A flood forecasting method based on spatio-temporal feature fusion coupled with deep learning

Abstract

The application relates to the technical field of flood forecasting, and proposes a coupled deep learning flood forecasting method based on space-time feature fusion, which comprises the following steps: collecting time feature data and space feature data of a flood forecasting research area; preprocessing the time feature data and the space feature data; constructing a coupled deep learning network for processing the time feature data and the space feature data into water depth grid data for predicting water depth; training the coupled deep learning network by using the preprocessed time feature data and space feature data, so as to obtain a trained coupled deep learning network; and applying the trained coupled deep learning network to flood forecasting.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of flood forecasting, and more particularly to a coupled deep learning flood forecasting method based on spatiotemporal feature fusion. BACKGROUND

[0002] Under the dual influence of climate change and urbanization, urban waterlogging disasters occur frequently, which not only brings inconvenience to the daily operation and life of the city, but also poses a serious threat to social life and property safety. Therefore, it is crucial to establish a fast and accurate flood forecasting method.

[0003] The prior art discloses a small watershed flood forecasting and warning method and terminal, comprising: obtaining historical hydrological data and watershed basic data, building a hydrology-hydrodynamic coupling initial model based on the MIKE series software and the watershed basic data, calibrating the hydrology-hydrodynamic coupling initial model through the historical hydrological data to obtain a calibrated hydrology-hydrodynamic coupling model; obtaining grid weather forecast data, evaporation data and preheating period rainfall data of each rainfall station in the research watershed, obtaining forecast period rainfall data according to the grid weather forecast data, and importing the preheating period rainfall data, evaporation data and forecast period rainfall data into the hydrology-hydrodynamic coupling model to generate and display flood forecasting and warning information.

[0004] However, the prior art mainly performs hydrodynamic calculation based on a physical model, and the construction process needs to rely on a large amount of data support and tedious modeling steps, which has low calculation efficiency and low calculation accuracy, and has obvious limitations in large-scale calculation and real-time forecasting, and can only perform small watershed flood forecasting, and cannot perform coupled deep learning flood forecasting based on spatiotemporal feature fusion. SUMMARY

[0005] The present application provides a coupled deep learning flood forecasting method based on spatiotemporal feature fusion with high calculation efficiency and high calculation accuracy to overcome the defects of low calculation efficiency and low calculation accuracy of the prior art.

[0006] To solve the above technical problems, the technical solution of the present application is as follows:

[0007] S1: collecting time feature data and space feature data of a flood forecasting research area;

[0008] S2: preprocessing the time feature data and the space feature data;

[0009] S3: constructing a coupled deep learning network for processing the time feature data and the space feature data into water depth grid data for predicting water depth;

[0010] S4: training the coupled deep learning network by using the preprocessed time feature data and space feature data to obtain a trained coupled deep learning network;

[0011] S5: using the trained coupled deep learning network for flood prediction.

[0012] The application further proposes a coupled deep learning flood prediction system based on spatiotemporal feature fusion, which is used to implement the above-mentioned coupled deep learning flood prediction method based on spatiotemporal feature fusion. The system comprises:

[0013] a collection module configured to collect time feature data and space feature data of a flood prediction research area;

[0014] a preprocessing module configured to preprocess the time feature data and space feature data;

[0015] a network construction module configured to construct a coupled deep learning network for processing the time feature data and space feature data into water depth grid data of predicted water depth;

[0016] a network training module configured to train the coupled deep learning network by using the preprocessed time feature data and space feature data to obtain a trained coupled deep learning network;

[0017] a flood prediction module configured to use the trained coupled deep learning network for flood prediction.

[0018] The application further proposes a computer device comprising a memory and a processor, wherein the memory stores computer readable instructions, and when the computer readable instructions are executed by the processor, the processor executes the steps of the coupled deep learning flood prediction method based on spatiotemporal feature fusion proposed by the application.

