Methods, apparatus and systems for predicting differentiable mixing floods with adaptive runoff generation mechanisms
The differentiable hybrid flood forecasting method based on adaptive runoff generation mechanism utilizes neural networks and physical constraints to adaptively discriminate runoff generation patterns and inference parameters, constructing a multi-mechanism runoff generation framework. This solves the applicability problem of cross-regional flood forecasting and achieves high-precision and stable flood forecasting results.
Patent Information
- Application Number
- CN202610933736.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-26
- Publication Date
- 2026-11-10
AI Technical Summary
Existing flood forecasting methods are not applicable in cross-regional and heterogeneous watersheds. Traditional physical models are insufficient in characterizing the spatial heterogeneity of watershed runoff generation mechanisms, data-driven models lack physical constraints and have poor generalization ability, and hybrid models lack adaptive adjustment capabilities.
A differentiable mixed flood forecasting method with an adaptive runoff generation mechanism is proposed. This method uses a runoff generation pattern recognition neural network to adaptively identify the dominant runoff generation pattern, and combines it with a parameter inference neural network to dynamically infer hydrological parameters under physical constraints. A multi-mechanism physical runoff generation framework is constructed, and an adaptive confluence module is introduced to achieve end-to-end differentiable computation.
It significantly improves the accuracy and generalization ability of flood forecasting across regions and heterogeneous watersheds, maintains good physical interpretability, and is suitable for real-time flood forecasting and hydrological analysis of watersheds with no or little data, thus having high engineering application value.
Smart Images

Figure CN123072145A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary fields of hydrology and water resources engineering, disaster prevention and mitigation and artificial intelligence, and specifically relates to a method, device and system for forecasting differentially mixed floods with an adaptive runoff generation mechanism. Background Technology
[0002] Flood disasters are among the most frequent, widespread, and destructive natural disasters globally. With the increasing frequency of extreme weather events, flood events are characterized by their suddenness and complex causative mechanisms, placing higher demands on watershed flood control and disaster reduction, water resource allocation, and the safe operation of water conservancy projects. Therefore, constructing high-accuracy, stable, and adaptable flood forecasting models is one of the core issues of long-term concern in the field of hydrology and water resources. Existing flood forecasting methods can be broadly classified into two categories: hydrological models based on physical mechanisms and data-driven statistical or deep learning models.
[0003] Physically based hydrological models typically simulate flow by conceptualizing or semi-physically describing hydrological processes such as rainfall, infiltration, runoff generation, and confluence, such as the Xin'anjiang model and the HBV model. These models offer good physical interpretability, with clearly defined parameters, and can achieve relatively stable forecasting results when data is readily available and the model assumptions match the actual hydrological processes in the watershed. However, these models often rely on a single or fixed runoff generation mechanism assumption. For example, the Xin'anjiang model, which assumes full-storage runoff generation, performs well in humid regions, but its simulation accuracy and stability significantly decrease in arid and semi-arid watersheds dominated by infiltration runoff. This insufficient characterization of the spatial heterogeneity of watershed runoff generation mechanisms limits the applicability of traditional physical models in large-scale, multi-type watersheds.
[0004] With the development of remote sensing observation and computing capabilities, data-driven flood forecasting methods, especially deep learning models (such as Long Short-Term Memory networks), can automatically extract nonlinear mapping relationships from large amounts of historical data, to some extent compensating for the errors caused by the simplification of traditional model structures. However, purely data-driven models usually lack explicit constraints on hydrophysical processes, their internal calculation processes are difficult to interpret, and model parameters lack clear physical meaning. When applied to data-scarce watersheds, cross-regional extrapolation, or extreme climate scenarios, the model's generalization ability and robustness often decrease significantly, and it is difficult to output physically meaningful intermediate hydrological state variables.
[0005] To leverage the strengths of both physical and data-driven models, a new class of hybrid hydrological models has emerged in recent years. These models incorporate machine learning methods to optimize the parameters or local processes of traditional hydrological models. While these methods have improved model performance to some extent, most hybrid models still rely on a single, fixed physical process structure, merely learning or modifying parameters. The model structure itself lacks adaptive adjustment capabilities, making it difficult to characterize the differences in runoff generation mechanisms across different climatic zones and underlying surface conditions. This structural rigidity is particularly pronounced in cross-basin, large-sample application scenarios.
[0006] Therefore, how to introduce a differentiable learning mechanism while maintaining hydrophysical consistency, so that the model can automatically characterize and adapt to different runoff generation mechanisms according to watershed characteristics and achieve end-to-end joint optimization, remains a key technical problem that urgently needs to be solved in the field of flood forecasting. Summary of the Invention
[0007] To address the aforementioned issues, this invention proposes a method, apparatus, and system for forecasting differentiable mixed floods with an adaptive runoff generation mechanism. While maintaining good physical interpretability, it significantly improves the accuracy and generalization ability of flood forecasting across regions and heterogeneous watersheds.
[0008] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:
[0009] In a first aspect, the present invention provides a method for forecasting differentiable mixing floods with an adaptive runoff generation mechanism, characterized in that it includes:
[0010] The acquired hydrological and meteorological forcing data and static attribute data of the target watershed are fed into a pre-trained differentiable mixing hydrological model to obtain the simulated flow at the outlet section of the target watershed.
