Deep learning-based flood inundation prediction disposal method and system

Abstract

The invention provides a flood inundation prediction processing method and system based on deep learning, and the method comprises the steps: constructing a rainfall space-time simulation database and a facility joint scheduling scheme library based on the actually measured rainfall data of a target region and a facility historical scheduling sequence; constructing a hydrodynamic mechanism model based on the target area confluence characteristic model, inputting the hydrodynamic mechanism model into a rainfall space-time simulation database, and obtaining flood inundation space-time distribution data; and then, constructing a two-dimensional rainfall matrix, an initial flood state numerical matrix and a two-dimensional inundation water depth matrix, obtaining a plurality of samples through a translation window, and constructing a training data set for training to obtain a flood inundation prediction model so as to predict flood inundation distribution of the target area. And flood inundation treatment is carried out on the target area through the facility joint scheduling scheme library. According to the method, complex urban confluence features are captured, space-time hydraulic features are extracted from local to overall so as to fully consider the physical consistency of output results, and the reliability of prediction treatment results is improved.

Description

Technical Field

[0001] This invention relates to the field of flood prediction and management technology, and in particular to a method and system for predicting and managing flood inundation based on deep learning. Background Technology

[0002] With the accelerating pace of urbanization, the actual physical spatial characteristics of cities have undergone significant changes, and the process of rainwater production and drainage is becoming increasingly complex. Especially under extreme weather conditions, urban drainage systems and river and lake systems may operate beyond their capacity, leading to disasters such as urban flooding. Therefore, improving the reliability of flood prediction and mitigation results has become a technical problem that needs further research.

[0003] Currently, existing technologies mainly employ data-driven flood prediction and response methods. These methods rely on a large amount of historical sample data and automatically extract the mapping relationship between input and output data through neural networks to obtain flood prediction and response results. However, in the process of achieving flood prediction and response through data-driven methods, existing technologies do not consider physical constraints such as the physical spatial layout of cities or the physical laws of rainwater runoff. This results in a lack of physical consistency in the output results of existing technologies, leading to poor reliability of prediction and response results. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a flood inundation prediction and response method and system based on deep learning. This method captures complex urban confluence characteristics and extracts spatiotemporal hydraulic features from local to global perspectives to fully consider the physical consistency of the output results, thereby improving the reliability of the prediction and response outcomes.

[0005] To achieve the above objectives, embodiments of the present invention provide a flood inundation prediction and management method based on deep learning, comprising: constructing a rainfall spatiotemporal simulation database and a facility joint scheduling scheme library based on pre-acquired measured rainfall data of the target area and historical facility scheduling sequences; constructing a hydrodynamic mechanism model based on a pre-acquired confluence characteristic model of the target area, and inputting the rainfall spatiotemporal simulation database into the hydrodynamic mechanism model to obtain flood inundation spatiotemporal distribution data; constructing a two-dimensional rainfall matrix, an initial flood state numerical matrix, and a two-dimensional inundation depth matrix based on the rainfall spatiotemporal simulation database, the target area confluence characteristic model, and the flood inundation spatiotemporal distribution data; acquiring several samples through a preset translation window based on the two-dimensional rainfall matrix, the initial flood state numerical matrix, and the two-dimensional inundation depth matrix, and constructing a training dataset based on the several samples; obtaining a flood inundation prediction model based on the training dataset, the facility joint scheduling scheme library, and the preset deep learning model; predicting the flood inundation distribution of the target area through the flood inundation prediction model to obtain the flood inundation prediction result, and managing the flood inundation of the target area through the facility joint scheduling scheme library based on the flood inundation prediction result.

[0006] This invention proposes a deep learning-based method for predicting and managing flood inundation. It constructs a spatiotemporal simulation database of rainfall and a joint facility scheduling scheme library using measured rainfall data and historical facility scheduling sequences in the target area. Combined with a hydrodynamic mechanism model built using the confluence characteristic model of the target area, it obtains spatiotemporal distribution data of flood inundation, thereby capturing complex urban confluence characteristics and achieving deep modeling of local hydraulic connections to global spatiotemporal features. Then, through data matrix processing, the data is organically fused to obtain a training dataset, and a flood inundation prediction model is constructed to predict the flood inundation distribution of the target area. Finally, based on the flood inundation prediction results, combined with the joint facility scheduling scheme library, flood inundation management is implemented in the target area. Therefore, by capturing the urban confluence characteristics of the target area and combining the physical constraints of the hydrodynamic mechanism model, the physical consistency of the output results is fully considered, improving the reliability of the prediction and management results.

[0007] Furthermore, based on the pre-acquired measured rainfall data of the target area and historical facility scheduling sequences, a rainfall spatiotemporal simulation database and a facility joint scheduling scheme library are constructed, including: acquiring measured rainfall data of the target area based on several rain gauges; randomly scaling the measured rainfall data of the target area using a preset scaling factor to obtain rainfall data of different magnitudes; dividing the target area into several matrix grids and generating gradient vectors at the intersection points of the matrix grids using a preset noise generation algorithm; mapping the gradient vectors to a preset rainfall intensity range based on a preset time step and a preset mapping algorithm to obtain the spatial pattern of rainfall data; constructing a rainfall spatiotemporal simulation database based on rainfall data of different magnitudes, the spatial pattern of rainfall data, and a preset historical rainfall center; acquiring scheduling sequence data of the target area based on the rainfall spatiotemporal simulation database; and randomly scaling and combining the scheduling sequence data of the target area using a preset scaling factor based on the preset facility capacity range of the target area to obtain a facility joint scheduling scheme library.

[0008] In the above scheme, measured rainfall is scaled up by multiple orders of magnitude, and noise generation and gradient mapping techniques are combined to simulate rainfall scenarios with different spatial distributions, generating a rainfall data spatial pattern that is close to reality and covers extreme weather conditions. At the same time, historical scheduling data is scaled to form a diversified scheduling scheme library, enabling the model to take into account the impact of different engineering control strategies during the learning process. This effectively solves the problem of insufficient or unevenly distributed training data, and helps to improve the accuracy of subsequent deep learning model predictions and the reliability of prediction and handling results.

[0009] Furthermore, a hydrodynamic mechanism model is constructed based on the pre-acquired target area confluence characteristic model, and the rainfall spatiotemporal simulation database is input into the hydrodynamic mechanism model to obtain flood inundation spatiotemporal distribution data. This includes: acquiring the hydrological characteristic model, one-dimensional river channel characteristic model, two-dimensional surface characteristic model, and drainage network characteristic model of the target area to obtain the target area confluence characteristic model; obtaining the hydrodynamic mechanism model by coupling the confluence characteristic model; inputting the rainfall spatiotemporal simulation database into the hydrodynamic mechanism model, and simulating the flood evolution process of the target area through the hydrodynamic mechanism model to obtain flood inundation spatiotemporal distribution data.