[0019] Compared with the prior art, the technical scheme of the application has the following beneficial effects:

[0020] The application collects time feature data and space feature data of a flood prediction research area, trains a coupled deep learning network by using the collected time feature data and space feature data, processes time series and space data of different dimensions into water depth grid data of predicted water depth by using the coupled deep learning network, and performs flood prediction based on the water depth grid data of predicted water depth, thereby solving the problems of data dependence and complex modeling of traditional physical models, without the need for water dynamic calculation based on physical models, and with high calculation efficiency, the application can perform flood prediction for large-scale cities. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A flowchart of the coupled deep learning flood prediction method based on spatiotemporal feature fusion proposed for Example 1.

[0022] Figure 2 A diagram of the flood prediction result of the trained coupled deep learning network proposed for Embodiment 1 is shown in the figure;

[0023] Figure 3 A diagram of the overall framework of the coupled deep learning flood prediction system based on spatio-temporal feature fusion proposed for Embodiment 2 is shown in the figure. DETAILED DESCRIPTION

[0024] The accompanying drawings are only for illustrative purposes and should not be construed as limiting the patent;

[0025] In order to better illustrate the embodiments, some components in the drawings may be omitted, enlarged or reduced, and do not represent the actual size of the product;

[0026] It is understandable that some well-known structures and their descriptions in the drawings may be omitted for those skilled in the art.

[0027] The technical solutions of the present application will be further described below in conjunction with the drawings and embodiments.

[0028] Embodiment 1

[0029] The present embodiment proposes a coupled deep learning flood prediction method based on spatio-temporal feature fusion, Figure 1 A flowchart of the coupled deep learning flood prediction based on spatio-temporal feature fusion of the present embodiment is shown in the figure, which includes the following steps:

[0030] S1: Collecting time feature data and spatial feature data of the flood prediction research area;

[0031] S2: Preprocessing the time feature data and spatial feature data;

[0032] S3: Constructing a coupled deep learning network for processing the time feature data and spatial feature data into water depth grid data for predicting water depth;

[0033] S4: Training the coupled deep learning network using the preprocessed time feature data and spatial feature data to obtain a trained coupled deep learning network;

[0034] S5: Using the trained coupled deep learning network for flood prediction.

[0035] In the implementation process, time characteristic data and space characteristic data of a flood forecasting research area are collected, and a coupled deep learning network is trained by using the collected time characteristic data and space characteristic data. Different dimensions of time series and space data are processed into water depth grid data of predicted water depth by using the coupled deep learning network, flood forecasting is performed based on the water depth grid data of predicted water depth, the problems of data dependence and complex modeling of a traditional physical model are solved, and without water dynamic calculation based on the physical model, the calculation efficiency is high, and large-scale cities can be subjected to flood forecasting.

[0036] In an optional embodiment, the collected time characteristic data of the flood forecasting research area includes rainfall data corresponding to a plurality of design return periods and rainfall durations;

[0037] The rainfall data includes a design rainfall intensity, and a calculation expression of the design rainfall intensity is:

[0038]

[0039] In the formula, q is the design rainfall intensity, indicating the rainfall flow per unit area of the flood forecasting research area; P is the design return period; t is the rainfall duration of a single rainfall; A1, C and n are preset values calculated according to the rainfall data of previous years of the flood forecasting research area.

[0040] As an example, the calculation expression of the design rainfall intensity of the optional embodiment is:

[0041]

[0042] The design return period is selected as 1 to 100 years, and the rainfall duration is selected as 2, 4 and 6 hours. The design return period and the rainfall duration are input into the area rainfall intensity formula, and the design rainfall intensity data corresponding to each design return period and rainfall duration can be obtained by calculation.

[0043] In an optional embodiment, the space characteristic data includes flood driving factor data and flood water depth data corresponding to a plurality of design return periods and rainfall durations;

[0044] The flood driving factor data includes terrain elevation data, a change rate of ground elevation, terrain channel depth of the flood forecasting research area, terrain humidity index data of the flood forecasting research area, building and road distribution of the flood forecasting research area, total impervious area of an upstream of the flood forecasting research area, and pipeline layout of the flood forecasting research area.