[0011] Flood forecasting based on simulated flow rates;
[0012] The differentiable mixing hydrological model includes, in sequence, a runoff pattern recognition neural network, a parameter inference neural network, a multi-mechanism physical runoff skeleton, and an adaptive confluence module, which are configured to perform the following steps on the received data:
[0013] By using a runoff pattern recognition neural network, the probability distribution of a target watershed under different runoff patterns is calculated based on static attribute data of the watershed, and the dominant runoff pattern of the target watershed is adaptively determined according to the principle of maximizing probability.
[0014] By using a parameter inference neural network, watershed static attribute data are mapped to physical hydrological parameters, and activation functions are used to constrain the physical hydrological parameters within a reasonable physical range.
[0015] Based on the identified dominant runoff generation patterns and physical hydrological parameters of the target watershed, the runoff is obtained by hourly step-by-hour calculation of the hydro-meteorological forcing data using a multi-mechanism physical runoff generation framework.
[0016] An adaptive confluence module is used to generate a confluence core with physical constraints based on the static properties of the watershed, and the runoff is converted into the simulated flow at the outlet section of the target watershed based on the confluence core.
[0017] Secondly, the present invention provides a differentiable mixing flood forecasting device with an adaptive runoff generation mechanism, comprising:
[0018] The simulated flow prediction module is used to input the acquired hydrological and meteorological forcing data and static attribute data of the target watershed into a pre-trained differentiable mixed hydrological model to obtain the simulated flow at the outlet section of the target watershed.
[0019] The flood forecasting module is used to forecast floods based on simulated flow rates;
[0020] The differentiable mixing hydrological model includes, in sequence, a runoff pattern recognition neural network, a parameter inference neural network, a multi-mechanism physical runoff skeleton, and an adaptive confluence module, which are configured to perform the following steps on the received data:
[0021] By using a runoff pattern recognition neural network, the probability distribution of a target watershed under different runoff patterns is calculated based on static attribute data of the watershed, and the dominant runoff pattern of the target watershed is adaptively determined according to the principle of maximizing probability.
[0022] By using a parameter inference neural network, watershed static attribute data are mapped to physical hydrological parameters, and activation functions are used to constrain the physical hydrological parameters within a reasonable physical range.
[0023] Based on the identified dominant runoff generation patterns and physical hydrological parameters of the target watershed, the runoff is obtained by hourly step-by-hour calculation of the hydro-meteorological forcing data using a multi-mechanism physical runoff generation framework.
[0024] An adaptive confluence module is used to generate a confluence core with physical constraints based on the static properties of the watershed, and the runoff is converted into the simulated flow at the outlet section of the target watershed based on the confluence core.
[0025] Thirdly, the present invention provides a differentiable mixing flood forecasting system with an adaptive flow generation mechanism, including a storage medium and a processor;
[0026] The storage medium is used to store instructions;
[0027] The processor is configured to operate according to the instructions to perform the method according to any one of the first aspects.
[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0029] This invention utilizes static attribute data of a watershed to adaptively identify the dominant runoff generation patterns of a target watershed through a runoff generation pattern recognition neural network. It then combines this with a parameter inference neural network to dynamically infer physical hydrological parameters under physical constraints, constructing a multi-mechanism physical runoff generation framework that includes infiltration-excess runoff, storage-saturation runoff, and mixed runoff. Simultaneously, an adaptive confluence module with physical consistency constraints is introduced to achieve end-to-end differentiability of simulated flow from runoff to the watershed outlet section. While maintaining good physical interpretability, this significantly improves the accuracy and generalization ability of flood forecasting across regions and heterogeneous watersheds. It can be effectively applied to real-time flood forecasting, flood control scheduling, and hydrological analysis of watersheds with little or no data, demonstrating high engineering application value and promising prospects for wider application.
[0030] Furthermore, this invention can greatly improve forecast accuracy by jointly training a differentiable hybrid hydrological model using a composite loss function that includes multi-scale error terms and physical constraint regularization terms. Attached Figure Description
[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:
[0032] Figure 1 This is a schematic diagram of the structure of a hybrid hydrological model with an adaptive runoff generation mode according to an embodiment of the present invention;
[0033] Figure 2 This is a comparison of a pure data-driven model and eight hybrid hydrological models according to one embodiment of the present invention;
[0034] Figure 3 This diagram illustrates the comparison of the flood simulation pass rates of a hybrid hydrological model and a pure data-driven model in data-free watersheds and data-free areas, according to an embodiment of the present invention. Detailed Implementation
[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0036] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.
[0037] Example 1
[0038] This invention provides a method for forecasting differentiable mixing floods with an adaptive runoff generation mechanism, comprising the following steps:
[0039] (1) Input the obtained hydrological and meteorological forcing data and static attribute data of the target watershed into the pre-trained differentiable mixed hydrological model to obtain the simulated flow at the outlet section of the target watershed;
[0040] (2) Flood forecasting based on simulated flow rates;
[0041] The differentiable mixing hydrological model includes, in sequence, a runoff pattern recognition neural network, a parameter inference neural network, a multi-mechanism physical runoff skeleton, and an adaptive confluence module, which are configured to perform the following steps on the received data:
[0042] By using a runoff pattern recognition neural network, the probability distribution of a target watershed under different runoff patterns is calculated based on static attribute data of the watershed, and the dominant runoff pattern of the target watershed is adaptively determined according to the principle of maximizing probability.
