Mountainous small watershed flood forecasting method and system

Abstract

The application discloses a mountainous small watershed flood forecasting method and system, and relates to the technical field of intelligent water conservancy. The method comprises the following steps: acquiring multi-source hydro-meteorological data of a mountainous small watershed, and obtaining a multi-source heterogeneous data set through preprocessing. A long short-term memory network based on residual correction is constructed, the measured flow at the last moment is superimposed with the current moment prediction increment, and the current predicted flow is obtained. A physical constraint module is constructed based on the watershed water balance differential equation, and a physically consistent loss function is established with the change rate of water storage as an auxiliary variable. The data fitting error and the physical loss function are weighted and fused to train a residual physical constraint neural network model. Real-time monitoring data is used to deduce the watershed water storage dynamics, and flood forecasting is output. The application breaks the prediction inertia barrier of high-frequency data, eliminates the prediction lag, and actively provides a flood control safety margin when a sudden flood peak occurs, and realizes the reconstructability of the physical mechanism.

Description

Technical Field

[0001] This invention relates to the field of smart water conservancy technology, and in particular to a method and system for forecasting floods in small watersheds in mountainous areas. Background Technology

[0002] How to achieve high-precision real-time flood forecasting under conditions of scarce data and complex monitoring environments has become a key technical challenge in the fields of hydrological science and disaster prevention and mitigation.

[0003] Currently, flood forecasting methods are mainly divided into two categories: process-driven hydrophysical models and data-driven deep learning models. Traditional distributed physical models (such as HEC-HMS) are highly dependent on high-precision topographic (DEM) and soil texture parameters, making accurate parameter calibration difficult in data-scarce mountainous areas. In recent years, pure data-driven models, represented by Long Short-Term Memory (LSTM) networks, have been widely used in hydrological forecasting, but they also face three major technical bottlenecks: First, the models lack physical constraints and are essentially "black box" systems, which may lead to anomalies in the prediction results that violate the water balance law (such as negative flow), resulting in poor physical interpretability. Second, when facing sudden flash floods, due to the strong autocorrelation of high-frequency sampled data (such as hourly data), pure deep learning models are prone to falling into the "inertia trap" based on data translation, resulting in severe "phase lag" in flood peak forecasts and systematic underestimation and underreporting at flood peak extremes. Third, they are prone to overfitting when dealing with noisy data (such as satellite rainfall errors) or extreme conditions, resulting in weak generalization ability. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for forecasting floods in small watersheds in mountainous areas, aiming to solve or improve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for flood forecasting in small watersheds in mountainous areas includes:

[0007] Multi-source hydrological and meteorological data of small watersheds in mountainous areas are acquired and preprocessed to obtain a multi-source heterogeneous dataset; the multi-source hydrological and meteorological data includes ground observation data, remote sensing data, and atmospheric reanalysis data;

[0008] A time-series deep learning network model based on residual correction is constructed. The time-series deep learning network model is used to extract the time-series features of multi-source heterogeneous datasets. The measured cross-sectional flow at the previous time step and the predicted cross-sectional flow increment at the current time step are subjected to residual superposition processing to calculate the predicted flow at the current time step. The time-series deep learning network model is a long short-term memory network model.

[0009] A physical constraint module is constructed based on the macroscopic water balance differential equation of the watershed, and the rate of change of water storage is inferred. The rate of change of water storage is used as an auxiliary state variable to establish a physical consistency loss function.

[0010] The data fitting error between the predicted flow rate and the measured cross-sectional flow rate at the current moment is calculated. The data fitting error is then weighted and fused with the physical consistency loss function to construct a total objective function. The total objective function is used to perform dual-drive training on the time-series deep learning network model to generate a residual physical constraint neural network model. The data fitting error is calculated using the mean square error function in a normalized dimensionless space.

[0011] The real-time monitoring data stream is input into the trained residual physical constraint neural network model to deduce the dynamic physical state of water storage in the basin and output real-time flood forecast results.

[0012] Optionally, the ground observation data includes rainfall and cross-sectional flow, and the remote sensing data includes satellite precipitation data, normalized difference vegetation index, and topographic data; the atmospheric reanalysis data includes shallow soil moisture, air temperature, and potential evaporation; the preprocessing involves resampling data at different resolutions and aligning them to a uniform temporal resolution.

