Construction and evaluation method of urban flood rapid prediction model based on hybrid deep learning

Abstract

The present invention belongs to the field of natural disasters, and discloses a hybrid deep learning urban flood rapid prediction model construction and evaluation method, which includes: collecting relevant data of the study area, and using the hydrodynamic model LISFLOOD‑FP to calculate the flood process data set. The flood process data set is divided into a training data set and a verification data set; on the basis of the CNN model, the GRU and multi-head attention modules are connected in series to construct an urban flood rapid prediction model, and the urban flood rapid prediction model is trained using the training data set; a suitable accuracy evaluation index is selected; and the time accuracy, spatial accuracy and robustness of the urban flood rapid prediction model are evaluated using the verification data set and the accuracy evaluation index. The urban flood rapid prediction model construction and evaluation method proposed in the present invention makes the urban flood prediction model more efficient and accurate, and the model construction and evaluation are more comprehensive.

Description

Technical Field

[0001] The present invention belongs to the field of natural disasters and relates to a hybrid deep learning method for constructing and evaluating a rapid urban flood prediction model. Background Art

[0002] Urban flooding is one of the major extreme natural disasters. Compared with the prediction of only the flood-prone points or the maximum flooding depth, the rapid and accurate prediction of the flood range and water depth of the entire area at different times during the entire flood process can help to reduce losses more effectively. Two-dimensional hydrodynamic models such as LISFLOOD-FP are widely used, but the model has low computational efficiency. With the development of artificial intelligence, artificial intelligence technology has also begun to be applied to the key field of urban flood prediction. For example, the convolutional neural network (CNN) establishes a corresponding mathematical model between input and output by analyzing a certain amount of relevant data (meteorological, hydrological, etc.). The model has high computational efficiency, but its accuracy is not high. In addition, the current evaluation of model accuracy often uses subjective evaluation of artificial vision combined with accuracy verification of some observation points. There is a lack of systematic and comprehensive model accuracy evaluation methods, and it is impossible to know the accuracy performance of the model at different times and locations, and whether the accuracy is stable. Summary of the invention

[0003] In response to the above problems, the present invention proposes a hybrid deep learning urban flood rapid prediction model construction and evaluation method, aiming to solve the problems of low computational efficiency and accuracy of urban flood prediction models in the existing technology, and lack of comprehensive model accuracy evaluation.

[0004] In a first aspect, the present invention provides a method for constructing and evaluating a rapid urban flood prediction model using hybrid deep learning, the method comprising:

[0005] Step S1: Collect relevant data of the study area and calculate the flood process dataset using the hydrodynamic model LISFLOOD-FP;

[0006] Step S2: dividing the flood process data set into a training data set and a verification data set;

[0007] Step S3: Based on the CNN model, the GRU and multi-head attention modules are connected in series to construct a rapid urban flood prediction model, and the rapid urban flood prediction model is trained using the training data set;

[0008] Step S4: Select appropriate accuracy evaluation indicators;

[0009] Step S5: using the validation data set and the accuracy evaluation index to evaluate the time accuracy of the urban flood rapid prediction model;

[0010] Step S6: using the validation data set and the accuracy evaluation index to evaluate the spatial accuracy of the urban flood rapid prediction model;

[0011] Step S7: thinning the training data set to evaluate the robustness of the urban flood rapid prediction model.

[0012] Furthermore, the relevant data of the study area are collected and the flood process data set is obtained by using the hydrodynamic model LISFLOOD-FP:

[0013] S11: Collect relevant data of the study area including digital elevation model (DEM) and boundary conditions of upstream points of the inflow river corresponding to different flood events;

[0014] S12: removing noise and obstacles from the digital elevation model in ARCGIS software to obtain digital elevation model data for driving the hydrodynamic model LISFLOOD-FP, wherein the noise and obstacles include trees and bridges;

[0015] S13: According to the boundary conditions and the digital elevation model data obtained in step S12, a flood process data set is calculated using the hydrodynamic model LISFLOOD-FP.