[0043] By using a parameter inference neural network, watershed static attribute data are mapped to physical hydrological parameters, and activation functions are used to constrain the physical hydrological parameters within a reasonable physical range.
[0044] Based on the identified dominant runoff generation patterns and physical hydrological parameters of the target watershed, the runoff is obtained by hourly step-by-hour calculation of the hydro-meteorological forcing data using a multi-mechanism physical runoff generation framework.
[0045] An adaptive confluence module is used to generate a confluence core with physical constraints based on the static properties of the watershed, and the runoff is converted into the simulated flow at the outlet section of the target watershed based on the confluence core.
[0046] The above scheme utilizes static attribute data of the watershed and adaptively identifies the dominant runoff generation patterns of the target watershed through a runoff generation pattern recognition neural network. It then combines this with a parameter inference neural network to dynamically infer physical hydrological parameters under physical constraints, constructing a multi-mechanism physical runoff generation framework that includes infiltration-excess runoff, storage-saturation runoff, and mixed runoff. Simultaneously, an adaptive confluence module with physical consistency constraints is introduced to achieve end-to-end differentiability of simulated flow from runoff to the watershed outlet section. While maintaining good physical interpretability, this significantly improves the accuracy and generalization ability of flood forecasting across regions and heterogeneous watersheds. It can be effectively applied to real-time flood forecasting, flood control scheduling, and hydrological analysis of watersheds with little or no data, demonstrating high engineering application value and promising prospects for wider application.
[0047] In one specific embodiment of the present invention, the formula used to determine the dominant flow generation mechanism is:
[0048] ,
[0049] In the formula, The dominant runoff generation mode of watershed b is represented by IE, the super-permeable runoff generation mode is represented by SE, the full-saturation runoff generation mode is represented by MIX, and the mixed runoff generation mode is represented by MIX. Represents the normalized static attribute vector of watershed b; This represents a flow pattern recognition neural network, which is a feedforward neural network structure containing two hidden layers. Its output layer uses the Softmax activation function to generate the probability of each pattern. Indicates finding The parameter value that yields the maximum value.
[0050] In one specific embodiment of the present invention, the calculation formula used when mapping watershed static attribute data to physical hydrological parameters is as follows:
[0051] ,
[0052] ,
[0053] In the formula, Represents a watershed The advantageous flow modes are: IE represents the super-permeable flow mode, SE represents the full-capacity flow mode, and MIX represents the mixed flow mode. This represents a parameter inference network, which is a feedforward neural network containing two hidden layers. Its output layer uses the sigmoid activation function to ensure that the output of the parameter inference network is between 0 and 1. Represents a watershed The normalized static attribute vector of the watershed; This indicates the dominant flow generation pattern output by the parameter inference network. Lower basin The set of coefficients for physical hydrological parameters; Indicating the dominant production flow model Lower basin The One physical hydrological parameter; for The elements in the text represent the dominant production flow pattern. Lower basin The parameters are used to infer the coefficients of the p-th physical hydrological parameter in the original output of the neural network. and These represent the dominant production flow model. Next The lower and upper boundaries of a physical hydrological parameter.
[0054] In one specific embodiment of the present invention, when the dominant runoff generation mode is the super-permeable runoff generation mode, the formula used to calculate the runoff is:
[0055] ,
[0056] ,
[0057] In the formula, The average infiltration capacity of the watershed; To stabilize the infiltration rate; The infiltration sensitivity coefficient; This represents the average soil water storage capacity of the watershed. This represents the current soil water storage level; This refers to runoff under the super-permeable flow pattern; Effective rainfall data comes from hydrometeorological forcing. The shape parameter represents the infiltration distribution.
[0058] In one specific embodiment of the present invention, when the dominant runoff generation model is the full runoff generation mode, the formula used to calculate the runoff is:
[0059] ,
[0060] ,
[0061] In the formula, Current soil water storage The corresponding soil single-point water storage capacity value; This represents the average soil water storage capacity of the watershed. The shape parameter for water storage distribution; To store runoff in the runoff generation mode; Effective rainfall data comes from hydrometeorological forcing.
[0062] In one specific embodiment of the present invention, when the dominant runoff generation module is in a mixed runoff generation mode, the formula used to calculate runoff is:
[0063] ,
[0064] ,
[0065] ,
[0066] ,
[0067] ,
[0068] In the formula, The average infiltration capacity of the watershed; To stabilize the infiltration rate; The infiltration sensitivity coefficient; This represents the average soil water storage capacity of the watershed. This represents the current soil water storage level; This refers to runoff under the super-permeable flow pattern; Effective rainfall data comes from hydrometeorological forcing. The shape parameter of the infiltration distribution. Current soil water storage The corresponding soil single-point water storage capacity value; The shape parameter for water storage distribution; To store runoff in the runoff generation mode; This refers to runoff under a mixed runoff generation pattern.