[0010] The aforementioned scheme integrates multiple physical features, including hydrological, river channel, surface, and pipe network characteristics, to accurately depict the confluence, evolution, and interaction of rainwater in the surface and underground pipe networks. By coupling these physical features into a hydrodynamic mechanism model, and inputting a spatiotemporal simulation database of rainfall into this model, physically consistent spatiotemporal distribution data of flood inundation can be output, providing a reliable data foundation for subsequent deep learning model training. Therefore, by capturing the urban confluence characteristics of the target area and combining this with the physical constraints of the hydrodynamic mechanism model, while fully considering the physical consistency of the output results, the reliability of prediction and response outcomes can be improved.

[0011] Furthermore, based on the spatiotemporal simulation database of rainfall, the confluence characteristic model of the target area, and the spatiotemporal distribution data of flood inundation, a two-dimensional rainfall matrix, an initial flood state numerical matrix, and a two-dimensional inundation depth matrix are constructed. This includes: performing inverse distance interpolation on the spatiotemporal simulation database of rainfall using a preset interpolation algorithm to obtain a rainfall spatial numerical matrix, and converting the rainfall spatial numerical matrix into a two-dimensional rainfall matrix; acquiring real-time monitoring water level data of rivers, surfaces, and drainage networks based on the confluence characteristic model of the target area; constructing flood state data using the real-time monitoring water level data of rivers, surfaces, and drainage networks; and applying a preset interpolation algorithm to the flood state data. The data is inversely interpolated to obtain an initial flood state numerical matrix; the spatiotemporal distribution data of flood inundation is converted into a two-dimensional inundation depth matrix; the conversion of the spatiotemporal distribution data of flood inundation into a two-dimensional inundation depth matrix includes: obtaining several hydrodynamic grids based on the spatiotemporal distribution data of flood inundation; identifying the center points of several hydrodynamic grids based on a preset regular grid; if there are at least two center points of hydrodynamic grids in the preset regular grid, the hydrodynamic grid that meets the preset inundation depth requirement is selected as the target inundation depth grid; if there is one center point of hydrodynamic grid in the preset regular grid, the hydrodynamic grid is used as the target inundation depth grid; and a two-dimensional inundation depth matrix is ​​constructed based on the target inundation depth grid.

[0012] In the above scheme, inverse distance interpolation is used to transform the spatiotemporal simulation database of rainfall and flood status data into a spatially continuous two-dimensional rainfall matrix and an initial flood status numerical matrix. Furthermore, the unstructured grid inundation data output by the hydrodynamic model is transformed into a regular grid two-dimensional inundation depth matrix. By aligning the rainfall data, water level data, and inundation depth data on the same spatial structure and resolution, it is beneficial for the subsequent deep learning model to directly extract the response feature data from the data matrix. At the same time, it also ensures the physical consistency of the input and output data, which helps to improve the reliability of the prediction and treatment results.

[0013] Furthermore, based on the two-dimensional rainfall matrix, the initial flood state numerical matrix, and the two-dimensional inundation depth matrix, several samples are obtained through a preset translation window, and a training dataset is constructed based on these samples. This includes: scaling down the initial flood state numerical matrix based on a preset scaling factor and the two-dimensional rainfall matrix to obtain the target flood state numerical matrix; constructing a hydrodynamic simulation dataset based on the two-dimensional rainfall matrix, the target flood state numerical matrix, and the two-dimensional inundation depth matrix; traversing the hydrodynamic simulation dataset using a preset translation window to generate several samples; and obtaining the training dataset based on a preset scaling factor and the several samples.

[0014] In the above scheme, the initial flood state numerical matrix is ​​reduced to be aligned with the resolution of the two-dimensional rainfall matrix. Then, a large number of samples with local spatiotemporal correlations are extracted from the limited matrix data using a preset translation window, and a training dataset is divided to expand the scale of the training data. This helps the subsequent deep learning model to better learn the local patterns and spatiotemporal dependencies in the evolution of floods, and also enhances the prediction ability of the subsequent deep learning model for unknown rainfall and flood scenarios, which helps to improve the reliability of the prediction and response results.

[0015] Furthermore, based on the training dataset, the facility joint scheduling scheme library, and the preset deep learning model, a flood inundation prediction model is obtained, including: inputting the training dataset into the preset deep learning model; extracting features from the training dataset using preset convolution and deconvolution algorithms to obtain the initial rainfall level spatiotemporal features; performing time-series modeling on the initial rainfall level spatiotemporal features using a preset spatiotemporal feature extraction algorithm to obtain the target rainfall level spatiotemporal features; extracting features from the facility joint scheduling scheme library using a preset multivariate time-series feature extraction algorithm to obtain the engineering facility control features; restoring the target rainfall level spatiotemporal features and the engineering facility control features to obtain the initial flood inundation prediction result; and training the preset deep learning model using the initial flood inundation prediction result and the target rainfall level spatiotemporal features until the loss function of the preset deep learning model meets the preset training requirements to obtain the flood inundation prediction model.

[0016] In the above scheme, convolution and deconvolution operations are used to extract spatial features of rainfall and water level, and then temporal modeling methods are used to capture their dynamic evolution. Simultaneously, multivariate temporal features are extracted for facility scheduling sequences. Finally, the two types of features are fused and restored to form the flood prediction result, and a flood prediction model with a loss function that meets the preset training requirements is trained. Therefore, the constructed flood prediction model can not only learn the nonlinear relationship between rainfall, water level, and flood depth data, but also incorporate human intervention factors by combining a facility joint scheduling scheme library. This effectively captures the urban confluence characteristics of the target area, and by combining the physical constraints of the hydrodynamic mechanism model, fully considers the physical consistency of the output results, improving the reliability of the prediction and response results.

[0017] This invention also provides a deep learning-based flood inundation prediction and management system, comprising: a first data acquisition module, used to construct a rainfall spatiotemporal simulation database and a facility joint scheduling scheme library based on pre-acquired measured rainfall data of the target area and historical facility scheduling sequences; a second data acquisition module, used to construct a hydrodynamic mechanism model based on a pre-acquired target area confluence characteristic model, and input the rainfall spatiotemporal simulation database into the hydrodynamic mechanism model to obtain flood inundation spatiotemporal distribution data; and a data matrix construction module, used to construct a two-dimensional rainfall matrix and an initial... The system includes a flood state numerical matrix and a two-dimensional inundation depth matrix; a training dataset acquisition module, which acquires several samples through a preset translation window based on the two-dimensional rainfall matrix, the initial flood state numerical matrix, and the two-dimensional inundation depth matrix, and constructs a training dataset based on these samples; a model training module, which obtains a flood inundation prediction model based on the training dataset, a facility joint scheduling scheme library, and a preset deep learning model; and a prediction and treatment module, which predicts the flood inundation distribution of the target area using the flood inundation prediction model, obtains the flood inundation prediction results, and performs flood inundation treatment on the target area based on the flood inundation prediction results through the facility joint scheduling scheme library.