[0016] Furthermore, the GRU and multi-head attention modules are connected in series on the basis of the CNN model to construct a rapid urban flood prediction model: the calculation formula of the GRU model is:

[0017] z t =σ(W z ·[h t-1 , x t ])

[0018] r t =σ(W r ·[h t-1 , x t ])

[0019]

[0020]

[0021] Among them, z t and r t are the outputs of the update gate and the reset gate, respectively. is a candidate hidden state, h t is the hidden state of the current time step; σ represents the Sigmoid function, which limits the value between 0 and 1; W is the weight matrix; h t-1 is the hidden state of the previous time step; x t is the input for the current time step.

[0022] Furthermore, the GRU and multi-head attention modules are connected in series on the basis of the CNN model to construct a rapid urban flood prediction model. Specifically, the calculation process of the multi-head attention module is to apply different learnable projection parameter matrices to the input Q, K, V, and then dot product with the attention. The specific calculation formula is:

[0023] Multihead(Q,K,V)=Concat(head1,...,head h )W O

[0024]

[0025]

[0026] Where Concat is the concatenation operation, and the parameter matrix W i Q ∈R L×D , is the projection, W M is the weight parameter of the projection, is to transform the attention matrix into a standard normal distribution; the above calculation is repeated h times, where h is the number of heads.

[0027] Furthermore, the selection of appropriate accuracy evaluation indicators is specifically: selecting root mean square error (RMSE), Nash coefficient (NSE) and mean difference (DIFF) as indicators for evaluating time accuracy; the specific calculation formulas are:

[0028]

[0029]

[0030]

[0031] Among them, y i is the predicted value, is the observed value, is the mean of the observed values, and N is the sample size.

[0032] Furthermore, the use of the validation data set and the accuracy evaluation index to evaluate the temporal accuracy of the urban flood rapid prediction model is specifically as follows: respectively extracting the output results of the urban flood rapid prediction model and the LISFLOOF-FP model at time t2 to obtain two raster data, and then calculating the accuracy index between the two raster data as the accuracy of the urban flood rapid prediction model at time t2; the same method is used to perform calculations at each time, and finally the temporal accuracy of the urban flood rapid prediction model at different times is obtained.

[0033] Furthermore, the use of the validation data set and the accuracy evaluation index to evaluate the spatial accuracy of the urban flood rapid prediction model is specifically as follows: respectively extracting the sequence values ​​of the output results of the urban flood rapid prediction model and the LISFLOOF-FP model at position (i, j) to obtain two arrays, and then calculating the accuracy index between the two arrays to obtain the accuracy at position (i, j); the same method is used to calculate each position, and finally the spatial accuracy of the urban flood rapid prediction model at different positions is obtained.

[0034] Furthermore, the thinning of the training data set to evaluate the robustness of the urban flood rapid prediction model is specifically as follows: thinning the training data set to 1 / M (M=2, 3, ... 10) respectively, and calculating the training time and accuracy of the urban flood rapid prediction model.

[0035] Furthermore, thinning the training data set to 1 / M specifically includes: retaining only the first data out of every M data.

[0036] In a second aspect, the present invention provides a hybrid deep learning urban flood rapid prediction model construction and evaluation system, comprising:

[0037] Data collection module: collect relevant data of the study area and use the hydrodynamic model LISFLOOD-FP to calculate the flood process data set;

[0038] Data processing module: dividing the flood process data set into a training data set and a verification data set;

[0039] Model training module: Based on the CNN model, the GRU and multi-head attention modules are connected in series to build a rapid urban flood prediction model, and the rapid urban flood prediction model is trained using the training data set;

[0040] Evaluation index setting module: select appropriate accuracy evaluation index;

[0041] Model building and evaluation module: using the verification data set and the accuracy evaluation index to evaluate the temporal accuracy and spatial accuracy of the urban flood rapid prediction model; thinning the training data set to evaluate the robustness of the urban flood rapid prediction model.

