CN-122021418-A - Urban waterlogging real-time simulation method and system based on Transformer and depth residual error UNet model

CN122021418ACN 122021418 ACN122021418 ACN 122021418ACN-122021418-A

Abstract

The invention discloses a city waterlogging real-time simulation method based on a Transformer and depth residual error UNet model, which comprises the following steps of collecting and processing basic geographic data and historical rainfall data of a research area, constructing a multi-reproduction-period storm event, generating a pipe network operation and surface water accumulation data set through a SWMM-LISFLOOD-FP coupling model, dividing a training set and a testing set, training the Transformer-ResUnet deep learning model, and realizing real-time simulation and performance verification of waterlogging based on actual or designed rainfall. The method creatively fuses the modeling capability of the transducer long sequence and the advantage of ResUnet spatial feature extraction, effectively solves the problems that the traditional hydrodynamic model is low in calculation efficiency and cannot meet the real-time early warning requirement, and simultaneously breaks through the problems that the traditional data-driven model is insufficient in combined modeling of the time-space features. The model output can analyze the water accumulation space-time evolution process under the complex urban topography in real time, provides minute-level prediction support for waterlogging risk dynamic early warning and emergency response decision, and remarkably improves the intelligent prevention and control level of urban storm disasters.

Inventors

CHENG LEI
QIU JINGYU
ZHOU LIHAO
YANG YUHAN

Assignees

武汉大学

Dates

Publication Date: 20260512
Application Date: 20260116

Claims (10)

1. The urban waterlogging real-time simulation method based on the Transformer and depth residual UNet model is characterized by comprising the following steps of: Acquiring rainfall data, topographic data, historical ponding data and mask information of a research area, wherein the mask information is used for distinguishing whether data input into a deep learning model represents a real position in the research area or not; the method comprises the steps of inputting acquired data into a trained deep learning model, outputting predicted future ponding conditions, wherein the deep learning model is embedded with a vision transducer-based attention module on a main structure of a deep residual error UNet network, the vision transducer-based attention module adopts a combination of image block embedding and a multi-stage attention mechanism, ponding is guided by topography in a first stage of the multi-stage attention mechanism, and is fused by rainfall driving in a second stage, and the fused feature sequence enters the deep residual error UNet network for processing, and the predicted future ponding conditions are output.
2. The method for real-time simulation of urban inland inut based on the transform and depth residual UNet model according to claim 1, characterized in that the execution of the multi-stage attention mechanism further comprises: In the first stage, a space encoder is adopted to realize terrain guidance ponding; in the second stage, a time encoder is adopted to realize rainfall driving fusion.
3. The method for real-time simulation of urban inland inut based on the transform and depth residual UNet model according to claim 2, characterized in that the first-stage execution comprises: Respectively carrying out image blocking on the ponding data and the topographic data to form two paths of image block sequences; And respectively carrying out self-attention calculation on the ponding image block sequence and the topographic image block sequence, and carrying out cross-attention calculation based on two paths of self-attention calculation results.
4. The method for real-time simulation of urban inland inut based on the transform and depth residual UNet model according to claim 3, wherein the second stage is performed by: image blocking is carried out on rainfall data to form an image block sequence; And performing self-attention calculation on the rainfall image block sequence, performing self-attention calculation on the output result of the space encoder, performing cross-attention calculation based on two paths of self-attention calculation results, and outputting the fused characteristic sequence.
5. The method for simulating urban inland inundation based on the transform and depth residual UNet model according to claim 1, wherein the training of the depth learning model comprises: collecting and processing basic geographic data and historical rainfall data of a research area, so as to design storm events with different reappearance periods; according to the collected data, simulating pipe network operation and ground water accumulation conditions of a research area under a rainfall design scene by utilizing an SWMM model and a LISFLOOD-FP model so as to construct a data set; training and verifying the deep learning model by using the constructed data set, and outputting the trained deep learning model.
6. The method for real-time simulation of urban inland inundation based on the Transformer and depth residual UNet models according to claim 5, wherein the steps of collecting and processing the basic geographical data and the historical rainfall data of the research area, thereby designing the rainstorm events with different reproduction periods, comprise: Acquiring digital elevation data, drainage system data, historical rainfall data of a research area and grid characteristic data obtained by calculation based on the digital elevation data, wherein the grid characteristic data comprises gradient, direction and flow gradient vectors; based on historical rainfall data, a annual maximum method is adopted to select historical rainfall, a P-III curve is used for fitting, related parameters in a Chicago rainfall type storm intensity formula are estimated through a least square method, comprehensive rainfall peak position coefficients are determined, the storm intensity formula is calculated, a corresponding rainfall type formula is deduced, accumulated rainfall and average rainfall in each period are calculated according to the rainfall type formula, and design storm scenes under different reproduction periods are obtained.
7. The method for simulating urban inland inundation based on the Transformer and depth residual UNet models according to claim 6, wherein the method for simulating pipe network operation and surface water accumulation of the research area in the design rainfall scene by using the SWMM model and the LISFLOOD-FP model comprises the following steps: inputting the designed rainfall scene into a SWMM model, extracting overflow amount time sequence of each node, and converting coordinates and overflow process of overflow nodes into a format required by LISFLOOD-FP model; Converting the elevation data into a format required by LISFLOOD-FP model; Inputting simulation conditions, running LISFLOOD-FP models to obtain the ground water accumulation conditions at different moments, and further constructing a data set suitable for the deep learning model.
8. The method for real-time simulation of urban inland inut based on the transform and depth residual UNet model according to claim 1, further comprising, after obtaining the predicted future water accumulation condition: And counting the ponding depth and the submerged area of each area, comparing the prediction result with the result generated by the hydrodynamic model, calculating an error, and comparing the change process line of the ponding submerged area of the deep learning model and the hydrodynamic model so as to test the performance of the model.
9. Urban waterlogging real-time simulation system based on a transducer and depth residual UNet model is characterized by comprising: the data acquisition module is used for acquiring rainfall data, topographic data, historical ponding data and mask information of the research area, wherein the mask information is used for distinguishing whether the data input into the deep learning model represents the real position in the research area or not; The real-time simulation module is used for inputting the acquired data into the trained deep learning model and outputting predicted future ponding conditions, wherein the deep learning model is embedded with a vision-transducer-based attention module on a main structure of a deep residual error UNet network, the vision-transducer-based attention module adopts a combination of image block embedding and a multi-stage attention mechanism, ponding is guided by topography in a first stage of the multi-stage attention mechanism, and is fused by rainfall driving in a second stage, and the fused feature sequence enters the deep residual error UNet network for processing and outputs the predicted future ponding conditions.
10. A non-transitory computer readable storage medium storing computer instructions that cause the computer to perform the method of real-time simulation of urban inland inundation based on the Transformer and depth residual UNet model of any one of claims 1 to 8.

