End-to-end drainage basin flood risk influence forecasting method based on physical constraint

Abstract

The invention discloses an end-to-end drainage basin flood risk influence forecasting method based on physical constraints, and particularly relates to the technical field of meteorological and hydrological forecasting and artificial intelligence, and the method comprises the steps: S1, obtaining multi-source spatio-temporal data of a target drainage basin, constructing a drainage basin graph structure after spatio-temporal alignment and meshing, and constructing a drainage basin graph structure; and S2, constructing an end-to-end deep learning network comprising an adaptive multi-modal feature extraction module and an influence forecasting module, and training the network by using a loss function fused with physical constraints to obtain a flood risk influence forecasting model, the physical constraints at least comprising a water balance constraint. According to the method, the physical mechanism is deeply embedded into the data driving model, the key problems of error accumulation, poor physical consistency and disaster sample scarcity of a traditional cascade forecasting mode are solved, and the forecasting precision and reliability are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of meteorological and hydrological forecasting and artificial intelligence, more specifically, the present application relates to an end-to-end basin flood risk impact forecasting method based on physical constraints. BACKGROUND

[0002] Basin flood disaster is one of the most common and destructive natural disasters in China and even the world, which seriously threatens people's life and property safety and social and economic stability. In recent years, affected by global climate change, extreme heavy rainfall events have occurred frequently and intensified, posing more severe challenges to the accuracy and timeliness of basin flood forecasting.

[0003] Traditional basin flood forecasting services mainly use the cascade mode of numerical weather prediction-hydrological and hydrodynamic models. This mode first predicts future precipitation through a numerical weather prediction model, and then uses the predicted precipitation as input to drive a distributed hydrological model or a hydrodynamic model to simulate the basin runoff process and flood evolution. Although this mode has a clear physical mechanism, it has significant bottlenecks in practical application. First, numerical weather prediction has limited ability to predict local heavy rainfall caused by small-scale convective systems, resulting in inherent bias in the prediction input. Second, the hydrological and hydrodynamic models are complex in structure, high in computational cost, and difficult to calibrate parameters, making it difficult to meet the high timeliness requirements of flood emergency response. Finally, there are multiple links in the cascade mode, resulting in error propagation and accumulation, affecting the final prediction accuracy.

[0004] With the rapid development of artificial intelligence technology, deep learning and other methods are increasingly widely used in meteorological and hydrological fields. In terms of precipitation nowcasting, models based on convolutional neural networks, recurrent neural networks, and generative adversarial networks have shown excellent performance and can effectively capture the spatiotemporal evolution characteristics of precipitation fields. In terms of flood forecasting, long short-term memory networks and graph neural networks have been tried for predicting flow or water level at basin outlet sections.

[0005] However, existing deep learning-based flood forecasting research still has obvious shortcomings. On the one hand, most research still follows the fragmented approach of predicting precipitation first and then simulating floods, failing to achieve end-to-end modeling from observation data to risk impact, resulting in error accumulation and prediction delay. On the other hand, data-driven deep learning models lack physical consistency in their prediction results, and their generalization ability and reliability are questionable when faced with extreme events beyond the training data. In addition, the scarcity of high-quality flood disaster samples also severely restricts the training effect and business applicability of data-driven models.

[0006] Therefore, in view of the above situation, the present application provides an end-to-end basin flood risk impact forecasting method based on physical constraints. SUMMARY

[0007] In order to overcome the above-mentioned defects of the prior art, the present application provides an end-to-end basin flood risk impact prediction method based on physical constraints to solve the problems raised in the above background art.