[0018] This invention proposes a deep learning-based flood inundation prediction and response system. It constructs a rainfall spatiotemporal simulation database and a joint facility scheduling scheme library using measured rainfall data and historical facility scheduling sequences in the target area. Combined with a hydrodynamic mechanism model built using the target area's confluence characteristic model, it obtains flood inundation spatiotemporal distribution data, thereby capturing complex urban confluence characteristics and achieving deep modeling of local hydraulic connections to global spatiotemporal features. Then, through data matrix processing, the data is organically fused to obtain a training dataset, and a flood inundation prediction model is constructed to predict the flood inundation distribution in the target area. Finally, based on the flood inundation prediction results and the joint facility scheduling scheme library, flood inundation response is implemented in the target area. Therefore, by capturing the urban confluence characteristics of the target area and combining the physical constraints of the hydrodynamic mechanism model, the system fully considers the physical consistency of the output results, improving the reliability of the prediction and response results.

[0019] Furthermore, the second data acquisition module is used to construct a hydrodynamic mechanism model based on the pre-acquired target area confluence characteristic model, and input the rainfall spatiotemporal simulation database into the hydrodynamic mechanism model to obtain flood inundation spatiotemporal distribution data. This includes: a confluence characteristic model acquisition unit, used to acquire the hydrological characteristic model, one-dimensional river channel characteristic model, two-dimensional surface characteristic model, and drainage network characteristic model of the target area to obtain the target area confluence characteristic model; a hydrodynamic mechanism model acquisition unit, used to obtain the hydrodynamic mechanism model by coupling the confluence characteristic model; and a flood evolution simulation unit, used to input the rainfall spatiotemporal simulation database into the hydrodynamic mechanism model, and simulate the flood evolution process of the target area through the hydrodynamic mechanism model to obtain flood inundation spatiotemporal distribution data.

[0020] Furthermore, the data matrix construction module is used to construct a two-dimensional rainfall matrix, an initial flood state numerical matrix, and a two-dimensional inundation depth matrix based on the rainfall spatiotemporal simulation database, the target area confluence characteristic model, and flood inundation spatiotemporal distribution data. This includes: a first matrix construction unit, used to perform inverse distance interpolation on the rainfall spatiotemporal simulation database using a preset interpolation algorithm to obtain a rainfall spatial numerical matrix, and then converting the rainfall spatial numerical matrix into a two-dimensional rainfall matrix; a real-time data acquisition unit, used to acquire real-time river monitoring water level data, real-time surface monitoring water level data, and real-time drainage network monitoring water level data based on the target area confluence characteristic model; a flood state data acquisition unit, used to construct flood state data using the real-time river monitoring water level data, real-time surface monitoring water level data, and real-time drainage network monitoring water level data; a second matrix construction unit, used to perform inverse distance interpolation on the flood state data using a preset interpolation algorithm to obtain an initial flood state numerical matrix; and a third matrix construction unit, used to convert the flood inundation spatiotemporal distribution data into a two-dimensional inundation depth matrix. Attached Figure Description

[0021] Figure 1 A flowchart illustrating the steps of a deep learning-based flood inundation prediction and mitigation method according to a certain embodiment of the present invention; Figure 2 A schematic diagram of a two-dimensional rainfall matrix of a target area for a flood inundation prediction and control method based on deep learning, provided in a certain embodiment of the present invention; Figure 3 A schematic diagram of the initial flood state numerical matrix of the target area for a flood inundation prediction and treatment method based on deep learning provided in a certain embodiment of the present invention; Figure 4 A schematic diagram of the two-dimensional inundation depth matrix transformation process of a flood inundation prediction and treatment method based on deep learning provided in a certain embodiment of the present invention; Figure 5 A schematic diagram of the deep learning network topology for a flood inundation prediction and disposal method based on deep learning, provided in a certain embodiment of the present invention; Figure 6 A schematic diagram of the CLNet model calculation process for a flood inundation prediction and disposal method based on deep learning provided in a certain embodiment of the present invention; Figure 7 This is a schematic diagram of the module structure of a flood inundation prediction and control system based on deep learning, provided in one embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1 See Figure 1 , Figure 1 This is a flowchart illustrating the steps of a deep learning-based flood inundation prediction and mitigation method according to a certain embodiment of the present invention. Figure 1 As shown in the figure, this invention proposes a flood inundation prediction and management method based on deep learning, including steps 101 to 106, each step of which is as follows: Step 101: Based on the pre-acquired measured rainfall data of the target area and the historical scheduling sequence of facilities, construct a rainfall spatiotemporal simulation database and a joint scheduling scheme library for facilities; Step 102: Construct a hydrodynamic mechanism model based on the pre-acquired target area confluence feature model, and input the rainfall spatiotemporal simulation database into the hydrodynamic mechanism model to obtain the spatiotemporal distribution data of flood inundation; Step 103: Based on the spatiotemporal simulation database of rainfall, the confluence characteristic model of the target area, and the spatiotemporal distribution data of flood inundation, construct a two-dimensional rainfall matrix, an initial flood state numerical matrix, and a two-dimensional inundation depth matrix; Step 104: Based on the two-dimensional rainfall matrix, the initial flood state numerical matrix, and the two-dimensional inundation depth matrix, a number of samples are obtained through a preset translation window, and a training dataset is constructed based on the samples. Step 105: Based on the training dataset, the facility joint scheduling scheme library, and the preset deep learning model, a flood inundation prediction model is obtained; Step 106: Predict the flood inundation distribution of the target area using the flood inundation prediction model, obtain the flood inundation prediction results, and based on the flood inundation prediction results, carry out flood inundation treatment of the target area through the facility joint scheduling scheme library.

[0024] One possible implementation involves real-time measurement of rainfall data within the target area. A spatiotemporal simulation database of rainfall is obtained by introducing proportional scaling and Berlin noise processing. Then, a joint facility scheduling scheme library is obtained by proportional scaling and random combination from the historical facility scheduling sequences of the target area. Next, a target area confluence characteristic model is acquired to construct a hydrodynamic mechanism model. The rainfall spatiotemporal simulation database is input into the hydrodynamic mechanism model to simulate the flood evolution process, obtaining spatiotemporal distribution data of flood inundation. In this embodiment, the target area confluence characteristic model includes: a hydrological characteristic model, a one-dimensional river channel characteristic model, a two-dimensional surface characteristic model, and a drainage network characteristic model. Next, the spatiotemporal simulation database of rainfall, the confluence feature model of the target area, and the spatiotemporal distribution data of flood inundation are characterized and matrixed to obtain a two-dimensional rainfall matrix, an initial flood state numerical matrix, and a two-dimensional inundation depth matrix. At this point, the training data for the deep learning model is ready. Next, several samples are obtained from the two-dimensional rainfall matrix, the initial flood state numerical matrix, and the two-dimensional inundation depth matrix through a preset translation window, and these samples are divided into several groups according to a preset ratio to obtain the training dataset. In this embodiment, the size of the preset translation window can be set according to actual conditions, and the preset ratio is similarly adjusted. Finally, the preset deep learning model is trained using the training dataset and the facility joint scheduling scheme library, allowing the preset deep learning model to fully learn the feature data in the training dataset and the facility joint scheduling scheme library, resulting in a flood inundation prediction model. This model is then used to predict the flood inundation distribution of the target area. Based on the flood inundation distribution prediction results, the corresponding facility joint scheduling scheme is called from the facility joint scheduling scheme library to handle the flood inundation in the target area.