[0045] As an example, the topographic elevation data is obtained using a Digital Elevation Model (DEM); the rate of change of surface elevation represents the actual surface water flow direction in the flood forecasting study area, i.e., the aspect (ASP); the topographic flume depth in the flood forecasting study area represents the water depth in the depressions (SDEPTH), which is obtained by calculating the difference between the flume outlet elevation and the topographic elevation; the topographic wetness index (TWI) data of the flood forecasting study area characterizes the influence of the topography of the flood forecasting study area on runoff direction and water storage, reflecting the drainage situation of surface water in the watershed; imperviousness (IMP) is calculated using the distribution of buildings and roads in the flood forecasting study area, and as an example, the imperviousness of buildings and roads in the flood forecasting study area is set to 100%; the total impervious area upstream of the flood forecasting study area (FLIMP) is calculated by weighted summation of imperviousness and flow rate.

[0046] In an optional embodiment, the step of collecting the flood depth data includes:

[0047] S1.1: Construct a one-dimensional pipe network model to simulate the flow and unsteady flow in the pipes of the one-dimensional pipe network. The one-dimensional pipe network model is established with the pipes in the flood forecasting study area as connecting lines and the connection points of two or more pipe segments as nodes.

[0048] The expression for calculating the flow rate in the pipes of a one-dimensional pipe network, as simulated by the one-dimensional pipe network model, is as follows:

[0049]

[0050]

[0051] In the formula, Q i,j Let C be the flow rate of the pipe segment between the i-th node and the j-th node in the pipeline network. i,j Let A be the flow coefficient of the pipe segment between the i-th node and the j-th node. i,j Let g be the flow area of ​​the pipe segment between the i-th node and the j-th node, g be the acceleration due to gravity, and Δh be the flow area of ​​the pipe segment between the i-th node and the j-th node. i,j h represents the head loss of the pipe segment between the i-th node and the j-th node. i and h j f represents the water head of the i-th node and the j-th node, respectively. i,j Let L be the friction coefficient of the pipe segment between the i-th node and the j-th node.i,j is the length of the pipe segment between the i th node and the j th node, D i,j is the diameter of the pipe segment between the i th node and the j th node;

[0052] The calculation expression of the one-dimensional pipe network model for simulating the unsteady flow in the pipe of the one-dimensional pipe network is:

[0053]

[0054]

[0055] wherein y is the water depth in the pipe segment between the i th node and the j th node, g is the gravity acceleration, c is the wave speed, x is the length of the pipe segment between the i th node and the j th node, and t is the time;

[0056] S1.2: constructing a two-dimensional ground surface model for simulating the water flow state of the ground surface, the two-dimensional ground surface model being established by using the flood driving factor data, for simulating the real ground surface of the flood forecasting research area, and the two-dimensional ground surface model being divided into a plurality of grid cells;

[0057] wherein the two-dimensional ground surface model simulates the water flow state of the ground surface by calculating the water flow state of each grid cell, and the calculation expression of the water flow state of each grid cell is:

[0058]

[0059]

[0060]

[0061] wherein h is the water depth of the grid cell, u is the velocity component in the x-axis direction of the ground surface, v is the velocity component in the y-axis direction of the ground surface, S fx is the x-axis component of the ground surface friction, S fy is the y-axis component of the ground surface friction;

[0062] S1.3: connecting the pipe of the one-dimensional pipe network in the one-dimensional pipe network model with the ground surface of the two-dimensional ground surface model according to the actual connection of the underground pipe and the ground surface of the flood forecasting research area, and regarding the connection point of the pipe and the ground surface as an overflow port, to realize the dynamic coupling of the one-dimensional pipe network model and the two-dimensional ground surface model;

[0063] wherein the calculation expression of the overflow amount Q out of the overflow port, which is simulated by using the dynamically coupled one-dimensional pipe network model and two-dimensional ground surface model, from the pipe to the ground surface through the overflow port, is:

[0064]

[0065] C = A H d is the overflow coefficient of the overflow pipe section, A out is the area of the overflow port, H in is the water head in the one-dimensional pipe network corresponding to the overflow port, H out is the water head of the ground surface corresponding to the overflow port;

