Hydrological ensemble forecasting method and system based on space-time adaptive generative adversarial-random differential disturbance and medium

Abstract

The invention provides a hydrological ensemble forecasting method and system based on space-time adaptive generative adversarial-random differential disturbance integration and a medium. The method comprises the steps of multi-source data fusion and standardization processing, spatial-temporal feature extraction, chaotic initial field generation, differential generative adversarial network construction, SDE disturbance field and physical constraint design, integrated forecast generation and statistical post-processing, and dynamic updating and set performance verification after a new observation value arrives. By coupling data driving and a physical mechanism, unified modeling of uncertainty propagation and spatial-temporal heterogeneity in hydrological forecasting is realized, and a high-resolution, reliable and uncertain set forecasting result is provided for hydrological forecasting.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrological forecasting, and in particular to a hydrological ensemble forecasting method, system and medium based on spatiotemporal adaptive generative adversarial-random differential perturbation integration. Background Art

[0002] Hydrological ensemble forecasting refers to the process of generating multiple possible forecast results to form a probabilistic forecast in order to cope with the uncertainty of the hydrological model, and introducing random disturbances, that is, adding the influence of random factors, to improve the accuracy and reliability of the forecast. The random disturbances selected by traditional ensemble forecasting methods include initial condition disturbances, parameter disturbances, or model structure disturbances. For example, the ensemble forecast system of the European Centre for Medium-Range Weather Forecasts (ECMWF) generates multiple forecast members by perturbing the initial conditions. In the GLUE model (generalized likelihood uncertainty estimation model), parameter sampling is introduced to generate the ensemble.

[0003] At present, in the process of combining random disturbances with traditional hydrological models, there is a lack of consideration for the location where the disturbance is applied and the spatial / temporal correlation, and the distribution and correlation of the disturbance are not considered, resulting in deviations in the uncertainty of the forecast. In particular, the introduction of random disturbances involves the selection of probability distributions. Traditional methods use Gaussian distributions, but actual hydrological processes have characteristics such as skewness and heteroscedasticity. The disturbance generation of existing technologies fails to maintain the spatial and temporal correlation structure, and does not reflect the statistical characteristics of extreme events / meteorology, making it difficult for ensemble forecast models to systematically and accurately deal with uncertainties. The reliability and accuracy of ensemble forecasts need to be improved. Summary of the invention

[0004] The purpose of the present invention is to provide a hydrological ensemble forecasting method based on spatiotemporal adaptive generative adversarial-random differential perturbations. By coupling data-driven and physical mechanisms, unified modeling of uncertainty propagation and spatiotemporal heterogeneity in hydrological forecasting is achieved, providing high-resolution and reliable uncertainty ensemble prediction results for hydrological forecasting.

[0005] According to a first aspect of the present invention, a hydrological ensemble forecasting method based on spatiotemporal adaptive generation of adversarial-random differential perturbations is proposed, comprising the following steps:

[0006] Step 1: Obtain multi-source observation data tensors, geographic information data, and ECMWF rainfall ensemble forecasts in the basin, fuse the multi-source data, align them to a 1km / 1h grid in time and space after mask cropping, and standardize each variable data to construct a three-dimensional tensor;

[0007] Step 2: Extract spatiotemporal features from the three-dimensional tensor, encode and output multi-scale features, where the spatiotemporal receptive field of the convolution kernel is dynamically adjusted through the expansion rate to match the nonlinear deformation of flood wave propagation;

[0008] Step 3: Construct a phase space reconstructor based on singular spectrum analysis to extract the Lyapunov exponents of the historical runoff series, and generate an initial field set that conforms to the chaotic dynamics characteristics through a noise-injected variational autoencoder to cover the distribution of hydrological initial values;

[0009] Step 4: Map the initial field and spatiotemporal features into the future water level prediction field, embed physical dynamic constraints, and construct a differential generative adversarial network:

[0010] The generator G is a deep residual network driven by fractional-order stochastic partial differential equations, which is used to generate future water level prediction fields;

[0011] The discriminator D is a three-channel hybrid discriminator in the space-time-frequency domain, including a space branch, a time branch, and a frequency domain branch, which is used to output the authenticity probability of the space-time field;

[0012] Step 5: Design the SDE disturbance field driven by Brownian motion in the forecast and apply physical constraint projection so that the introduced controlled random disturbance conforms to the laws of physics;

[0013] Step 6: Generate a prediction field for each disturbance sample, and calculate the discreteness of each grid point set after correcting the system deviation and spatiotemporal covariance of the prediction field, and output the risk probability map accordingly;

[0014] Step 7: Update the initial field according to the newly arrived observations, and then update the weights of the set members through system resampling to achieve dynamic update;

[0015] Step 8: Use the continuous ranked probability score CRPS and the Brier score BS to evaluate the ensemble performance, where: the continuous ranked probability score CRPS is used to measure the overall distance between the ensemble forecast distribution and the observed value. Ideally, CRPS approaches 0. For the prediction calibration of binary events, the Brier score BS is calculated for evaluation.

[0016] As an optional embodiment, in step 1, the obtained meteorological satellite observation data tensor is expressed as:

[0017] T, H, and W stand for time, height, and width, respectively;

[0018] Obtain geographic information tensor G, including geographic longitude and latitude, geographic elevation, and vegetation coverage;

[0019] ECMWF rainfall ensemble forecast data, expressed as:

[0020] Then, the basin boundary binary mask M is used to remove irrelevant areas, and the mask is cropped and aligned to a unified spatiotemporal network of 1 km × 1 h grid to focus on the hydrological process in the key areas.

[0021] Finally, the spatiotemporal mean and standard deviation of each variable are used to standardize the data of each variable individually, construct a three-dimensional tensor, and provide a standardized spatiotemporal grid for the model data.

[0022] As an optional embodiment, in step 2, spatiotemporal feature extraction is performed on the three-dimensional tensor, and multi-scale features are encoded and output, wherein the spatiotemporal receptive field of the convolution kernel is dynamically adjusted by the expansion rate to match the nonlinear deformation of flood wave propagation, including:

[0023] The three-dimensional tensor is subjected to feature extraction through a three-dimensional dilated convolutional network followed by a temporal attention mechanism to capture the coupling relationship between the basin topography and the spatiotemporal distribution of rainfall, and output multi-scale features to describe the spatiotemporal correlation between basin topography and rainfall. The spatiotemporal receptive field of the convolution kernel is dynamically adjusted through the expansion rate, and the spatiotemporal coverage is expanded layer by layer to capture the long-range dependency of the upstream and downstream of the basin to match the nonlinear deformation of flood wave propagation. The attention weight α is calculated in the time dimension. t :

[0024]

[0025] In the formula, q t is the query vector for the current time step, K is the historical feature matrix, and d is the feature dimension.

