Deep learning flood routing prediction method coupling flood forecasting and flooding simulation

Abstract

The invention discloses a deep learning flood routing prediction method coupling flood forecasting and submerging simulation, which comprises the following steps: 1, acquiring underlying surface data and historical flood data of a research area, and determining a precision evaluation index; 2, setting an input time period and a forecast period, dividing the historical flood data into a deep learning flood forecast training set and a deep learning flood forecast verification set, and constructing a deep learning flood forecast model; 3, driving the two-dimensional hydrodynamic model to generate a historical submerging process data set; 4, taking the historical flood data as input, taking the historical submerging process data set as output, and training a deep learning submerging simulation model; 5, connecting the deep learning flood forecasting model and the deep learning flooding simulation model in series to form a deep learning flood flooding rapid prediction framework; 6, evaluating the precision of submerging prediction; and S7, analyzing efficiency improvement and an effective prediction period. According to the invention, a precious'rescue time 'window is provided for urban flood control, personnel evacuation and emergency material dispatching.

Description

Technical Field

[0001] This invention belongs to the field of natural disaster early warning, and specifically relates to a deep learning flood evolution prediction method that couples flood forecasting and inundation simulation. Background Technology

[0002] Floods are among the most destructive natural disasters globally, posing a serious threat to personal safety, the economy, and water resource management. Therefore, rapid and effective forecasting of their evolution is of great significance. Traditional two-dimensional hydrodynamic models, such as LISFLOOD-FP, can simulate the spatiotemporal evolution of flood extent and depth, but they primarily rely on real-time flood discharge monitoring data and have low computational efficiency, making them unsuitable for high-resolution rapid forecasting.

[0003] With the development of deep learning, on the one hand, deep learning models such as Convolutional Neural Networks (CNNs) can establish input-output mapping relationships by analyzing meteorological, hydrological, and historical flood data, enabling rapid inundation simulation. This is currently the main method in rapid flood inundation prediction research, but it primarily focuses on real-time simulation of the flood inundation process rather than advance forecasting with an effective lead time. On the other hand, deep learning models such as Recurrent Neural Networks (RNNs), such as Gated Recurrent Units (GRUs), can capture temporal evolution information from data, providing possibilities for flood forecasting. However, existing research has not yet seen any examples of combining the two to achieve effective forecasting of future flood inundation processes. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based flood evolution prediction method that couples flood forecasting and inundation simulation, thereby improving the forecast timeliness and computational efficiency of future flood inundation predictions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The deep learning-based flood evolution prediction method that couples flood forecasting and inundation simulation as described in this invention includes the following steps: Step S1: Obtain underlying surface data and historical flood data for the study area, and determine accuracy assessment indicators; Step S2: Set the input time period and forecast period, divide the historical flood data into a deep learning flood forecast training set and a deep learning flood forecast validation set, construct a deep learning flood forecast model (M1) using a gated recurrent unit (GRU) model, and verify the flood forecast accuracy; Step S3: Using underlying surface data and historical flood data of the study area, drive the two-dimensional hydrodynamic model (LISFLOOD-FP) to generate a dataset of historical inundation processes; Step S4: Take historical flood data as input and the historical inundation process dataset as output, divide the deep learning inundation simulation training set and the deep learning inundation simulation validation set, and train the deep learning inundation simulation model (M2). Step S5: Connect the deep learning flood forecasting model (M1) and the deep learning inundation simulation model (M2) in series to form a deep learning flood inundation rapid prediction framework; Step S6: Select sampling points and evaluate the accuracy of flood prediction based on the deep learning flood simulation validation set and the accuracy evaluation index; Step S7: Calculate the running time of the deep learning-based rapid flood inundation prediction framework, compare it with the traditional two-dimensional hydrodynamic model (LISFLOOD-FP), and analyze the efficiency improvement and effective prediction period.

[0006] Furthermore, in step S1, the specific steps of acquiring underlying surface data and historical flood data of the study area and determining accuracy evaluation indicators are as follows: S1.1: Collect the digital elevation model (DEM) of the study area and the upstream boundary conditions of the inflow channels corresponding to multiple flood events; S1.2: Preprocess the digital elevation model (DEM) using geographic information technology, including eliminating terrain noise, correcting the interference of bridges, trees and artificial / natural structures on the elevation data of the study area, and generating standardized digital elevation model (DEM) terrain data; S1.3: The selected accuracy evaluation metrics include, but are not limited to, root mean square error (RMSE) and Nash efficiency coefficient (NSE), used to evaluate the accuracy of flood forecasting and inundation prediction. The specific calculation formulas are as follows: in, yᵢ For predicted values, ŷᵢ For the observed values, ȳ The mean of the observed values, N This represents the sample size.