[0066] S1.4: for each design return period and rainfall duration, input the time characteristic data to the dynamically coupled one-dimensional pipe network model and two-dimensional ground surface model, the one-dimensional pipe network model calculates the flow and unsteady flow in the pipe in the one-dimensional pipe network according to the time characteristic data and the calculation expression of unsteady flow, obtains the overflow amount of the overflow port, determines the boundary condition of the calculation expression of the water flow state of the two-dimensional ground surface model according to the overflow amount of the overflow port, calculates the water flow state of the ground surface under the boundary condition, obtains the water depth data of the ground surface composed of the water depth of each grid cell, determines the boundary condition of the calculation expression of the unsteady flow of the one-dimensional pipe network model according to the water depth data of the ground surface, calculates the water flow state in the one-dimensional pipe network under the boundary condition, and obtains the updated overflow amount of the overflow port;

[0067] repeat the calculation of the overflow amount of the overflow port of the one-dimensional pipe network model and the water depth data of the ground surface of the two-dimensional ground surface model until a preset termination condition is reached, stop the calculation, and obtain the flood water depth data for each design return period and rainfall duration; the flood water depth data includes the water depth data of the ground surface when the calculation is stopped.

[0068] As an exemplary illustration, the one-dimensional pipe network of the one-dimensional pipe network model is established by using the pipe layout (PIPE) of the pipe in the flood forecasting research area in the waterlogging driving factor data as the connection line and the connection point of two or more pipe sections as the node, and the PIPE also represents the laying density and volume of the one-dimensional pipe network;

[0069] As an exemplary illustration, the DEM, ASP, SDEPTH, TWI, IMP and FLIMP of the waterlogging driving factor data are taken as the DEM, ASP, SDEPTH, TWI, IMP and FLIMP data of each grid of the two-dimensional ground surface model, so as to establish the two-dimensional ground surface model;

[0070] As an exemplary illustration, the flood water depth data is grid data with a plurality of grid cells, and the water depth data of the ground surface corresponding to each grid cell is recorded on each grid cell. The waterlogging situation in the grid range can be reflected through the water depth data of the ground surface recorded on each grid cell, and the grid cell with non-zero water depth data of the ground surface indicates that the flood forecasting research area corresponding to the grid cell is in a waterlogged state.

[0071] In an optional embodiment, the time feature data and the space feature data are normalized for preprocessing, and the calculation expression of the normalization preprocessing is:

[0072]

[0073] In the formula, x * is the normalized data, x is the time feature data or the space feature data to be processed, x min is the minimum value in x, and x max is the maximum value in x.

[0074] As an exemplary illustration, the normalized data value is between 0 and 1.

[0075] In an optional embodiment, the coupling deep learning network is composed of a SegNet network and a GRU network, the encoder of the SegNet network comprises eight convolutional layers, and each convolutional layer of the encoder of the SegNet network corresponds to a convolutional layer for decoding of a decoder;

[0076] The fully connected layer of the GRU network is connected to the first convolutional layer for decoding of the decoder of the SegNet network, and the output layer of the decoder of the SegNet network is sequentially provided with a global average pooling layer, a flattening layer, a fully connected layer, and a regression layer.

[0077] As an exemplary illustration, the SegNet is a convolutional neural network based on an encoder-decoder structure, which can classify each pixel in an input image into different attribute categories, the encoder of the SegNet comprises eight convolutional layers, each layer corresponding to a decoding layer; the convolutional layer adopts batch normalization and a ReLU activation function, followed by a maximum pooling layer to gradually reduce the resolution of the feature map; the decoder adopts a convolutional layer, batch normalization, a ReLU activation function, and an up-sampling layer to gradually restore the resolution of the feature map, the "skip connection" mechanism introduced by the SegNet connects the feature map of the encoder to the corresponding position of the decoder, solving the problem of possible detail loss in the training process, and the SegNet is good at processing large-scale space feature data and can realize efficient feature extraction;

[0078] The GRU (Gated Recurrent Unit network) is an improved recurrent neural network, which introduces a gating mechanism to control the information transmission in the network. The GRU mainly has two gates, namely an update gate and a reset gate. The GRU effectively prevents overfitting and improves network performance on the basis of simplifying network design. The update gate z t is used to control the state information h t-1 of the previous moment to be brought into the current state h tthe extent to which the state information at the previous moment is brought in; the greater the value of the reset gate r t controlling the state information h at the previous moment t-1 how many are written into the current candidate set The smaller the value of the reset gate, the smaller the extent to which the information of the previous state is written in, and the GRU allows the input of different time-dimension sequences to extract the time characteristics of the sequence. As an exemplary illustration, in the present optional embodiment, the GRU is used to extract the time evolution of the time characteristic data, and the SegNet is used to identify the spatial characteristic data. The coupled deep learning network is composed of the SegNet network and the GRU network, and the time characteristic data and the spatial characteristic data are combined by using the coupled deep learning network.