[0013] Optionally, the formulas for calculating the cross-sectional flow increment and the predicted flow are:

[0014] in, express The predicted cross-sectional flow increment at any given time, in units of ; A nonlinear mapping function representing a long short-term memory network; express The input feature vector at time step 1 is dimensionless. Represents the state of memory units in the Long Short-Term Memory (LSTM) network; dimensionless. Represents network weight parameters, which are dimensionless; express The predicted flow rate calculated at each time step, in units of ; express The measured cross-sectional flow rate received at time -1, in units of .

[0015] Optionally, the formula for the macroscopic water balance differential equation of the watershed is:

[0016]

[0017] in, This indicates the water storage capacity of the basin, in mm. This represents the areal rainfall, expressed in mm / h. This represents the actual evaporation rate, expressed in mm / h. This indicates the runoff depth rate, with units of mm / h.

[0018] Optionally, the physical consistency loss function unifies the calculation dimensions by converting the predicted flow rate into runoff depth rate, and the calculation formula is as follows:

[0019]

[0020] in, Represents the loss of physical consistency; it is dimensionless. Represents the total number of time steps, dimensionless; express The average isal rainfall input in the basin at time , in units of ; Indicates potential evapotranspiration input, in units of ; This represents the evapotranspiration conversion factor learned automatically by the network, and Dimensionless; Indicates the target watershed catchment area, in units of ; A comprehensive unit conversion factor for converting flow rate into runoff depth rate across the entire watershed, used to unify spatial and temporal scales; express The predicted flow rate calculated at each time step, in units of ; This represents the predicted rate of change in water storage output by the model, in units of... .

[0021] Optionally, the formula for calculating the overall objective function is:

[0022]

[0023] in, This represents the overall objective function, which is dimensionless. This represents the data fitting error calculated in a normalized dimensionless space, and is dimensionless. This represents the physical constraint weight hyperparameter, with a value of 0.01≤λ≤0.1, and is dimensionless. The physical consistency loss function is represented by the forward computation and residual differentiation, which are performed in the real physical dimension space after the network output is denormalized.

[0024] Optionally, the inference of the dynamic physical state of water storage in the basin specifically includes: using the rate of change of water storage for physical mapping; when the inferred rate of change of water storage is positive, it characterizes the initial loss mechanism of soil moisture absorption and water storage in the early stage of rainfall; when the inferred rate of change of water storage is negative, it characterizes the decline physical process of water receding from the soil and replenishing the river channel after the flood peak.

[0025] This invention also provides a flood forecasting system for small watersheds in mountainous areas, comprising:

[0026] The data preprocessing module is used to acquire multi-source hydrological and meteorological data of small watersheds in mountainous areas, and to obtain multi-source heterogeneous datasets through preprocessing; the multi-source hydrological and meteorological data includes ground observation data, remote sensing data and atmospheric reanalysis data;

[0027] The traffic prediction module is used to construct a time-series deep learning network model based on residual correction. The time-series deep learning network model is used to extract the time-series features of multi-source heterogeneous datasets. The measured cross-sectional traffic at the previous time step is superimposed with the predicted cross-sectional traffic increment at the current time step using residual processing to calculate the predicted traffic at the current time step. The time-series deep learning network model is a long short-term memory network model.

[0028] The physical constraint module is used to construct the physical constraint module based on the macroscopic water balance differential equation of the watershed, and to infer the rate of change of water storage. The rate of change of water storage is used as an auxiliary state variable to establish a physical consistency loss function.

[0029] The dual-drive training module is used to calculate the data fitting error between the predicted flow and the measured cross-sectional flow at the current moment. The data fitting error is weighted and fused with the physical consistency loss function to construct the overall objective function. The overall objective function is used to perform dual-drive training on the time-series deep learning network model to generate a residual physical constraint neural network model. The data fitting error is calculated using the mean square error function in a normalized dimensionless space.

[0030] The flood forecasting module is used to input real-time monitoring data streams into a trained residual physical constraint neural network model to deduce the dynamic physical state of water storage in the basin and output real-time flood forecast results.