[0042] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0043] (1) The present invention combines the CNN model with GRU and multi-head attention: The gated recurrent unit (GRU) is a recurrent neural network (RNN) variant commonly used to process time series data. Compared with other RNN models, it has fewer parameters, lower computational complexity, and faster speed. Most importantly, the gating mechanism of GRU allows it to effectively handle long-term dependency problems and helps the model better understand the features in time series data; multi-head attention can focus on features at different locations at the same time, allowing the model to consider information from multiple time steps and different geographical locations in the time series at the same time; these combinations will enhance the model's ability to understand and capture spatiotemporal information, thereby improving the accuracy and efficiency of the model.

[0044] (2) The present invention proposes a time-space-robustness accuracy assessment framework to comprehensively evaluate the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0046] Figure 1 It is a general flow chart of a method for constructing and evaluating a rapid urban flood prediction model based on hybrid deep learning provided by an embodiment of the present invention;

[0047] Figure 2 It is the digital elevation model data for driving the hydrodynamic model LISFLOOD-FP provided by the embodiment of the present invention;

[0048] Figure 3 is a schematic diagram of time accuracy evaluation of a rapid urban flood prediction model based on hybrid deep learning provided by an embodiment of the present invention;

[0049] Figure 4 is a schematic diagram of spatial accuracy evaluation of a rapid urban flood prediction model based on hybrid deep learning provided by an embodiment of the present invention;

[0050] Figure 5 is a schematic diagram of the time accuracy evaluation results of the urban flood rapid prediction model based on hybrid deep learning provided by an embodiment of the present invention;

[0051] Figure 6 is a schematic diagram of spatial accuracy evaluation results of a rapid urban flood prediction model based on hybrid deep learning provided by an embodiment of the present invention;

[0052] Figure 7It is a schematic diagram of the robustness evaluation results of the urban flood rapid prediction model based on hybrid deep learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Obviously, the specific embodiments described herein are only used to explain the present invention and are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0054] Example 1

[0055] The hybrid deep learning urban flood rapid prediction model construction and evaluation method provided in Example 1 of the present invention is as follows: Figure 1-4 As shown, the flow chart specifically includes:

[0056] Step S1: Collect relevant data of the study area and use the hydrodynamic model LISFLOOD-FP to calculate the flood process dataset.

[0057] Specifically, the research area of ​​the present invention is Carlisle, UK, and the relevant data of the research area include digital elevation model (DEM) and boundary conditions of upstream points of the inlet river corresponding to different flood events. The DEM data is a digital elevation model (DEM) provided by the School of Geographical Sciences of the University of Bristol, with a spatial resolution of 5 meters, from the UK Environment Agency (EA).

[0058] Furthermore, the DEM data was subjected to noise and obstacles removal in ARCGIS software, including trees and bridges; that is, trees and bridges in the DEM data were manually delineated by visual interpretation, and then replaced with the DEM values ​​of the surrounding areas without noise and obstacles to obtain Figure 2 The digital elevation model data used to drive the hydrodynamic model LISFLOOD-FP is shown. Then, according to the boundary conditions and DEM, the hydrodynamic model LISFLOOD-FP is used to calculate the flood process dataset.

[0059] Step S2: dividing the flood process data set into a training data set and a verification data set;

[0060] Specifically, the 2005 catastrophic flood event in Carlisle, UK was used as a validation dataset, and the remaining data were used as training datasets.

[0061] Step S3: Based on the CNN model, the GRU and multi-head attention modules are connected in series to construct a rapid urban flood prediction model, and the rapid urban flood prediction model is trained using the training data set.

[0062] The construction process of the urban flood rapid prediction model is specifically as follows:

[0063] The calculation formula of the GRU model is:

[0064] z t =σ(W Z ·[h t-1 , x t ])

[0065] r t =σ(W r ·[h t-1 , x t ])

[0066]

[0067]

[0068] Where z t and r t are the outputs of the update gate and the reset gate, respectively. is a candidate hidden state, h t is the hidden state of the current time step; σ represents the Sigmoid function, which limits the value between 0 and 1; W is the weight matrix; h t-1 is the hidden state of the previous time step; x t is the input for the current time step.