[0079] In an optional embodiment, the preprocessed time characteristic data and spatial characteristic data are divided into a training set and a test set according to a preset ratio. The training set is used to train the coupled deep learning network, and the performance of the coupled deep learning network is evaluated based on the test set by using an evaluation standard to obtain an evaluation value.

[0080] The process of training the coupled deep learning network by using the training set is as follows:

[0081] The preprocessed spatial characteristic data in the training set is input into the SegNet network of the coupled deep learning network. The encoder of the SegNet network converts the preprocessed spatial characteristic data into a tensor.

[0082] The preprocessed time characteristic data in the training set is input into the GRU network of the coupled deep learning network. The GRU network converts the preprocessed time characteristic data into a one-dimensional time sequence, and then converts the one-dimensional time sequence into a tensor with the same height and width as the tensor obtained by the encoder of the SegNet network.

[0083] The tensor obtained by the GRU network is spliced with the tensor obtained by the SegNet network to obtain a new tensor. The new tensor is input into the first convolutional layer of the decoder of the SegNet network for further processing. After the processing of the decoder, the global average pooling layer, the flattening layer, the fully connected layer and the regression layer of the SegNet network, the water depth raster data of the predicted water depth is output at the output layer of the SegNet network. The water depth raster data of the predicted water depth records the spatial distribution of the flood, and each grid of the water depth raster data of the predicted water depth records the predicted water depth value of the grid.

[0084] The predicted water depth value recorded by each grid of the water depth raster data of the predicted water depth is used to determine the training loss function MSE. The loss function MSE is iteratively solved, and the iteration is ended when the number of iterations reaches a preset value to obtain the trained coupled deep learning network.

[0085] The expression of the loss function MSE is:

[0086]

[0087] In the formula, n represents the number of samples participating in the calculation, is the predicted water depth value of the i-th sample, Y i represents the corresponding actual water depth value.

[0088] As an exemplary illustration, the tensor corresponding to the preprocessed spatial feature data records the spatial features of the flood forecasting research area, and the tensor corresponding to the preprocessed temporal feature data records the temporal features of the flood forecasting research area;

[0089] As an exemplary illustration, the preprocessed temporal feature data and spatial feature data are divided into a training set and a test set according to a ratio of 9:1;

[0090] As an exemplary illustration, the network weights are updated using the back propagation algorithm and the adaptive moment estimation optimizer to iteratively solve the loss function, the momentum parameters coupled with the deep learning network are set to 0.9 and 0.999 respectively, the initial learning rate for training is set to 0.01, the batch size is set to 8, and the total number of training iterations is 100.

[0091] In an optional embodiment, the performance of the coupled deep learning network is evaluated using four evaluation criteria, including the mean absolute error MAE, the root mean square error RMSE, the Nash efficiency coefficient NSE, and the Kling-Gupta efficiency coefficient KGE. The expressions of the four evaluation criteria are respectively:

[0092]

[0093]

[0094]

[0095]

[0096] In the formula, n is the number of samples for verification; Q abs and Q sim respectively represent the actual water depth value and the predicted water depth value, represents the average value of the actual water depth data; r is the Pearson correlation coefficient, a represents the ratio of the average value of the predicted water depth data to the average value of the actual water depth data, and β represents the deviation of the actual water depth value and the predicted water depth value.

[0097] As an exemplary illustration, Figure 2For the trained coupled deep learning network, the closer the evaluation criteria MAE and RMSE are to 0, and the closer the NSE and KGE are to 1, the better the prediction performance of the network. In this embodiment, the coupled deep learning network has an MAE value of 0.0085, an RMSE value of 0.0306, an NSE value of 0.9627, and a KGE value of 0.6949, which are used as the trained coupled deep learning network. The values of MAE, RMSE, NSE, and KGE indicate that the trained coupled deep learning network can accurately predict the flood depth of the grid cells in the study area, and prove the accuracy and effectiveness of the trained coupled deep learning network.