[0031] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0032] This invention discloses a method and system for flood forecasting in small watersheds in mountainous areas. The method includes acquiring multi-source hydrological and meteorological data from small watersheds in mountainous areas, and preprocessing this data to obtain a multi-source heterogeneous dataset. A long short-term memory network based on residual correction is constructed, and the measured flow rate at the previous moment is superimposed with the predicted increment at the current moment to obtain the current predicted flow rate. A physical constraint module is constructed based on the watershed water balance differential equation, and a physical consistency loss function is established using the rate of change in water storage as an auxiliary variable. The data fitting error is weighted and fused with the physical loss function to train and generate a residual physical constraint neural network model. Real-time monitoring data is used to deduce the watershed water storage dynamics and output a flood forecast. This invention breaks through the prediction inertia barrier of high-frequency data, eliminates forecast lag, and proactively provides flood control safety margins during sudden flood peaks, achieving an interpretable reconstruction of the physical mechanism. Attached Figure Description

[0033] 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 introduced 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.

[0034] Figure 1 This is a flowchart of the flood forecasting method for small watersheds in mountainous areas according to the present invention;

[0035] Figure 2 This is a schematic diagram of the model architecture modules in this embodiment;

[0036] Figure 3 This is a comparison diagram of the phase lag diagnosis and prediction process of different models in this embodiment for dealing with typical sudden flood peaks;

[0037] Figure 4 This is an interpretability analysis diagram of the watershed water storage dynamics derived from the model in this embodiment. Detailed Implementation

[0038] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] The purpose of this invention is to provide a method and system for forecasting floods in small watersheds in mountainous areas, aiming to solve or improve at least one of the above-mentioned technical problems.

[0040] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] As a first aspect, such as Figure 1 As shown, this invention provides a method for flood forecasting in small watersheds in mountainous areas, comprising:

[0042] Multi-source hydrological and meteorological data of small watersheds in mountainous areas are acquired and preprocessed to obtain a multi-source heterogeneous dataset. The multi-source hydrological and meteorological data includes ground observation data, remote sensing data, and atmospheric reanalysis data. The ground observation data includes rainfall and cross-sectional flow, and the remote sensing data includes satellite precipitation data, normalized difference vegetation index, and topographic data. The atmospheric reanalysis data includes shallow soil moisture, air temperature, and potential evaporation. The preprocessing involves resampling data at different resolutions and aligning them to a uniform temporal resolution.

[0043] A time-series deep learning network model based on residual correction is constructed. This model extracts time-series features from multi-source heterogeneous datasets. The measured cross-sectional flow rate from the previous time step is superimposed with the predicted cross-sectional flow rate increment at the current time step using residual processing to calculate the predicted flow rate at the current time step. The time-series deep learning network model is a Long Short-Term Memory (LSTM) network model. This residual superposition processing is used to eliminate the phase lag problem in peak flood forecasting using purely data-driven models in current technologies.

[0044] The formulas for calculating the cross-sectional flow increment and the predicted flow are as follows:

[0045]

[0046]

[0047] in, express The predicted cross-sectional flow increment at any given time, in units of ; A nonlinear mapping function representing a long short-term memory network; express The input feature vector at time step 1 is dimensionless. Represents the state of memory units in the Long Short-Term Memory (LSTM) network; dimensionless. Represents network weight parameters, which are dimensionless; express The predicted flow rate calculated at each time step, in units of ; express The measured cross-sectional flow rate received at time -1, in units of .

[0048] A physical constraint module is constructed based on the differential equation of macroscopic water balance in the watershed, and the rate of change of water storage is inferred. The rate of change of water storage is used as an auxiliary state variable to establish a physical consistency loss function.

[0049] The formula for the macroscopic water balance differential equation of the watershed is as follows:

[0050]

[0051] in, This indicates the water storage capacity of the basin, in mm. This represents the areal rainfall, expressed in mm / h. This represents the actual evaporation rate, expressed in mm / h. This indicates the runoff depth rate, with units of mm / h.

[0052] The data fitting error between the predicted flow rate and the measured cross-sectional flow rate at the current moment is calculated. The data fitting error is then weighted and fused with the physical consistency loss function to construct a total objective function. The total objective function is used to perform dual-drive training on the time-series deep learning network model to generate a residual physical constraint neural network model. The data fitting error is calculated using the mean square error function in a normalized dimensionless space.