[0007] Furthermore, in step S2, verifying the accuracy of flood forecasts specifically involves: S2.1: Based on the research objectives and requirements, the input time period is set to a hour, and the forecast period is set to b hours. Then, the input of the deep learning flood forecasting model (M1) is the flow monitoring value of hour a, and the output is the flow prediction value of hour b; a and b are natural numbers greater than 0. S2.2: According to a set ratio, the historical flood data is divided into the deep learning flood forecast training set and the deep learning flood forecast validation set; S2.3: Construct the deep learning flood forecasting model (M1) using the gated recurrent unit (GRU) model. The specific calculation formula is as follows: in, and These are the outputs for updating the door and resetting the door, respectively. It is a candidate hidden state. It is the hidden state at the current time step; This represents the Sigmoid function, which restricts values ​​to between 0 and 1; It is a weight matrix; It is the hidden state of the previous time step; This is the input for the current time step; S2.4: Calculate the accuracy of flood forecast results using the deep learning flood forecast validation set to evaluate the prediction performance of the deep learning flood forecast model (M1) under different lead times.

[0008] Further, in step S4, training the deep learning flooding simulation model (M2) specifically involves: The deep learning flooding simulation model (M2) is constructed by cascading a Long Short-Term Memory (LSTM) module onto a Convolutional Neural Network (CNN) model. The specific calculation formula for this module is as follows: in, These represent the forget gate, input gate, and output gate, respectively. This is the unit state; It is in a hidden state; and These are the weight matrix and the bias term, respectively. It is the Sigmoid activation function. It is the hyperbolic tangent activation function.

[0009] Furthermore, in step S5, the construction of the deep learning-based rapid flood inundation prediction framework specifically includes: The deep learning flood forecasting model (M1) reads flood monitoring data to make flood predictions and provides the prediction results to the deep learning inundation simulation model (M2) for rapid flood inundation prediction.

[0010] Further, in step S6, the assessment of the accuracy of the flooding prediction specifically includes: During the validation process, several key control points, i.e. sampling points, within the study area were selected. The predicted water depth sequence and the actual water depth sequence of these points throughout the entire flood process were extracted. The root mean square error (RMSE) and Nash efficiency coefficient (NSE) were calculated to evaluate the predictive performance of the deep learning-based rapid flood inundation prediction framework.

[0011] This invention represents a leap from "near-term monitoring" to "early forecasting," extending the lead time for flood warnings. Existing technologies largely rely on real-time rainfall or water level monitoring for immediate inundation simulations, lacking the ability to predict future trends. This invention, by constructing a flood forecasting module based on a gated cyclic unit (GRU) model, can predict flow boundary conditions several hours (e.g., 4 hours) in advance and couple this prediction with a deep learning-based inundation simulation model. This deep learning-based alternative model enables the system to quickly generate inundation evolution maps before the flood arrives, providing a valuable "rescue time" window for urban flood control, evacuation, and emergency material dispatch. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0013] Figure 1 This is a flowchart of the deep learning flood evolution prediction method that couples flood forecasting and inundation simulation as described in an embodiment of the present invention.

[0014] Figure 2 This is a schematic diagram of the structure for training a deep learning flooding simulation model (CNN-LSTM) in an embodiment of the present invention.

[0015] Figure 3 This is the flood forecast verification result in the embodiment of the present invention.

[0016] Figure 4 This is a comparison chart of the predicted results and the actual flood at different times in the embodiments of the present invention.

[0017] Figure 5 This is a distribution diagram of key sampling control points in an embodiment of the present invention.

[0018] Figure 6 This is a diagram showing the accuracy evaluation results based on control points in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the specific embodiments described herein are only for explaining the invention and represent a part of the embodiments, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] Example: The deep learning-based flood evolution prediction method that couples flood forecasting and inundation simulation provided in this invention embodiment, such as... Figure 1 As shown, taking the flooding incident in Carlisle, UK as an example, the specific steps and verification results are as follows: Step S1: Obtain underlying surface data and historical flood data for the study area, and determine accuracy assessment indicators.

