CN-122022042-A - Urban waterlogging prediction method and device based on deep learning, storage medium and electronic equipment

CN122022042ACN 122022042 ACN122022042 ACN 122022042ACN-122022042-A

Abstract

The application discloses a city waterlogging prediction method and device based on deep learning, a storage medium and electronic equipment. The method comprises the steps of obtaining grid-level space characteristics and rainfall characteristics of a target area, simulating maximum submerged water depth, X-direction maximum flow rate and Y-direction maximum flow rate of each grid unit of the target area based on a coupling hydrodynamic model, constructing a waterlogging simulation data set, performing fine tuning training on an optimized multi-gate expert hybrid model through the waterlogging simulation data set to generate an urban waterlogging prediction model, and inputting the grid-level space characteristics and the rainfall characteristics of the area to be detected into the urban waterlogging prediction model to predict the maximum submerged water depth, the X-direction maximum flow rate and the Y-direction maximum flow rate of each grid unit in the area to be detected. The method solves the problem of low reliability of the prediction result in the prediction of urban inland inundation in the prior art.

Inventors

CHU WENHAO
ZHANG CHUNXIAO
HU YUQIAN
SUN KANG

Assignees

中国地质大学（北京）

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (9)

1. A city waterlogging prediction method based on deep learning is characterized by comprising the following steps: acquiring grid-level space characteristics and rainfall characteristics of a target area; simulating the maximum submerged water depth, the X-direction maximum flow velocity and the Y-direction maximum flow velocity of each grid unit of the target area under different rainfall characteristics based on a coupling hydrodynamic model; Constructing a waterlogging simulation data set according to the grid-level space characteristics, the rainfall characteristics, the maximum submerged water depth, the X-direction maximum flow velocity and the Y-direction maximum flow velocity; Performing fine tuning training on the optimized multi-gate expert mixed model through the waterlogging simulation data set to generate an urban waterlogging prediction model; inputting grid-level space characteristics and rainfall characteristics of an area to be detected into the urban waterlogging prediction model to predict the maximum submerged depth, the X-direction maximum flow rate and the Y-direction maximum flow rate of each grid unit in the area to be detected; The grid-level spatial features are used for describing spatial substrate data on each grid unit, and the coupled hydrodynamic model is an integrated model formed by coupling a SWMM model and a LISFLOOD-FP model through a data interface.
2. The method of claim 1, wherein the acquiring grid-level spatial features and rainfall features of the target area comprises: setting a plurality of different values of the sedimentation depth on the same grid unit to generate a plurality of different sedimentation densities; Determining an elevation of the target area on each grid cell, land utilization data, water impermeability data, drainage network data, the silting density as the grid level spatial signature of the target area; simulating a storm process with fixed duration based on a storm intensity formula, and forming a plurality of rainfall features by setting different reproduction periods and rain peak coefficients; The elevation is data extracted from a digital elevation model, the drainage pipe network data comprise but are not limited to pipe diameter, gradient, burial depth and connection relation, the sedimentation density is the density of the sedimentation depth, the rainfall characteristic is a series of rainfall data collected according to a preset time interval in the fixed time period, and each rainfall data collected at the time interval comprises rainfall intensity, accumulated rainfall, rainfall duration, the reproduction period and the rain peak coefficient.
3. The method of claim 1, wherein simulating the maximum submerged depth, X-direction maximum flow rate, and Y-direction maximum flow rate for each grid cell of the target area under different ones of the rainfall characteristics based on the coupled hydrodynamic model comprises: And simulating water flow, water level and node overflow of each grid unit in the target area under different rainfall characteristics based on the SWMM model in the coupled hydrodynamic model, inputting overflow quantity as a point source boundary into the LISFLOOD-FP model, and calculating the maximum submerged water depth, the X-direction maximum flow rate and the Y-direction maximum flow rate of each grid unit in the target area under different rainfall characteristics through the LISFLOOD-FP model.