[0053] The physical consistency loss function converts the predicted flow rate into runoff depth rate to unify the calculation dimensions, and the calculation formula is as follows:

[0054]

[0055] in, Represents the loss of physical consistency; it is dimensionless. Represents the total number of time steps, dimensionless; express The average isal rainfall input in the basin at time , in units of ; Indicates potential evapotranspiration input, in units of ; This represents the evapotranspiration conversion factor learned automatically by the network, and Dimensionless; Indicates the target watershed catchment area, in units of ; A comprehensive unit conversion factor for converting flow rate into runoff depth rate across the entire watershed, used to unify spatial and temporal scales; express The predicted flow rate calculated at each time step, in units of ; This represents the predicted rate of change in water storage output by the model, in units of... .

[0056] The formula for calculating the overall objective function is as follows:

[0057]

[0058] in, This represents the overall objective function, which is dimensionless. This represents the data fitting error calculated in a normalized dimensionless space, and is dimensionless. The physical consistency loss function is represented by the forward computation and residual differentiation, which are performed in the real physical dimension space after the network output is denormalized. The physical constraint weight hyperparameter is defined as λ, which has a value of 0.01 ≤ λ ≤ 0.1 and is dimensionless. By adjusting the physical constraint weight hyperparameter λ, the constraint strength of the physical consistency loss function in the overall objective function is controlled, so that the residual physical constraint neural network model generates a predictive surplus response at the moment of sudden flood peak. The predictive surplus response provides a flood control safety margin to avoid underestimating the risk of underreporting.

[0059] The real-time monitoring data stream is input into the trained residual physical constraint neural network model to deduce the dynamic physical state of water storage in the basin and output real-time flood forecast results.

[0060] Specifically, the inference of the dynamic physical state of water storage in the basin includes: using the rate of change of water storage for physical mapping; when the inferred rate of change of water storage is positive, it characterizes the initial loss mechanism of soil moisture absorption and water storage in the early stage of rainfall; when the inferred rate of change of water storage is negative, it characterizes the decline physical process of water receding from the soil and replenishing the river channel after the flood peak.

[0061] Based on the above technical solution, the following is provided: Figures 1-4 The example shown.

[0062] Example 1

[0063] In this embodiment, the main purpose is to establish a "data-physical" dual-drive architecture to achieve flood warning, specifically including the following steps:

[0064] S1. Acquisition and Preprocessing of Multi-Source Heterogeneous Data: Surface rain gauge and hydrological station observation data of the small watersheds in the mountainous areas to be forecasted were acquired. Given the sparse distribution of single-source stations in mountainous areas, GPM IMERG remote sensing rainfall was further introduced to assist in spatial distribution characteristics, MODIS normalized vegetation index was used to reflect underlying surface dynamics, and shallow soil moisture (reflecting previous dryness and wetness), air temperature, and potential evaporation data were extracted from the ERA5-Land reanalysis dataset. GIS technology was used to extract all the above multi-source heterogeneous data into watershed mean surface data, and the data were uniformly resampled and aligned to a 1-hour time resolution on the time axis to construct a standardized time-series feature input dataset.

[0065] S2. Constructing a Long Short-Term Memory Network Model Based on Residual Correction: Establishing a prediction architecture of "baseline + residual". Using a Long Short-Term Memory (LSTM) network as the core feature extractor, the LSTM network no longer directly maps absolute flow, but instead extracts it through input feature vectors. and the memory state of the previous moment Predicting the abrupt change in flow increment caused by meteorological factors within the target step (i.e., micro-jump increment). ).

[0066] S3. Eliminate forecast phase lag: Assimilate the measured cross-sectional flow that was observed and incorporated in the previous time step. Treated as a hydrological baseline (baseline), it is added to the micro-jump flow increment predicted by the network in step S2. The predicted flow rate at the current moment is obtained by superimposing the values. This step effectively prevents the historical data translation inertia that data-driven models are prone to fall into through an explicit residual compensation architecture, significantly reducing forecast delays.

[0067] S4. Constructing Physical Constraints and Calculating Consistency Loss: A parallel auxiliary output branch node is added to the later part of the neural network model structure to infer and predict the rate of change of macroscopic watershed water storage. Introducing the catchment area of ​​the watershed and conversion constants The volumetric flow rate directly predicted by the model ( Accurately convert to basin-wide runoff depth rate with uniform water depth meaning. Time-series surface rainfall evaporation By substituting these equations into the calculus-integrated macroscopic water balance equation and calculating the residuals on both sides, a rigorous physical consistency loss can be constructed. .