[0098] As an example, the time feature data and the space feature data of another area that needs to be predicted are input into the trained coupled deep learning network, and the water depth grid data of the predicted water depth can also be obtained.

[0099] Embodiment 2

[0100] This embodiment proposes a coupled deep learning flood prediction system based on spatiotemporal feature fusion, which is used to implement the coupled deep learning flood prediction method based on spatiotemporal feature fusion proposed in embodiment 1.

[0101] Figure 3 This is the overall framework diagram of the coupled deep learning flood prediction system based on spatiotemporal feature fusion in this embodiment.

[0102] The coupled deep learning flood prediction system based on spatiotemporal feature fusion comprises:

[0103] The acquisition module is configured to acquire time feature data and space feature data of a flood prediction study area.

[0104] The preprocessing module is configured to preprocess the time feature data and the space feature data.

[0105] The network construction module is configured to construct a coupled deep learning network for processing the time feature data and the space feature data into water depth grid data of predicted water depth.

[0106] The network training module is configured to train the coupled deep learning network using the preprocessed time feature data and space feature data, and obtain a trained coupled deep learning network.

[0107] The flood prediction module is configured to use the trained coupled deep learning network for flood prediction.

[0108] It can be understood that the system of this embodiment is applied to the method of embodiment 1 described above, and the options in embodiment 1 described above are also applicable to this embodiment, so they will not be described again here.

[0109] Embodiment 3

[0110] The embodiment provides a computer device, comprising a memory and a processor, wherein the memory stores computer readable instructions, and the computer readable instructions are executed by the processor to enable the processor to perform the steps of the flood forecasting method based on spatio-temporal feature fusion and coupled deep learning provided in the embodiment 1.

[0111] It can be understood that the computer device of the embodiment is applied to the method of the above-mentioned embodiment 1, and the optional items in the above-mentioned embodiment 1 are also applicable to the embodiment, and thus will not be repeatedly described herein.

[0112] The same or similar reference numerals correspond to the same or similar components;

[0113] The terms describing the positional relationship in the drawings are only used for exemplary illustration, and should not be understood as a limitation to the patent;

[0114] Obviously, the above embodiments of the present application are merely exemplary for clearly illustrating the present application, and are not intended to limit the implementation modes of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, it is not necessary and also impossible to enumerate all the implementation modes. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the claims of the present application.

Claims

1. A flood forecasting method based on spatio-temporal feature fusion coupled deep learning, characterized in that, The method comprises the following steps: S1: collecting time characteristic data and space characteristic data of a flood forecasting research area; S2: preprocessing the time characteristic data and the space characteristic data; S3: constructing a coupled deep learning network for processing the time characteristic data and the space characteristic data into water depth grid data of predicted water depth; S4: training the coupled deep learning network by using the preprocessed time characteristic data and space characteristic data to obtain a trained coupled deep learning network; S5: using the trained coupled deep learning network for flood forecasting; the collected time characteristic data of the flood forecasting research area comprises rainfall data corresponding to a plurality of design return periods and rainfall durations; The rainfall data comprises a design rainfall intensity, and a calculation expression of the design rainfall intensity is: In the formula, q is the rainfall intensity for design, representing the rainfall flow per unit area of the flood forecasting research area; P is the design return period; t is the rainfall duration of a single rainfall; , and are preset values calculated according to the rainfall data of previous years of the flood forecasting research area; the spatial feature data include the flood driving factor data and the flood water depth data corresponding to a plurality of design return periods and rainfall durations; The flood driving factor data comprises terrain elevation data, a change rate of ground elevation, terrain channel depth of the flood forecasting research area, terrain humidity index data of the flood forecasting research area, building and road distribution of the flood forecasting research area, total impervious area of an upstream of the flood forecasting research area, and pipeline layout of the flood forecasting research area; The coupled deep learning network is composed of a SegNet network and a GRU network, an encoder of the SegNet network comprises eight convolutional layers, and each convolutional layer of the encoder of the SegNet network corresponds to a convolutional layer for decoding of a decoder; A fully connected layer of the GRU network is connected to a first convolutional layer for decoding of the decoder of the SegNet network, and an output layer of the decoder of the SegNet network is sequentially provided with a global average pooling layer, a flattening layer, a fully connected layer and a regression layer.