[0026] As an optional embodiment, in step 3, a phase space reconstructor based on singular spectrum analysis is constructed to extract the Lyapunov exponents of the historical runoff sequence, and an initial field set that conforms to the chaotic dynamic characteristics is generated through a noise-injected variational autoencoder to cover the distribution of hydrological initial values, including:

[0027] For the historical runoff series {Q t} Apply SSA decomposition to decompose the runoff sequence into trend term, period term and noise term, intercept the dominant component to reconstruct the phase space, the decomposition dimension K is determined by the embedding theorem, K≥2D+1K≥2D+1, D is the system degree of freedom, and reconstruct the phase space Estimation of the maximum Lyapunov exponent λ max ;

[0028] Variational autoencoder outputs hidden variable distribution And through the KL divergence constraint z is close to the standard normal distribution;

[0029] The decoder adds chaotic noise ξ to generate the initial field h0=g φ(z+∈ξ), the power spectrum of ξ and the maximum Lyapunov index λ max Matching ensures that the initial field set covers the true uncertainty range, that is:

[0030] As an optional embodiment, in step 4, the design of the generator G is as follows:

[0031] The residual network driven by fractional-order SDE is expressed as:

[0032]

[0033] In the formula, is the fractional Brownian motion, σ t is the adaptive diffusion coefficient, the fractional order is used to control the long-range correlation of the noise, f θ (h, t) represents the residual term, which is designed by physical heuristics of mass conservation term;

[0034] The design of the discriminator D is as follows:

[0035] Spatial branch: Use wavelet anisotropic Gabor convolution kernel to detect water surface related texture;

[0036] Time branch: Use CRF-LSTM model to model state transition probability;

[0037] Frequency branch: HHT decomposes the predicted field into intrinsic mode functions and calculates the energy entropy S E , identify whether the frequency domain energy distribution characteristics are reasonable disturbances.

[0038] As an optional embodiment, in step 5, a Brownian motion-driven SDE disturbance field is designed in the forecast, and a physical constraint projection is applied so that the introduced controlled random disturbance conforms to the physical laws, including:

[0039] The SDE disturbance field driven by the Brownian motion is expressed as follows:

[0040]

[0041] In the formula, B t , B′ t is the spatially correlated Brownian field, λ is the attenuation factor, which is set to the inverse of the flood wave propagation time, and σ is used to control the disturbance intensity; the spatial correlation is controlled by the kernel function, and its length scale is proportional to the basin radius;

[0042] The physical constraint projection includes converting η t Projection to the mass conservation equation get:

[0043]

[0044] Where g(·) is the conservation constraint equation and J is its Jacobian matrix;

[0045] The mass conservation equation is discretized into a system of linear equations, and the projection matrix is ​​quickly solved based on QR decomposition.

[0046] As an optional embodiment, in step 6, a prediction field is generated for each disturbance sample, and the prediction field is corrected for system deviation and spatiotemporal covariance to calculate the discreteness of each grid point set, and a risk probability map is outputted accordingly, including:

[0047] First, for each perturbation sample Generate a set of prediction fields:

[0048]

[0049] Among them, the initial field Sampling from the latent space of the variational autoencoder output, perturbation Obey the SDE solution path;

[0050] Then, for the obtained prediction field ensemble, fractional bit mapping is applied to the ensemble mean to correct the systematic error, and the covariance matrix is ​​updated through ensemble adjustment Kalman filtering to preserve the spatial structure of the ensemble discreteness;

[0051] Finally, the discreteness of each grid point set is calculated and converted into uncertainty quantification, and the preset warning threshold y th , calculate the probability of exceeding Forecast and output as risk probability map.

[0052] As an optional embodiment, in step 7, the initial field is updated according to the newly arrived observation value, and the system resamples and updates the set member weights to achieve dynamic update, including:

[0053] When new observations Upon arrival, based on new observations Update the initial field using particle filtering:

[0054]

[0055] After system resampling, the weights of set members are updated to avoid particle degradation, maintain set diversity, and achieve dynamic updates.

[0056] According to a second aspect of the present invention, a computer-readable storage medium is further provided, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the aforementioned method are implemented.

[0057] According to a third aspect of the present invention, a computer system is provided, comprising:

[0058] one or more processors;

[0059] The memory stores operable instructions, which, when executed by the one or more processors, enable the one or more processors to perform operations, wherein the operations include the steps of the aforementioned method.

[0060] Combined with the implementation of the above-mentioned hydrological ensemble forecasting method based on spatiotemporal adaptive generation of adversarial-stochastic differential perturbations, its significant advantages are:

[0061] (1) Integrate meteorological satellite observation data, geographic information data and ECMWF rainfall ensemble forecast data, focus on the hydrological process in key areas based on multi-source information data through mask clipping and spatiotemporal alignment, and eliminate the impact of data dimension and numerical differences through standardization, so as to provide a reliable data basis for subsequent model training, improve the performance and performance of the model, and enhance the forecast accuracy;

[0062] (2) A three-dimensional dilated convolutional network is further used to extract spatiotemporal features in combination with a temporal attention mechanism, which can effectively capture the coupling relationship between the basin topography and the spatiotemporal distribution of rainfall. The receptive field of the convolution kernel is dynamically adjusted to match the nonlinear deformation of flood wave propagation, which conforms to the complex spatiotemporal changes of topography and rainfall in the hydrological process and improves the forecast accuracy.

[0063] (3) In the method of the present invention, singular spectrum analysis and variational autoencoder are used to generate an initial field set that conforms to the chaotic dynamics characteristics, quantify the uncertainty of hydrological initial conditions, cover a wider distribution of hydrological initial values, conform to the chaotic characteristics of the hydrological system, and better handle the uncertainty in the forecast;

[0064] (4) The present invention introduces physical constraints into the generative adversarial network. In the differential generative adversarial network, the generator combines fractional-order stochastic partial differential equations and physical heuristic design, and the discriminator evaluates the prediction field from multiple channels. The combination of the two can embed physical dynamic constraints when generating the prediction field, thereby improving the rationality and reliability of the prediction field.

[0065] It should be understood that all combinations of the aforementioned concepts and the additional concepts described in more detail below can be considered as part of the inventive subject matter of the present disclosure as long as such concepts are not mutually inconsistent. In addition, all combinations of the claimed subject matter are considered as part of the inventive subject matter of the present disclosure.

