Flood forecasting method, device and system coupling physical mechanism and deep learning

Abstract

The invention discloses a physical mechanism and deep learning coupled flood forecasting method, device and system. The flood forecasting method comprises the steps of obtaining DEM data, hydro meteorological data and drainage basin static attribute data of a target drainage basin; based on the DEM data, dividing a target drainage basin into a plurality of sub-drainage basins, extracting a river network, and calculating static attribute data of the river network, the river network comprising a plurality of river channels; inputting the hydrometeorological data, the drainage basin static attribute data and the river network static attribute data of each sub-drainage basin into a pre-trained distributed hybrid hydrological model to obtain simulated flow of a hydrological observation station, and completing flood forecasting; the distributed hybrid hydrological model is coupled with a physical mechanism and deep learning. On the premise that the distributed hybrid hydrological model has good interpretability, high-precision daily runoff and flood simulation can be achieved, and high calculation efficiency is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of flood forecasting, and specifically relates to a flood forecasting method, device and system that couples physical mechanisms and deep learning. Background Art

[0002] Floods, one of the most destructive natural disasters, not only pose a serious threat to human safety but also have significant impacts on infrastructure and socioeconomic development. Flood forecasting models, as crucial non-structural tools for flood management, play an irreplaceable role in protecting life and property, promoting the sustainable use of water resources, and supporting the construction and operation of water conservancy infrastructure. Therefore, accurately predicting flood events, particularly in data-deficient basins, and implementing effective disaster prevention and mitigation measures in advance are crucial for mitigating the negative impacts of floods.

[0003] Currently, flood forecasting models are mainly divided into two categories: process-driven models and data-driven models. Process-driven models are based on physical mechanisms, have clear physical interpretability, and can output additional state variables such as soil moisture and evapotranspiration. However, due to the complexity of the hydrological system, especially in the context of changes in hydrological conditions caused by climate change and human activities, it is very challenging to accurately describe the rainfall-runoff process through precise physical equations. Data-driven models are trained through big data and generally perform well in simulation accuracy. They can also predict flows in ungauged basins through regionalized training. However, such models lack physical interpretability, cannot output additional state variables through flow calibration alone, and are highly dependent on large amounts of training data. Currently, there is a lack of publicly available large-sample flood datasets, which severely limits the application of data-driven models in operational flood forecasting.

[0004] In recent years, hybrid hydrological models, which combine the advantages of process-driven and data-driven models, have become a hot topic. By encoding physical mechanisms and neural networks within a unified deep learning framework and utilizing backpropagation algorithms for simultaneous parameter calibration, hybrid models not only achieve simulation performance comparable to data-driven models but also maintain good physical interpretability. Furthermore, hybrid models can simultaneously simulate key hydrological variables such as evapotranspiration, soil moisture, and baseflow, providing more comprehensive information support for flood forecasting. However, existing hybrid hydrological models still lack the ability to account for spatial heterogeneity in precipitation and underlying surfaces. In particular, the development of distributed hybrid hydrological models faces two major challenges: First, distributed hybrid models require multiple recursive iterations to calibrate parameters, and channel (reach) confluence methods that consider upstream and downstream flood evolution significantly increase the model's computational burden. Second, how to rationally determine sub-basin runoff parameters based on the static properties of the basin's underlying surface remains a key issue that needs to be addressed. Summary of the Invention

[0005] In response to the above problems, the present invention proposes a flood forecasting method, device and system that couples physical mechanisms and deep learning, aiming to effectively solve the difficulties of hybrid models in computational efficiency and accuracy.

[0006] In order to achieve the above technical objectives and the above technical effects, the present invention is implemented through the following technical solutions:

[0007] In a first aspect, the present invention provides a flood forecasting method that couples physical mechanisms and deep learning, comprising:

[0008] Obtain DEM data, hydrological and meteorological data, and static attribute data of the target watershed;

[0009] Based on the DEM data, the target watershed is divided into a plurality of sub-watersheds, and a river network is extracted, and static attribute data of the river network is calculated, wherein the river network includes a plurality of river channels;

[0010] The hydrological and meteorological data of each sub-basin, the static attribute data of the basin, and the static attribute data of the river network are input into a pre-trained distributed hybrid hydrological model to obtain the simulated flow of the hydrological observation station and complete the flood forecast; the distributed hybrid hydrological model couples physical mechanisms and deep learning.

[0011] In combination with the first aspect, optionally, the distributed hybrid hydrological model includes a sub-basin runoff calculation model and a confluence calculation model that are sequentially arranged;

[0012] The mathematical expression of the sub-basin runoff calculation model is:

[0013] ,

[0014] The mathematical expression of the confluence calculation model is:

[0015] ,

[0016] Where, represents the outflow of sub-basin i at time t; HM represents the lumped hydrological model; represents the hydrological and meteorological data of sub-basin i at time t; represents the parameters of the lumped hydrological model of sub-basin i; represents the static attribute data of sub-basin i; FNN1 represents the feedforward neural network used to learn the mapping relationship between the static attribute data of the sub-basin and the parameters of the lumped hydrological model; W FNN1 Represents the parameters of FNN1; It represents the discharge of sub-basin i at time t after the outflow through the river channel or river section converges at the hydrological observation station; RM is the river channel confluence model or river section confluence model; are the parameters of the river channel confluence model or river reach confluence model; represents the static attribute data of a river channel or a river section; FNN2 represents a feedforward neural network used to learn the parameter mapping relationship between the static attribute data of a river channel or a river section and the river channel confluence model or the river section confluence model; W FNN2 Represents the parameters of FNN2; represents the simulated flow at the hydrological observation station at time t.