2. The spatio-temporal feature fusion-based coupled deep learning flood forecasting method according to claim 1, characterized in that, The step of collecting the flood water depth data comprises: S1.1: constructing a one-dimensional pipe network model for simulating flow and unsteady flow in a pipe of a one-dimensional pipe network, the one-dimensional pipe network of the one-dimensional pipe network model being established by taking the pipe of the flood forecasting research area as a connection line and taking a connection point of two or more pipe sections as a node; A calculation expression of the one-dimensional pipe network model for simulating flow in a pipe of a one-dimensional pipe network is: wherein Qi,jis the flow rate of the pipe segment between the ith node and the jth node in the pipe network, Ci,jis the flow coefficient of the pipe segment between the ith node and the jth node, Ai,jis the flow area of the pipe segment between the ith node and the jth node, g is the acceleration due to gravity, hi,jis the head loss of the pipe segment between the ith node and the jth node, and hiand hjare the heads of the ith node and the jth node, respectively, fi,jis the friction factor of the pipe segment between the ith node and the jth node, li,jis the length of the pipe segment between the ith node and the jth node, di,jis the diameter of the pipe segment between the ith node and the jth node; A calculation expression of the one-dimensional pipe network model for simulating unsteady flow in a pipe of a one-dimensional pipe network is: In the formula, y is water depth in a pipe section between an i th node and a j th node, g is gravity acceleration, c is wave speed, x is length of the pipe section between the i th node and the j th node, and t is time; S1.2: constructing a two-dimensional ground model for simulating a water flow state of a ground surface, the two-dimensional ground model being established by using the flood driving factor data, for simulating a real ground surface of the flood forecasting research area, and the two-dimensional ground model being divided into a plurality of grid units; In the formula, y is water depth in a pipe section between an i th node and a j th node, g is gravity acceleration, c is wave speed, x is length of the pipe section between the i th node and the j th node, and t is time; wherein is the water depth of the grid cell, is the velocity component in the x-axis direction of the surface, is the velocity component in the y-axis direction of the surface, is the x-axis component of the surface friction, is the y-axis component of the surface friction; S1.3: According to the actual connection of the underground pipeline and the ground surface in the flood forecasting research area, the pipeline in the one-dimensional pipe network model is connected with the ground surface in the two-dimensional ground surface model, and the connection point of the pipeline and the ground surface is regarded as an overflow port, so as to realize dynamic coupling of the one-dimensional pipe network model and the two-dimensional ground surface model; Wherein, the dynamic coupling one-dimensional pipe network model and two-dimensional ground surface model are used to simulate the overflow amount of water flow from the pipe to the ground surface through the overflow port The calculation expression is: wherein is the overflow coefficient for the overflow pipe section, is the area of the overflow, is the head in the one-dimensional pipe network corresponding to the overflow, is the head of the ground surface corresponding to the overflow; S1.4: For each design return period and rainfall duration, time characteristic data is input into the dynamically coupled one-dimensional pipe network model and two-dimensional ground surface model, the one-dimensional pipe network model calculates the flow and unsteady flow in the pipeline in the one-dimensional pipe network according to the time characteristic data and the calculation expression of the unsteady flow, obtains the overflow amount of the overflow port, determines the boundary condition of the calculation expression of the water flow state of the two-dimensional ground surface model according to the overflow amount of the overflow port, calculates the water flow state of the ground surface under the boundary condition corresponding to the overflow amount of the overflow port, obtains the water depth data of the ground surface composed of the water depth of each grid cell, determines the boundary condition of the calculation expression of the unsteady flow of the one-dimensional pipe network model according to the water depth data of the ground surface, and calculates the water flow state in the one-dimensional pipe network under the boundary condition corresponding to the water depth data of the ground surface to obtain updated overflow amount of the overflow port; The overflow amount of the overflow port of the one-dimensional pipe network model and the water depth data of the ground surface of the two-dimensional ground surface model are repeatedly calculated until a preset termination condition is reached, the calculation is stopped, and the flood water depth data for each design return period and rainfall duration is obtained; the flood water depth data includes the water depth data of the ground surface when the calculation is stopped.