[0066] The foregoing and other aspects, embodiments and features of the present invention can be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the present invention, such as the features and / or beneficial effects of the exemplary embodiments, will be apparent from the following description or learned from the practice of the specific embodiments according to the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in various figures may be represented by the same reference numeral. For clarity, not every component is labeled in every figure. Embodiments of various aspects of the present invention will now be described by way of example and with reference to the accompanying drawings, in which:

[0068] Figure 1 It is a flow chart of the hydrological ensemble forecasting method based on spatiotemporal adaptive generation of adversarial-random differential perturbations proposed in accordance with the present invention. DETAILED DESCRIPTION

[0069] In order to better understand the technical content of the present invention, specific embodiments are given and described as follows in conjunction with the accompanying drawings.

[0070] {Example 1}

[0071] Combined with Figure 1 As shown, the hydrological ensemble forecasting method based on spatiotemporal adaptive generation of adversarial-random differential perturbations according to an embodiment of the present invention includes the following steps:

[0072] Step 1: Obtain multi-source observation data tensors, geographic information data, and ECMWF rainfall ensemble forecasts in the basin, fuse the multi-source data, align them to a 1km / 1h grid in time and space after mask cropping, and standardize each variable data to construct a three-dimensional tensor;

[0073] Step 2: Extract spatiotemporal features from the three-dimensional tensor, encode and output multi-scale features, where the spatiotemporal receptive field of the convolution kernel is dynamically adjusted through the expansion rate to match the nonlinear deformation of flood wave propagation;

[0074] Step 3: Construct a phase space reconstructor based on singular spectrum analysis to extract the Lyapunov exponents of the historical runoff series, and generate an initial field set that conforms to the chaotic dynamics characteristics through a noise-injected variational autoencoder to cover the distribution of hydrological initial values;

[0075] Step 4: Map the initial field and spatiotemporal features into the future water level prediction field, embed physical dynamic constraints, and construct a differential generative adversarial network:

[0076] The generator G is a deep residual network driven by fractional-order stochastic partial differential equations, which is used to generate future water level prediction fields;

[0077] The discriminator D is a three-channel hybrid discriminator in the space-time-frequency domain, including a space branch, a time branch, and a frequency domain branch, which is used to output the authenticity probability of the space-time field;

[0078] Step 5: Design the SDE disturbance field driven by Brownian motion in the forecast and apply physical constraint projection so that the introduced controlled random disturbance conforms to the laws of physics;

[0079] Step 6: Generate a prediction field for each disturbance sample, and calculate the discreteness of each grid point set after correcting the system deviation and spatiotemporal covariance of the prediction field, and output the risk probability map accordingly;

[0080] Step 7: Update the initial field according to the newly arrived observations, and then update the weights of the set members through system resampling to achieve dynamic update;

[0081] Step 8: Use the continuous ranked probability score CRPS and the Brier score BS to evaluate the ensemble performance, where: the continuous ranked probability score CRPS is used to measure the overall distance between the ensemble forecast distribution and the observed value. Ideally, CRPS approaches 0. For the prediction calibration of binary events, the Brier score BS is calculated for evaluation.

[0082] As an optional embodiment, in step 1, the obtained meteorological satellite observation data tensor is expressed as:

[0083] T, H, and W stand for time, height, and width, respectively;

[0084] Obtain geographic information tensor G, including geographic longitude and latitude, geographic elevation, and vegetation coverage;

[0085] ECMWF rainfall ensemble forecast data, expressed as:

[0086] Then, the basin boundary binary mask M is used to remove irrelevant areas, and the mask is cropped and aligned to a unified spatiotemporal network of 1 km × 1 h grid to focus on the hydrological process in the key areas.

[0087] Finally, the spatiotemporal mean and standard deviation of each variable are used to standardize the data of each variable individually, construct a three-dimensional tensor, and provide a standardized spatiotemporal grid for the model data.

[0088] As an optional embodiment, in step 2, spatiotemporal feature extraction is performed on the three-dimensional tensor, and multi-scale features are encoded and output, wherein the spatiotemporal receptive field of the convolution kernel is dynamically adjusted by the expansion rate to match the nonlinear deformation of flood wave propagation, including:

[0089] The three-dimensional tensor is subjected to feature extraction through a three-dimensional dilated convolutional network followed by a temporal attention mechanism to capture the coupling relationship between the basin topography and the spatiotemporal distribution of rainfall, and output multi-scale features to describe the spatiotemporal correlation between basin topography and rainfall. The spatiotemporal receptive field of the convolution kernel is dynamically adjusted through the expansion rate, and the spatiotemporal coverage is expanded layer by layer to capture the long-range dependency of the upstream and downstream of the basin to match the nonlinear deformation of flood wave propagation. The attention weight α is calculated in the time dimension. t :

[0090]

[0091] In the formula, q t is the query vector for the current time step, K is the historical feature matrix, and d is the feature dimension.

[0092] As an optional embodiment, in step 3, a phase space reconstructor based on singular spectrum analysis is constructed to extract the Lyapunov exponents of the historical runoff sequence, and an initial field set that conforms to the chaotic dynamic characteristics is generated through a noise-injected variational autoencoder to cover the hydrological initial value distribution, including:

[0093] For the historical runoff series {Q t} Apply SSA decomposition to decompose the runoff sequence into trend term, period term and noise term, intercept the dominant component to reconstruct the phase space, the decomposition dimension K is determined by the embedding theorem, K≥2D+1K≥2D+1, D is the system degree of freedom, and reconstruct the phase space Estimation of the maximum Lyapunov exponent λ max ;

[0094] Variational autoencoder outputs hidden variable distribution And through the KL divergence constraint z is close to the standard normal distribution;

[0095] The decoder adds chaotic noise ξ to generate the initial field h0=g φ (z+∈ξ), the power spectrum of ξ and the maximum Lyapunov index λ max Matching ensures that the initial field set covers the true uncertainty range, that is:

[0096] As an optional embodiment, in step 4, the design of the generator G is as follows:

[0097] The residual network driven by fractional-order SDE is expressed as:

[0098]

[0099] In the formula, is the fractional Brownian motion, σ t is the adaptive diffusion coefficient, the fractional order is used to control the long-range correlation of the noise, fθ (h, t) represents the residual term, which is designed by physical heuristics of mass conservation term;

[0100] As an optional embodiment, the design of the discriminator D is as follows:

[0101] Spatial branch: Use wavelet anisotropic Gabor convolution kernel to detect water surface related texture;

[0102] Time branch: Use CRF-LSTM model to model state transition probability;

[0103] Frequency branch: HHT decomposes the predicted field into intrinsic mode functions and calculates the energy entropy S E , identify whether the frequency domain energy distribution characteristics are reasonable disturbances.