[0017] In combination with the first aspect, optionally, the sub-basin The river course to the downstream hydrological observation station is recorded as the river course When using the Muskingum confluence calculation method, each river channel needs to be Divided into River sections of equal length, river section index satisfy , the outflow of the sub-basin is calculated from upstream to downstream, and for the upstream river section , where the inflow is the outflow of the corresponding sub-basin i ,reach The outflow is calculated as follows:

[0018] ,

[0019] Where, and Respectively represent river channels Middle River Section exist and The outflow of time, and Represents sub-basins exist and The outflow of time, Indicates river channel The river section exist Lateral inflow at the moment; for the downstream river section , , whose inflow is equal to the adjacent upstream reach The outflow of the river section The outflow is calculated as follows:

[0020] ;

[0021] Where, and Respectively represent river channels Middle River Section exist and The outflow of the moment;

[0022] Danghe section When , calculate according to the following formula Moment sub-basin The flow at the hydrological observation station after the outflow passes through the river section :

[0023] ;

[0024] Where, for Moment sub-basin The outflow is the flow at the hydrological observation station after the confluence of the river section;

[0025] Parameters C1, C2, C3 and C4 are calculated using the following formulas:

[0026] ,

[0027] Where, Indicates river channel The river section The storage time constant, Indicates river channel The river section The weight factor of Δt represents the calculation time step; if the river section Located in a sub-basin Inside, here , is the total number of sub-basins, then The runoff will be distributed to the river section as lateral inflow , Indicates river channel The river section exist The specific calculation formula for the lateral inflow at the moment is as follows:

[0028] ,

[0029] Where, Represents a subbasin Middle River Section The proportionality coefficient, Represents a subbasin At the moment The outflow, Represents a subbasin Middle River Section Average catchment area controlled, Represents a subbasin area;

[0030] The parameters and Adaptive learning by feedforward neural network FNN2, whose value is determined by the river section The specific calculation formula is as follows:

[0031] ,

[0032] Where, Indicates river channel The river section Static attribute data.

[0033] In combination with the first aspect, optionally, the sub-basin The river course to the downstream hydrological observation station is recorded as the river course When the diffusion wave confluence method is used, the movement of flood in the river channel is described by the convection-diffusion equation, and combined with the initial conditions and boundary conditions, we get Moment sub-basin The flow calculation formula of the outflow at the hydrological observation station is as follows:

[0034] ,

[0035] Where, for Moment sub-basin The outflow is the flow at the hydrological observation station after the confluence of the river section. Indicates river channel exist The inflow at time i is also the outflow of subbasin i; Indicates that river channel i is The impulse response function corresponding to time is, For the river length; Indicates the maximum time step considering historical inflow, which is a hyperparameter;

[0036] parameter and Adaptive learning by feedforward neural network FNN2, whose value is determined by the river The specific calculation formula is as follows:

[0037] ,

[0038] Where, Indicates river channel Static attribute data.

[0039] In combination with the first aspect, optionally, the loss function Loss of the confluence calculation model is:

[0040] ,

[0041] ,

[0042] Where, is the number of the hydrological observation station, N is the total number of hydrological observation stations, NSE is the Nash efficiency coefficient, Indicates the moment, represents the length of the time series, Represents a subbasin The observed flow rate, Represents a subbasin The simulated flow rate, Represents a subbasin The average value of the observed flow.

[0043] In combination with the first aspect, optionally, when the convolutional neural network confluence method is adopted, each river channel has a unique static attribute corresponding to a different convolution kernel, Moment sub-basin The flow at the hydrological observation station after the outflow passes through the river confluence The calculation formula is:

[0044] ,

[0045] Where, Indicates river channel The corresponding convolution kernel weights, Indicates the position of each weight, , represents the total number of weights, Indicates river channel Static attribute data of Represents a subbasin exist The outflow of time.

[0046] In combination with the first aspect, optionally, the loss function Loss of the confluence calculation model is:

[0047] ,

[0048] ,

[0049] Where λ1 and Represent the coefficients of each component respectively, and the sum of the two is 1; and Respectively represent river channels The corresponding position in the convolution kernel and The weight of represents the total number of sub-basins; is the activation function; Represents the index of the convolution kernel with the largest weight, ; Indicates river channel The first-order difference of the weights in the corresponding convolution kernel; is the number of the hydrological observation station, N is the total number of hydrological observation stations, and NSE represents the Nash efficiency coefficient.

[0050] In combination with the first aspect, optionally, the training process of the distributed hybrid hydrological model includes:

[0051] In the pre-training stage, the distributed hybrid hydrological model is pre-trained using daily-scale hydrological and meteorological data, watershed static attribute data, and river network static attribute data to obtain a pre-trained model;

[0052] In the fine-tuning stage, the pre-trained model is trained using hourly hydrological and meteorological data, watershed static attribute data, and river network static attribute data to obtain the final distributed hybrid hydrological model.

[0053] In a second aspect, the present invention provides a flood forecasting device that couples physical mechanisms and deep learning, comprising:

[0054] Data acquisition module, used to obtain DEM data, hydrological and meteorological data and static attribute data of the target watershed;

[0055] A sub-basin division module is used to divide the target basin into a plurality of sub-basins based on the DEM data, extract the river network, and calculate the static attribute data of the river network, wherein the river network includes a plurality of river channels;

[0056] The forecasting module is used to input the hydrological and meteorological data of each sub-basin, the static attribute data of the basin, and the static attribute data of the river network into a pre-trained distributed hybrid hydrological model to obtain the simulated flow of the hydrological observation station and complete flood forecasting; the distributed hybrid hydrological model couples physical mechanisms and deep learning.