[0069] In one specific embodiment of the present invention, the adaptive merging module employs a physically constrained one-dimensional convolutional neural network, and the formula for calculating the simulated flow rate is:
[0070] ,
[0071] ,
[0072] In the formula, Represents a watershed The length is The confluence core; The spool parameter mapping network is a feedforward neural network containing two hidden layers. Its output layer uses the Softmax activation function to ensure that the output of the parameter inference network is between 0 and 1, and that the sum is 1. Represents a watershed The normalized static attribute vector of the watershed; express The basin is Simulated flow rate at the outlet section at any given time; for The element represents the time lag step. Corresponding confluence weights; For the basin exist The flow of time.
[0073] In one specific embodiment of the present invention, the composite loss function used in the training process of the differentiable hybrid hydrological model includes a multi-scale error term and a physical constraint regularization term. The expression of the composite loss function is as follows:
[0074] ,
[0075] ,
[0076] ,
[0077] In the formula, For composite loss function, For the target loss item; As a single-peak shape constraint term, it forces the watershed to... The confluence nucleus exhibits a single-peak pattern of first rising and then falling. For balance coefficient, This represents the total number of river basins. This indicates the calculation of the root mean square error. Represents a watershed The observed flow sequence at the outlet section; Represents a watershed The simulated flow sequence at the outlet section consists of several constitute; It is a smoothing factor; This is the offset; For the basin The peak index of the confluence core; Indicates the length of the bus core; It is an activation function; for The element represents the time lag step. Corresponding confluence weights; for The element represents the time lag step. The corresponding confluence weight.
[0078] In the above scheme, a differentiable hybrid hydrological model is jointly trained by a composite loss function that includes multi-scale error terms and physical constraint regularization terms. While maintaining good physical interpretability, this invention significantly improves the accuracy and generalization ability of flood forecasting across regions and heterogeneous watersheds.
[0079] To further illustrate the applicability, stability, and engineering feasibility of the differentially mixed hydrological flood forecasting method with adaptive runoff generation mechanism proposed in this invention under complex hydrological conditions, this embodiment provides a detailed explanation of the specific application process of the method in cross-regional flood forecasting tasks based on large-sample, multi-climate zone actual hydrological data.
[0080] In practical flood forecasting operations, the following technical challenges are frequently encountered: significant differences in climate conditions and underlying surface characteristics among different watersheds make it difficult to universally apply a single runoff generation model; many watersheds lack long-term or continuous flow observation data, making traditional parameter calibration methods difficult to implement; purely data-driven models are prone to structural failure or extreme misjudgments when applied across regions; and hourly-scale flood forecasting places higher demands on model structural stability and physical consistency. To address these issues, this embodiment applies the method proposed in this invention to a large sample set of watersheds covering multiple hydrological and climatic zones to verify its practical effectiveness in flood forecasting tasks in watersheds with available data, watersheds without data, and watersheds spanning different climatic zones.
[0081] This embodiment selects a large sample watershed set covering multiple hydro-climate zones in North America as the application object. This large sample watershed set contains hundreds to thousands of independent watersheds, with watershed areas ranging from tens to thousands of square kilometers, covering various hydrological types such as humid, semi-humid, semi-arid, and arid regions.
[0082] For the aforementioned watershed, the following data were obtained as input and validation data for the differentiable mixing hydrological model:
[0083] Hourly-scale hydrometeorological forcing data: including hourly precipitation, temperature and potential evapotranspiration data, used to drive calculations of hydrological runoff generation and confluence processes.
[0084] Static watershed attribute data includes, but is not limited to: topographic attributes such as watershed area, average slope, river network density, and channel classification; soil attributes such as soil thickness, soil texture, and available water content; climate background attributes such as annual average precipitation and climate aridity / humidity index; and land use types and the distribution characteristics of water conservancy projects. These static attributes are used to characterize differences in watershed runoff patterns and serve as inputs to the differentiable mixing hydrological model in this invention.
[0085] Flow observation data: For watersheds with observation conditions, hourly outlet cross-sectional flow data are obtained for training of differentiable mixing hydrological models and verification of application effects.
[0086] Based on historical hourly flow sequences of each watershed, flood events are automatically identified and constructed. Specifically, this includes: statistically analyzing the historical flow distribution characteristics of each watershed; using a relative threshold method to identify significant flood events and ensuring sufficient time intervals between adjacent flood events; determining the start and end times of each flood event and removing events with severe data gaps or abnormal patterns; and setting a preheating period before each flood event to initialize state variables such as soil moisture content within the differentiable mixing hydrological model. Through these steps, a large-sample flood dataset containing tens of thousands of flood events is constructed, serving as the application basis for the method in this embodiment of the invention.