[0104] It should be understood that in the design of the generative adversarial network (GAN) in an embodiment of the present invention, the generator aims to generate samples that the discriminator misjudges, and the discriminator is to accurately distinguish between real and generated samples. Both can be optimized through adversarial training, and the training process is guided by a loss function.

[0105] As an optional embodiment, the goal of the generator G is to generate a future water level prediction field that can deceive the discriminator D. Since the generator is based on a deep residual network driven by a fractional-order stochastic partial differential equation, its loss function should be related to the difference between the generated field and the real field. Taking into account physical dynamic constraints, the loss term can be constructed in combination with physical conditions such as mass conservation. For example, the generated field needs to satisfy the mass conservation equation. If it does not satisfy it, a corresponding penalty will be generated, which is reflected in the loss function. As an optional embodiment, in order to make the generated field closer to the real distribution, the cross entropy loss related to the discriminator output can be further used.

[0106] In the example of the present invention, assuming that the real sample label is 1, the predicted probability of the generated sample output by the discriminator is P fake , the loss function L of the generator C Contains -log(1-P fake ) item, which encourages the generator to generate samples that are closer to the real ones, making it difficult for the discriminator to distinguish, thereby minimizing the loss.

[0107] At the same time, the discriminator D has three branches: space, time, and frequency, which can evaluate the authenticity of the prediction field from multiple dimensions. The loss function of the training process comprehensively considers the judgment results of each branch on the real samples and generated samples. For the spatial branch, the wavelet anisotropic Gabor convolution kernel is used to detect the water surface texture. If the real sample and the generated sample have a large difference in texture features from the actual situation, a large loss will be generated; the time branch uses the CRF-LSTM model to model the state transition probability. When the state transition of the prediction field does not conform to the actual situation, the loss will increase; the frequency branch uses HHT decomposition and energy entropy to judge the rationality of the frequency domain energy distribution. If it is unreasonable, it will also cause losses.

[0108] As an example of the present invention, assuming that the real sample label is 1, the generated sample label is 0, and the predicted probability of the discriminator outputting the real sample is P real , the predicted probability of generating sample output is P fake , then the discriminator loss function L D Contains -log(P real )-log(1-P fake ) item, by minimizing the loss function, the discriminator can more accurately distinguish between real samples and generated samples.

[0109] As an optional embodiment, in step 5, a Brownian motion-driven SDE disturbance field is designed in the forecast, and a physical constraint projection is applied so that the introduced controlled random disturbance conforms to the physical laws, including:

[0110] The SDE disturbance field driven by the Brownian motion is expressed as follows:

[0111]

[0112] In the formula, B t , B′ t is the spatially correlated Brownian field, λ is the attenuation factor, which is set to the inverse of the flood wave propagation time, and σ is used to control the disturbance intensity; the spatial correlation is controlled by the kernel function, and its length scale is proportional to the basin radius;

[0113] The physical constraint projection includes converting η t Projection to the mass conservation equation get:

[0114]

[0115] Where g(·) is the conservation constraint equation and J is its Jacobian matrix;

[0116] The mass conservation equation is discretized into a system of linear equations, and the projection matrix is ​​quickly solved based on QR decomposition.

[0117] As an optional embodiment, in step 6, a prediction field is generated for each disturbance sample, and the prediction field is corrected for system deviation and spatiotemporal covariance to calculate the discreteness of each grid point set, and a risk probability map is outputted accordingly, including:

[0118] First, for each perturbation sample Generate a set of prediction fields:

[0119]

[0120] Among them, the initial field Sampling from the latent space of the variational autoencoder output, perturbation Obey the SDE solution path;

[0121] Then, for the obtained prediction field ensemble, fractional bit mapping is applied to the ensemble mean to correct the systematic error, and the covariance matrix is ​​updated through ensemble adjustment Kalman filtering to preserve the spatial structure of the ensemble discreteness;

[0122] Finally, the discreteness of each grid point set is calculated and converted into uncertainty quantification, and the preset warning threshold y th , calculate the probability of exceeding Forecast and output as risk probability map.

[0123] As an optional embodiment, in step 7, the initial field is updated according to the newly arrived observation value, and the system resamples and updates the set member weights to achieve dynamic update, including:

[0124] When new observations Upon arrival, based on new observations Update the initial field using particle filtering:

[0125]

[0126] After system resampling, the weights of set members are updated to avoid particle degradation, maintain set diversity, and achieve dynamic updates.

[0127] {Example 2}

[0128] In this example, we further elaborate on the implementation process of the hydrological ensemble forecasting method based on spatiotemporal adaptive generation of adversarial-random differential perturbations in combination with the above-mentioned embodiment.

[0129] 1. Multivariate data fusion and standardization, which is used to integrate meteorological, geographic and forecast data, eliminate data heterogeneity (such as resolution and dimensional differences), and provide standardized space-time grids for model input.

[0130] Input acquisition and input include the aforementioned meteorological satellite observation data tensor (time × height × width), geographic information tensor (geographic location, elevation, vegetation cover, etc.), and ECMWF rainfall ensemble forecast.

[0131] On the one hand, through the watershed mask M∈{0,1} H×W Crop irrelevant areas and align them to a unified spatiotemporal grid of 1km×1h; it should be understood that the watershed mask in this embodiment is usually selected as a binary mask of the watershed boundary to eliminate irrelevant areas (such as neighboring mountains) and focus on the hydrological process in key areas.

[0132] In this example, the watershed mask (matrix) can be generated by GIS hydrological analysis (such as DEM-based runoff accumulation) or rasterized using official watershed boundary vector maps.

[0133] In the spatiotemporal alignment process, satellite data (e.g., 0.1° × 0.1° grid) and geographic data (DEM, etc.) are interpolated to a 1 km grid, and bilinear interpolation is used to maintain spatial continuity.

[0134] The temporal alignment is aggregated to 1 hour resolution via sliding window averaging.

[0135] On the other hand, the independent variables are standardized by standard deviation and variance: where μ χ , σ χ are the spatiotemporal mean and standard deviation of each variable, respectively, to avoid numerical differences dominating model training.

[0136] In the specific operation process, each meteorological variable (such as rainfall, temperature) is standardized independently, μ χ The value is the historical average for the same period.

[0137] The spatiotemporal mean can be obtained by calculating the mean of the same period in history (such as the same month in the past 30 years) through a sliding window. The standard deviation is the historical standard deviation of each variable, which is used to eliminate dimensional differences and stabilize model training. The standard deviation calculation of the same period is Winsorized for extreme values ​​(such as truncated at the 99% quantile).