[0008] To achieve the above-mentioned purpose, the present application provides the following technical solution: an end-to-end basin flood risk impact prediction method based on physical constraints, specifically comprising the following steps: S1, obtaining multi-source spatio-temporal data of a target basin, the multi-source spatio-temporal data comprising dynamic meteorological and hydrological data and static underlying surface data; S2, inputting the multi-source spatio-temporal data into a pre-trained flood risk impact prediction model to obtain a flood risk impact prediction result of the target basin in a future preset time period, the prediction result comprising water levels of key cross sections and / or inundation ranges of the basin, and an economic loss or population impact evaluation map generated based on superimposition of the inundation ranges and hazard-bearing body data, wherein; The flood risk impact prediction model is constructed and trained in the following manner: S2.1, constructing an end-to-end deep learning network architecture comprising a feature extraction module and an impact prediction module; S2.2, training the end-to-end deep learning network architecture using a training sample set, and imposing physical law constraints on the model prediction results through a physical constraint loss function during the training process, so as to minimize both data fitting errors and degrees of violation of physical laws, thereby obtaining the flood risk impact prediction model.

[0009] Preferably, in the S1 step, when obtaining the multi-source spatio-temporal data of the target basin, the data is fused and processed, specifically comprising: spatio-temporally aligning and gridding the dynamic meteorological and hydrological data and the static underlying surface data to construct a unified spatio-temporal grid data set, wherein the dynamic meteorological and hydrological data at least comprises radar reflectivity sequences, satellite precipitation estimation sequences and antecedent soil moisture data, and the static underlying surface data at least comprises digital elevation models, river network data and land use type data; abstracting the basin as a graph structure, wherein nodes represent sub-catchment areas or river cross sections, and edges represent water flow connectivity, and mapping features in the spatio-temporal grid data set as node attributes and edge attributes of the graph structure.

[0010] Preferably, in the S2.1 step, the feature extraction module is an adaptive multi-modal feature extraction module for synchronously extracting multi-scale dynamic features of a precipitation system and static features of a basin underlying surface; The adaptive multi-modal feature extraction module adopts a multi-branch hybrid network architecture, comprising: a multi-scale spatio-temporal Transformer branch for processing radar echo sequences, which extracts evolution characteristics of precipitation fields at different spatio-temporal scales through a multi-head self-attention mechanism; a graph convolutional network branch for processing a basin graph structure, which extracts spatial topological features related to terrain, river network and land use; a feature fusion layer that uses a cross-attention mechanism to adaptively weight and fuse dynamic features output by the multi-scale spatio-temporal Transformer branch and static features output by the graph convolutional network branch, and outputs a unified multi-modal feature tensor.

[0011] Preferably, in the S2.1 step, the impact prediction module outputs a key section water level prediction sequence and a basin inundation area map in parallel according to the multi-modal feature tensor; wherein the key section water level prediction is predicted by a spatio-temporal graph neural network model that performs spatio-temporal calculation on the basin graph structure in combination with a gated recurrent unit; the basin inundation area map is predicted by an improved U-Net++ semantic segmentation model, and the encoder input end of the model fuses the multi-modal feature tensor and static underlying surface data.

[0012] Preferably, in the S2.2 step, the physical constraint loss function at least includes a water balance constraint term, and specifically, the physical constraint loss function further includes a momentum conservation constraint term and / or an energy conservation constraint term; the physical constraint loss function has an expression as follows: ; wherein, is a water balance constraint loss, which is calculated based on model-predicted basin storage changes, input total precipitation and predicted total outflow; is a momentum conservation constraint loss, which is calculated based on a simplified Saint-Venant equation set; is an energy conservation constraint loss; , and are weight coefficients of each constraint term.

[0013] Preferably, in the S2.2 step, the construction of the training sample set includes generating synthetic flood samples with the aid of a physical model, and the specific steps are as follows: based on disaster records of historical typical flood events, calibrate and verify the hydrological and hydrodynamic model; using the calibrated and verified hydrological and hydrodynamic model, simulate the flood evolution of historical heavy precipitation processes to generate high spatio-temporal resolution simulated inundation area and water level data; The simulation data and the measured data jointly constitute a training sample set.

[0014] Preferably, on the basis of generating the synthetic flood sample with the aid of the physical model, the sample is further expanded by using a physically enhanced generative adversarial network: The physically enhanced generative adversarial network generates the synthetic flood sample under the condition of the precipitation sequence and the static feature, and introduces a physical constraint discriminator, so as to ensure that the generated sample meets the water balance and the water flow movement law.

