early warning system, characterized in that, A method for implementing a small watershed flood forecasting and early warning system as described in any one of claims 1-9 includes: The data acquisition and preprocessing module is used to collect multi-source data from the study area and perform preprocessing. A physics-guided neural network model building module is used to embed hydrodynamic physical equations as constraints into the network training process. The dynamic adaptability evaluation index construction module is used to calculate the initial values ​​of the rainfall-runoff response consistency index, model structure suitability index, and forecast uncertainty index. The reinforcement learning agent construction module is used to start with three initial values ​​as the initial state, execute a loop correction mechanism, output correction actions to iteratively correct the three indicators in turn until the convergence condition is met, and obtain the final converged values ​​of the three indicators. The flood forecasting and early warning generation module is used to configure the parameters of the physical-guided neural network model based on the three final converged index values, execute flood forecasting using a hierarchical progressive strategy, and generate graded early warning information based on the flood forecasting results combined with dynamic disaster-bearing body information.

Images