Description

Urban waterlogging real-time simulation method and system based on Transformer and depth residual error UNet model Technical Field The invention relates to the technical field of urban waterlogging simulation, in particular to a real-time urban waterlogging simulation method and system based on a transform and depth residual error UNet model. Background With the rapid development of the urban progress, urban surface hardening and underlying surface types become more complex, and urban inland inundation has become a serious problem seriously threatening the life safety of urban infrastructure and residents. The urban water accumulation and flood disasters frequently occur due to high-intensity short-duration rainfall events, so that the development of the efficient and accurate urban water-logging real-time simulation technology has important practical significance for improving the urban disaster prevention and reduction capability. The current waterlogging simulation method mainly adopts a mixed modeling system for coupling a one-dimensional pipe network model and a two-dimensional earth surface flooding model. Wherein, one-dimensional models such as SWMM (storm flood management model) are good at simulating the hydraulic characteristics of a pipe network system, and two-dimensional models such as LISFLOOD-FP can finely describe the spatial distribution of surface runoffs. The coupling mode can give consideration to hydraulic interaction between a pipe network and the earth surface, but the double numerical solution process of the coupling mode can generate huge calculation load, single simulation time is long, and timeliness requirements of real-time early warning and emergency response are difficult to meet. In recent years, the deep learning technology provides a new solution for waterlogging simulation. The conventional research attempts to replace the traditional numerical calculation by using a data driving model such as a convolutional neural network and the like, but has obvious defects in practical application, and mainly shows that the conventional network architecture is difficult to capture the long-range dependence characteristics of rainfall space-time sequences and the high-dimensional spatial characteristics of urban terrains at the same time, so that the dynamic simulation precision of the storm evolution process is insufficient. How to construct a deep learning model with higher efficiency and higher precision has become a key subject for breaking through the technical bottleneck of real-time simulation of urban inland inundation. Disclosure of Invention In order to overcome the defects of the prior art, the invention provides a real-time urban waterlogging simulation method based on a transducer and depth residual error UNet model, which effectively solves the problems that the traditional hydrodynamic model is low in calculation efficiency and cannot meet real-time early warning requirements by fusing the advantages of the transducer long sequence modeling capability and the depth residual error Unet spatial feature extraction, and simultaneously breaks through the problems that the traditional data driving model is insufficient in combined modeling of the time-space features. According to an aspect of the present invention, there is provided a real-time simulation method for urban inland inut based on a transform and depth residual UNet model, including: Acquiring rainfall data, topographic data, historical ponding data and mask information of a research area, wherein the mask information is used for distinguishing whether data input into a deep learning model represents a real position in the research area or not; the method comprises the steps of inputting acquired data into a trained deep learning model, outputting predicted future ponding conditions, wherein the deep learning model is embedded with a vision transducer-based attention module on a main structure of a deep residual error UNet network, the vision transducer-based attention module adopts a combination of image block embedding and a multi-stage attention mechanism, ponding is guided by topography in a first stage of the multi-stage attention mechanism, and is fused by rainfall driving in a second stage, and the fused feature sequence enters the deep residual error UNet network for processing, and the predicted future ponding conditions are output. As a further technical solution, the performing of the multi-stage attention mechanism further includes: In the first stage, a space encoder is adopted to realize terrain guidance ponding; in the second stage, a time encoder is adopted to realize rainfall driving fusion. As a further technical solution, the executing process of the first stage includes: Respectively carrying out image blocking on the ponding data and the topographic data to form two paths of image block sequences; And respectively carrying out self-attention calculation on the ponding image block sequence and t