[0021] Specifically, Carlisle, UK, was selected as the study area, which is influenced by rivers such as the Eden and Caldew. 15-minute interval flow monitoring data from three upstream sections of the study area were collected, along with a 5-meter resolution digital elevation model (DEM). The DEM data was preprocessed using geographic information technology, and the effects of vegetation, bridges, and other noise and obstacles were manually delineated and removed to obtain topographic data for the two-dimensional hydrodynamic model (LISFLOOD-FP). Root mean square error (RMSE) and Nash efficiency coefficient (NSE) were selected as metrics to evaluate the accuracy of the deep learning-based rapid flood inundation prediction framework.

[0022] Historical flood data for the study area were collected, including seven general floods and one catastrophic flood with a return period of 170 years. To test the performance of the deep learning-based framework for rapid flood inundation prediction, the seven general floods were amplified to generate corresponding synthetic flood events, ultimately resulting in seven synthetic flood events and one real flood event.

[0023] Step S2: Set the input time period and forecast period, divide the historical flood data into a deep learning flood forecast training set and a deep learning flood forecast validation set, and use the gated recurrent unit (hereinafter referred to as: GRU) model to construct a deep learning flood forecast model (hereinafter referred to as: M1) to verify the flood forecast accuracy.

[0024] Beneficially or exemplaryly, the input time period is set to 2 hours and the forecast period is set to 4 hours. The flow monitoring data obtained in step S1 is divided into a deep learning flood forecast training set and a deep learning flood forecast validation set to construct M1, which contains 2 GRU layers.

[0025] By inputting historical flow data from a deep learning flood forecast validation set into M1, the flow rate changes for the next 1 to 4 hours are predicted, and the error between the predicted results and the actual flow rate is analyzed and calculated. Figure 3 As shown, the results indicate that the flow rates at upstream points 1, 2, and 3 all exhibited clear fluctuations throughout the event, but the characteristics varied at each point. The peak flow rate at upstream point 1 was relatively high and occurred later, approximately 40 hours after the flood, with a step-like increase during the rise. The peak flow rate at upstream point 2 was lower, with a more gradual increase, peaking approximately 30 hours after the flood. The flow rate at upstream point 3 fluctuated dramatically 20 hours prior, then rose sharply, reaching its peak within the next 5 hours. Consistent with these dynamic characteristics, the prediction error was mainly concentrated near several inflection points in the flow rate evolution, particularly in the region near the peak flow rate, such as... Figure 3 The area shown is a circular region indicated by the dashed line.

[0026] The RMSE and NSE values ​​calculated for upstream points 1, 2, and 3 show that prediction accuracy generally decreases with increasing lead time. At upstream point 1, the RMSE for a 1-hour lead time decreases from 41.04m. 3 / s increased to 111.76m for a 4-hour forecast period. 3 / s, while the NSE value decreased slightly from 0.987 to 0.983. The RMSE value at upstream point 2 was 3.45m. 3 / s to 6.76m 3 The NSE value fluctuated between 0.989 and 0.976, decreasing from 0.989 to 0.976. In contrast, the performance degradation at upstream point 3 was more significant: as the forecast duration increased from 1 hour to 4 hours, the RMSE decreased from 10.75 m / s. 3 / s increased to 33.83m 3 The NSE value dropped significantly from 0.997 to 0.801. These results indicate that M1 maintains high accuracy in short-term forecasts, while errors gradually accumulate in long-term forecasts. Overall, M1 successfully captures the characteristics of flow evolution over time during flood events, and its predicted hydrological curves largely retain the fluctuation patterns observed in the baseline data. Short-term forecasts (1-2 hours) are particularly strong, with NSE values ​​consistently above 0.97. When the forecast timeframe is extended to 3-4 hours, typical smoothing and hysteresis effects begin to appear, especially near peak flow, leading to increased errors, but the accuracy remains reliable.

[0027] Step S3: Use data from the study area to drive LISFLOOD-FP and generate a dataset of historical inundation processes: Using LISFLOOD-FP, input the flood process data and underlying surface data obtained in step S1 to simulate and generate the corresponding historical inundation process dataset, including 7 synthetic flood inundation process datasets and 1 real flood inundation process dataset, with a time step of 15 minutes.

[0028] Step S4: Take historical flood data as input and historical inundation process dataset as output, divide the deep learning inundation simulation training set and deep learning inundation simulation validation set, and train the deep learning inundation simulation model CNN-LSTM (hereinafter referred to as: M2).