[0138] Second, spatiotemporal feature extraction, through the 3D dilated convolutional network (3D Dilated ConvNet) nested temporal attention mechanism, capture the spatiotemporal heterogeneity of basin topography and precipitation, and obtain the coupling relationship between topography (such as elevation, slope) and spatiotemporal distribution of rainfall. Deformable convolution layer (Deformable Convolution) can be further introduced to adaptively adjust the receptive field, match the nonlinear deformation of flood wave propagation, model the nonlinear dynamic characteristics of flood wave propagation, and obtain multi-scale features that express the spatiotemporal correlation between basin topography and rainfall. Where S is the number of multi-scale feature layers, which represents the number of multi-scale feature maps output by the dilated convolution, and is used to capture the associations of different spatial ranges.

[0139] In this embodiment, the number of multi-scale feature layers S can be set according to the watershed area, for example, for a small watershed, S=3, corresponding to 1km, 3km, and 5km receptive fields).

[0140] In the specific processing process, the 3D dilated convolutional network is expressed as:

[0141]

[0142] Followed by a sequential attention gate:

[0143]

[0144] Among them, q t is the query vector for the current time step, K is the historical feature matrix, and d is the feature dimension, which determines the query vector q in the temporal attention mechanism. t and the dimension of the historical feature matrix K. In this example, the feature dimension d is 64-256, which is adjusted according to the number of input feature channels and needs to satisfy d≤C, where C is the number of input channels.

[0145] It should be understood that the attention weight α is calculated in the time dimension t , for example, increasing the weight during periods of sudden rainfall to suppress the impact of drought periods.

[0146] In the design of 3D dilated convolutional networks, the receptive field of the convolution kernel is dynamically adjusted based on the dilation rate, and the watershed characteristics are obtained by expanding the receptive field. For example, the dilation rate sequence is set to [1,2,4][1,2,4] to expand the spatiotemporal coverage layer by layer and capture the long-range dependencies between the upstream and downstream of the watershed.

[0147] In an embodiment of the present invention, the longer the time dimension of the convolution kernel is, the longer the temporal correlation is captured.

[0148] As an optional embodiment, the network uses a deformable convolution layer DC (Deformable Convolution) to adaptively adjust the receptive field to match the nonlinear deformation of flood wave propagation, where the spatial offset Δp of the deformable convolution layer is the offset of the convolution kernel sampling point obtained through learning, so that the convolution kernel adaptively matches the terrain or water flow deformation. The spatial offset Δp characterizes the terrain deformation caused by flood erosion and can be generated by a learnable deformation kernel, which is initialized to a zero-mean Gaussian distribution.

[0149] Δp=W d *

[0150] Where W drepresents a learnable deformation kernel.

[0151] This spatial offset can be learned from historical data. For example, in a curved area of ​​a river, the convolution kernel automatically bends to fit the actual water flow path.

[0152] 3. Chaotic initial field generation is used to quantify the uncertainty of hydrological initial conditions (such as soil moisture and groundwater level) and generate a diverse set of initial fields that conform to the characteristics of chaotic systems. The specific processing includes phase space reconstruction and noise injection VAE generation.

[0153] Spatial reconstruction: historical runoff series {Q t}Use singular spectrum analysis SSA decomposition to decompose the runoff sequence into trend term, periodic term and noise term, intercept the dominant component to reconstruct the phase space, the decomposition dimension K is determined by the embedding theorem, K≥2D+1K≥2D+1, D is the system degree of freedom, and reconstruct the phase space Estimation of the maximum Lyapunov exponent λ max , through λ max Quantify the chaos of the system, reflect the sensitivity of the initial conditions, and constrain the range of noise injection; in an optional embodiment, the Wolf algorithm can be used to estimate the historical runoff sequence, and the typical flood event λ max ≈0.1-0.3;

[0154] Noise injection VAE generation: Variational autoencoder outputs latent variable distribution And through the KL divergence constraint z is close to the standard normal distribution; the decoder adds chaotic noise ξ when decoding to generate the initial field h0 = g φ (z+∈ξ), the power spectrum of ξ and the maximum Lyapunov index λ max Matching ensures that the initial field set covers the true uncertainty range, that is:

[0155] Among them, the noise intensity coefficient ∈ is used to control the amplitude of chaotic noise when VAE generates the initial field to match the chaotic characteristics of the system. In this example, the noise intensity coefficient ∈ is based on λ max Adjustment, the empirical formula is: ∈=[0.1-0.12]*λ max *T, T represents the forecast period, which is usually 1 year.

[0156] In an embodiment of the present invention, the dimension of the VAE latent space is determined by the number of principal components of the SSA reconstructed phase space, ensuring that the initial field covers more than 90% of the variance contribution rate.

[0157] In this embodiment, the reconstruction decomposition dimension K of SSA represents the number of principal components retained in the phase space, and a typical value that can be taken according to the above principle is 8-12.

[0158] 4. The differential generative adversarial network is constructed to map the initial field and spatiotemporal features into the future water level prediction field, while embedding physical dynamic constraints, including the design of the generator G and the three-channel discriminator D.

[0159] Generator G: A residual network driven by fractional-order SDE:

[0160]

[0161] In the formula, is fractional Brownian motion, and the fractional order α is used to control the long-range correlation of noise (i.e., the long-range memory of noise, which affects the persistence of disturbance), fit the soil water transport, for example, take 0.6-0.8 to simulate the sub-diffusion characteristics of soil water transport (empirical fitting), match and fit the soil water transport process; f θ (h, t) represents the residual term, which is designed by the physical heuristic of the mass conservation term (mass conservation term

[0162] Parameter σ t is the adaptive diffusion coefficient, which is used to adjust the temporal function of the random perturbation intensity to match the growth of forecast uncertainty.

[0163] The discriminator D is designed to jointly evaluate the rationality of the time, space and frequency domains of the prediction field, using a three-channel design of time, space and frequency, as follows:

[0164] Spatial branch: Use the wavelet anisotropic Gabor convolution kernel to detect water surface related texture. The Gabor convolution kernel is expressed as follows:

[0165]

[0166] In the formula, parameters a and b are used to control spatial resolution and directional sensitivity respectively (a is the attenuation on the x-axis, b is the attenuation on the y-axis). For example, if a < b is designed, it means enhancing the response along the river channel. Taking a = 0.5 and b = 2.0 as an example, it matches the length-width ratio of the river channel.

[0167] The frequency f0 represents the center frequency of the Gabor wavelet, determines the scale of texture detection, matches the wavelength of the typical flood front, and can be set according to the width of the typical flood front. For example, if the front width is about 2km, then f0 = 0.0010.001cycles / m, matches the width of the river channel, and corresponds to a half-wavelength of 1km.