[0069] The calculation process of the multi-head attention module is to apply different learnable projection parameter matrices to the input Q, K, V, and then dot product with the attention. This process is repeated h times, where h is the number of heads. As shown in the following formula:

[0070] Multihead(Q,K,V)=Concat(head1,...,head h )W O

[0071]

[0072]

[0073] Where Concat is the concatenation operation, and the parameter matrix W i Q ∈R L×D , is the projection, W Mis the weight parameter of the projection, It is to transform the attention matrix into a standard normal distribution.

[0074] Specifically, the model is trained on a computer equipped with an NVIDIA GeForce RTX 3090 graphics card.

[0075] Step S4: Select a suitable accuracy evaluation indicator.

[0076] Specifically, the root mean square error (RMSE), Nash coefficient (NSE) and mean difference (DIFF), which are widely used to evaluate model performance, are selected as indicators for evaluating time accuracy. Their calculation formulas are:

[0077]

[0078]

[0079]

[0080] where y i is the predicted value of the present invention, is the observed value, is the mean of the observed values, and N is the sample size.

[0081] Step S5: using the validation data set and the accuracy evaluation index to evaluate the time accuracy of the urban flood rapid prediction model;

[0082] Specifically, the time accuracy of the urban flood rapid prediction model constructed by the present invention and the LISFLOOF-FP model are compared using the validation data set. Figure 3 As shown. The output results of the urban flood rapid prediction model trained by the present invention and the LISFLOOF-FP model at time t2 are extracted respectively to obtain two grid data, and then the accuracy index between the two grid data is calculated as the accuracy of the model at time t2. The same method is used to calculate at each moment, and finally the time accuracy of the urban flood rapid prediction model trained in step S3 at different moments is obtained.

[0083] Step S6: using the validation data set and the accuracy evaluation index to evaluate the spatial accuracy of the urban flood rapid prediction model;

[0084] Specifically, similar to the time accuracy calculation, such as Figure 4As shown, the sequence values ​​of the output results of the urban flood rapid prediction model trained by the present invention and the LISFLOOF-FP model at the position (i, j) are extracted respectively to obtain two arrays, and then the accuracy index between the two arrays is calculated to obtain the accuracy at the position (i, j). The same method is used to calculate each position, and finally the spatial accuracy of the urban flood rapid prediction model at different positions is obtained.

[0085] Step S7: thinning the training data set to evaluate the robustness of the urban flood rapid prediction model.

[0086] Specifically, the training data set is thinned to 1 / 2, 1 / 3, 1 / 4, ..., 1 / 10 (taking 1 / 10 as an example, only the first data out of every 10 data is retained), and the training time and accuracy of the urban flood rapid prediction model are calculated.

[0087] Example 2

[0088] The construction and evaluation results of the urban flood rapid prediction model based on hybrid deep learning provided in Example 2 of the present invention are as follows: Figure 5-7 The present invention proposes a time-space-robustness accuracy evaluation framework. In order to more intuitively demonstrate the accuracy and robustness of the urban flood rapid prediction model, the urban flood rapid prediction model trained by the present invention is compared with the CNN model. The specific comparison results are shown in Figure 6-7 shown.