[0087] Based on the differentiable mixing flood forecasting method proposed in this embodiment of the invention, a differentiable mixing hydrological model with an end-to-end differentiable structure is constructed. For example... Figure 1 As shown, the differentiable mixing hydrological model includes a runoff pattern recognition neural network, a parameter inference neural network, a multi-mechanism physical runoff skeleton, and an adaptive confluence module arranged sequentially; the differentiable mixing hydrological model includes the following core functional modules:
[0088] Watershed static attribute data are input into a runoff pattern recognition neural network to determine dominant runoff patterns. The specific formula is as follows:
[0089] ,
[0090] In the formula, The dominant runoff generation mode of watershed b is represented by IE, the super-permeable runoff generation mode is represented by SE, the full-saturation runoff generation mode is represented by MIX, and the mixed runoff generation mode is represented by MIX. Represents the normalized static attribute vector of watershed b; This represents a neural network for generating flow patterns. The network uses a network architecture with two hidden layers, with 256 and 128 neurons in the two hidden layers, respectively. Its output layer uses the Softmax activation function to generate the probability of each pattern. Indicates finding The parameter value that yields the maximum value.
[0091] Furthermore, a parameter inference neural network is used to map the static attribute data of the watershed to physical hydrological parameters, and the physical hydrological parameters are constrained within a physically reasonable range by an activation function. The formulas used include:
[0092] ,
[0093] ,
[0094] In the formula, Represents a watershed The advantageous flow modes are: IE represents the super-permeable flow mode, SE represents the full-capacity flow mode, and MIX represents the mixed flow mode. This represents a parameter inference network, which is a feedforward neural network containing two hidden layers with 256 and 128 neurons respectively. Its output layer uses the sigmoid activation function to ensure that the output of the parameter inference network is between 0 and 1. Represents a watershed The normalized static attribute vector of the watershed; This indicates the dominant flow generation pattern output by the parameter inference network. Lower basin The set of coefficients for physical hydrological parameters; Indicating the dominant production flow model Lower basin The One physical hydrological parameter; for The elements in the text represent the dominant production flow pattern. Lower basin The parameters are used to infer the coefficients of the p-th physical hydrological parameter in the original output of the neural network. and These represent the dominant production flow model. Next The lower and upper boundaries of a physical hydrological parameter.
[0095] Based on the dominant runoff generation model and physical-hydrological parameters determined in the above two steps, runoff calculation is performed. When the dominant runoff generation model is the excess runoff generation model (IE), the formula used to calculate runoff is:
[0096] ,
[0097] ,
[0098] In the formula, The average infiltration capacity of the watershed; To stabilize the infiltration rate; The infiltration sensitivity coefficient; This represents the average soil water storage capacity of the watershed. This represents the current soil water storage level; This refers to runoff under the super-permeable flow pattern; Effective rainfall data comes from hydrometeorological forcing. The shape parameter represents the infiltration distribution.
[0099] When the dominant runoff pattern is the full-storage runoff pattern (SE), the runoff is calculated using the water storage capacity area distribution curve. The formula used to calculate the runoff is:
[0100] ,
[0101] ,
[0102] In the formula, Current soil water storage The corresponding soil single-point water storage capacity value; This represents the average soil water storage capacity of the watershed. The shape parameter for water storage distribution; To store runoff in the runoff generation mode; Effective rainfall data comes from hydrometeorological forcing.
[0103] When the dominant runoff pattern is a mixed runoff pattern (MIX), the formula used to calculate runoff is:
[0104] ,
[0105] ,
[0106] ,
[0107] ,
[0108] ,
[0109] In the formula, The average infiltration capacity of the watershed; To stabilize the infiltration rate; The infiltration sensitivity coefficient; This represents the average soil water storage capacity of the watershed. This represents the current soil water storage level; This refers to runoff under the super-permeable flow pattern; Effective rainfall data comes from hydrometeorological forcing. The shape parameter of the infiltration distribution. Current soil water storage The corresponding soil single-point water storage capacity value; The shape parameter for water storage distribution; To store runoff in the runoff generation mode; This refers to runoff under a mixed runoff generation pattern.
[0110] After the runoff calculation is completed, the static attributes of the watershed are further input into the adaptive flow-collecting module. The adaptive flow-collecting module uses a physically constrained flow-collecting parameter mapping network. The formula for calculating the simulated flow at the outlet section is as follows:
[0111] ,
[0112] ,
[0113] In the formula, Represents a watershed The length is The confluence core, The spool parameter mapping network is a feedforward neural network containing two hidden layers. Its output layer uses the Softmax activation function to ensure that the output of the parameter inference network is between 0 and 1, and that the sum is 1. Represents a watershed The normalized static attribute vector of the watershed; express The basin is Simulated flow rate at the outlet section at any given time; for The element represents the time lag step. Corresponding confluence weights; For the basin exist The flow of time.
[0114] Furthermore, a composite loss function is constructed based on the simulated flow rate at the outlet section, the measured flow rate at the outlet section, and physical constraints. :
[0115] ,
[0116] ,
[0117] ,
[0118] In the formula, For composite loss function, For the target loss item; As a single-peak shape constraint term, it forces the watershed to... The confluence nucleus exhibits a single-peak pattern of first rising and then falling. For balance coefficient, This represents the total number of river basins. This indicates the calculation of the root mean square error. Represents a watershed The observed flow sequence at the outlet section; Represents a watershed The simulated flow sequence at the outlet section consists of several constitute; It is a smoothing factor; This is the offset; For the basin The peak index of the confluence core; Indicates the length of the bus core; It is an activation function used to ensure that the output value is greater than 0; for The element represents the time lag step. Corresponding confluence weights; for The element represents the time lag step. The corresponding confluence weight.