[0015] The application further provides an end-to-end basin flood risk impact prediction system based on a physical constraint, comprising: one or more processors; a storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method described above.

[0016] The application further provides an electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described above when executing the program.

[0017] The application further provides a computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the method described above.

[0018] The technical effects and advantages of the application are as follows: 1. The application realizes direct and rapid mapping from multi-source observation data to flood risk impact indicators through an end-to-end deep learning architecture, effectively avoids the error accumulation problem caused by multiple links in the traditional cascade mode, and in actual business application verification, the Nash efficiency coefficient of the key section water level prediction is improved from 0.72 of the traditional method to 0.89, the root mean square error is reduced by about 35%, and at the same time, the prediction publishing time is shortened from 1-2h of the traditional cascade mode to 5-10min, which saves valuable decision-making time for flood control emergency response; 2. The application effectively guides the extrapolation behavior of the model in the data sparse or missing area by deepening the water balance, momentum conservation and energy conservation and other physical constraints into the model loss function as soft constraints, fundamentally prevents the model from generating physically impossible prediction results, and significantly improves the generalization ability and prediction reliability of the model when facing super-historical extreme precipitation events not covered by the training data; 3、The application proposes a multi-level sample expansion scheme combining physical model assisted generation of synthetic flood samples and physical enhancement of generative adversarial networks, a large number of high-quality training samples with physical consistency are generated from limited historical disaster records, the amount of labeled flood samples required for model training is reduced by about 60%, which is especially suitable for flood risk warning business in small and medium-sized river basins and areas without data, and solves the fundamental restriction of disaster sample scarcity on the performance of deep learning models; 4、The application breaks through the limitation of traditional prediction which only stays in hydrological and hydrodynamic elements, further realizes the fusion analysis with disaster body data based on the output water level and inundation range, and synchronously generates direct economic loss estimation map, population affected distribution map and other impact assessment products by superimposing the inundation range map with population density, land use, infrastructure distribution and other data, so as to convert the prediction information into intuitive disaster pre-assessment and provide key decision support for precise disaster prevention and mitigation; 5、The application has significant improvement in Nash efficiency coefficient, root mean square error and other key indicators, for inundation range prediction, the critical success index reaches more than 0.78 on average, the false alarm rate is reduced by about 25%, stable prediction performance is shown for floods of different magnitudes, and the calculation efficiency is high, which has the feasibility of deployment and application in the conventional business system. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The overall method step flowchart of the application.

[0020] Figure 2 The flood risk impact prediction model construction and training step flowchart of the application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0022] The application provides an end-to-end flood risk impact prediction method based on physical constraints, as shown in the accompanying Figure 1 The method realizes end-to-end rapid mapping from multi-source observation data to flood risk impact indicators by constructing a deep learning model integrating physical mechanisms, and specifically includes the following steps: S1, data preparation and preprocessing This step mainly completes the collection and fusion processing of multi-source data. First, long time series multi-source data of the target river basin are collected, including dynamic meteorological and hydrological data and static underlying surface data, wherein; Dynamic meteorological and hydrological data include: The radar reflectivity sequence is preferably set with a spatiotemporal resolution of 1 km × 1 km and a time interval of 6 minutes. Satellite precipitation estimation products, such as GPM IMERG, have a spatiotemporal resolution of 0.1 degrees × 0.1 degrees and a time interval of 30 minutes. The preliminary soil moisture data were obtained from CLDAS land surface assimilation data, with a resolution of 0.0625 degrees × 0.0625 degrees; Static underlying surface data includes: A digital elevation model with a resolution of 30 meters is used for terrain feature extraction. River network and water system data, 1:50000 scale vector data; Land use type data, 10-meter resolution classification product; In the data fusion process, dynamic meteorological and hydrological data and static underlying surface data are first spatiotemporally aligned and gridded to construct a unified spatiotemporal grid dataset. In specific implementation, all data are unified to the WGS84 coordinate system and UTC time reference, and the watershed is divided into a regular spatiotemporal grid of 1 km × 1 km with a time step of 1 hour. Appropriate interpolation algorithms are used for different source data: bilinear interpolation is used for radar and satellite data, and kriging interpolation is used for station data to generate a gridded dataset covering the entire watershed. Secondly, based on the digital elevation model data, the D8 algorithm is used to extract the water flow direction and divide the sub-catchment areas. The watershed is then abstracted into a graph structure G=(V,E), where the node set V represents the outlet or river cross-section of the sub-catchment area (a cross-section is set every 3 kilometers on the main river), and the edge set E represents the water flow connectivity. Features such as elevation, slope, and land use type in the spatiotemporal grid dataset are mapped to the node attributes of the graph structure, and hydraulic features such as river length, slope, and roughness are mapped to the edge attributes, forming graph structure data that can be processed by graph neural networks.