[0025] This invention proposes a deep learning-based method for predicting and managing flood inundation. It constructs a spatiotemporal simulation database of rainfall and a joint facility scheduling scheme library using measured rainfall data and historical facility scheduling sequences in the target area. Combined with a hydrodynamic mechanism model built using the confluence characteristic model of the target area, it obtains spatiotemporal distribution data of flood inundation, thereby capturing complex urban confluence characteristics and achieving deep modeling of local hydraulic connections to global spatiotemporal features. Then, through data matrix processing, the data is organically fused to obtain a training dataset, and a flood inundation prediction model is constructed to predict the flood inundation distribution of the target area. Finally, based on the flood inundation prediction results, combined with the joint facility scheduling scheme library, flood inundation management is implemented in the target area. Therefore, by capturing the urban confluence characteristics of the target area and combining the physical constraints of the hydrodynamic mechanism model, the physical consistency of the output results is fully considered, improving the reliability of the prediction and management results.

[0026] A preferred scheme involves constructing a rainfall spatiotemporal simulation database and a facility joint scheduling scheme library based on pre-acquired measured rainfall data of the target area and historical facility scheduling sequences. The scheme includes: acquiring measured rainfall data of the target area from several rain gauges; randomly scaling the measured rainfall data of the target area using a preset scaling factor to obtain rainfall data of different magnitudes; dividing the target area into several matrix grids and generating gradient vectors at the intersections of the matrix grids using a preset noise generation algorithm; mapping the gradient vectors to a preset rainfall intensity range based on a preset time step and a preset mapping algorithm to obtain the spatial pattern of the rainfall data; constructing a rainfall spatiotemporal simulation database based on rainfall data of different magnitudes, the spatial pattern of the rainfall data, and a preset historical rainfall center; acquiring scheduling sequence data of the target area based on the rainfall spatiotemporal simulation database; and randomly scaling and combining the scheduling sequence data of the target area using a preset scaling factor based on the preset facility capacity range of the target area to obtain a facility joint scheduling scheme library.

[0027] One preferred implementation involves collecting measured rainfall data from 24 large-scale rainfall stations in the target area from April 2021 to August 2025, and multiplying the measured rainfall data of each rainfall station by a random scaling factor within the range of [0.5, 3] to generate a series of rainfall data with historically known rainfall centers but different magnitudes, i.e., rainfall data of different magnitudes. Then, the target area is divided into several rectangular grids, and the gradient vector in the interval [0, 1] at the intersection of each grid is generated using the Berlin noise method. At each time step, the gradient vector is repeatedly mapped to the preset rainfall intensity range through existing mature interpolation, scaling and offset operations to obtain the spatial pattern of rainfall data. Finally, a spatiotemporal simulation database of rainfall is constructed by combining rainfall data of different magnitudes, the spatial pattern of rainfall data and historical rainfall centers. In this embodiment, the preset rainfall intensity range can be set according to the actual situation of the target area. Then, the scheduling sequence data corresponding to the rainfall events with large rainfall in the target area are obtained from the spatiotemporal simulation database of rainfall. Each scheduling sequence data is multiplied by a random scaling factor in the interval [0.5, 3], and the upper or lower limit of the facility capacity is taken for the data that exceeds the control capacity range after scaling. Finally, a joint facility scheduling scheme library is obtained by random combination.

[0028] In the above scheme, measured rainfall is scaled up by multiple orders of magnitude, and noise generation and gradient mapping techniques are combined to simulate rainfall scenarios with different spatial distributions, generating a rainfall data spatial pattern that is close to reality and covers extreme weather conditions. At the same time, historical scheduling data is scaled to form a diversified scheduling scheme library, enabling the model to take into account the impact of different engineering control strategies during the learning process. This effectively solves the problem of insufficient or unevenly distributed training data, and helps to improve the accuracy of subsequent deep learning model predictions and the reliability of prediction and handling results.

[0029] A preferred approach involves constructing a hydrodynamic mechanism model based on a pre-acquired target area confluence characteristic model, and inputting a rainfall spatiotemporal simulation database into the hydrodynamic mechanism model to obtain flood inundation spatiotemporal distribution data. This includes: acquiring a hydrological characteristic model, a one-dimensional river channel characteristic model, a two-dimensional surface characteristic model, and a drainage network characteristic model of the target area to obtain a target area confluence characteristic model; obtaining a hydrodynamic mechanism model by coupling the confluence characteristic model; inputting the rainfall spatiotemporal simulation database into the hydrodynamic mechanism model, and simulating the flood evolution process of the target area through the hydrodynamic mechanism model to obtain flood inundation spatiotemporal distribution data.

[0030] One preferred implementation involves acquiring a confluence characteristic model of the target area to construct a hydrodynamic mechanism model, inputting a rainfall spatiotemporal simulation database into the hydrodynamic mechanism model to simulate the flood evolution process, and obtaining spatiotemporal distribution data of flood inundation. Specifically, this involves acquiring a hydrological characteristic model, a one-dimensional river channel characteristic model, a two-dimensional surface characteristic model, and a drainage network characteristic model of the target area. In this embodiment, the hydrological characteristic model is used to simulate the flow process from the non-built-up area sub-basin to the main stream outlet; the one-dimensional river channel characteristic model is used to simulate the river flood propagation process based on the river topology and hydraulic structure parameters of the target area; the two-dimensional surface characteristic model is used to simulate the evolution process of urban surface floods on an irregularly dense grid; and the drainage network characteristic model is used to simulate the underground drainage process based on the network topology, cross-section, pumping stations, and scheduling procedures of the target area. The specific process of coupling the confluence characteristic model is as follows: using the hydrological characteristic model as a reference, the one-dimensional river channel characteristic model and the two-dimensional surface characteristic model are coupled... The surface feature model is coupled with the embankment boundary conditions to simulate the propagation process of river floods on the surface after breaching or overflowing the embankment. The one-dimensional river feature model and the drainage network feature model are coupled at the drainage outlet to simulate the backflow or flooding effect of the river water level on the drainage network. The two-dimensional surface feature model and the drainage network feature model are coupled through rainwater inlets and surface units to simulate the interaction between drainage network and surface overflow. Thus, the coupled confluence feature model yields the hydrodynamic mechanism model. Finally, the spatiotemporal simulation database of rainfall is input into the hydrodynamic mechanism model, and measured water level and flow data and urban waterlogging reporting data are used as comparative data to simulate the flood evolution process in the target area and obtain the spatiotemporal distribution data of flood inundation.