[0068] S5. Dual-Driven Model Joint Training and Business Application: Addressing the Error of Conventional Data Fitting loss of physical consistency In conjunction with other methods, to avoid computational bias caused by differences in physical dimensions, and to limit data fitting loss. Computation is performed in a normalized dimensionless space to accelerate convergence; while physical consistency loss is used. The forward computation and residual differentiation must be performed within the true physical dimension space after the network output has been denormalized to ensure that the water balance law holds absolutely. This is achieved by setting physical constraint weight hyperparameters. The overall objective function is constructed, and the model is iteratively trained using the backpropagation optimization algorithm. Finally, the real-time monitoring data stream is input into the trained residual physical constraint neural network model to output high-precision flood control forecast flow in real time and assist in the deduction of internal hydrological water storage dynamics.

[0069] Example 2

[0070] This embodiment focuses on parameter optimization and model verification of the proposed method in the Guanshan River Basin, a typical small watershed in a mountainous area. The catchment area of ​​this watershed is approximately 318 km². 2 It is a typical "data-deficient" area characterized by high-intensity runoff, short confluence time, and a long-term lack of supporting internal hydrological observation elements.

[0071] In this embodiment, the total data time span is selected as 2010 to 2025. When setting and optimizing the overall objective function, the core step is to conduct sensitivity experiments on the key hyperparameter (physical constraint weight λ) that controls the balance between data mining capabilities and physical consistency.

[0072] The experimental procedure set the λ interval to [0, 1.0]. The results showed that its effect exhibited a typical inverted U-shaped variation:

[0073] When λ < 0.01 (data-dominated region), the constraint effect of the physical loss term is too weak, and the network degenerates into an approximately pure data statistical driven model, which is prone to overfitting under noise interference.

[0074] When λ > 0.1 (physical dominance region), the excessively strong rigid physical balance limits the neural network's ability to flexibly correct time series residuals, resulting in a rigid overall forecast bias (underfitting).

[0075] Through multiple rounds of verification, this embodiment determined the globally optimal equilibrium point parameter to be λ = 0.05. Under this configuration, in the unseen test set from 2022 to 2025, the model not only overcame the aforementioned defects at both ends, but also achieved an excellent global Nash efficiency coefficient (NSE) of 0.9907.

[0076] Example 3

[0077] This embodiment focuses on demonstrating the effectiveness of the model constructed in the above embodiments in responding to real extreme flood events (selecting the largest sudden flash flood caused by a rainstorm in 2023, with a measured peak flood value of 251.0m). 3 In practical application of detection and reporting, the three core technological advantages demonstrated are: eliminating lag, providing security margins, and breaking the black box of deep learning.

[0078] Breaking the inertia of data translation and completely eliminating forecast phase lag from a fundamental mechanism: Given the extremely strong smooth autocorrelation of hourly high-frequency flash flood data, conventional pure data-driven models are prone to falling into the trap of "over-smoothing of historical data" (for example, in this embodiment, the peak time of the conventional pure LSTM model is severely lagging by up to 9 hours). Even inertial benchmark models that rely solely on the flow rate of the previous moment will inevitably be "half a beat behind" at the most critical moment of the sudden flood peak due to physical time delay. The residual physical constraint forecast model of this invention, by forcing the network to strip away the background base, shifts the prediction target to the "flow increment (abrupt change information)" driven by meteorology, successfully blocking the inertial interference of data translation and accurately synchronizing the inflection point of the flood peak's rise and fall on the time axis, thus completely eliminating phase lag from the underlying mechanism.

[0079] Overcoming the risk of underreporting extreme values ​​and proactively providing valuable flood control safety margins: Due to the aforementioned physical time lag, traditional models inevitably produce severe systematic negative biases (i.e., underreporting) when predicting flood peak values. In this field measurement, the pure LSTM model underestimated the extreme value by -57.53%, and the inertial reference model produced a systematic underestimation of -11.95%, which poses a significant risk of underreporting disasters in flood control operations. In contrast, the model of this invention, guided by the soft constraints of its internal macroscopic water balance equation, possesses a highly sensitive response capability to extreme heavy rainfall inputs. The measured flood peak reached 251.0 m. 3 At a speed of ( / s), this model captured a height of up to 268.4 m. 3 The invention achieves a qualitative leap from "passive underreporting of negative deviations" to "active and moderate overreporting of positive deviations," adhering to the bottom line of "rather overestimate risks than underreport disasters," and providing extremely valuable engineering flood control safety margins for frontline flood control emergency rescue and personnel evacuation.