[0057] In a third aspect, the present invention provides a flood forecasting system that couples physical mechanisms and deep learning, including a storage medium and a processor;

[0058] The storage medium is used to store instructions;

[0059] The processor is configured to operate according to the instructions to execute the method according to any one of the first aspects.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] The present invention proposes a distributed hybrid hydrological model that can couple physical mechanisms and deep learning. The distributed hybrid hydrological model can achieve high-precision daily runoff and flood simulation while ensuring good interpretability, and can improve the model's calculation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive work, among which:

[0063] Figure 1 This is a process diagram of a flood forecasting method based on coupled physical mechanisms and deep learning according to an embodiment of the present invention.

[0064] Figure 2 A schematic diagram of the Fenshui River Basin according to an embodiment of the present invention;

[0065] Figure 3 This is a schematic diagram of model performance comparison in a time division strategy according to an embodiment of the present invention;

[0066] Figure 4 This is a schematic diagram of model performance comparison in a time-space partitioning strategy according to an embodiment of the present invention;

[0067] Figure 5 The daily flow and flood model curve of the prediction model at the GX station in one embodiment of the present invention;

[0068] Figure 6 This is a diagram showing the loss curve and time consumption of a prediction model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0070] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features specified as "first" and "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but this must be based on the fact that ordinary technicians in this field can implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0071] Example 1

[0072] An embodiment of the present invention provides a flood forecasting method that couples physical mechanisms and deep learning, including the following steps:

[0073] (1) Obtain DEM data, hydrological and meteorological data, and static attribute data of the target watershed;

[0074] (2) Based on the DEM data, the target watershed is divided into a number of sub-watersheds, and a river network is extracted, and static attribute data of the river network is calculated, wherein the river network includes a number of river channels;

[0075] (3) The hydrological and meteorological data of each sub-basin, the static attribute data of the basin, and the static attribute data of the river network are input into a pre-trained distributed hybrid hydrological model to obtain the simulated flow of the hydrological observation station and complete the flood forecast; the distributed hybrid hydrological model couples physical mechanisms and deep learning.

[0076] In the above scheme, a distributed hybrid hydrological model that can couple physical mechanisms and deep learning is proposed. This distributed hybrid hydrological model can achieve high-precision daily runoff and flood simulation while ensuring good interpretability, and can improve the model's computational efficiency.

[0077] In a specific implementation of the embodiment of the present invention, the distributed hybrid hydrological model includes a sub-basin runoff calculation model and a confluence calculation model that are sequentially arranged;

[0078] The mathematical expression of the sub-basin runoff calculation model is:

[0079] ,

[0080] The mathematical expression of the confluence calculation model is:

[0081] ,

[0082] Where, represents the outflow of sub-basin i at time t; HM represents the lumped hydrological model; represents the hydrological and meteorological data of sub-basin i at time t; represents the parameters of the lumped hydrological model of sub-basin i; represents the static attribute data of sub-basin i; FNN1 represents the feedforward neural network used to learn the mapping relationship between the static attribute data of the sub-basin and the parameters of the lumped hydrological model; W FNN1 Represents the parameters of FNN1; It represents the discharge of sub-basin i at time t after the outflow through the river channel or river section converges at the hydrological observation station; RM is the river channel confluence model or river section confluence model; are the parameters of the river channel confluence model or river reach confluence model; represents the static attribute data of a river channel or a river section; FNN2 represents a feedforward neural network used to learn the parameter mapping relationship between the static attribute data of a river channel or a river section and the river channel confluence model or the river section confluence model; W FNN2 Represents the parameters of FNN2; represents the simulated flow at the hydrological observation station at time t.

[0083] In a specific implementation of the embodiment of the present invention, the sub-basin The river course to the downstream hydrological observation station is recorded as the river course When using the Muskingum confluence calculation method, each river channel needs to be Divided into River sections of equal length, river section index satisfy , the outflow of the sub-basin is calculated from upstream to downstream, and for the upstream river section , where the inflow is the outflow of the corresponding sub-basin i ,reach The outflow is calculated as follows:

[0084] ,

[0085] Where, and Respectively represent river channels Middle River Section exist and The outflow of time, and Represents sub-basins exist and The outflow of time, Indicates river channel The river section exist Lateral inflow at the moment; for the downstream river section , , whose inflow is equal to the adjacent upstream reach The outflow of the river section The outflow is calculated as follows:

[0086] ;

[0087] Where, and Respectively represent river channels Middle River Section exist and The outflow of the moment;

[0088] Danghe section When , calculate according to the following formula Moment sub-basin The flow at the hydrological observation station after the outflow passes through the river section :

[0089]

[0090] Where, for Moment sub-basin The outflow is the flow at the hydrological observation station after the confluence of the river section;

[0091] Parameters C1, C2, C3 and C4 are calculated using the following formulas:

[0092] ,

[0093] Where, Indicates river channel The river section The storage time constant, Indicates river channel The river section The weight factor of Δt represents the calculation time step; if the river section Located in a sub-basin Inside, here , is the total number of sub-basins, then The runoff will be distributed to the river section as lateral inflow , Indicates river channel The river section exist The specific calculation formula for the lateral inflow at the moment is as follows:

[0094] ,

[0095] Where, Represents a subbasin Middle River Section The proportionality coefficient, Represents a subbasin At the moment The outflow (i.e., flow production), Represents a subbasin Middle River Section Average catchment area controlled, Represents a subbasin area;

[0096] The parameters and Adaptive learning by feedforward neural network FNN2, whose value is determined by the river section The specific calculation formula is as follows:

[0097] ,

[0098] Where, Indicates river channel The river section Static attribute data.