[0089] Figure 5The accuracy of the CNN model and the urban flood rapid prediction model trained by the present invention is shown in time. We divide the entire flood process into four time nodes: early, growing, peak, and receding. From the calculation results of RMSE, before the early stage of the flood, and from early to growing, the performance of the two models is very close, both changing from about 0.2m to close to 0, and the performance is good. However, with the development of the flood, after the growing node, the prediction errors of the two models began to rise, the error of CNN quickly increased to more than 0.2m, and the error of the urban flood rapid prediction model trained by the present invention was between 0 and 0.2m. After the peak node, a significant gap began to appear. The prediction error of the CNN model reached 0.4m, and the error of the urban flood rapid prediction model trained by the present invention was closer to 0.2m. After the receding node, the flood began to subside, and the RMSE of the two models remained at around 0.2m. The calculation results of NSE and DIFF are similar to RMSE. The fitting effects of the two models reach the best after the early node and begin to deteriorate after the growing node. In addition, the CNN model shows obvious differences from the urban flood rapid prediction model trained by the present invention in the growing-peak-receding stage.

[0090] Figure 6 The spatial performance of the accuracy of the CNN model and the urban flood rapid prediction model trained by the present invention is demonstrated. We use the union of the flood ranges of all water depth files of the 2005 flood event output by LISFLOOF-FP as the flood range. From the calculation results of RMSE, except for the Eden River, the prediction errors of the urban flood rapid prediction model trained by the present invention are all below 0.4m, which is 1 to 2 levels better than the CNN model overall. Similarly, the calculation results of NSE and DIFF also show a significant gap. Except for the surrounding areas, the NSE of the urban flood rapid prediction model trained by the present invention is mainly between 0.8 and 0.1, which almost achieves perfect fit, and DIFF is mainly between -0.3 and 0.1. However, the accuracy of the CNN model lags behind the urban flood rapid prediction model trained by the present invention again by 1 to 2 levels.

[0091] Figure 7The robustness test results of the CNN model and the urban flood rapid prediction model trained by the present invention are shown in Figure 2. When the complete 15-minute interval flood process data is used for training, the NSEs of the two models are 0.93 and 0.86, respectively, showing a significant gap, and the training time used for the urban flood rapid prediction model trained by the present invention is reduced by 13% (45s vs. 52s), which means that the urban flood rapid prediction model trained by the present invention has higher efficiency while having more accurate results. However, as the sparsity experiment proceeds, the data available for training decreases, and the accuracy and training time of the urban flood rapid prediction model trained by the present invention both show a fluctuating downward trend. When the data is sparsely extracted to 1 / 2, the NSE of the urban flood rapid prediction model trained by the present invention is reduced to 0.87, and the corresponding training time is 7s. The difference is that although the training time of the CNN model also shows a fluctuating downward trend, its accuracy continues to be stable between 0.84 and 0.87, showing good robustness. This phenomenon occurs because when the data is sparse, some data points are deleted or reduced, which may cause some multi-head attentions to fail to capture enough information, and thus the model cannot effectively use incomplete data to perform multi-head attention. Similarly, when the data becomes sparse, the GRU module will also be more challenged because some time points in the sequence may have been deleted or missing, causing the GRU module to be ineffective when processing sparse data because it cannot capture timing information very effectively. It is worth noting that although the accuracy of the urban flood rapid prediction model trained by the present invention will decrease as the training data decreases, it is still ahead of the CNN model and the prediction results are still reliable.

[0092] The above results prove that the present invention improves efficiency while further improving accuracy, and can therefore provide a scientific basis and technical support for understanding and responding to urban flood disasters, help solve possible social and ecological environmental problems, and promote sustainable urban development.

[0093] Example 3

[0094] According to another aspect of an embodiment of the present invention, a hybrid deep learning urban flood rapid prediction model construction and evaluation system is provided, including:

[0095] Data collection module: collect relevant data of the study area and use the hydrodynamic model LISFLOOD-FP to calculate the flood process data set;

[0096] Data processing module: dividing the flood process data set into a training data set and a verification data set;

[0097] Model training module: Based on the CNN model, the GRU and multi-head attention modules are connected in series to build a rapid urban flood prediction model, and the rapid urban flood prediction model is trained using the training data set;

[0098] Evaluation index setting module: select appropriate accuracy evaluation index;

[0099] Model construction and evaluation module: using the validation data set and the accuracy evaluation index to evaluate the temporal accuracy of the urban flood rapid prediction model; using the validation data set and the accuracy evaluation index to evaluate the spatial accuracy of the urban flood rapid prediction model; thinning the training data set to evaluate the robustness of the urban flood rapid prediction model.