[0119] During the training phase of the differentiable mixed hydrological model, a regional joint training approach was adopted to simultaneously optimize model parameters on flood event samples from multiple watersheds. A composite loss function, including simulated flow error terms and physical constraint regularization terms, was constructed to jointly train the differentiable mixed hydrological model. Based on the backpropagation algorithm, the parameters of the differentiable mixed hydrological model were updated according to the loss function until the accuracy met the expected standard.
[0120] To further verify the technical advantages of the differentiable mixing flood forecasting method with adaptive runoff generation mechanism proposed in this embodiment of the invention compared with the pure data-driven flood forecasting model, a comparative application was conducted between the differentiable mixing flood forecasting method with adaptive runoff generation mechanism and the data-driven model without explicit physical constraints under the same data conditions and application scenarios. The data-driven model uses hydrological and meteorological forcing data as the main input and generates flood forecasting results by learning the mapping relationship between historical inputs and flow rates, but does not explicitly introduce runoff generation mechanisms, physical parameter constraints, or descriptions of confluence processes.
[0121] (1) Comparative application results under the condition of available watershed data, such as Figure 2 As shown. Given historical flow observation data for the target watershed, both models can be trained using historical samples and generate flood forecast results. Comparative application results show that the data-driven model can fit the overall trend of the flood process within the training data coverage. In contrast, the method of this invention, under the same data conditions, can maintain the structural consistency of the flood process through physical runoff generation and confluence constraints. The simulation results exhibit more stable evolutionary characteristics in the flood peak, rise and fall processes, and recession stages, reducing non-physical oscillations caused by purely statistical fitting. Figure 2In this context, LSTM represents the data-driven model, HM-ISM-CNN is the differentiable mixed hydrological model proposed in this embodiment, and the other seven (HM-ISM-UH, HM-MIX-CNN, HM-MIX-UH, HM-SE-CNN, HM-SE-UH, HM-IE-CNN, HM-IE-UH) are comparative models. HM represents the mixed hydrological model, ISM represents the adaptive runoff generation mode, IE, SE and MIX represent the infiltration runoff generation mode, the full-storage runoff generation mode and the mixed runoff generation mode, respectively, UH represents the traditional method of river confluence based on unit hydrograph, and CNN represents the river confluence method proposed in this embodiment, which adaptively outputs the confluence kernel based on watershed static attribute data and performs convolution calculation (corresponding to the adaptive confluence module in this embodiment). Figure 2 The results show the success rate of each model in flood simulations in different climate zones under regionalized training conditions.
[0122] (2) Figure 3 The difference in the pass rate of flood simulation between the differentiable hybrid hydrological model HM-ISM-CNN and the pure data-driven model in the application scenarios of data-free watersheds and data-free areas proposed in this embodiment of the invention is denoted as ΔQR. Figure 3 (a) and Figure 3 (b) shows the spatial distribution of ΔQR and its corresponding bar chart statistics in the data-free watershed scenario; Figure 3 (c) and Figure 3 (d) shows the spatial distribution of ΔQR and its corresponding bar chart statistics in the data-free region scenario. The comparative application results under the data-free watershed condition are as follows: Figure 3 As shown in (a)-(b) of the paper, in application scenarios where no historical flow observation data is available for the target watershed, the data-driven model, lacking effective supervision information, relies heavily on the distribution of training samples for its forecast results, easily leading to problems such as abnormal flood response amplitude or time scale mismatch. Comparative application results show that, under data-free watershed conditions, the flood process generated by the data-driven model deviates from the actual hydrological response patterns of the watershed in some cases, making it difficult to guarantee the engineering reliability of the forecast results. The method of this invention, in this scenario, can automatically identify runoff generation patterns and configure physical parameters based on the static attributes of the watershed, and can still output physically reasonable flood hydrographs even without flow observation constraints. These comparative results demonstrate that the method of this invention has stronger generalization ability and stability compared to pure data-driven models in data-free watershed applications.
[0123] (3) Comparative application results under cross-regional application conditions, such as Figure 3As shown in (c)-(d) of the diagram. In cross-regional application scenarios, the hydrological and climatic conditions of the target watershed differ significantly from those of the model training area. Data-driven models, due to their internal mapping relationships being highly dependent on the distribution of training samples, are prone to structural failures in this scenario, such as significant shifts in flood peak locations or overly smoothed flood processes. Comparative application results demonstrate that the method of this invention can dynamically adjust the runoff generation structure and confluence characteristics based on the static attributes of the watershed, effectively mitigating model mismatch problems caused by regional differences. Compared to data-driven models, the method in this embodiment generates more continuous and stable flood processes under cross-regional conditions, significantly reducing extreme anomalies and demonstrating higher engineering applicability.