[0023] S2, Model Construction and Training, as attached Figure 2 As shown, S2.1 Network architecture construction: Construct an end-to-end deep learning network architecture that includes a feature extraction module and an impact prediction module. The feature extraction module is an adaptive multimodal feature extraction module that adopts a multi-branch hybrid network architecture. In practical implementation, this module includes: Multi-scale spatiotemporal Transformer branch: The input is the radar reflectivity sequence of the past 6 hours, which is processed by a 6-layer Transformer encoder. Each layer contains 8 attention heads. The multi-head self-attention mechanism is used to extract the evolution features of the precipitation field at different spatiotemporal scales, such as from convective cells to mesoscale systems. The graph convolution network branch for processing the basin graph structure adopts a two-layer graph convolution network structure, the first layer uses a ReLU activation function, and the second layer uses a Tanh activation function, and spatial topological features related to terrain, river network density and land use pattern are extracted through neighborhood information aggregation; The feature fusion layer adopts a cross-attention mechanism to interact the dynamic features output by the multi-scale spatio-temporal Transformer branch and the static features output by the graph convolution network branch, to realize adaptive weighted fusion and output a unified multi-modal feature tensor; The influence prediction module outputs a key section water level prediction sequence and a basin inundation range map in parallel according to the multi-modal feature tensor, wherein the key section water level prediction is predicted by a spatio-temporal graph neural network model, which combines graph convolution operation (to capture spatial dependence) and gated recurrent unit (to capture time dynamics) on the basin graph structure to perform hydrological spatio-temporal calculation and output water level time series data of each key node in the future 1-6 hours, and the basin inundation range map is realized by an improved U-Net++ semantic segmentation model, the encoder of which includes five down-sampling stages, and the input end fuses the multi-modal feature tensor and static underlying surface data such as land use and terrain, and the decoder fuses multi-scale features through dense skip connection to finally output a binary inundation range probability map in the future 1-6 hours; S2.2, model training and constraints The network is trained using the training sample set, and during the training process, a physical constraint loss function is used to constrain the model prediction results according to physical laws, and the expression of the physical constraint loss function is: ; Wherein: is the water balance constraint loss, which is calculated based on the model-predicted change in basin storage (estimated by the change in inundation volume), the total input precipitation and the predicted total outflow, and is specifically expressed as: wherein, The Penman-Monteith formula is used for estimation; is the momentum conservation constraint loss, which is calculated based on the simplified Saint-Venant equation set, and is specifically expressed as: wherein, is the river channel conductivity, which is determined by the river channel geometry and roughness; is the energy conservation constraint loss, which is calculated based on the Bernoulli equation to calculate the energy loss along the way; , and are the weight coefficients of each constraint term, and the optimal values are determined by grid search, and the typical values are​ = 0.5, = 0.2, = 0.1; The construction of the training sample set includes generating synthetic flood samples with the aid of a physical model. The specific steps are as follows: based on the disaster records of typical historical flood events, a selected hydrological and hydrodynamic model (HEC-RAS two-dimensional model) is finely calibrated and verified to ensure simulation accuracy (Nash efficiency coefficient > 0.75), and the hydrological and hydrodynamic model calibrated and verified is used to simulate flood evolution for a series of heavy rainfall processes in history to generate high spatio-temporal resolution simulated inundation range and water level data. The simulated data generated by the physical model and the measured disaster data are combined to form a training sample set, and 50 flood processes are preferably included; Preferably, on the basis of generating synthetic flood samples with the aid of a physical model, a physically enhanced generative adversarial network is further used to expand the samples. The generative adversarial network takes the rainfall sequence and static features as conditional input to generate synthetic flood samples, and by introducing a physical constraint discriminator, it ensures that the generated samples meet the water balance and water flow movement rules. In the training process, the AdamW optimizer is used, the initial learning rate is set to 1x10 -4 , the batch size is 8, the cosine annealing learning rate scheduling is adopted, the training loss, validation loss and satisfaction degree of each physical constraint are monitored during the training process, the early stopping mechanism is enabled when the validation loss does not decrease for 10 consecutive periods, and the entire training process takes about 24 hours to complete on a workstation equipped with NVIDIA V100 GPU.