[0099] In a specific implementation of the embodiment of the present invention, the sub-basin The river course to the downstream hydrological observation station is recorded as the river course When the diffusion wave (DW) confluence method is used, the movement of flood in the river channel is described by the convection-diffusion equation, and combined with the initial conditions and boundary conditions, a well-defined problem is formed, where the convection-diffusion equation is:

[0100] ,

[0101] The mathematical expressions of the initial conditions and boundary conditions are:

[0102] ,

[0103] Where, represents the flow at any location in the river; x is a spatial variable, i.e., any spatial location; t is a time variable, C represents the wave velocity, and D represents the diffusion coefficient;

[0104] The convection-diffusion equation is solved using Laplace transform to obtain the following general solution:

[0105] ,

[0106] Where, is the impulse response function of the convection-diffusion equation under initial and boundary conditions, represents the inflow of the river at time t - τ, , Indicates the maximum time length, which is a hyperparameter;

[0107] Finally got Moment sub-basin The flow calculation formula of the outflow at the hydrological observation station is as follows:

[0108] ,

[0109] Where, for Moment sub-basin The outflow is the flow at the hydrological observation station after the confluence of the river section. Indicates river channel exist The inflow at time i is also the outflow of subbasin i; Indicates that river channel i is The impulse response function corresponding to time is, For the river length; Indicates the maximum time step considering historical inflow, which is a hyperparameter;

[0110] parameter and Adaptive learning is performed by the feedforward neural network FNN2, and its value is determined by the static attribute data of the river channel i. The specific calculation formula is as follows:

[0111] ,

[0112] Where, Indicates river channel Static attribute data.

[0113] In a specific implementation of the embodiment of the present invention, the loss function Loss of the confluence calculation model is:

[0114] ,

[0115] ,

[0116] Where, is the number of the hydrological observation station, N is the total number of hydrological observation stations, NSE is the Nash efficiency coefficient, Indicates the moment, represents the length of the time series, Represents a subbasin The observed flow rate, Represents a subbasin The simulated flow rate, Represents a subbasin The average value of the observed flow.

[0117] In a specific implementation of the present invention, when a convolutional neural network (CNN) confluence method is used, each river channel has unique static properties, corresponding to a different convolution kernel. Moment sub-basin The flow at the hydrological observation station after the outflow passes through the river confluence The calculation formula is:

[0118] ,

[0119] Where, Indicates river channel The corresponding convolution kernel weights, Indicates the position of each weight, , represents the total number of weights, Indicates river channel Static attribute data of Represents a subbasin exist The flow rate at the time.

[0120] In a specific implementation of the embodiment of the present invention, the loss function Loss of the confluence calculation model is:

[0121] ,

[0122] ,

[0123] Where λ1 and Represent the coefficients of each component respectively, and the sum of the two is 1; and Respectively represent river channels The corresponding position in the convolution kernel and The weight of represents the total number of sub-basins; is the activation function; Represents the index of the convolution kernel with the largest weight, ; Indicates river channel The first-order difference of the weights in the corresponding convolution kernel.

[0124] In a specific implementation of the embodiment of the present invention, the training process of the distributed hybrid hydrological model includes:

[0125] In the pre-training stage, the distributed hybrid hydrological model is pre-trained using daily-scale hydrological and meteorological data, watershed static attribute data, and river network static attribute data to obtain a pre-trained model;

[0126] In the fine-tuning stage, the pre-trained model is trained using hourly hydrological and meteorological data, watershed static attribute data, and river network static attribute data to obtain the final distributed hybrid hydrological model.

[0127] The following combination Figure 1 A specific implementation method is described in detail for the flood forecasting method in the embodiment of the present invention.

[0128] Step 1. Prepare input data. For the target watershed, collect daily or hourly precipitation and evapotranspiration data from hydrological observatories within the target watershed, daily flow data from hydrological observatories (i.e., hydrological stations), and hourly flood records (flood records are hourly flow data). Also, obtain high-precision digital elevation model (DEM) data (resolution must be within 100 meters) and watershed static attribute data. This watershed static attribute data covers key attributes such as soil type, vegetation cover, and terrain characteristics, providing foundational data support for subsequent model construction.

[0129] Step 2: Prepare input data. Based on the location of the hydrological observation station at the basin outlet and DEM data, extract the basin boundary and divide the river network. The basin is further divided into several sub-basins (this process is implemented using existing technologies). Combining the sub-basin boundaries and the data collected in Step 1, extract the average precipitation and evapotranspiration time series for each sub-basin, and obtain the corresponding basin static attribute data. Furthermore, based on the DEM data, calculate the static attribute data of the river network from each sub-basin outlet to the hydrological observation station. This static attribute data includes average channel length, slope, elevation, meandering ratio, and upstream catchment area, providing spatial heterogeneity information for model construction.