[0124] The comparative results show that, compared to data-driven flood forecasting models without explicit physical constraints, the differentiable hybrid flood forecasting method with adaptive runoff generation mechanism proposed in this invention exhibits more stable forecasting behavior under different data availability and regional conditions. By introducing differentiable physical runoff generation and confluence mechanisms into the data-driven learning process and achieving adaptive adjustment of model structure and parameters, this invention effectively compensates for the shortcomings of pure data-driven models in terms of generalization ability, physical consistency, and engineering reliability, further enhancing its application value in practical flood forecasting operations.
[0125] Example 2
[0126] This invention provides a differentiable mixing flood forecasting device with an adaptive runoff generation mechanism, comprising:
[0127] The simulated flow prediction module is used to input the acquired hydrological and meteorological forcing data and static attribute data of the target watershed into a pre-trained differentiable mixed hydrological model to obtain the simulated flow at the outlet section of the target watershed.
[0128] The flood forecasting module is used to forecast floods based on simulated flow rates;
[0129] The differentiable mixing hydrological model includes, in sequence, a runoff pattern recognition neural network, a parameter inference neural network, a multi-mechanism physical runoff skeleton, and an adaptive confluence module, which are configured to perform the following steps on the received data:
[0130] By using a runoff pattern recognition neural network, the probability distribution of a target watershed under different runoff patterns is calculated based on static attribute data of the watershed, and the dominant runoff pattern of the target watershed is adaptively determined according to the principle of maximizing probability.
[0131] By using a parameter inference neural network, watershed static attribute data are mapped to physical hydrological parameters, and activation functions are used to constrain the physical hydrological parameters within a reasonable physical range.
[0132] Based on the identified dominant runoff generation patterns and physical hydrological parameters of the target watershed, the runoff is obtained by hourly step-by-hour calculation of the hydro-meteorological forcing data using a multi-mechanism physical runoff generation framework.
[0133] An adaptive confluence module is used to generate a confluence core with physical constraints based on the static properties of the watershed, and the runoff is converted into the simulated flow at the outlet section of the target watershed based on the confluence core.
[0134] The rest are the same as in Example 1.
[0135] Example 3
[0136] Based on the same invention as Embodiment 1, a differentiable mixing flood forecasting system with an adaptive flow generation mechanism is provided, characterized in that it includes a storage medium and a processor;
[0137] The storage medium is used to store instructions;
[0138] The processor is configured to operate according to the instructions to execute the method according to any one of Embodiment 1.
[0139] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0140] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0141] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0142] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0143] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
[0144] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A method for forecasting differentiable mixing floods with an adaptive runoff generation mechanism, characterized in that, include: The acquired hydrological and meteorological forcing data and static attribute data of the target watershed are fed into a pre-trained differentiable mixing hydrological model to obtain the simulated flow at the outlet section of the target watershed. Flood forecasting based on simulated flow rates; The differentiable mixing hydrological model includes, in sequence, a runoff pattern recognition neural network, a parameter inference neural network, a multi-mechanism physical runoff skeleton, and an adaptive confluence module, which are configured to perform the following steps on the received data: By using a runoff pattern recognition neural network, the probability distribution of a target watershed under different runoff patterns is calculated based on static attribute data of the watershed, and the dominant runoff pattern of the target watershed is adaptively determined according to the principle of maximizing probability. By using a parameter inference neural network, watershed static attribute data are mapped to physical hydrological parameters, and activation functions are used to constrain the physical hydrological parameters within a reasonable physical range. Based on the identified dominant runoff generation patterns and physical hydrological parameters of the target watershed, the runoff is obtained by hourly step-by-hour calculation of the hydro-meteorological forcing data using a multi-mechanism physical runoff generation framework. An adaptive confluence module is used to generate a confluence core with physical constraints based on the static properties of the watershed, and the runoff is converted into the simulated flow at the outlet section of the target watershed based on the confluence core.
2. The method for forecasting differentiable mixing floods with an adaptive runoff generation mechanism according to claim 1, characterized in that: The formula used to determine the dominant production flow pattern is: , In the formula, The dominant runoff generation mode of watershed b is represented by IE, the super-permeable runoff generation mode is represented by SE, the full-saturation runoff generation mode is represented by MIX, and the mixed runoff generation mode is represented by MIX. Represents the normalized static attribute vector of watershed b; This represents a flow pattern recognition neural network, which is a feedforward neural network structure containing two hidden layers. Its output layer uses the Softmax activation function to generate the probability of each pattern. Indicates finding The parameter value that yields the maximum value.
3. The method for forecasting differentiable mixing floods with an adaptive runoff generation mechanism according to claim 1, characterized in that: The calculation formula used when mapping watershed static attribute data to physical hydrological parameters is as follows: , , In the formula, Represents a watershed The advantageous flow modes are: IE represents the super-permeable flow mode, SE represents the full-capacity flow mode, and MIX represents the mixed flow mode. This represents a parameter inference network, which is a feedforward neural network containing two hidden layers and an output layer that uses... The activation function is used to ensure that the output of the parameter inference network is between 0 and 1; Represents a watershed The normalized static attribute vector of the watershed; This indicates the dominant flow generation pattern output by the parameter inference network. Lower basin The set of coefficients for physical hydrological parameters; Indicating the dominant production flow model Lower basin The One physical hydrological parameter; for The elements in the text represent the dominant production flow pattern. Lower basin The parameters are used to infer the coefficients of the p-th physical hydrological parameter in the original output of the neural network. and These represent the dominant production flow model. Next The lower and upper boundaries of a physical hydrological parameter.