[0024] After deploying the trained model in the business system, when a forecast is needed: real-time acquisition of multi-source spatio-temporal data of the target basin, standardized processing according to the aforementioned data preparation and preprocessing method, input of the processed data into the trained flood risk impact prediction model, direct output of the flood risk impact prediction results of the target basin in the future preset period by the model, including the water level sequence of the key section and / or the inundation range map of the basin, further, the inundation range map can be superimposed and analyzed with disaster-bearing body data (such as population density, land use type, infrastructure distribution) to generate economic loss or population impact assessment map, providing more direct decision support for disaster prevention and mitigation; In actual business application, the model can complete the flood risk prediction for the next 6 hours within 2 minutes, greatly improving the timeliness of flood control emergency response, and the prediction results are visualized on the WebGIS platform to provide intuitive technical support for flood control decision-making.

[0025] Technical effects: The application realizes rapid and accurate prediction of the influence of flood risk of a river basin by a deep learning architecture from end to end, deep integration of physical constraints into the model training process, and innovative sample expansion technology. Compared with traditional methods, the application has the following outstanding advantages: Through the end-to-end architecture, the error accumulation of traditional cascade prediction is avoided, and the Nash efficiency coefficient of the key section water level prediction is improved from 0.72 to 0.89. The introduction of physical constraints improves the generalization ability of the model and the physical rationality of the prediction results, and the prediction ability is significantly improved under extreme precipitation scenarios. The innovative sample expansion method effectively solves the problem of disaster sample scarcity. It is particularly suitable for flood risk warning of small and medium-sized river basins and ungauged areas, and the prediction and release time is shortened from 1-2 hours to 5-10 minutes.

[0026] The above only describes the preferred embodiments of the application and is not used to limit the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A physical constraint-based end-to-end catchment flood risk impact forecasting method, characterized in that: Specifically comprising the following steps: S1, obtaining multi-source spatio-temporal data of a target basin, the multi-source spatio-temporal data comprising dynamic meteorological and hydrological data and static underlying surface data; S2, inputting the multi-source spatio-temporal data into a pre-trained flood risk impact prediction model to obtain a flood risk impact prediction result of the target basin in a future preset time period, the prediction result comprising water levels of key sections and / or a flooded area of the basin, and an economic loss or population impact evaluation map generated based on superimposition of the flooded area and hazard-bearing body data, wherein; The flood risk impact prediction model is constructed and trained in the following manner: S2.1, constructing an end-to-end deep learning network architecture comprising a feature extraction module and an impact prediction module; S2.2, training the end-to-end deep learning network architecture using a training sample set, and imposing a physical constraint loss function on the model prediction result during the training process to constrain the model prediction result by physical laws, so as to minimize both data fitting error and degree of violation of physical laws, thereby obtaining the flood risk impact prediction model.