[0130] Step 3: Construction of a distributed hybrid hydrological model. The construction of the distributed hybrid hydrological model needs to ensure differentiability and be fully programmed in a deep learning framework to ensure gradient calculation, thereby realizing the calibration of distributed hybrid hydrological model parameters. The model architecture mainly consists of two core modules: one is the sub-basin runoff calculation module, and the other is the river (river section) confluence calculation module. The sub-basin runoff calculation module adopts a traditional lumped hydrological model (HM), such as the Xinanjiang model (China), the HBV model (Sweden), the Tank model (Japan), the Sactomento model (USA), etc. In order to adapt to the characteristics of different sub-basins, a feedforward neural network (FNN1) is introduced to establish a mapping relationship between the static properties of each sub-basin and the corresponding HM physical parameters, so as to achieve refined and differentiated parameter configuration. Subsequently, the runoff results of each sub-basin will be input into the river (river section) confluence calculation module, and the final simulated flow will be obtained by simulating the propagation and convergence process of water flow in the river network. Assume that the basin is divided into several sub-basins, and the sub-basin number i satisfies , the sub-basin The river course to the downstream hydrological observation station is recorded as the river course The above specific process can be expressed as follows:

[0131] (1);

[0132] Where, (m 3 / s) indicates Moment sub-basin The yield; HM represents the lumped hydrological model; express Moment sub-basin Hydrological and meteorological data (including precipitation, evapotranspiration, etc.); Represents a subbasin Parameters of the lumped hydrological model; represents the static attribute data of sub-basin i, which needs to be normalized; FNN1 represents the feedforward neural network used to learn the mapping relationship between the static attribute data of the sub-basin and the parameters of the lumped hydrological model; W FNN1 Represents the parameters of FNN1; (m 3 / s) represents the discharge of sub-basin i at time t after the outflow passes through the river channel or river section and reaches the hydrological observation station; RM is the river channel confluence model or river section confluence model; are the parameters of the river channel confluence model or river reach confluence model; represents the static attribute data of a river channel or river section, which needs to be normalized; FNN2 represents a feedforward neural network used to learn the parameter mapping relationship between the static attribute data of a river channel or river section and the river channel confluence model or river section confluence model; W FNN2 Represents the parameters of FNN2; (m 3 / s) represents the simulated flow rate at the hydrological observation station at time t, which is sum.

[0133] In this implementation, three methods were developed for calculating river channel (reach) confluence. These include confluence methods based on Muskingum (MK), diffusion wave (DW), and convolutional neural network (CNN). A feedforward neural network (FNN2) was used to map the static properties of the river channel at each sub-basin to a specific section and the physical parameters of the corresponding confluence method.

[0134] For the MK confluence method, each channel needs to be Divided into River sections of equal length, each section is fixed at 2km, river section index satisfy The outflow of the sub-basin is from the upstream ( ) to downstream ( ) to perform confluence calculations in sequence. , where the inflow is the outflow of the corresponding sub-basin i ,reach The outflow is calculated as follows:

[0135] (2);

[0136] Where, and Represent the river sections in river channel i exist and The outflow of time, and Respectively represent the sub-basin i in and The outflow of time, represents the river section in river channel i exist For the downstream river section ( ), whose inflow is equal to that of the adjacent upstream reach The outflow of the river section The outflow is calculated as follows:

[0137] (3);

[0138] Danghe section Calculate according to the following formula:

[0139] (4);

[0140] The parameters C1, C2, C3 and C4 in formulas (2), (3) and (4) are calculated using the following formula:

[0141] (5);

[0142] Where, Indicates river channel The river section Storage time constant (parameter), Indicates river channel The river section The weight factor (parameter) is Δt, which represents the calculation time step. Located in a sub-basin Inside, here , then the sub-basin The runoff will be distributed to the river section as lateral inflow , the specific calculation formula is as follows:

[0143] (6);

[0144] in, Represents a subbasin Middle River Section The proportionality coefficient, Represents a subbasin At the moment The flow rate (i.e. outflow), Represents a subbasin Middle River Section Average catchment area controlled, Represents a subbasin Area. Parameters in the MK confluence method and Adaptive learning by feedforward neural network FNN2, whose value is determined by the river section The specific calculation formula is as follows:

[0145] (7);

[0146] In the formula Represents the static attribute data of river segment j in river channel i.

[0147] The DW confluence method is based on the concept of unit lines and implements the flow routing of each sub-basin to a specific section in parallel. When the downstream section of the river is a free downstream boundary that is not affected by factors such as backwater effects, the movement of flood water in the river can be described by the convection-diffusion equation (8). Combined with the initial and boundary conditions given by equation (9), a well-defined problem is formed:

[0148] (8);

[0149] (9);

[0150] in, represents the flow at any location in the river, x is a spatial variable, i.e., any spatial location, t is a time variable, C represents the wave velocity, and D represents the diffusion coefficient. The above equation can be solved using Laplace transform to obtain the following general solution:

[0151] (10);

[0152] Where, is the impulse response function of Equation (8) under the initial and boundary conditions of Equation (9), represents the inflow of the river at time t - τ, , Indicates the maximum time length, which is a hyperparameter. Specifically, for any sub-basin The outflow at the hydrological observation station can be calculated using the following formula:

[0153] (11);

[0154] Where, represents the inflow of channel i at time t - τ (also the outflow of subbasin i), represents the impulse response function corresponding to the river channel i, where x is the length of the river channel i. and Adaptive learning is performed by the feedforward neural network FNN2, and its value is determined by the static properties of the river channel i. The specific calculation formula is as follows:

[0155] (12);

[0156] In the above formula Represents the static attribute data of river channel i.

[0157] For the CNN confluence method, the invention is inspired by the similarity between the concepts of CNN and unit line structures. The core of one-dimensional CNN is to use a fixed-size convolution kernel to perform convolution operations along the specified dimension with a defined stride, thereby effectively extracting features from the input data. Although the size of the convolution kernel remains unchanged, its weights can be updated through training iterations. Given a time series of inflow into a river channel, with a stride of 1, the CNN confluence method assumes that the convolution kernel function is similar to a unit line, allowing the outflow of the river channel to be calculated by convolution. In the present invention, we extend its application to river confluence, where each river channel has unique static properties, corresponding to a different convolution kernel. The weights of the convolution kernel are determined by mapping these properties through a feedforward neural network FNN2, as follows:

[0158] (13);

[0159] In the formula, in the formula, Indicates river channel The corresponding convolution kernel weights, Indicates the position of each weight, , represents the total number of weights, Indicates river channel Static attribute data of Represents a subbasin exist The flow rate at the time, In order to establish a connection between the convolution kernel and the unit line and enhance the interpretability of the model, Additional physical constraints are imposed. Specifically, the activation function of the output layer of FNN2 is set to Softmax (a universal function in mathematics), ensuring that falls within the range of 0 to 1, and the sum is 1. In addition, a regularization constraint is added to the loss function of the neural network to ensure that the shape of the convolution kernel imitates the unimodal form of the unit line.