4. The method for forecasting differentiable mixing floods with an adaptive runoff generation mechanism according to claim 1, characterized in that, When the dominant runoff pattern is the over-permeability runoff pattern, the formula used to calculate runoff is: , , In the formula, The average infiltration capacity of the watershed; To stabilize the infiltration rate; The infiltration sensitivity coefficient; This represents the average soil water storage capacity of the watershed. This represents the current soil water storage level; This refers to runoff under the super-permeable flow pattern; Effective rainfall data comes from hydrometeorological forcing. The shape parameter represents the infiltration distribution.
5. The method for forecasting differentiable mixing floods with an adaptive runoff generation mechanism according to claim 1, characterized in that: When the dominant runoff generation model is the full runoff generation model, the formula used to calculate runoff is: , , In the formula, Current soil water storage The corresponding soil single-point water storage capacity value; This represents the average soil water storage capacity of the watershed. The shape parameter for water storage distribution; To store runoff in the runoff generation mode; Effective rainfall data comes from hydrometeorological forcing.
6. The method for forecasting differentiable mixing floods with an adaptive runoff generation mechanism according to claim 1, characterized in that: When the dominant runoff generation module is a mixed runoff generation mode, the formula used to calculate runoff is: , , , , , In the formula, The average infiltration capacity of the watershed; To stabilize the infiltration rate; The infiltration sensitivity coefficient; This represents the average soil water storage capacity of the watershed. This represents the current soil water storage level; This refers to runoff under the super-permeable flow pattern; Effective rainfall data comes from hydrometeorological forcing. The shape parameter of the infiltration distribution. Current soil water storage The corresponding soil single-point water storage capacity value; The shape parameter for water storage distribution; To store runoff in the runoff generation mode; This refers to runoff under a mixed runoff generation pattern.
7. The method for forecasting differentiable mixing floods with an adaptive runoff generation mechanism according to claim 1, characterized in that, The adaptive merging module uses a physically constrained merging parameter mapping network, and the formula for calculating the simulated flow rate at the outlet section is as follows: , , In the formula, Represents a watershed The length is The confluence core; This is a confluence parameter mapping network, which is a feedforward neural network containing two hidden layers and its output layer uses... The activation function is used to ensure that the output of the parameter inference network is between 0 and 1, and that the sum is 1; Represents a watershed The normalized static attribute vector of the watershed; express The basin is Simulated flow rate at the outlet section at any given time; for The element represents the time lag step. Corresponding confluence weights; For the basin exist The flow of time.
8. A method for forecasting differentiable mixing floods with an adaptive runoff generation mechanism according to claim 7, characterized in that: The differentiable hybrid hydrological model uses a composite loss function during training that includes a multi-scale error term and a physical constraint regularization term. The expression for the composite loss function is as follows: , , , In the formula, For composite loss function, For the target loss item; As a single-peak shape constraint term, it forces the watershed to... The confluence nucleus exhibits a single-peak pattern of first rising and then falling. For balance coefficient, This represents the total number of river basins. This indicates the calculation of the root mean square error. Represents a watershed The observed flow sequence at the outlet section; Represents a watershed The simulated flow sequence at the outlet section consists of several constitute; It is a smoothing factor; This is the offset; For the basin The peak index of the confluence core; Indicates the length of the bus core; It is an activation function; for The element represents the time lag step. Corresponding confluence weights; for The element represents the time lag step. The corresponding confluence weight.
9. A differentiable mixing flood forecasting device with an adaptive runoff generation mechanism, characterized in that, include: The simulated flow prediction module is used to input the acquired hydrological and meteorological forcing data and static attribute data of the target watershed into a pre-trained differentiable mixed hydrological model to obtain the simulated flow at the outlet section of the target watershed. The flood forecasting module is used to forecast floods based on simulated flow rates; The differentiable mixing hydrological model includes, in sequence, a runoff pattern recognition neural network, a parameter inference neural network, a multi-mechanism physical runoff skeleton, and an adaptive confluence module, which are configured to perform the following steps on the received data: By using a runoff pattern recognition neural network, the probability distribution of a target watershed under different runoff patterns is calculated based on static attribute data of the watershed, and the dominant runoff pattern of the target watershed is adaptively determined according to the principle of maximizing probability. By using a parameter inference neural network, watershed static attribute data are mapped to physical hydrological parameters, and activation functions are used to constrain the physical hydrological parameters within a reasonable physical range. Based on the identified dominant runoff generation patterns and physical hydrological parameters of the target watershed, the runoff is obtained by hourly step-by-hour calculation of the hydro-meteorological forcing data using a multi-mechanism physical runoff generation framework. An adaptive confluence module is used to generate a confluence core with physical constraints based on the static properties of the watershed, and the runoff is converted into the simulated flow at the outlet section of the target watershed based on the confluence core.
10. A differentiable mixing flood forecasting system with an adaptive runoff generation mechanism, characterized in that, Including storage media and processor; 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.