[0168] The direction parameter φ is set to be consistent with the direction of the river channel in the basin.

[0169] Time branch: CRF-LSTM joint modeling is used. LSTM models time dependency. CRF (conditional random field) introduces state transition constraints (such as water level cannot change suddenly). CRF-LSTM model is used to model state transition probability p(Yt |Y t-1：t-τ );

[0170] Frequency branch: HHT decomposes the predicted field into intrinsic mode function IMF and calculates the energy entropy S E , identify whether the frequency domain energy distribution characteristics are reasonable disturbances.

[0171] It should be understood that in this example, the process of HHT decomposition includes:

[0172] Where K represents the number of layers of the intrinsic mode function (IMF), which affects the frequency domain energy distribution analysis. The number of layers K is automatically determined through the screening process, and usually the first 5-8 IMFs are retained;

[0173] Then, calculate the energy entropy S E :

[0174] S E =-Σp k logp k

[0175] If the energy entropy S E If it exceeds the preset value, it means that the frequency energy is dispersed, and it is judged as unreasonable disturbance, otherwise it is reasonable disturbance.

[0176] Fifth, Brownian motion-driven SDE perturbations and physical constraints are introduced to introduce controlled random perturbations into the forecast to reflect the uncertainty of the model structure while ensuring that the perturbation field conforms to physical laws (such as conservation of mass).

[0177] As an optional implementation, the SDE disturbance field driven by Brownian motion is introduced as follows:

[0178]

[0179] In the formula, B t , B′ t is the spatially correlated Brownian field, and λ is the attenuation factor, which is used to control the attenuation speed of the disturbance and match the scale of the flood wave propagation time τ. In this embodiment, it is set to the inverse of the flood wave propagation time, that is, λ = 1 / τ. If τ is set to 12h, the corresponding attenuation factor λ is approximately equal to 0.083.

[0180] The parameter σ is used to control the perturbation intensity and can be adjusted dynamically according to the set discreteness requirements.

[0181] Spatial correlation is controlled by a kernel function whose length scale is proportional to the basin radius; the sum function is expressed as follows:

[0182]

[0183] Among them, the length scale is proportional to the radius of the flow domain and is used to control the spatial smoothness of the disturbance field. In this example, the length scale The value is 1 / 5 of the basin radius.

[0184] As an optional embodiment, the physical constraint includes converting η to t Projection to the mass conservation equation get:

[0185]

[0186] Where g(·) is the conservation constraint equation and J is its Jacobian matrix;

[0187] In which, by solving the mass conservation equation Discretize into a system of linear equations and quickly solve the projection matrix based on QR decomposition

[0188] In this example, the Jacobian matrix is ​​used as the derivative matrix of the physical constraint equation g(η)=0 for the projection solution.

[0189] Among them, the Lagrange multiplier in the mass conservation equation is optimized through adaptive step size to avoid over-smoothing.

[0190] 6. Ensemble forecast generation and statistical post-processing, including generating a forecast field for each disturbance sample, then correcting the forecast field for systematic bias and spatiotemporal covariance, calculating the ensemble discreteness of each grid point, and outputting a risk probability map based on this.

[0191] First, for each perturbation sample Generate a set of prediction fields:

[0192]

[0193] Among them, the initial field Sampling from the latent space of the variational autoencoder output, perturbation Obey the SDE solution path.

[0194] In this example, the number of integrated members K represents the number of parallel sampling of the initial field and the disturbance field, which affects the coverage of uncertainty. It is usually in the range of 50-100 and can be weighed based on the calculation accuracy and the required computing resources.

[0195] Then, for the obtained prediction field ensemble, fractional bit mapping is applied to the ensemble mean to correct the systematic error, and the covariance matrix is ​​updated through the ensemble adjusted Kalman filter to preserve the spatial structure of the ensemble discreteness, including:

[0196] Bias correction: ensemble mean Apply a quantile map:

[0197]

[0198] Among them, F obs , F ens are the cumulative distribution functions of observations and forecasts, respectively;

[0199] Space-time covariance correction: Update the covariance matrix by ensemble-adjusting the Kalman filter:

[0200] Σ post =Σ prior -KH∑ prior .

[0201] Finally, the discreteness of each grid point set is calculated and converted into uncertainty quantification, and the preset warning threshold y th (For example, the flood warning water level is determined by the historical warning water level of the hydrological station, or by frequency analysis (such as the 20-year return water level) to calculate the probability of exceeding Forecast and output as risk probability map.

[0202] 7. Dynamic update, which is used to perform data assimilation using particle filtering when new observations are received.

[0203] As an optional implementation, when a new observation When it arrives, the initial field is updated through the particle filter:

[0204]

[0205] Update ensemble member weights after resampling.

[0206] In this example, after updating the weights according to the newly arrived observations, systematic resampling is used to avoid particle degradation and maintain the diversity of the set.

[0207] 8. Verification and evaluation: CRPS and BS are used to evaluate the performance of the set, including:

[0208] The continuous ranking probability score CRPS is used to measure the overall distance between the ensemble forecast distribution and the observed value. Ideally, CRPS approaches 0.

[0209] The prediction calibration for binary events, such as whether the water level exceeds the warning line, can be evaluated by calculating the Brier score BS.

[0210] Among them, the continuous ranking probability score CRPS is expressed as:

[0211]

[0212] The Brier score BS is expressed as:

[0213]

[0214] Among them, p t Represents the predicted probability, binary event indicator function o t ∈{0, 1} indicates whether it actually occurred, and is used to mark whether a binary event such as a flood event occurred.

[0215] like If it actually occurs, the flag value is 1, otherwise the flag value is 0.

[0216] In an embodiment of the present invention, combined with the implementation process of steps 1 to 8 above, it can be seen that the multi-scale features output by the dilated convolution capture the spatial heterogeneity from local terrain (such as river curvature) to regional terrain gradient (such as the overall slope of the basin), providing physical constraints for the spatial distribution of the chaotic initial field, and avoiding the generation of random noise and the initial water level that is inconsistent with the terrain (for example, high probability floods occur at the ridge); at the same time, the multi-scale features output by the dilated convolution encode the impact of short-term rainfall fluctuations on soil moisture through low-level features at a small dilation rate, and reflect seasonal snowmelt or groundwater recharge trends through high-level features at a large dilation rate. These time domain information is subsequently compressed into latent variables through the VAE encoder to guide the retention of historical evolution patterns when generating the initial field.

[0217] For example, if the feature extraction layer outputs a tensor Taking 64 channels as an example, in the generation of the chaotic initial field:

[0218] VAE encoder input: Average pooling along the time axis

[0219] VAE encoder hidden variable z dimension: d z =20, satisfying d z Much smaller than CHW to enforce learning of compact representations.