[0160] In the distributed hybrid hydrological model of the present invention, the loss function Loss of the MK and DW confluence methods is as follows:

[0161] (14);

[0162] (15);

[0163] Where, is the number of the hydrological observation station, N is the total number of hydrological observation stations, NSE is the Nash efficiency coefficient, Indicates the moment, represents the length of the time series, represents the observed flow rate, Indicates the simulated flow, Represents the average value of the observed flow. For the Loss of the CNN confluence method, a regularization term is added, and the specific formula is as follows:

[0164] (16);

[0165] (17);

[0166] Where λ1 and Represent the coefficients of each component respectively, and the sum of the two is 1; and Respectively represent river channels The corresponding position in the convolution kernel and The weight of represents the total number of sub-basins; is the activation function; Represents the index of the convolution kernel with the largest weight, ; Indicates river channel The first-order difference of the weights in the corresponding convolution kernel.

[0167] Step 4: The runoff generation of the above sub-basins and the confluence of the river channels or river sections are calculated in sequence. After obtaining the simulated flow of the hydrological observation station, the gradient is calculated by backpropagation according to the loss function Loss defined in equation (14) or (16), and the neural network weight W is updated with a specific step size η. Finally, under the constraints of physical information, the parameter update is achieved in an "end-to-end" manner. The specific formula is as follows:

[0168] (18)

[0169] (19)

[0170] Where η is the number of iterations of model training; θ is the physical parameter.

[0171] Step 5: A two-stage strategy of pre-training and fine-tuning. In actual flood forecasting, many basins lack continuous hourly precipitation and flow observation data. Usually, only daily data is available, and hourly records are only available during specific flood events. Traditional methods usually first perform an initial calibration of model parameters on a daily scale to obtain the initial state variables required to predict flood events. Then, certain runoff parameters are scaled based on expert experience, and specific parameters are adjusted using hourly flood event data. However, this method is highly dependent on expert experience. The distributed hybrid hydrological model proposed in this invention adopts a two-stage strategy of pre-training and fine-tuning, making full use of daily data and hourly flood event data. In the pre-training stage, the daily outflow of the sub-basin is calculated using daily hydro-meteorological data and sub-basin static attribute data. Assuming that the daily flow is constant, the daily outflow is converted to hourly outflow. The hourly outflow of the basin is then obtained through river channel (river section) confluence calculation. Finally, the average of the 24-hour predicted values is used as the daily flow prediction result to obtain the pre-trained model. During the fine-tuning stage, based on the pre-trained model, some model parameters in the sub-basin runoff calculation model are scaled to adapt to hourly runoff calculation, and the model is further calibrated using observed flood event data and initial state variables derived from the pre-trained model.

[0172] Example verification.

[0173] 1. Invention area and data.

[0174] This example takes the Fenshui River Basin in Hunan Province as an example. Figure 2 , and validated the proposed distributed hybrid hydrological model. The Fenshui River Basin is a typical humid basin with an area of approximately 10,305 square kilometers. It has a subtropical monsoon climate, an average annual temperature of 18.0°C, and an average annual precipitation of 1,561.2 mm. Affected by atmospheric circulation, precipitation shows significant seasonality and spatial variability. There are multiple hydrological observation stations in the basin, including Ganxi Station (GX) in the lower reaches of the mainstream, Longjiashan Station (LJS) in the middle reaches, and Wulipai Station (WLP), Kengkou Station (KK), and Anren Station (AR) in the upper reaches. In addition, there are 35 automatic telemetered precipitation stations and 1 evaporation station distributed in the basin. The hydrological data are derived from the Hydrological Yearbook and the Hunan Provincial Hydrological Bureau, covering daily precipitation, evapotranspiration, and flow records from January 1, 1980, to December 31, 1985, and from January 1, 2000, to December 31, 2006, as well as hourly flood data during the flood season.

[0175] 2. Data preprocessing.

[0176] Based on 90-meter-resolution DEM data, the river network of the Fenshui River Basin was extracted and divided into 81 discrete subbasins. Using the MK confluence method, the river network was further discretized into 487 segments, each approximately 2 kilometers long. Precipitation data from 35 meteorological stations were interpolated using kriging to obtain the average precipitation for each subbasin. Because there was only one evaporation station, the evapotranspiration input for all subbasins was initially set to a uniform value. However, in this distributed hybrid hydrological model, a neural network was used to determine a specific evapotranspiration coefficient appropriate to the properties of each subbasin, implicitly capturing the spatial heterogeneity of evapotranspiration. Furthermore, the model requires input of static properties of subbasins and river segments. Fourteen surface-averaged properties related to topography, soils, and vegetation were calculated for the 81 subbasins, and five static properties of the river channel (such as channel length and slope) were derived from the DEM to compensate for the lack of measured river channel data.

[0177] 3. Model construction and comparison.