4. The method of claim 1, wherein the fine-tuning the optimized multi-gate expert hybrid model through the waterlogging simulation dataset to generate a city waterlogging prediction model comprises: Setting a plurality of expert sub-networks of the multi-expert hybrid model as networks with three different structures, and setting three gating networks for the multi-expert hybrid model, wherein each gating network corresponds to one prediction task, and the gating networks are used for carrying out weighted summation on output characteristics of the plurality of expert sub-networks; Performing fine tuning training on the set multi-expert mixed model based on the waterlogging simulation data set, and updating parameters of the multi-expert mixed model according to a joint loss function in the training process; And determining the model generated after the fine tuning training is completed as the urban waterlogging prediction model.
5. The method of claim 4, wherein said arranging the plurality of expert subnetworks of the multi-gate expert hybrid model into three distinct classes of networks comprises: setting a first type of topography expert sub-network which takes a convolution layer and a pooling layer as core structures to extract the related features of the grid-level space features and the water depth accumulation; setting a second class of water flow dynamics expert sub-network which takes a circulating layer and an attention layer of a circulating neural network as core structures so as to extract the correlation characteristics of the rainfall characteristic and the flow velocity; setting a third type of mixed expert sub-network which takes the convolution layer and the circulation layer of the circulation neural network as core structures so as to extract the coupling characteristics of water depth and flow velocity; a plurality of the expert subnetworks of the multi-expert hybrid model are configured to include the first type of terrain expert subnetwork, the second type of hydrodynamic expert subnetwork, and the third type of hybrid expert subnetwork.
6. The method of claim 4, wherein updating parameters of the multi-gate expert hybrid model according to a joint loss function during training comprises: The task loss and physical loss are calculated according to the following formulas: Wherein, the For the purpose of the task loss, In order for the physical loss to be a concern, For the maximum submerged water depth predicted by the multi-expert hybrid model, The maximum submerged water depth marked in the waterlogging simulation data set is used for the maximum submerged water depth, An X-direction maximum flow rate predicted for the multi-expert hybrid model, The maximum flow velocity in the X direction marked in the waterlogging simulation data set is obtained, A predicted Y-direction maximum flow rate for the multi-expert hybrid model, The maximum flow velocity in the Y direction marked in the waterlogging simulation data set is E which is a mathematical expectation, 、 And All represent weight coefficients; Determining a weighted sum of the task loss and the physical loss as the joint loss function; and updating parameters of the multi-expert hybrid model based on the joint loss function.
7. Urban waterlogging prediction device based on deep learning, characterized by comprising: The acquisition unit is used for acquiring grid-level space characteristics and rainfall characteristics of the target area; the simulation unit is used for simulating the maximum submerged water depth, the X-direction maximum flow velocity and the Y-direction maximum flow velocity of each grid unit of the target area under different rainfall characteristics based on a coupling hydrodynamic model; The construction unit is used for constructing a waterlogging simulation data set according to the grid-level space characteristics, the rainfall characteristics, the maximum submerged water depth, the X-direction maximum flow velocity and the Y-direction maximum flow velocity; the training unit is used for carrying out fine tuning training on the optimized multi-expert mixed model through the waterlogging simulation data set so as to generate an urban waterlogging prediction model; The prediction unit is used for inputting grid-level space characteristics and rainfall characteristics of the region to be detected into the urban waterlogging prediction model so as to predict the maximum submerged depth, the X-direction maximum flow velocity and the Y-direction maximum flow velocity of each grid unit in the region to be detected; The grid-level spatial features are used for describing spatial substrate data on each grid unit, and the coupled hydrodynamic model is an integrated model formed by coupling a SWMM model and a LISFLOOD-FP model through a data interface.
8. A computer-readable storage medium, characterized in that the computer-readable storage medium comprises a stored program, wherein the program, when executed by a processor, performs the method of any one of claims 1 to 6.
9. An electronic device comprising a memory and a processor, characterized in that the memory has stored therein a computer program, the processor being arranged to execute the method according to any of the claims 1 to 6 by means of the computer program.