[0031] The aforementioned scheme integrates multiple physical features, including hydrological, river channel, surface, and pipe network characteristics, to accurately depict the confluence, evolution, and interaction of rainwater in the surface and underground pipe networks. By coupling these physical features into a hydrodynamic mechanism model, and inputting a spatiotemporal simulation database of rainfall into this model, physically consistent spatiotemporal distribution data of flood inundation can be output, providing a reliable data foundation for subsequent deep learning model training. Therefore, by capturing the urban confluence characteristics of the target area and combining this with the physical constraints of the hydrodynamic mechanism model, while fully considering the physical consistency of the output results, the reliability of prediction and response outcomes can be improved.

[0032] A preferred embodiment, based on a spatiotemporal rainfall simulation database, a target area confluence characteristic model, and flood inundation spatiotemporal distribution data, constructs a two-dimensional rainfall matrix, an initial flood state numerical matrix, and a two-dimensional inundation depth matrix. This includes: performing inverse distance interpolation on the spatiotemporal rainfall simulation database using a preset interpolation algorithm to obtain a rainfall spatial numerical matrix, and converting the rainfall spatial numerical matrix into a two-dimensional rainfall matrix; acquiring real-time river channel water level data, real-time surface water level data, and real-time drainage network water level data based on the target area confluence characteristic model; constructing flood state data using the real-time river channel water level data, real-time surface water level data, and real-time drainage network water level data; and applying a preset interpolation algorithm to the flood state data. The initial flood state numerical matrix is ​​obtained by inverse distance interpolation of the flood state data; the spatiotemporal distribution data of flood inundation is converted into a two-dimensional inundation depth matrix; the conversion of the spatiotemporal distribution data of flood inundation into a two-dimensional inundation depth matrix includes: obtaining several hydrodynamic grids based on the spatiotemporal distribution data of flood inundation; identifying the center points of several hydrodynamic grids based on a preset regular grid; if there are at least two center points of hydrodynamic grids in the preset regular grid, the hydrodynamic grid that meets the preset inundation depth requirement is selected as the target inundation depth grid; if there is one center point of hydrodynamic grid in the preset regular grid, the hydrodynamic grid is used as the target inundation depth grid; and a two-dimensional inundation depth matrix is ​​constructed based on the target inundation depth grid.

[0033] One preferred implementation method is described in [reference]. Figure 2 , Figure 2 This is a schematic diagram of a two-dimensional rainfall matrix of a target area for a deep learning-based flood inundation prediction and mitigation method according to a certain embodiment of the present invention; as shown below. Figure 2 As shown, time-series rainfall data from various rain gauge stations were obtained from a spatiotemporal rainfall simulation database. The inverse distance weighting (IDW) interpolation algorithm was used to transform the time-series rainfall data from each rain gauge station into a two-dimensional rainfall matrix with a resolution close to the hydrodynamic grid in the spatiotemporal distribution data of flood inundation. The specific formula is as follows: In the formula, Let be the spatial numerical matrix of rainfall at time t. Let W1 be the rainfall at time t for the nth rain gauge station, and H1 be the width and height of the rainfall value matrix, respectively. For example... Figure 2 As shown, the two-dimensional rainfall matrix represents the spatiotemporal distribution characteristics of rainfall in the target area. The x index represents the width index of the rainfall numerical matrix, and the y index represents the height index of the rainfall numerical matrix. The two-dimensional rainfall matrix enables subsequent deep learning models to capture the hydraulic connections and flood volume differences between grid catchment units, thereby improving the accuracy of urban surface runoff simulation.

[0034] See Figure 3 , Figure 3 This is a schematic diagram of the initial flood state numerical matrix of a target area for a flood inundation prediction and control method based on deep learning, provided in a certain embodiment of the present invention; as shown. Figure 3 As shown, real-time monitoring water levels are obtained from key river sections, pipeline nodes, and lakes / reservoirs within the target area using a target area confluence characteristic model, and these are used to construct flood status data. Similarly, an inverse distance weighted interpolation algorithm is employed to generate an initial flood status numerical matrix, the specific formula of which is as follows: ; In the formula, Let be the initial state-space numerical matrix of the flood at time t. Let be the water level value at time t for the m-th water level station, and let W2 and H2 be the width and height of the numerical matrix, respectively. It is a positive integer. For example... Figure 3 As shown, the initial flood state numerical matrix represents the spatiotemporal distribution characteristics of water depth in the target area. The x index represents the width index of the rainfall numerical matrix, and the y index represents the height index of the rainfall numerical matrix. The initial flood state numerical matrix can reflect the overall water storage capacity of the study area.

[0035] See Figure 4 , Figure 4 A schematic diagram illustrating the two-dimensional inundation depth matrix transformation process of a deep learning-based flood inundation prediction and mitigation method according to a certain embodiment of the present invention; as shown. Figure 4 As shown, due to the irregular grid pattern in the spatiotemporal distribution data of flood inundation, it is necessary to convert the data into a regular two-dimensional inundation depth matrix. Specifically, this involves obtaining several hydrodynamic grids from the spatiotemporal distribution data of flood inundation and identifying the center point of each grid. Figure 4 The elements a, b, c, d, e, f, and g are used. When multiple hydrodynamic grids overlap with a preset regular grid, the maximum inundation depth of the hydrodynamic grid whose center point is located within the preset regular grid is taken as the inundation depth value of that element, serving as the target inundation depth grid. After all hydrodynamic grids have completed the regular grid transformation, a two-dimensional inundation depth matrix is ​​obtained. In this embodiment, the preset regular grid is consistent with the grid of the deep learning model; in this embodiment, the Hydro-Y-Net grid is used for interpretation.

[0036] In the above scheme, inverse distance interpolation is used to transform the spatiotemporal simulation database of rainfall and flood status data into a spatially continuous two-dimensional rainfall matrix and an initial flood status numerical matrix. Furthermore, the unstructured grid inundation data output by the hydrodynamic model is transformed into a regular grid two-dimensional inundation depth matrix. By aligning the rainfall data, water level data, and inundation depth data on the same spatial structure and resolution, it is beneficial for the subsequent deep learning model to directly extract the response feature data from the data matrix. At the same time, it also ensures the physical consistency of the input and output data, which helps to improve the reliability of the prediction and treatment results.

[0037] A preferred approach involves acquiring several samples through a preset translation window based on a two-dimensional rainfall matrix, an initial flood state numerical matrix, and a two-dimensional inundation depth matrix, and constructing a training dataset based on these samples. This includes: scaling down the initial flood state numerical matrix using a preset scaling factor and the two-dimensional rainfall matrix to obtain a target flood state numerical matrix; constructing a hydrodynamic simulation dataset based on the two-dimensional rainfall matrix, the target flood state numerical matrix, and the two-dimensional inundation depth matrix; traversing the hydrodynamic simulation dataset using a preset translation window to generate several samples; and acquiring a training dataset based on a preset scaling factor and the samples.