[0029] Beneficially or exemplaryly, the eight flood process data obtained in step S1 are matched one-to-one with the eight historical inundation process datasets obtained in step S3, and the seven synthetic flood inundation process datasets are divided into a deep learning inundation simulation training set, while the real flood data are divided into a deep learning inundation simulation verification set.

[0030] The M2 constructed in this embodiment is based on a convolutional neural network (CNN) module (hereinafter referred to as CNN) connected in series with a long short-term memory network (LSTM) module (hereinafter referred to as LSTM). The structural diagram of M2 is shown below. Figure 2 As shown; where CNN ( Figure 2 One-dimensional convolutional layers are used to extract spatial features of terrain and water depth at each time step, while LSTM is used to capture the temporal dependence of flood evolution. LSTM units effectively solve the gradient vanishing problem in long-sequence training by introducing cell states and gating mechanisms (forget gate, input gate, output gate), thus enabling accurate prediction of flood inundation distribution at the next time step based on the current state. The key calculation formulas for LSTM are as follows: in, These represent the forget gate, input gate, and output gate, respectively. This is the unit state; It is in a hidden state; and These are the weight matrix and the bias term, respectively. It is the Sigmoid activation function. It is the hyperbolic tangent activation function.

[0031] The above coupled model was trained using the training dataset, and the parameters were optimized using the Adam optimizer to obtain M2.

[0032] Step S5: Connect M1 and M2 in series to form a deep learning-based rapid flood inundation prediction framework.

[0033] Beneficially or exemplaryly, M1 reads flood monitoring data to make flood predictions, and provides the prediction results to M2 for rapid flood inundation prediction, such as... Figure 4 As shown, the flood rise and peak periods are selected, and the predicted water depths for 1 hour, 2 hours, 3 hours, and 4 hours are compared with the actual water depths. Figure 4 It can be seen that the overall prediction effect is good. In terms of flood depth, the prediction results are close to the actual water depth at different stages of the flood. In terms of flood range, the prediction results for the first two hours are close to the actual water depth, while the prediction results for the 3rd and 4th hours are slightly insufficient, which is specifically reflected in the western river area.

[0034] Step S6: Select sampling points and evaluate the accuracy of flooding prediction based on the deep learning flooding simulation verification dataset and accuracy evaluation metrics.

[0035] Beneficially or exemplaryly, in order to systematically evaluate the prediction performance of a deep learning-based fast flood inundation prediction framework under different flood events, such as Figure 5 As shown, this embodiment selected 18 key control points (sampling points) within the study area. These points are distributed in major flood-prone areas (such as riverbanks and low-lying areas). By coupling M1 and M2, flood evolution predictions were performed for different forecast periods, and a preliminary accuracy assessment was conducted based on the flood depth sampled from the control points. The results are as follows. Figure 6 As shown. Overall, CNN-LSTM performed well across all forecast periods, with an average RMSE of 0.33m and an average NSE of 0.67. The accuracy of both metrics showed a decreasing trend with the extension of the forecast period, but the increase in RMSE was not significant, and NSE remained greater than 0.5, indicating that the deep learning-based rapid flood inundation prediction framework can still effectively capture the spatiotemporal evolution of floods.

[0036] Step S7: Statistically analyze the running time of the deep learning flood flood rapid prediction framework, compare it with LISFLOOD-FP, and analyze the efficiency improvement and effective prediction period.

[0037] Beneficially or exemplaryly, this embodiment takes approximately 80 minutes to complete a flood evolution simulation using the LISFLOOD-FP model on an NVIDIA GTX 3090 hardware platform. In contrast, on the same device, the deep learning-based M2 training takes approximately 1 minute and the prediction takes approximately 20 seconds; if M1 is further superimposed (training takes approximately 30 seconds and prediction takes approximately 10 seconds), the entire system can achieve an effective forecast of the flood evolution process over the next 4 hours within approximately 2 minutes.

[0038] These results demonstrate that the method of this invention, by cascading M1 and M2, can not only achieve flood evolution prediction with a 4-hour lead time, but also maintain high prediction accuracy even when there are prediction fluctuations in upstream boundary conditions. By rapidly and accurately capturing the flood evolution process, it helps to reduce the negative impact of floods on urban socio-economic and ecological environments, contributing to the construction of resilient cities and the promotion of sustainable development.