[0220] In an embodiment of the present invention, multi-scale features are simultaneously used to drive the spatiotemporal adaptation of SDE-physical constraints, wherein the high-frequency components (such as the center of a local rainstorm) are adapted through spatial branch Gabor wavelet parameters to control the intensity of the SDE disturbance in heavy rainfall areas (for example, +0.02 mm / h), while suppressing the disturbance amplitude in drought areas.

[0221] In the design of the differential generative adversarial network, Gabor filters of different scales (such as f0 = 0.001 corresponding to 1 km wavelength) are used on the spatial branch to detect false flood fronts that do not match the feature pyramid; the temporal attention weight α of the LSTM state vector and the feature extraction layer is used on the temporal branch.t Cross-validation: If the phase difference between the two rainfall processes is greater than 1 hour, the penalty term is triggered; the HHT energy entropy of the frequency domain branch is compared with the power spectral density (PSD) of the multi-scale features on the frequency identification branch:

[0222] Legal disturbance:

[0223] Illegal disturbance: When the entropy ratio is abnormal (such as >2.0), it is judged as unreasonable.

[0224] {Example 3}

[0225] According to an embodiment of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the aforementioned method are implemented.

[0226] {Example 4}

[0227] According to an embodiment of the present invention, a computer system is provided, comprising:

[0228] one or more processors;

[0229] The memory stores operable instructions, which, when executed by the one or more processors, enable the one or more processors to perform operations, wherein the operations include the steps of the aforementioned method.

[0230] {Example 5}

[0231] According to the method of the above embodiment, we take a tributary of the Yangtze River (basin area 500km 2 ) The ensemble forecast of water level in the next 72 hours is further elaborated with the goal of 1km spatial resolution and 1hour temporal resolution.

[0232] 1. Data input processing:

[0233] Satellite rainfall was interpolated to a 1km grid (inverse distance weighted method), and DEM was upscaled to 1km (bicubic interpolation);

[0234] Construct a 3D input tensor, Time 72h, height 200m (latitude) and width 200m (longitude);

[0235] Normalization parameters:

[0236] Rainfall: μ R =2.1 2.1mm / h,σ R =4.5mm / h (historical summer average)

[0237] Elevation: μ DEM =356m,σDEM =128m.

[0238] 2. Spatiotemporal Feature Extraction Network

[0239] 3D variable convolutional network design:

[0240] The structure is 4 layers of dilated convolution, with dilation rates of [1, 2, 4, 8] and 64 output channels per layer;

[0241] Deformable offset learning rate: 0.01 (Adam optimizer);

[0242] Temporal attention design: time window = 24, number of attention heads 4, hidden layer dimension 128.

[0243] 3. Initial Field Generation (VAE Configuration)

[0244] SSA phase space reconstruction:

[0245] After decomposition of historical runoff, the first 10 principal components (cumulative variance 98.3%) were retained;

[0246] Estimate the Lyapunov exponent λmax = 0.25 (hourly level);

[0247] Noise Injection VAE:

[0248] Latent space dimension d z =20, noise intensity∈=0.15;

[0249] The initial field set size K = 50, satisfying:

[0250] 4. Differentiable Generative Adversarial Networks

[0251] Generator G (SDE driven): residual network depth 15 layers, each layer hidden unit 256

[0252] Discriminator D:

[0253] Spatial branch: 8-directional Gabor filter (a=0.5, b=2.0, f0=0.0012);

[0254] Time branch: 3-layer BiLSTM (128 hidden units) + CRF state transfer matrix

[0255] Frequency domain branch: HHT energy entropy threshold S E th =2.3.

[0256] 5. SDE disturbance field parameter configuration

[0257] Disturbance intensity σ = 0.08m, attenuation rate λ = 0.1;

[0258] Physical constraint mass conservation projection is iterated twice, with a relative tolerance of 10-4;

[0259] 6. Post-collection statistical processing

[0260] Quantile mapping: non-parametric quantile matching is used, and the training period is 2010-2020 daily data set. Kalman filter adjustment: localization radius 50km, expansion factor 1.05;

[0261] 7. Dynamic Update and Verification

[0262] Assimilation cycle: integrate the latest water level station observations every hour (23 stations)

[0263] Verification indicators:

[0264] CRPS target: <0.15m (72-hour forecast)

[0265] Brier score threshold yth = warning water level 34.5m.

[0266] After 72 hours, the actual water level is 34.8m, CRPS = 0.12m

[0267] BS=0.09 (for prediction of super-alert events).

[0268] Although the present invention has been disclosed as above with preferred embodiments, it is not intended to limit the present invention. A person with ordinary knowledge in the technical field to which the present invention belongs may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be determined by the definition of the claims.

Claims

1. A hydrological ensemble forecasting method based on spatiotemporal adaptive generative adversarial-random differential perturbation, characterized in that: The following steps are involved: Step 1: Obtain multi-source observation data tensors, geographic information data, and ECMWF rainfall ensemble forecasts in the basin, fuse the multi-source data, align them to a 1km / 1h grid in time and space after mask cropping, and standardize each variable data to construct a three-dimensional tensor; Step 2: Extract spatiotemporal features from the three-dimensional tensor, encode and output multi-scale features, where the spatiotemporal receptive field of the convolution kernel is dynamically adjusted through the expansion rate to match the nonlinear deformation of flood wave propagation; Step 3: Construct a phase space reconstructor based on singular spectrum analysis to extract the Lyapunov exponents of the historical runoff series, and generate an initial field set that conforms to the chaotic dynamics characteristics through a noise-injected variational autoencoder to cover the distribution of hydrological initial values; Step 4: Map the initial field and spatiotemporal features into the future water level prediction field, embed physical dynamic constraints, and construct a differential generative adversarial network: The generator G is a deep residual network driven by fractional-order stochastic partial differential equations, which is used to generate future water level prediction fields; The discriminator D is a three-channel hybrid discriminator in the space-time-frequency domain, including a space branch, a time branch, and a frequency domain branch, which is used to output the probability of the authenticity of the space-time field; Step 5: Design the SDE disturbance field driven by Brownian motion in the forecast and apply physical constraint projection so that the introduced controlled random disturbance conforms to the physical laws; Step 6: Generate a prediction field for each disturbance sample, and calculate the discreteness of each grid point set after correcting the system deviation and spatiotemporal covariance of the prediction field, and output the risk probability map accordingly; Step 7: Update the initial field according to the newly arrived observations, and then update the weights of the set members through system resampling to achieve dynamic update; Step 8: Use the continuous ranked probability score CRPS and the Brier score BS to evaluate the ensemble performance, where: the continuous ranked probability score CRPS is used to measure the overall distance between the ensemble forecast distribution and the observed value. Ideally, CRPS approaches 0. For the prediction calibration of binary events, the Brier score BS is calculated for evaluation.