[0038] In one preferred implementation, before constructing the hydrodynamic simulation dataset, the resolution of the initial flood state numerical matrix is ​​first reduced proportionally to the rainfall numerical matrix to obtain the target flood state numerical matrix. This reflects the consistency of the initial flood state within the coverage area of ​​the same drainage facility and facilitates the subsequent fusion calculation of different features by the deep learning model. Then, the two-dimensional rainfall matrix, the target flood state numerical matrix, and the two-dimensional inundation depth matrix are used as the hydrodynamic simulation dataset. Multiple samples are generated from the hydrodynamic simulation dataset using a preset translation window. In this embodiment, the preset translation window can be translated hourly. Finally, several samples are randomly divided into a training set, a validation set, and a test set according to a preset ratio to obtain the training dataset.

[0039] In the above scheme, the initial flood state numerical matrix is ​​reduced to be aligned with the resolution of the two-dimensional rainfall matrix. Then, a large number of samples with local spatiotemporal correlations are extracted from the limited matrix data using a preset translation window, and a training dataset is divided to expand the scale of the training data. This helps the subsequent deep learning model to better learn the local patterns and spatiotemporal dependencies in the evolution of floods, and also enhances the prediction ability of the subsequent deep learning model for unknown rainfall and flood scenarios, which helps to improve the reliability of the prediction and response results.

[0040] A preferred approach involves obtaining a flood inundation prediction model based on a training dataset, a facility joint scheduling scheme library, and a pre-defined deep learning model. The process includes: inputting the training dataset into the pre-defined deep learning model; extracting features from the training dataset using pre-defined convolution and deconvolution algorithms to obtain the initial rainfall level spatiotemporal features; performing time-series modeling on the initial rainfall level spatiotemporal features using a pre-defined spatiotemporal feature extraction algorithm to obtain the target rainfall level spatiotemporal features; extracting features from the facility joint scheduling scheme library using a pre-defined multivariate time-series feature extraction algorithm to obtain engineering facility control features; performing feature restoration on the target rainfall level spatiotemporal features and engineering facility control features to obtain the initial flood inundation prediction result; and training the pre-defined deep learning model using the initial flood inundation prediction result and the target rainfall level spatiotemporal features until the loss function of the pre-defined deep learning model meets pre-defined training requirements, thereby obtaining the flood inundation prediction model.

[0041] One preferred implementation method is described in [reference]. Figure 5 , Figure 5 A schematic diagram of the deep learning network topology for a flood inundation prediction and control method based on deep learning, provided in a certain embodiment of the present invention; as shown. Figure 5 As shown, the training dataset is input into a preset deep learning model. In this embodiment, the preset deep learning model adopts an encoder-decoder network architecture, including an encoder, a decoder, and model training. Its network topology is as follows. Figure 5 As shown, firstly, a combination of preset convolution and deconvolution algorithms is used to scale and expand the spatiotemporal data of rainfall and water level in the training dataset. Then, channel normalization and the LeakyReLU activation function are applied to improve the model's nonlinear expressiveness and generalization ability, resulting in initial spatiotemporal features of rainfall and water level. In this embodiment, the spatiotemporal data of rainfall is denoted as... The spatiotemporal data of water level are recorded as follows: The preset convolution algorithm and preset deconvolution algorithm are explained using convolutional neural networks (CNN) and deconvolutional neural networks (DCNN); then, the preset spatiotemporal feature extraction algorithm is used to perform time series modeling of the spatiotemporal features of the initial rainfall water level, extract global hydrodynamic feature associations, and obtain the spatiotemporal features of the target rainfall water level. The preset spatiotemporal feature extraction algorithm is explained using convolutional long short-term memory network (ConvLSTM). See Figure 6 , Figure 6 This is a schematic diagram of the CLNet model calculation process for a flood inundation prediction and control method based on deep learning, provided in one embodiment of the present invention; as shown below. Figure 6As shown, facility regulation multivariate time series data, denoted as D, is obtained from the facility joint scheduling scheme library. Then, a preset multivariate time series feature extraction algorithm is used to extract the deep features of the facility regulation multivariate time series data tensor to obtain the engineering facility regulation features. In this embodiment, the preset spatiotemporal feature extraction algorithm is explained using a CLNet model that combines longitudinal convolution of the time axis, linear merging of feature maps, horizontal convolution, fully connected layers, and reconstruction (reshape) process steps.

[0042] Finally, the spatiotemporal characteristics of the target rainfall level and the control characteristics of engineering facilities are used to restore the features, resulting in the initial flood inundation prediction. Then, the mean square error (MSE) of the initial flood inundation prediction and the mean square error (MSE) of the measured flood inundation are calculated, and a loss function is constructed using the same weighting coefficients. The specific expression is as follows: ; In the formula, , The weights for the two loss components are 0.5 for each; n is the number of matrix grids. This represents the predicted inundation depth for the i-th grid. B represents the submerged water depth value of the i-th grid calculated by the hydrodynamic model; B is the number of water level monitoring sections. This represents the predicted water level at the q-th cross-section. The value of the water level at the q-th cross section is calculated using hydrodynamics.

[0043] In this embodiment, the measured flood inundation result is composed of time-series rainfall data from each rain gauge obtained from the rainfall spatiotemporal simulation database and real-time monitoring water levels obtained from key river sections, pipeline nodes, and lakes and reservoirs within the target area based on the target area confluence characteristic model.

[0044] In the above scheme, convolution and deconvolution operations are used to extract spatial features of rainfall and water level, and then temporal modeling methods are used to capture their dynamic evolution. Simultaneously, multivariate temporal features are extracted for facility scheduling sequences. Finally, the two types of features are fused and restored to form the flood prediction result, and a flood prediction model with a loss function that meets the preset training requirements is trained. Therefore, the constructed flood prediction model can not only learn the nonlinear relationship between rainfall, water level, and flood depth data, but also incorporate human intervention factors by combining a facility joint scheduling scheme library. This effectively captures the urban confluence characteristics of the target area, and by combining the physical constraints of the hydrodynamic mechanism model, fully considers the physical consistency of the output results, improving the reliability of the prediction and response results.

[0045] Example 2 See Figure 7 , Figure 7This is a schematic diagram of the module structure of a flood inundation prediction and control system based on deep learning, provided in one embodiment of the present invention. Figure 7 As shown, this embodiment of the invention also provides a flood inundation prediction and response system based on deep learning, comprising: a first data acquisition module 201, used to construct a rainfall spatiotemporal simulation database and a facility joint scheduling scheme library based on pre-acquired measured rainfall data of the target area and historical facility scheduling sequences; a second data acquisition module 202, used to construct a hydrodynamic mechanism model based on a pre-acquired target area confluence characteristic model, and input the rainfall spatiotemporal simulation database into the hydrodynamic mechanism model to obtain flood inundation spatiotemporal distribution data; and a data matrix construction module 203, used to construct a two-dimensional rainfall matrix and an initial... The system includes: an initial flood state numerical matrix and a two-dimensional inundation depth matrix; a training dataset acquisition module 204, which acquires several samples through a preset translation window based on the two-dimensional rainfall matrix, the initial flood state numerical matrix, and the two-dimensional inundation depth matrix, and constructs a training dataset based on these samples; a model training module 205, which obtains a flood inundation prediction model based on the training dataset, a facility joint scheduling scheme library, and a preset deep learning model; and a prediction and treatment module 206, which predicts the flood inundation distribution of the target area using the flood inundation prediction model, obtains the flood inundation prediction results, and performs flood inundation treatment on the target area based on the flood inundation prediction results through the facility joint scheduling scheme library.