2. The method of claim 1, wherein: In the S1 step, when obtaining the multi-source spatio-temporal data of the target basin, the data is fused and processed, specifically comprising: spatio-temporal alignment and gridding of the dynamic meteorological and hydrological data and the static underlying surface data to construct a unified spatio-temporal grid data set, wherein the dynamic meteorological and hydrological data at least comprises radar reflectivity sequences, satellite precipitation estimation sequences and antecedent soil moisture data, and the static underlying surface data at least comprises digital elevation model, river network data and land use type data; abstracting the basin as a graph structure, wherein the nodes represent sub-catchment areas or river section, and the edges represent water flow connection relationship, and mapping the features in the spatio-temporal grid data set to node attributes and edge attributes of the graph structure.

3. The method of claim 1, wherein: In the S2.1 step, the feature extraction module is an adaptive multi-modal feature extraction module for synchronously extracting multi-scale dynamic features of the precipitation system and static features of the underlying surface of the basin; The adaptive multi-modal feature extraction module adopts a multi-branch hybrid network architecture, comprising: a multi-scale spatio-temporal Transformer branch for processing radar echo sequences, which extracts evolution features of the precipitation field at different spatio-temporal scales through a multi-head self-attention mechanism; a graph convolutional network branch for processing the basin graph structure, which extracts spatial topological features related to terrain, river network and land use; a feature fusion layer which adopts a cross-attention mechanism to adaptively weight and fuse the dynamic features output by the multi-scale spatio-temporal Transformer branch and the static features output by the graph convolutional network branch, and outputs a unified multi-modal feature tensor.

4. The method of claim 1, wherein: In the S2.1 step, the impact prediction module outputs a key section water level prediction sequence and a basin flooded area map in parallel according to the multi-modal feature tensor; wherein the key section water level prediction is predicted by a spatio-temporal graph neural network model which performs spatio-temporal calculation on the basin graph structure in combination with a gated recurrent unit; the basin flooded area map is predicted by an improved U-Net++ semantic segmentation model, and the encoder input end of the model fuses the multi-modal feature tensor and the static underlying surface data.

5. The method of claim 1, wherein: In the S2.2 step, the physical constraint loss function at least includes a water balance constraint term, and specifically, the physical constraint loss function further includes a momentum conservation constraint term and / or an energy conservation constraint term; Physical constraint loss function The expression is: ； wherein, is the water balance constraint loss, calculated based on the model predicted change in basin storage, total input precipitation, and predicted total outflow. The momentum conservation constraint loss is calculated based on the simplified Saint-Venant equations. Loss for energy conservation constraint; , and are weight coefficients for each constraint term.

6. The method of claim 1, wherein: In the S2.2 step, the construction of the training sample set includes generating synthetic flood samples with the aid of a physical model, and the specific steps are as follows: Based on the disaster records of historical typical flood events, the hydrological and hydrodynamic model is calibrated and verified; Using the calibrated and verified hydrological and hydrodynamic model, the flood evolution of historical heavy rainfall processes is simulated to generate high spatio-temporal resolution simulated inundation range and water level data; The simulated data and the measured data are combined to form the training sample set.

7. The method of claim 6, wherein: On the basis of generating synthetic flood samples with the aid of a physical model, the sample is further expanded by using a physically enhanced generative adversarial network; The physically enhanced generative adversarial network generates synthetic flood samples under the condition of precipitation sequence and static features, and introduces a physical constraint discriminator to ensure that the generated samples meet the water balance and water flow movement rules. 8.A physical constraint based end-to-end catchment area flood risk impact forecasting system, characterized in that: Comprise: One or more processors; Storage devices for storing one or more programs; When one or more programs are executed by one or more processors, one or more processors implement the method as claimed in any one of claims 1-7.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that: The processor executes the program to implement the method as claimed in any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The program is executed by the processor to implement the method as claimed in any one of claims 1-7.

Images