[0080] Opening the "black box" of deep learning to achieve interpretable reconstruction of its internal hydrophysical mechanisms: Deep learning has long been criticized as a "black box" lacking physical interpretability. This embodiment extracts auxiliary feature nodes (i.e., the rate of change of water storage dS / dt) implicitly and autonomously inferred by the model during runtime without any artificial external rules or hard constraints, for source tracing and mechanism deconstruction. Time-series diagnostics are clearly visible: In the early stage of heavy rainfall, before the river flow has increased significantly, the dS / dt curve derived by the network rapidly shows a steep positive peak, reconstructing the physical process in hydrology where "early rainfall is preferentially absorbed and stored by dry soil (initial loss mechanism does not generate runoff)"; while when the flood peak passes and enters the receding stage, the feature curve automatically and smoothly enters the negative range, perfectly confirming the natural decay mechanism where "the water stored in the vadose zone of the watershed is slowly released and replenishes the river in the form of interflow." This spontaneous dynamic deduction confirms that the present invention truly achieves logical self-consistency of the internal neural network parameters with the law of conservation of mass in the natural water cycle, endowing the deep learning model with extremely strong hydrophysical interpretability.

[0081] Example 4

[0082] This embodiment verifies the independent contributions and synergistic effects of each core module of the present invention through ablation experiments. Under the same watershed dataset and experimental conditions, the following four sets of comparative experimental schemes were set up: (a) a pure LSTM baseline model without residual correction and physical constraints; (b) an LSTM model with only residual correction and no physical constraint loss; (c) an LSTM model with only physical constraint loss and no residual correction; and (d) the complete model of the present invention, which integrates both residual correction and physical constraints.

[0083] Experimental results show that: compared with scheme (a), scheme (b) significantly reduces the phase lag at the flood peak from several hours to less than 1 hour, confirming the independent effectiveness of the residual correction mechanism in eliminating forecast lag; compared with scheme (a), although the global Nash efficiency coefficient of scheme (c) is improved, there is still a significant phase lag problem, indicating that applying physical constraints alone cannot solve the inertial defects of data translation; while the complete model scheme (d) of this invention eliminates phase lag and provides flood control safety margin at the same time, and its comprehensive performance is significantly better than any single module scheme, confirming that there is a non-simple superposition synergistic effect between residual correction and physical constraints.

[0084] Example 5

[0085] To verify the applicability and generalization ability of the method of this invention under different watershed conditions, this embodiment selected several typical small mountain watersheds with different geomorphological features, climate types, and watershed areas for migration verification. Experimental results show that the method of this invention exhibits robust forecasting performance under different watershed conditions: in each verification watershed, the Nash efficiency coefficient remains at a high level, the phase lag at the flood peak is effectively eliminated, and the water storage change rate curves inferred by the model can reasonably reflect the runoff generation and confluence physical processes of each watershed. These results indicate that the dual-drive architecture of residual correction and physical constraints proposed in this invention has good cross-watershed migration capability. Its core mechanism does not rely on parameter experience of a specific watershed and can be extended to real-time flood forecasting operations in small and medium-sized mountain watersheds across my country.

[0086] In summary, it can be seen that the present invention achieves the following:

[0087] Breaking the "Lag Trap" and Eliminating Forecast Lag: This invention introduces a residual learning mechanism into the neural network architecture, transforming the model's prediction target from "absolute flow" to "flow increment (abrupt change information)" influenced by meteorological factors. This mechanism effectively cuts off the strong inertial benchmark limitation caused by high-frequency data, which assumes that "the flow at the previous moment is equal to the forecast at this moment," thus fundamentally eliminating the severe phase lag problem commonly found in traditional pure data-driven models for flood peak forecasting.