[0046] This invention proposes a deep learning-based flood inundation prediction and response system. It constructs a rainfall spatiotemporal simulation database and a joint facility scheduling scheme library using measured rainfall data and historical facility scheduling sequences in the target area. Combined with a hydrodynamic mechanism model built using the target area's confluence characteristic model, it obtains flood inundation spatiotemporal distribution data, thereby capturing complex urban confluence characteristics and achieving deep modeling of local hydraulic connections to global spatiotemporal features. Then, through data matrix processing, the data is organically fused to obtain a training dataset, and a flood inundation prediction model is constructed to predict the flood inundation distribution in the target area. Finally, based on the flood inundation prediction results and the joint facility scheduling scheme library, flood inundation response is implemented in the target area. Therefore, by capturing the urban confluence characteristics of the target area and combining the physical constraints of the hydrodynamic mechanism model, the system fully considers the physical consistency of the output results, improving the reliability of the prediction and response results.

[0047] Furthermore, the second data acquisition module 202 is used to construct a hydrodynamic mechanism model based on the pre-acquired target area confluence characteristic model, and input the rainfall spatiotemporal simulation database into the hydrodynamic mechanism model to obtain flood inundation spatiotemporal distribution data. This includes: a confluence characteristic model acquisition unit 301, used to acquire the hydrological characteristic model, one-dimensional river channel characteristic model, two-dimensional surface characteristic model, and drainage network characteristic model of the target area to obtain the target area confluence characteristic model; a hydrodynamic mechanism model acquisition unit 302, used to obtain a hydrodynamic mechanism model by coupling the confluence characteristic model; and a flood evolution simulation unit 303, used to input the rainfall spatiotemporal simulation database into the hydrodynamic mechanism model, and simulate the flood evolution process of the target area through the hydrodynamic mechanism model to obtain flood inundation spatiotemporal distribution data.

[0048] Furthermore, the data matrix construction module 203 is used to construct a two-dimensional rainfall matrix, an initial flood state numerical matrix, and a two-dimensional inundation depth matrix based on the rainfall spatiotemporal simulation database, the target area confluence characteristic model, and flood inundation spatiotemporal distribution data. It includes: a first matrix construction unit 401, used to perform inverse distance interpolation on the rainfall spatiotemporal simulation database based on a preset interpolation algorithm to obtain a rainfall spatial numerical matrix, and convert the rainfall spatial numerical matrix into a two-dimensional rainfall matrix; a real-time data acquisition unit 402, used to acquire real-time river monitoring water level data, real-time surface monitoring water level data, and real-time drainage network monitoring water level data based on the target area confluence characteristic model; a flood state data acquisition unit 403, used to construct flood state data using the real-time river monitoring water level data, real-time surface monitoring water level data, and real-time drainage network monitoring water level data; a second matrix construction unit 404, used to perform inverse distance interpolation on the flood state data using a preset interpolation algorithm to obtain an initial flood state numerical matrix; and a third matrix construction unit 405, used to convert the flood inundation spatiotemporal distribution data into a two-dimensional inundation depth matrix.

[0049] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0050] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the described specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0051] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

Claims

1. A flood inundation prediction and management method based on deep learning, characterized in that, include: Based on the pre-acquired measured rainfall data of the target area and the historical scheduling sequence of facilities, a spatiotemporal simulation database of rainfall and a joint scheduling scheme library for facilities are constructed. A hydrodynamic mechanism model is constructed based on the pre-acquired target area confluence feature model, and the rainfall spatiotemporal simulation database is input into the hydrodynamic mechanism model to obtain the spatiotemporal distribution data of flood inundation; Based on the aforementioned rainfall spatiotemporal simulation database, the target area confluence characteristic model, and the flood inundation spatiotemporal distribution data, a two-dimensional rainfall matrix, an initial flood state numerical matrix, and a two-dimensional inundation depth matrix are constructed. Based on the two-dimensional rainfall matrix, the initial flood state numerical matrix, and the two-dimensional inundation depth matrix, several samples are obtained through a preset translation window, and a training dataset is constructed based on these samples. Based on the training dataset, the facility joint scheduling scheme library, and the preset deep learning model, a flood inundation prediction model is obtained; The flood inundation distribution of the target area is predicted by the flood inundation prediction model, and the flood inundation prediction results are obtained. Based on the flood inundation prediction results, the flood inundation treatment of the target area is carried out through the facility joint scheduling scheme library.

2. The flood inundation prediction and control method based on deep learning as described in claim 1, characterized in that, The process involves constructing a rainfall spatiotemporal simulation database and a joint facility scheduling scheme library based on pre-acquired measured rainfall data of the target area and historical facility scheduling sequences, including: Based on several rain gauge stations, obtain measured rainfall data for the target area; By randomly scaling the measured rainfall data of the target area using a preset scaling factor, rainfall data of different magnitudes can be obtained. The target region is divided into several matrix grids, and the gradient vectors of the intersection points of the matrix grids are generated by a preset noise generation algorithm. Based on a preset time step and a preset mapping algorithm, the gradient vector is mapped to a preset rainfall intensity range to obtain the spatial pattern of rainfall data; Based on rainfall data of different magnitudes, the spatial pattern of rainfall data, and the preset historical rainfall centers, a spatiotemporal simulation database of rainfall is constructed. Based on the aforementioned spatiotemporal simulation database of rainfall, obtain the scheduling sequence data for the target area; Based on the preset facility capacity range of the target area, the scheduling sequence data of the target area is randomly scaled and randomly combined using the preset scaling factor to obtain a facility joint scheduling scheme library.

3. The flood inundation prediction and control method based on deep learning as described in claim 1, characterized in that, The hydrodynamic mechanism model is constructed based on the pre-acquired target area confluence characteristic model, and the spatiotemporal simulation database of rainfall is input into the hydrodynamic mechanism model to obtain the spatiotemporal distribution data of flood inundation, including: Obtain the hydrological characteristic model, one-dimensional river channel characteristic model, two-dimensional surface characteristic model, and drainage network characteristic model of the target area to obtain the confluence characteristic model of the target area. By coupling the aforementioned confluence characteristic model, a hydrodynamic mechanism model is obtained; The rainfall spatiotemporal simulation database is input into the hydrodynamic mechanism model, and the flood evolution process of the target area is simulated through the hydrodynamic mechanism model to obtain the spatiotemporal distribution data of flood inundation.