[0100] The above embodiments only express the preferred implementation mode of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A hybrid deep learning method for rapid urban flood prediction model construction and evaluation, characterized by: The following steps are involved: Step S1: Collect relevant data of the study area and calculate the flood process dataset using the hydrodynamic model LISFLOOD-FP; The relevant data of the study area are collected and the flood process data set is calculated using the hydrodynamic model LISFLOOD-FP, which is specifically: S11: Collect relevant data of the study area including digital elevation model (DEM) and boundary conditions of upstream points of the inlet river corresponding to different flood events; S12: removing noise and obstacles from the digital elevation model in ARCGIS software to obtain digital elevation model data for driving the hydrodynamic model LISFLOOD-FP, wherein the noise and obstacles include trees and bridges; S13: According to the boundary conditions and the digital elevation model data obtained in step S12, a flood process data set is calculated using the hydrodynamic model LISFLOOD-FP; Step S2: dividing the flood process data set into a training data set and a verification data set; Step S3: Based on the CNN model, the GRU and multi-head attention modules are connected in series to construct a rapid urban flood prediction model, and the rapid urban flood prediction model is trained using the training data set; The specific method of constructing a rapid urban flood prediction model by connecting GRU and multi-head attention modules in series on the basis of the CNN model is as follows: The calculation formula of the GRU model is: z t =σ(W Z ·[h t-1 ,x t ]) r t =σ(W r ·[h t-1 ,x t ]) Among them, z t and r t are the outputs of the update gate and the reset gate, respectively. is a candidate hidden state, h t is the hidden state of the current time step; σ represents the Sigmoid function, which limits the value between 0 and 1; W is the weight matrix; h t-1 is the hidden state of the previous time step; x t is the input of the current time step; The specific method of constructing a rapid urban flood prediction model by connecting GRU and multi-head attention modules in series on the basis of the CNN model is as follows: the calculation process of the multi-head attention module is to apply different learnable projection parameter matrices to the input Q, K, V, and then dot product with the attention. The specific calculation formula is: Multihead(Q,K,V)=Concat(head1,…,head h )W O head i =Attention(QW i Q ,KW i K ,VW i V ) Where Concat is the concatenation operation, and the parameter matrix W i Q ∈R L×D , W i K ∈R L×D ,W i V ∈R L×D is the projection, W M is the weight parameter of the projection, is to transform the attention matrix into a standard normal distribution; the above calculation will be repeated h times, where h is the number of heads; Step S4: Select appropriate accuracy evaluation indicators; Step S5: using the validation data set and the accuracy evaluation index to evaluate the time accuracy of the urban flood rapid prediction model; Step S6: using the validation data set and the accuracy evaluation index to evaluate the spatial accuracy of the urban flood rapid prediction model; Step S7: thinning the training data set to evaluate the robustness of the urban flood rapid prediction model.

2. The method for constructing and evaluating a hybrid deep learning urban flood rapid prediction model according to claim 1 is characterized in that: The selection of appropriate accuracy evaluation indicators is specifically: selecting root mean square error (RMSE), Nash coefficient (NSE) and mean difference (DIFF) as indicators for evaluating time accuracy; the specific calculation formulas are: Among them, y i is the predicted value, is the observed value, is the mean of the observed values, and N is the sample size.

3. The method for constructing and evaluating a rapid urban flood prediction model based on hybrid deep learning according to claim 2 is characterized in that: The use of the validation data set and the accuracy evaluation index to evaluate the time accuracy of the urban flood rapid prediction model is specifically as follows: respectively extracting the output results of the urban flood rapid prediction model and the LISFLOOF-FP model at time t2 to obtain two raster data, and then calculating the accuracy index between the two raster data as the accuracy of the urban flood rapid prediction model at time t2; The same method is used to calculate at each moment, and finally the time accuracy of the urban flood rapid prediction model at different moments is obtained.