[0178] In this example, the runoff generation component of a distributed hybrid hydrological model adopts the Xin'an River model structure, focusing on evaluating the effectiveness of three river channel (reach) confluence methods (MK, DW, and CNN) in simulating daily flows and flood events. Two validation strategies were adopted: a temporal partitioning strategy and a spatiotemporal partitioning strategy to evaluate the model's performance at both gauged and ungauged stations. The definition of "ungauged station" requires clarification of its practical meaning. Theoretically, the model should be evaluated at all ungauged river reaches. However, due to the infinite number of ungauged river reaches in a real watershed, "ungauged station" in this paper refers to the scenario where historical data for some of the four internal hydrological observation stations (LJS, WLP, KK, and AR) within the basin are unknown, and only test period data are used for model driving and validation.

[0179] In the temporal partitioning strategy, five scenarios were designed, each using data from one to five hydrological observatories to train the model and evaluate its effectiveness at each station. In the spatial-temporal partitioning strategy, four scenarios were designed, each using a different number of gauging stations for model training and evaluating the model's performance at ungauged stations. In the spatial-temporal partitioning strategy, station GX, located at the basin outlet, was always designated as the gauging station with historical data to align with the actual situation.

[0180] In addition, a lumped Xin'anjiang model (L-XAJ) was constructed for comparison. This model was implemented in PyTorch, with parameters updated via backpropagation. The upstream basin of each station was treated as an independent basin, and average precipitation and evapotranspiration were calculated. The L-XAJ model was calibrated for five basins, and its performance was tested.

[0181] All models were calibrated using data from January 1, 1980, to December 31, 1985, and tested using data from January 1, 2000, to December 31, 2006, with a one-year warm-up period. To account for modeling uncertainty, each model was trained 10 times using 10 different random seeds, while keeping all other hyperparameters constant. In this paper, the distributed hybrid hydrological model is named "DHFM-Routing-X," where "Routing" represents the MK, DW, and CNN channel (reach) confluence methods, and "X" represents the number of gauging stations used for training (1, 2, 3, 4, and 5).

[0182] 4. Result analysis.

[0183] Figure 3 and Figure 4 The NSE index results of several models in the time division strategy and the space-time division strategy are shown respectively. Figure 5 The actual simulation curve is shown. Figure 5 In the data, Obs represents the observed value, Obs.(Day) represents daily flow simulation, Obs.(Flood) represents flood simulation, Q represents flow, Day represents day, and Hour represents hour. The results show that the proposed distributed hybrid hydrological model can effectively simulate daily flow and flood events and is suitable for daily flow and flood simulation in data-free basins. Figure 6 The computational efficiency of the model was demonstrated, among which the DW and CNN river (river section) confluence methods significantly improved the computational efficiency and achieved second-level calculations, further verifying the efficiency and practicality of the model in practical applications.

[0184] Example 2

[0185] Based on the same inventive concept as Example 1, an embodiment of the present invention provides a flood forecasting device that couples physical mechanisms and deep learning, including:

[0186] Data acquisition module, used to obtain DEM data, hydrological and meteorological data and static attribute data of the target watershed;

[0187] A sub-basin division module is used to divide the target basin into a plurality of sub-basins based on the DEM data, extract the river network, and calculate the static attribute data of the river network, wherein the river network includes a plurality of river channels;

[0188] The forecast module is used to input the hydrological and meteorological data of each sub-basin, the static attribute data of the basin and the static attribute data of the river network into the pre-trained distributed hybrid hydrological model to obtain the simulated flow of the hydrological observation station and complete the flood forecast.

[0189] The rest are the same as in Example 1.

[0190] Example 3

[0191] An embodiment of the present invention provides a flood forecasting system that couples physical mechanisms and deep learning, including a storage medium and a processor;

[0192] The storage medium is used to store instructions;

[0193] The processor is configured to operate according to the instruction to execute the method according to any one of embodiment 1

[0194] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0195] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0196] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0197] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1The steps for the function specified in one or more boxes.

[0198] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which are all protected by the present invention.

[0199] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A flood forecasting method that couples physical mechanisms and deep learning, characterized in that: include: Obtain DEM data, hydrological and meteorological data, and static attribute data of the target watershed; Based on the DEM data, the target watershed is divided into a plurality of sub-watersheds, and a river network is extracted, and static attribute data of the river network is calculated, wherein the river network includes a plurality of river channels; The hydrological and meteorological data of each sub-basin, the static attribute data of the basin, and the static attribute data of the river network are input into a pre-trained distributed hybrid hydrological model to obtain the simulated flow of the hydrological observation station and complete the flood forecast; the distributed hybrid hydrological model couples physical mechanisms and deep learning.

2. The flood forecasting method of coupling physical mechanism and deep learning according to claim 1 is characterized by: The distributed hybrid hydrological model includes a sub-basin runoff calculation model and a confluence calculation model that are set in sequence; The mathematical expression of the sub-basin runoff calculation model is: ， The mathematical expression of the confluence calculation model is: ， Where, express Moment sub-basin outflow; represents a lumped hydrological model; represents the sub-basin at time t hydrometeorological data; Represents a subbasin Parameters of the lumped hydrological model; Represents a subbasin Basin static attribute data; A feedforward neural network for learning the mapping relationship between the sub-basin static attribute data and the parameters of the lumped hydrological model; express Parameters; It represents the discharge of sub-basin i at time t after the outflow through the river channel or river section converges at the hydrological observation station; RM is the river channel confluence model or river section confluence model; are the parameters of the river channel confluence model or river reach confluence model; Represents static attribute data of a river channel or river section; A feedforward neural network for learning the mapping relationship between static attribute data of a river channel or a river section and parameters of a river channel confluence model or a river section confluence model; express Parameters; represents the simulated flow at the hydrological observation station at time t.