4. The flood inundation prediction and control method based on deep learning as described in claim 3, characterized in that, The construction of a two-dimensional rainfall matrix, an initial flood state numerical matrix, and a two-dimensional inundation depth matrix based on the rainfall spatiotemporal simulation database, the target area confluence characteristic model, and flood inundation spatiotemporal distribution data includes: The rainfall spatiotemporal simulation database is subjected to inverse distance interpolation based on a preset interpolation algorithm to obtain a rainfall spatial numerical matrix, and the rainfall spatial numerical matrix is ​​then converted into a two-dimensional rainfall matrix. Based on the confluence characteristic model of the target area, real-time monitoring water level data of river channels, real-time monitoring water level data of surface water, and real-time monitoring water level data of drainage pipe networks are obtained. Flood status data is constructed by using real-time water level data from the river channel, real-time water level data from the surface, and real-time water level data from the drainage network. The flood status data is inversely interpolated using the preset interpolation algorithm to obtain an initial flood status numerical matrix; The spatiotemporal distribution data of the flood inundation is converted into a two-dimensional inundation depth matrix.

5. The flood inundation prediction and control method based on deep learning as described in claim 4, characterized in that, The spatiotemporal distribution data of the flood inundation is converted into a two-dimensional inundation depth matrix, including: Based on the aforementioned spatiotemporal distribution data of flood inundation, several hydrodynamic grids were obtained; Based on a preset rule grid, identify the center points of several hydrodynamic grids; If there are at least two center points of the hydrodynamic grid in the preset rule grid, then the hydrodynamic grid that meets the preset requirement for inundation depth is selected as the target inundation depth grid. If there is a center point of the hydrodynamic grid in the preset rule grid, then the hydrodynamic grid is used as the target flooding depth grid; Based on the target inundation depth grid, a two-dimensional inundation depth matrix is ​​constructed.

6. The flood inundation prediction and control method based on deep learning as described in claim 4, characterized in that, Based on the two-dimensional rainfall matrix, the initial flood state numerical matrix, and the two-dimensional inundation depth matrix, several samples are obtained through a preset translation window, and a training dataset is constructed based on these samples, including: Based on a preset scaling factor and the two-dimensional rainfall matrix, the initial flood state numerical matrix is ​​reduced to obtain the target flood state numerical matrix; Based on the two-dimensional rainfall matrix, the target flood state numerical matrix, and the two-dimensional inundation depth matrix, a hydrodynamic simulation dataset is constructed. Based on a preset translation window, the hydrodynamic simulation dataset is traversed to generate several samples; A training dataset is obtained based on a preset ratio and several of the aforementioned samples.

7. The flood inundation prediction and control method based on deep learning as described in claim 1, characterized in that, Based on the training dataset, the facility joint scheduling scheme library, and the preset deep learning model, a flood inundation prediction model is obtained, including: The training dataset is input into a preset deep learning model, and features are extracted from the training dataset using a preset convolution algorithm and a preset deconvolution algorithm to obtain the spatiotemporal features of the initial rainfall level. By using a preset spatiotemporal feature extraction algorithm, the spatiotemporal features of the initial rainfall level are modeled in time series to obtain the spatiotemporal features of the target rainfall level; By using a pre-defined multivariate time-series feature extraction algorithm, feature extraction is performed on the facility joint scheduling scheme library to obtain the engineering facility control features; The spatiotemporal characteristics of the target rainfall level and the control characteristics of engineering facilities are restored to obtain the initial flood inundation prediction results; The preset deep learning model is trained using the initial flood inundation prediction results and the spatiotemporal characteristics of the target rainfall level until the loss function of the preset deep learning model meets the preset training requirements, thus obtaining the flood inundation prediction model.

8. A flood inundation prediction and control system based on deep learning, characterized in that, Implementing a deep learning-based flood inundation prediction and mitigation method as described in any one of claims 1 to 7, comprising: The first data acquisition module is used to construct a rainfall spatiotemporal simulation database and a facility joint scheduling scheme library based on the pre-acquired measured rainfall data of the target area and the historical scheduling sequence of facilities. The second data acquisition module is used to construct a hydrodynamic mechanism model based on the pre-acquired target area confluence feature model, and input the rainfall spatiotemporal simulation database into the hydrodynamic mechanism model to obtain flood inundation spatiotemporal distribution data; The data matrix construction module is used to construct a two-dimensional rainfall matrix, an initial flood state numerical matrix, and a two-dimensional inundation depth matrix based on the rainfall spatiotemporal simulation database, the target area confluence characteristic model, and the flood inundation spatiotemporal distribution data. The training dataset acquisition module is used to acquire several samples through a preset translation window based on the two-dimensional rainfall matrix, the initial flood state numerical matrix, and the two-dimensional inundation depth matrix, and to construct a training dataset based on the several samples. The model training module is used to obtain a flood inundation prediction model based on the training dataset, the facility joint scheduling scheme library, and the preset deep learning model. The prediction and treatment module is used to predict the flood inundation distribution of the target area through the flood inundation prediction model, obtain the flood inundation prediction result, and, based on the flood inundation prediction result, carry out flood inundation treatment of the target area through the facility joint scheduling scheme library.

9. A flood inundation prediction and control system based on deep learning as described in claim 8, characterized in that, The second data acquisition module is used to construct a hydrodynamic mechanism model based on a pre-acquired target area confluence characteristic model, and input the rainfall spatiotemporal simulation database into the hydrodynamic mechanism model to obtain flood inundation spatiotemporal distribution data, including: The confluence feature model acquisition unit is used to acquire the hydrological feature model, one-dimensional river channel feature model, two-dimensional surface feature model and drainage network feature model of the target area, and obtain the confluence feature model of the target area. The hydrodynamic mechanism model acquisition unit is used to obtain the hydrodynamic mechanism model by coupling the confluence characteristic model; The flood evolution simulation unit is used to input the rainfall spatiotemporal simulation database into the hydrodynamic mechanism model, and to simulate the flood evolution process of the target area through the hydrodynamic mechanism model to obtain the spatiotemporal distribution data of flood inundation.

10. A flood inundation prediction and control system based on deep learning as described in claim 8, characterized in that, The data matrix construction module is used to construct a two-dimensional rainfall matrix, an initial flood state numerical matrix, and a two-dimensional inundation depth matrix based on the rainfall spatiotemporal simulation database, the target area confluence characteristic model, and flood inundation spatiotemporal distribution data, including: The first matrix construction unit is used to perform inverse distance interpolation on the rainfall spatiotemporal simulation database based on a preset interpolation algorithm to obtain a rainfall spatial numerical matrix, and to convert the rainfall spatial numerical matrix into a two-dimensional rainfall matrix; The real-time data acquisition unit is used to acquire real-time monitoring water level data of the river channel, real-time monitoring water level data of the surface, and real-time monitoring water level data of the drainage network based on the confluence characteristic model of the target area. The flood status data acquisition unit is used to construct flood status data by real-time monitoring water level data of the river channel, real-time monitoring water level data of the surface, and real-time monitoring water level data of the drainage network. The second matrix construction unit is used to perform inverse distance interpolation on the flood state data using the preset interpolation algorithm to obtain an initial flood state numerical matrix. The third matrix construction unit is used to convert the spatiotemporal distribution data of flood inundation into a two-dimensional inundation depth matrix.

Images