2. The hydrological ensemble forecasting method based on spatiotemporal adaptive generative adversarial-random differential perturbation according to claim 1 is characterized in that: In step 1, the obtained meteorological satellite observation data tensor is expressed as: T, H, and W stand for time, height, and width, respectively; Obtain geographic information tensor G, including geographic longitude and latitude, geographic elevation, and vegetation coverage; ECMWF rainfall ensemble forecast data, expressed as: Then, the basin boundary binary mask M is used to remove irrelevant areas, and the mask is cropped and aligned to a unified spatiotemporal network of 1 km × 1 h grid to focus on the hydrological process in the key areas. Finally, the spatiotemporal mean and standard deviation of each variable are used to standardize the data of each variable individually, construct a three-dimensional tensor, and provide a standardized spatiotemporal grid for the model data.

3. The hydrological ensemble forecasting method based on spatiotemporal adaptive generative adversarial-random differential perturbation according to claim 1 is characterized in that: In step 2, the spatiotemporal features of the three-dimensional tensor are extracted, and multi-scale features are encoded and output, wherein the spatiotemporal receptive field of the convolution kernel is dynamically adjusted by the expansion rate to match the nonlinear deformation of flood wave propagation, including: The three-dimensional tensor is subjected to feature extraction through a three-dimensional dilated convolutional network followed by a temporal attention mechanism to capture the coupling relationship between the basin topography and the spatiotemporal distribution of rainfall, and output multi-scale features to describe the spatiotemporal correlation between basin topography and rainfall. The spatiotemporal receptive field of the convolution kernel is dynamically adjusted through the dilation rate, and the spatiotemporal coverage is expanded layer by layer to capture the long-range dependency of the upstream and downstream of the basin to match the nonlinear deformation of flood wave propagation. The attention weight α is calculated in the time dimension. t : In the formula, q t is the query vector for the current time step, K is the historical feature matrix, and d is the feature dimension.

4. The hydrological ensemble forecasting method based on spatiotemporal adaptive generative adversarial-random differential perturbation according to claim 1 is characterized in that: In step 3, a phase space reconstructor based on singular spectrum analysis is constructed to extract the Lyapunov exponents of the historical runoff sequence, and an initial field set that conforms to the chaotic dynamics characteristics is generated through a noise-injected variational autoencoder, covering the distribution of hydrological initial values, including: For the historical runoff series {Q t } Apply SSA decomposition to decompose the runoff sequence into trend term, period term and noise term, intercept the dominant component to reconstruct the phase space, the decomposition dimension K is determined by the embedding theorem, K≥2D+1K≥2D+1, D is the system degree of freedom, and reconstruct the phase space Estimation of the maximum Lyapunov exponent λ max ; Variational autoencoder outputs hidden variable distribution And through the KL divergence constraint z is close to the standard normal distribution; The decoder adds chaotic noise ξ to generate the initial field h0=g φ (z+∈ξ), the power spectrum of ξ and the maximum Lyapunov index λ max Matching ensures that the initial field set covers the true uncertainty range, that is:

5. The hydrological ensemble forecasting method based on spatiotemporal adaptive generative adversarial-random differential perturbation according to claim 1 is characterized in that: In step 4, the design of the generator G is as follows: The residual network driven by fractional-order SDE is expressed as: In the formula, is the fractional Brownian motion, σ t is the adaptive diffusion coefficient, the fractional order is used to control the long-range correlation of the noise, f θ (h, t) represents the residual term, which is designed by physical heuristics of mass conservation term; The design of the discriminator D is as follows: Spatial branch: Use wavelet anisotropic Gabor convolution kernel to detect water surface related texture; Time branch: Use CRF-LSTM model to model state transition probability; Frequency branch: HHT decomposes the predicted field into intrinsic mode functions and calculates the energy entropy S E , identify whether the frequency domain energy distribution characteristics are reasonable disturbances.

6. The hydrological ensemble forecasting method based on spatiotemporal adaptive generative adversarial-random differential perturbation according to claim 1 is characterized in that: In step 5, a Brownian motion-driven SDE disturbance field is designed in the forecast, and a physical constraint projection is applied so that the introduced controlled random disturbance conforms to the physical laws, including: The SDE disturbance field driven by the Brownian motion is expressed as follows: In the formula, B t , B′ t is the spatially correlated Brownian field, λ is the attenuation factor, which is set to the inverse of the flood wave propagation time, and σ is used to control the disturbance intensity; the spatial correlation is controlled by the kernel function, and its length scale is proportional to the basin radius; The physical constraint projection includes converting η t Projection to the mass conservation equation get: Where g(·) is the conservation constraint equation and J is its Jacobian matrix; The mass conservation equation is discretized into a system of linear equations, and the projection matrix is ​​quickly solved based on QR decomposition.

7. The hydrological ensemble forecasting method based on spatiotemporal adaptive generative adversarial-random differential perturbation according to claim 1 is characterized in that: In step 6, a prediction field is generated for each disturbance sample, and the prediction field is corrected for system deviation and spatiotemporal covariance to calculate the discreteness of each grid point set, and a risk probability map is output based on the discreteness, including: First, for each perturbation sample Generate a set of prediction fields: Among them, the initial field Sampling from the latent space of the variational autoencoder output, perturbation Obey the SDE solution path; Then, for the obtained prediction field ensemble, fractional bit mapping is applied to the ensemble mean to correct the systematic error, and the covariance matrix is ​​updated through ensemble adjustment Kalman filtering to preserve the spatial structure of the ensemble discreteness; Finally, the discreteness of each grid point set is calculated and converted into uncertainty quantification, and the preset warning threshold y th , calculate the probability of exceeding Forecast and output as risk probability map.

8. The hydrological ensemble forecasting method based on spatiotemporal adaptive generative adversarial-random differential perturbation according to claim 1 is characterized in that: In step 7, the initial field is updated according to the newly arrived observation value, and the system resamples and updates the weights of the set members to achieve dynamic update, including: When new observations Upon arrival, based on new observations Update the initial field using particle filtering: After system resampling, the weights of set members are updated to avoid particle degradation, maintain set diversity, and achieve dynamic updates.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer system, characterized in that: include: one or more processors; A memory storing operable instructions, wherein when the instructions are executed by the one or more processors, the one or more processors are caused to perform operations, wherein the operations include the steps of the method according to any one of claims 1 to 8.

Images