[0039] The above embodiments merely illustrate preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A deep learning-based flood evolution prediction method that couples flood forecasting and inundation simulation, characterized in that, Includes the following steps: Step S1: Obtain underlying surface data and historical flood data for the study area, and determine accuracy assessment indicators; Step S2: Set the input time period and forecast period, divide the historical flood data into a deep learning flood forecast training set and a deep learning flood forecast validation set, construct a deep learning flood forecast model using a gated recurrent unit model, and verify the flood forecast accuracy; Step S3: Using the underlying surface data and historical flood data of the study area, drive the two-dimensional hydrodynamic model to generate a dataset of historical inundation processes; Step S4: Take historical flood data as input and the historical inundation process dataset as output, divide the deep learning inundation simulation training set and the deep learning inundation simulation validation set, and train the deep learning inundation simulation model. Step S5: Connect the deep learning flood forecasting model and the deep learning inundation simulation model in series to form a deep learning flood inundation rapid prediction framework; Step S6: Select sampling points and evaluate the accuracy of flood prediction based on the deep learning flood simulation validation set and the accuracy evaluation index; Step S7: Calculate the running time of the deep learning flood inundation rapid prediction framework and compare it with the two-dimensional hydrodynamic model to analyze the efficiency improvement and effective prediction period.

2. The deep learning flood evolution prediction method coupled with flood forecasting and inundation simulation as described in claim 1, characterized in that, In step S1, determining the accuracy evaluation index includes the following steps: S1.1: Collect digital elevation models of the study area and upstream boundary conditions of the inflowing rivers corresponding to multiple flood events; S1.2: Preprocess the digital elevation model using geographic information technology to generate standardized digital elevation model terrain data; S1.3: The selected accuracy evaluation indicators, including root mean square error and Nash efficiency coefficient, are used to evaluate the accuracy of flood forecasting and inundation prediction. The specific calculation formulas are as follows: Where RMSE is the root mean square error and NSE is the Nash efficiency coefficient. yᵢ For predicted values, ŷᵢ For the observed values, ȳ The mean of the observed values, N This represents the sample size.

3. The deep learning flood evolution prediction method coupled with flood forecasting and inundation simulation as described in claim 1 or 2, characterized in that, In step S2, verifying the accuracy of flood forecasts specifically involves: S2.1: Based on the research objectives and requirements, the input time period is set to a hour, and the forecast period is set to b hours. Then, the input of the deep learning flood forecast model is the flow monitoring value of hour a, and the output is the flow prediction value of hour b; a and b are natural numbers greater than 0. S2.2: According to a set ratio, the historical flood data is divided into the deep learning flood forecast training set and the deep learning flood forecast validation set; S2.3: Construct the deep learning flood forecasting model using the gated recurrent unit model. The calculation formula is as follows: in, and These are the outputs for updating the door and resetting the door, respectively. It is a candidate hidden state. It is the hidden state at the current time step; This represents the Sigmoid function, which restricts values ​​to between 0 and 1; It is a weight matrix; It is the hidden state of the previous time step; This is the input for the current time step; S2.4: Calculate the accuracy of flood forecast results using the deep learning flood forecast validation set to evaluate the prediction performance of the deep learning flood forecast model under different lead times.

4. The deep learning flood evolution prediction method coupled with flood forecasting and inundation simulation according to claim 1 or 2, characterized in that, In step S4, the deep learning flooding simulation model is constructed by connecting a long short-term memory network module in series with the convolutional neural network model. The specific calculation formula for this module is as follows: in, These represent the forget gate, input gate, and output gate, respectively. This is the unit state; It is in a hidden state; and These are the weight matrix and the bias term, respectively. It is the Sigmoid activation function. It is the hyperbolic tangent activation function.

5. The deep learning flood evolution prediction method coupled with flood forecasting and inundation simulation according to claim 1 or 2, characterized in that, In step S5, the framework for rapid prediction of flood inundation using deep learning is as follows: The deep learning flood forecasting model reads flood monitoring data to make flood predictions and provides the prediction results to the deep learning inundation simulation model for rapid flood inundation prediction.

6. The deep learning flood evolution prediction method coupled with flood forecasting and inundation simulation according to claim 1 or 2, characterized in that, In step S6, the accuracy of the flooding prediction assessment is: Several key control points within the study area were selected, and the predicted and actual water depth sequences for these points throughout the entire flood process were extracted. The root mean square error and Nash efficiency coefficient were then calculated.

Images