Description

Urban waterlogging prediction method and device based on deep learning, storage medium and electronic equipment Technical Field The application relates to the field of artificial intelligence, in particular to a city waterlogging prediction method and device based on deep learning, a storage medium and electronic equipment. Background Under the influence of global climate change and rapid urban transformation, urban storm waterlogging incidents are frequent and hazard is aggravated. Moreover, the problem of fouling caused by insufficient maintenance of the drainage pipe network is increasingly remarkable, so that the hydraulic transmission and drainage capacity of the drainage system is seriously weakened and the risk of waterlogging is further increased. In order to solve the problems of high calculation cost and long time consumption when using a storm runoff management model (SWMM) and a two-dimensional shallow water equation hydrodynamic model (LISFLOOD-FP) to simulate urban inland inundation in the traditional method, the prior art proposes to rapidly predict the urban inland inundation by using a neural network model, however, the traditional method tends to be capable of decoupling modeling by taking water depth and flow velocity as independent variables, and can not describe the inherent nonlinear coupling relationship of the two through joint learning, and also lacks explicit physical consistency constraint. The absence of such physical mechanisms results in models that are prone to non-physical predictions, such as high flow rates that are still output under very shallow water conditions, against basic hydraulic wisdom. In addition, the hydraulic condition can be changed by different sedimentation levels of the drainage pipe network, a nonlinear effect is introduced, and the coupling relation of water depth and flow velocity is remodeled, so that the prior art lacks accurate evaluation of the performance change of the drainage pipe network under different sedimentation levels. In view of the above problems, no effective solution has been proposed at present. Disclosure of Invention The embodiment of the application provides a city waterlogging prediction method and device based on deep learning, a storage medium and electronic equipment, and aims to at least solve the problem that the reliability of a prediction result is low when city waterlogging is predicted in the prior art. According to one aspect of the embodiment of the application, a city waterlogging prediction method based on deep learning is provided, and the city waterlogging prediction method comprises the steps of obtaining grid-level space characteristics and rainfall characteristics of a target area, simulating maximum inundation water depths, X-direction maximum flow rates and Y-direction maximum flow rates of each grid unit of the target area under different rainfall characteristics based on a coupling hydrodynamic model, constructing a waterlogging simulation data set according to the grid-level space characteristics, the rainfall characteristics, the maximum inundation water depths, the X-direction maximum flow rates and the Y-direction maximum flow rates, performing fine tuning training on an optimized multi-gate expert hybrid model through the waterlogging simulation data set to generate a city waterlogging prediction model, inputting the grid-level space characteristics and the rainfall characteristics of an area to be detected into the city waterlogging prediction model to predict the maximum inundation water depths, the X-direction maximum flow rates and the Y-direction maximum flow rates of each grid unit in the area to be detected, wherein the grid-level space characteristics are used for describing space base data on each grid unit, and the coupling hydrodynamic model is an integrated model formed by coupling a SWMM model and a LISFLOOD-FP model through a data interface. Optionally, in the embodiment of the application, the acquiring the grid-level spatial characteristics and the rainfall characteristics of the target area comprises setting a plurality of different values of sedimentation depths on the same grid unit to generate a plurality of different sedimentation densities, determining the elevation, land utilization data, water impermeability data, drainage pipe network data and sedimentation densities of the target area on each grid unit as the grid-level spatial characteristics of the target area, simulating a storm process of a fixed time period based on a storm intensity formula, and forming a plurality of rainfall characteristics by setting different reproduction periods and rain peak coefficients, wherein the elevation is data extracted from a digital elevation model, the drainage pipe network data comprises but is not limited to pipe diameters, gradients, burial depths and connection relations, the sedimentation densities are densities of the sedimentation depths, the rainfall characteristics are a series of