3. The spatio-temporal feature fusion-based coupled deep learning flood forecasting method according to claim 1, characterized in that, The time characteristic data and the space characteristic data are normalized and preprocessed, and the calculation expression of the normalized and preprocessed data is: In the formula, x * is the normalized data, x is the time feature data or spatial feature data to be processed, x min is the minimum value in x , x max is the maximum value in x .

4. The spatio-temporal feature fusion-based coupled deep learning flood forecasting method according to claim 1, characterized in that, The preprocessed time characteristic data and space characteristic data are divided into a training set and a test set according to a preset proportion, the coupled deep learning network is trained using the training set, and the performance of the coupled deep learning network is evaluated based on the test set using an evaluation standard to obtain an evaluation value; The process of training the coupled deep learning network using the training set is: The preprocessed space characteristic data in the training set is input into the SegNet network of the coupled deep learning network, and the encoder of the SegNet network converts the preprocessed space characteristic data into a tensor; The preprocessed time characteristic data in the training set is input into the GRU network of the coupled deep learning network, the GRU network converts the preprocessed time characteristic data into a one-dimensional time sequence, and then converts the one-dimensional time sequence into a tensor consistent with the height and width of the tensor obtained by the encoder of the SegNet network; After the tensor obtained by the GRU network is spliced with the tensor obtained by the SegNet network, a new tensor is obtained, the new tensor is input into the first convolutional layer of the decoder of the SegNet network for further processing, and after the processing of the decoder, the global average pooling layer, the flattening layer, the fully connected layer and the regression layer of the SegNet network, the water depth grid data of the predicted water depth is output at the output layer of the SegNet network, the water depth grid data of the predicted water depth records the spatial distribution of the flood, and each grid of the water depth grid data of the predicted water depth records the predicted water depth value of the grid; A prediction water depth value of each grid record of the water depth grid data of the prediction water depth is used to determine a training loss function MSE, the loss function MSE is solved iteratively, and when the number of iterations reaches a preset value, the iteration is ended to obtain a trained coupled deep learning network; An expression of the loss function MSE is: In the formula, n represents the number of samples participating in the calculation, is a predicted water depth value of the i-th sample, represents a corresponding actual water depth value.

5. The spatio-temporal feature fusion-based coupled deep learning flood forecasting method according to claim 4, characterized in that, Four evaluation criteria are used to evaluate the performance of the coupled deep learning network, the four evaluation criteria include: mean absolute error MAE, root mean square error RMSE, Nash efficiency coefficient NSE and Kling-Gupta efficiency coefficient KGE, and expressions of the four evaluation criteria are respectively: wherein n is the number of samples for validation; Q abs and Q sim respectively represent actual and predicted water depth values, represents the mean of actual water depth data; r is the Pearson correlation coefficient, α represents the ratio of the mean of predicted water depth data to the mean of actual water depth data, β represents the deviation of actual and predicted water depth values.

6. A spatiotemporal feature fusion-based coupled deep learning flood forecasting system for implementing the spatiotemporal feature fusion-based coupled deep learning flood forecasting method of any one of claims 1-5. Comprise: The acquisition module is configured to acquire time characteristic data and space characteristic data of a flood forecasting research area; The preprocessing module is configured to preprocess the time characteristic data and the space characteristic data; The network construction module is configured to construct a coupled deep learning network for processing the time characteristic data and the space characteristic data into water depth grid data of the prediction water depth; The network training module is configured to train the coupled deep learning network by using the preprocessed time characteristic data and the space characteristic data to obtain a trained coupled deep learning network; The flood forecasting module is configured to use the trained coupled deep learning network for flood forecasting.

7. A computer device comprising a memory and a processor, the memory having stored therein computer readable instructions, characterized in that, The computer readable instructions, when executed by the processor, cause the processor to perform the steps of the flood forecasting method based on spatiotemporal feature fusion according to any one of claims 1-5.

Images