4. The method for constructing and evaluating a hybrid deep learning urban flood rapid prediction model according to claim 3 is characterized in that: The method of evaluating the spatial accuracy of the urban flood rapid prediction model using the verification data set and the accuracy evaluation index is specifically as follows: respectively extracting the sequence values ​​of the output results of the urban flood rapid prediction model and the LISFLOOF-FP model at the position (i, j) to obtain two arrays, and then calculating the accuracy index between the two arrays to obtain the accuracy at the position (i, j); the same method is used to calculate each position, and finally the spatial accuracy of the urban flood rapid prediction model at different positions is obtained.

5. The method for constructing and evaluating a hybrid deep learning urban flood rapid prediction model according to claim 4 is characterized in that: The method of thinning the training data set and evaluating the robustness of the urban flood rapid prediction model is specifically as follows: thinning the training data set to 1 / M (M=2, 3, ... 10) respectively, and calculating the training time and accuracy of the urban flood rapid prediction model.

6. The method for constructing and evaluating a hybrid deep learning urban flood rapid prediction model according to claim 5 is characterized in that: Specifically, thinning the training data set to 1 / M includes retaining only the first data out of every M data.

7. A hybrid deep learning urban flood rapid prediction model construction and evaluation system, characterized by: include: Data collection module: collect relevant data of the study area and use the hydrodynamic model LISFLOOD-FP to calculate the flood process data set; The relevant data of the study area are collected and the flood process data set is calculated using the hydrodynamic model LISFLOOD-FP, which is specifically: S11: Collect relevant data of the study area including digital elevation model (DEM) and boundary conditions of upstream points of the inlet river corresponding to different flood events; S12: removing noise and obstacles from the digital elevation model in ARCGIS software to obtain digital elevation model data for driving the hydrodynamic model LISFLOOD-FP, wherein the noise and obstacles include trees and bridges; S13: According to the boundary conditions and the digital elevation model data obtained in step S12, a flood process data set is calculated using the hydrodynamic model LISFLOOD-FP; Data processing module: dividing the flood process data set into a training data set and a verification data set; Model training module: Based on the CNN model, the GRU and multi-head attention modules are connected in series to build a rapid urban flood prediction model, and the rapid urban flood prediction model is trained using the training data set; The specific method of constructing a rapid urban flood prediction model by connecting GRU and multi-head attention modules in series on the basis of the CNN model is as follows: The calculation formula of the GRU model is: z t =σ(W Z ·[h t-1 ,x t ]) r t =σ(W r ·[h t-1 ,x t ]) Among them, z t and r t are the outputs of the update gate and the reset gate, respectively. is a candidate hidden state, h t is the hidden state of the current time step; σ represents the Sigmoid function, which limits the value between 0 and 1; W is the weight matrix; h t-1 is the hidden state of the previous time step; x t is the input of the current time step; The specific method of constructing a rapid urban flood prediction model by connecting GRU and multi-head attention modules in series on the basis of the CNN model is as follows: the calculation process of the multi-head attention module is to apply different learnable projection parameter matrices to the input Q, K, V, and then dot product with the attention. The specific calculation formula is: Multihead(Q,K,V)=Concat(head1,…,head h )W O head i =Attention(QW i Q ,KW i K ,VW i V ) Where Concat is the concatenation operation, and the parameter matrix W i Q ∈R L×D , W i K ∈R L×D ,W i V ∈R L×D is the projection, W M is the weight parameter of the projection, is to transform the attention matrix into a standard normal distribution; the above calculation will be repeated h times, where h is the number of heads; Evaluation index setting module: select appropriate accuracy evaluation index; Model building and evaluation module: using the validation data set and the accuracy evaluation index to evaluate the temporal accuracy and spatial accuracy of the urban flood rapid prediction model; The training data set is thinned to evaluate the robustness of the urban flood rapid prediction model.

Images