[0088] Constructing water balance constraints to provide valuable flood control safety margins: This invention innovatively embeds the macroscopic water balance equation of the watershed as a soft constraint into the loss function, and forces the network to optimize within a solution space that follows physical conservation laws under unified time and spatial dimensions. This architecture endows the forecast model with sensitive response capabilities to rainfall inputs, and can generate a moderate forecast surplus (e.g., +7%) during extreme flood peaks. It completely eliminates the fatal flaw of "systematic underestimation" in traditional pure LSTM models and inertial models, achieving proactive defense and providing valuable safety margins for flood control safety scheduling.

[0089] By opening the "black box" system and achieving interpretability of hydrological dynamics, this invention adds an auxiliary physical state variable—the rate of change in water storage—allowing the model to autonomously infer watershed water storage dynamics that conform to physical laws without the need for externally imposed hard rules or complex parameter calibration. The model accurately reconstructs the entire alternating process of hydrological evolution: from the initial stage of rainfall—"soil initial loss, moisture absorption, water storage, and no runoff (manifested as a sharp positive rate of change)"—to the subsequent peak flood—"slow receding and replenishing water through interflow (manifested as a negative rate of change)." This successfully achieves the organic integration of high-precision data mining and physical interpretability.

[0090] As a second aspect, the present invention also provides a flood forecasting system for small watersheds in mountainous areas, comprising:

[0091] The data preprocessing module is used to acquire multi-source hydrological and meteorological data of small watersheds in mountainous areas, and to obtain multi-source heterogeneous datasets through preprocessing; the multi-source hydrological and meteorological data includes ground observation data, remote sensing data and atmospheric reanalysis data;

[0092] The traffic prediction module is used to construct a time-series deep learning network model based on residual correction. The time-series deep learning network model is used to extract the time-series features of multi-source heterogeneous datasets. The measured cross-sectional traffic at the previous time step is superimposed with the predicted cross-sectional traffic increment at the current time step using residual processing to calculate the predicted traffic at the current time step. The time-series deep learning network model is a long short-term memory network model.

[0093] The physical constraint module is used to construct the physical constraint module based on the macroscopic water balance differential equation of the watershed, and to infer the rate of change of water storage. The rate of change of water storage is used as an auxiliary state variable to establish a physical consistency loss function.

[0094] The dual-drive training module is used to calculate the data fitting error between the predicted flow and the measured cross-sectional flow at the current moment. The data fitting error is weighted and fused with the physical consistency loss function to construct the overall objective function. The overall objective function is used to perform dual-drive training on the time-series deep learning network model to generate a residual physical constraint neural network model. The data fitting error is calculated using the mean square error function in a normalized dimensionless space.

[0095] The flood forecasting module is used to input real-time monitoring data streams into a trained residual physical constraint neural network model to deduce the dynamic physical state of water storage in the basin and output real-time flood forecast results.

[0096] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0097] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for flood forecasting in small watersheds in mountainous areas, characterized in that, include: Multi-source hydrological and meteorological data of small watersheds in mountainous areas are acquired and preprocessed to obtain a multi-source heterogeneous dataset; the multi-source hydrological and meteorological data includes ground observation data, remote sensing data, and atmospheric reanalysis data; A time-series deep learning network model based on residual correction is constructed. The time-series deep learning network model is used to extract the time-series features of multi-source heterogeneous datasets. The measured cross-sectional flow at the previous time step and the predicted cross-sectional flow increment at the current time step are superimposed with residuals to calculate the predicted flow at the current time step. The temporal deep learning network model is a long short-term memory network model; A physical constraint module is constructed based on the macroscopic water balance differential equation of the watershed, and the rate of change of water storage is inferred. The rate of change of water storage is used as an auxiliary state variable to establish a physical consistency loss function. The data fitting error between the predicted flow rate and the measured cross-sectional flow rate at the current moment is calculated. The data fitting error is then weighted and fused with the physical consistency loss function to construct a total objective function. The total objective function is used to perform dual-drive training on the time-series deep learning network model to generate a residual physical constraint neural network model. The data fitting error is calculated using the mean square error function in a normalized dimensionless space. The real-time monitoring data stream is input into the trained residual physical constraint neural network model to deduce the dynamic physical state of water storage in the basin and output real-time flood forecast results. The formulas for calculating the cross-sectional flow increment and the predicted flow are as follows: in, express The predicted cross-sectional flow increment at any given time, in units of ; A nonlinear mapping function representing a long short-term memory network; express The input feature vector at time step 1 is dimensionless. Represents the state of memory units in the Long Short-Term Memory (LSTM) network; dimensionless. Represents network weight parameters, which are dimensionless; express The predicted flow rate calculated at each time step, in units of ; express The measured cross-sectional flow rate received at time -1, in units of ; The physical consistency loss function converts the predicted volumetric flow rate into runoff depth rate to unify the calculation dimensions. The calculation formula is as follows: in, Represents the loss of physical consistency; it is dimensionless. Represents the total number of time steps, dimensionless; express The average isal rainfall input in the basin at time , in units of ; Indicates potential evapotranspiration input, in units of ; This represents the evapotranspiration conversion factor learned automatically by the network, and Dimensionless; Indicates the target watershed catchment area, in units of ; A comprehensive unit conversion factor for converting flow rate into runoff depth rate across the entire watershed, used to unify spatial and temporal scales; express The predicted flow rate calculated at each time step, in units of ; This represents the predicted rate of change in water storage output by the model, in units of... .