3. The flood forecasting method of coupling physical mechanism and deep learning according to claim 2 is characterized by: Sub-basin The river course to the downstream hydrological observation station is recorded as the river course When using the Muskingum confluence calculation method, each river channel needs to be Divided into River sections of equal length, river section index satisfy , the outflow of the sub-basin is calculated from upstream to downstream, and for the upstream river section , where the inflow is the outflow of the corresponding sub-basin i ,reach The outflow is calculated as follows: ， Where, and Respectively represent river channels Middle River Section exist and The outflow of time, and Represents sub-basins exist and The outflow of time, Indicates river channel The river section exist Lateral inflow at the moment; for the downstream river section , , whose inflow is equal to the adjacent upstream reach The outflow of the river section The outflow is calculated as follows: ； Where, and Respectively represent river channels Middle River Section exist and The outflow of the moment; Danghe section When , calculate according to the following formula Moment sub-basin The flow at the hydrological observation station after the outflow passes through the river section : ； Where, for Moment sub-basin The outflow is the flow at the hydrological observation station after the confluence of the river section; Parameters C1, C2, C3 and C4 are calculated using the following formulas: ， Where, Indicates river channel The river section The storage time constant, Indicates river channel The river section The weight factor of Δt represents the calculation time step; if the river section Located in a sub-basin Inside, here , is the total number of sub-basins, then The runoff will be distributed to the river section as lateral inflow , Indicates river channel The river section exist The specific calculation formula for the lateral inflow at the moment is as follows: ， Where, Represents a subbasin Middle River Section The proportionality coefficient, Represents a subbasin At the moment The outflow, Represents a subbasin Middle River Section Average catchment area controlled, Represents a subbasin area; The parameters and Adaptive learning by feedforward neural network FNN2, whose value is determined by the river section The specific calculation formula is as follows: ， Where, Indicates river channel The river section Static attribute data.

4. The flood forecasting method of coupling physical mechanism and deep learning according to claim 2 is characterized by: Sub-basin The river course to the downstream hydrological observation station is recorded as the river course When the diffusion wave confluence method is used, the movement of flood in the river channel is described by the convection-diffusion equation, and combined with the initial conditions and boundary conditions, we get Moment sub-basin The flow calculation formula of the outflow at the hydrological observation station is as follows: ， Where, for Moment sub-basin The outflow is the flow at the hydrological observation station after the confluence of the river section. Indicates river channel exist The inflow at time i is also the outflow of subbasin i; Indicates that river channel i is The impulse response function corresponding to time is, For the river length; Indicates the maximum time step considering historical inflow, which is a hyperparameter; parameter and Adaptive learning by feedforward neural network FNN2, whose value is determined by the river The specific calculation formula is as follows: ， Where, Indicates river channel Static attribute data.

5. The flood forecasting method of coupling physical mechanism and deep learning according to claim 3 or 4, characterized in that: The loss function Loss of the confluence calculation model is: ， ， Where, is the number of the hydrological observation station, N is the total number of hydrological observation stations, NSE is the Nash efficiency coefficient, Indicates the moment, represents the length of the time series, Represents a subbasin The observed flow rate, Represents a subbasin The simulated flow rate, Represents a subbasin The average value of the observed flow.

6. The flood forecasting method of coupling physical mechanism and deep learning according to claim 2 is characterized by: When using the convolutional neural network confluence method, each river channel has unique static properties, corresponding to a different convolution kernel. Moment sub-basin The flow at the hydrological observation station after the outflow passes through the river confluence The calculation formula is: ， Where, Indicates river channel The corresponding convolution kernel weights, Indicates the position of each weight, , represents the total number of weights, Indicates river channel Static attribute data of Represents a subbasin exist The outflow of time.

7. The flood forecasting method of coupling physical mechanism and deep learning according to claim 6 is characterized by: The loss function Loss of the confluence calculation model is: ， ， Where, and Represent the coefficients of each component respectively, and the sum of the two is 1; and Respectively represent river channels The corresponding position in the convolution kernel and The weight of represents the total number of sub-basins; is the activation function; Represents the index of the convolution kernel with the largest weight, ; Indicates river channel The first-order difference of the weights in the corresponding convolution kernel; is the number of the hydrological observation station, N is the total number of hydrological observation stations, and NSE represents the Nash efficiency coefficient.

8. The flood forecasting method of coupling physical mechanism and deep learning according to claim 1 is characterized by: The training process of the distributed hybrid hydrological model includes: In the pre-training stage, the distributed hybrid hydrological model is pre-trained using daily-scale hydrological and meteorological data, watershed static attribute data, and river network static attribute data to obtain a pre-trained model; In the fine-tuning stage, the pre-trained model is trained using hourly hydrological and meteorological data, watershed static attribute data, and river network static attribute data to obtain the final distributed hybrid hydrological model.

9. A flood forecasting device that couples physical mechanisms and deep learning, characterized in that: include: Data acquisition module, used to obtain DEM data, hydrological and meteorological data and static attribute data of the target watershed; A sub-basin division module is used to divide the target basin into a plurality of sub-basins based on the DEM data, extract the river network, and calculate the static attribute data of the river network, wherein the river network includes a plurality of river channels; The forecasting module is used to input the hydrological and meteorological data of each sub-basin, the static attribute data of the basin, and the static attribute data of the river network into a pre-trained distributed hybrid hydrological model to obtain the simulated flow of the hydrological observation station and complete flood forecasting; the distributed hybrid hydrological model couples physical mechanisms and deep learning.

10. A flood forecasting system that couples physical mechanisms and deep learning, characterized by: including storage media and processors; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the method according to any one of claims 1-8.

Images