2. The flood forecasting method for small watersheds in mountainous areas according to claim 1, characterized in that, The ground observation data includes rainfall and cross-sectional flow; the remote sensing data includes satellite precipitation data, normalized vegetation index, and topographic data; the atmospheric reanalysis data includes shallow soil moisture, air temperature, and potential evaporation; the preprocessing involves resampling data at different resolutions and aligning them to a uniform temporal resolution.

3. The flood forecasting method for small watersheds in mountainous areas according to claim 1, characterized in that, The formula for the macroscopic water balance differential equation of the watershed is: in, This indicates the water storage capacity of the basin, in mm. This represents the areal rainfall, expressed in mm / h. This represents the actual evaporation rate, expressed in mm / h. This indicates the runoff depth rate, with units of mm / h.

4. The flood forecasting method for small watersheds in mountainous areas according to claim 1, characterized in that, The formula for calculating the overall objective function is as follows: in, This represents the overall objective function, which is dimensionless. This represents the data fitting error calculated in a normalized dimensionless space, and is dimensionless. This represents the physical constraint weight hyperparameter, with a value of 0.01≤λ≤0.1, and is dimensionless. The physical consistency loss function is represented by the forward computation and residual differentiation, which are performed in the real physical dimension space after the network output is denormalized.

5. The flood forecasting method for small watersheds in mountainous areas according to claim 1, characterized in that, The inference of the dynamic physical state of water storage in the basin specifically includes: using the rate of change of water storage for physical mapping. When the inferred rate of change of water storage is positive, it characterizes the initial loss mechanism of soil moisture absorption and water storage in the early stage of rainfall; when the inferred rate of change of water storage is negative, it characterizes the decline physical process of water receding from the middle reaches of the soil and replenishing the river channel after the flood peak.

6. A flood forecasting system for small watersheds in mountainous areas, using the method described in any one of claims 1-5, characterized in that, include: The data preprocessing module is used to acquire multi-source hydrological and meteorological data of small watersheds in mountainous areas, and to obtain multi-source heterogeneous datasets through preprocessing; the multi-source hydrological and meteorological data includes ground observation data, remote sensing data and atmospheric reanalysis data; The traffic prediction module is used to construct a time-series deep learning network model based on residual correction. The time-series deep learning network model is used to extract the time-series features of multi-source heterogeneous datasets. The measured cross-sectional traffic at the previous time step is superimposed with the predicted cross-sectional traffic increment at the current time step using residual processing to calculate the predicted traffic at the current time step. The time-series deep learning network model is a long short-term memory network model. The physical constraint module is used to construct the physical constraint module based on the macroscopic water balance differential equation of the watershed, and to infer the rate of change of water storage. The rate of change of water storage is used as an auxiliary state variable to establish a physical consistency loss function. The dual-drive training module is used to calculate the data fitting error between the predicted flow and the measured cross-sectional flow at the current moment. The data fitting error is weighted and fused with the physical consistency loss function to construct the overall objective function. The overall objective function is used to perform dual-drive training on the time-series deep learning network model to generate a residual physical constraint neural network model. The data fitting error is calculated using the mean square error function in a normalized dimensionless space. The flood forecasting module is used to input real-time monitoring data streams into a trained residual physical constraint neural network model to deduce the dynamic physical state of water storage in the basin and output real-time flood forecast results.

Images