Regional extreme rainfall forecasting and early warning method based on cGAN

Abstract

A regional extreme rainfall forecasting and early warning method based on a conditional generative adversarial network (cGAN) comprises the following steps: preprocessing multi-source meteorological data, including standardization and space consistency alignment, so as to ensure the availability of the data; and through a coding module of the cGAN, extracting and fusing interannual daily average features and horizontal features, constructing a multi-scale feature pyramid, and enhancing the characterization capability of a rainfall spatial mode. Forecasting data are generated through a cGAN generator and a condition discriminator, rainfall collaborative factors are introduced into the discriminator to serve as physical constraint conditions, and network parameters are optimized through an adversarial loss function and physical consistency loss. The forecasting data is post-processed through a graph neural network (GNN), and spatial dependency modeling and optimization correction are carried out by taking a meteorological station observation value as a true value, so that the forecasting precision and reliability of an extreme rainfall event are remarkably improved.

Description

Technical Field

[0001] The present invention relates to precipitation modeling and prediction technology, and in particular to a regional extreme precipitation forecasting and early warning method based on cGAN. Background Art

[0002] There are many different methods for precipitation modeling and prediction, including conceptual models, physical models, numerical methods, and statistical methods. In recent years, artificial intelligence technology, especially deep generative models (DGMs), has been widely studied in the field of rainfall forecasting. Generative adversarial networks (GANs), as a typical deep generative model, convert complex likelihood functions into a neural network framework, enabling the model to fit the target distribution by optimizing parameters. The core idea of ​​GANs is to enable the generator to generate samples that are closer to real data through adversarial training between the generator and the discriminator.

[0003] In the field of precipitation forecasting, Jing et al. designed a GAN-based precipitation forecasting model, in which the generator adopts a ConvLSTM structure and is equipped with two discriminators for radar extrapolation. The generator loss function of the model combines the mean square error (MSE) and the mean absolute error (MAE), while the discriminator uses the binary cross entropy (BCE) loss. The research team limited the radar reflectivity between 0 and 75 decibels and successfully predicted high-resolution radar data for the next 1.5 hours. In addition, Ravuri et al. proposed another GAN-based precipitation prediction model, DGMR, which uses ConvGRU as a generator and is equipped with two discriminators to capture spatial and temporal patterns respectively. The DGMR model relies only on radar observation data, successfully achieves high-resolution precipitation prediction, and outperforms other models in the critical success index (CSI) evaluation. The algorithm specifically targets heavy precipitation events and uses loss functions including hinge loss, in which the discriminator uses hinge loss and the generator combines hinge loss with MAE.

[0004] The GAN method has achieved remarkable success in the fields of super-resolution and statistical downscaling, especially in short-term forecasting tasks, effectively alleviating the problem of smoothing and intensity loss of forecast results over time. For example, Ravuri et al. (2021) used GANs to make probabilistic forecasts of global precipitation, which not only improved the spatial resolution of the forecast results, but also retained the structural characteristics of extreme precipitation events. Price and Rasp (2022) used conditional GAN ​​(cGAN) to downscale numerical weather forecasts (NWP), significantly improving the clarity of small-scale precipitation structures and reducing forecast uncertainty.

[0005] In summary, the application of GANs in precipitation modeling and prediction demonstrates its unique advantages in capturing complex precipitation patterns and extreme events, providing a new technical approach for high-resolution precipitation forecasting and accurate prediction of extreme weather events.

[0006] Although GANs have shown great potential in precipitation forecast correction and downscaling, current precipitation forecast research based on GANs rarely considers precipitation data from meteorological stations. However, precipitation data from meteorological stations can usually reflect the actual surface precipitation more accurately, especially in operational precipitation forecasting, which is crucial for responding to emergencies caused by heavy precipitation. In addition, the existing models have low reliance on multi-source data and physical constraints, which may limit their effective capture of incremental information brought by the fusion of different information sources, as well as the physical consistency with real rainfall, thus affecting the forecast accuracy.

[0007] It should be noted that the information disclosed in the above background technology section is only used for understanding the background of the present application, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the invention

[0008] The main purpose of the present invention is to overcome the defects existing in the above-mentioned background technology and provide a regional extreme precipitation forecasting and warning method based on cGAN.

[0009] To achieve the above object, the present invention adopts the following technical solutions: A regional extreme precipitation forecasting and warning method based on conditional generative adversarial network (cGAN) includes the following steps: S1. Preprocessing of multi-source meteorological data: standardizing multi-source meteorological data and aligning spatial consistency of data with different resolutions, wherein the multi-source meteorological data includes rainfall and its synergistic factors; S2. Feature extraction and fusion: The preprocessed data is input into the encoding module of the conditional generative adversarial network (cGAN), and the interannual daily average features and anomaly features are extracted through a multi-branch structure. Multi-source features are fused based on the attention mechanism, and a multi-scale feature pyramid is constructed to enhance the representation ability of rainfall spatial modality. S3. Generate forecast data: Input the fused features into the generator of the conditional generative adversarial network (cGAN), and generate forecast data through the decoding module; Use the conditional discriminator to supervise the generated data, and the discriminator explicitly introduces the rainfall coordination factor as a physical constraint condition, and optimizes the network parameters through the adversarial loss function and the physical consistency loss, so that the generated data conforms to the laws of meteorological physics; S4. Graph neural network post-processing: A topological map is constructed based on the coordinates of meteorological stations and grid points. The meteorological station observations are taken as true values. The graph neural network (GNN) is used to perform spatial dependency modeling and optimization correction on the generated forecast data to improve the forecast accuracy of extreme precipitation events.

[0010] A computer-readable storage medium stores a computer program, which implements the method when executed by a processor.

[0011] A computer program product comprises a computer program, wherein when the computer program is executed by a processor, the method described above is implemented.

[0012] The present invention has the following beneficial effects: The present invention proposes a regional extreme precipitation forecasting and early warning method based on conditional generative adversarial network cGAN, which integrates multi-source data, strengthens the model's modeling ability of physical laws, and significantly improves the forecast accuracy and reliability of extreme precipitation events. The fusion of multi-source data brings more information gain to the model, enabling the model to more effectively capture incremental information between different information sources and enhance the consistency between forecast data and actual precipitation. The present invention uses the cGAN model to force the model to learn the potential nonlinear relationship between rainfall and synergy factors by taking meteorological synergy factors as conditional information, which not only improves the accuracy of the prediction, but also enhances the physical consistency and reliability of the forecast results. Furthermore, the present invention uses the powerful modeling ability of graph neural network (GNN) for complex spatial dependencies, combines the topological structure between meteorological stations and grid points, and uses the neighboring information propagation mechanism to dynamically update node features, effectively capturing local details and global spatial change laws. This method performs post-processing correction on the preliminary prediction results of cGAN, significantly reduces the system error, enhances the model's ability to characterize extreme precipitation events, effectively alleviates the over-smoothing phenomenon in the prediction of heavy precipitation events, and improves the precision and reliability of the forecast results. Compared with the traditional GAN ​​model, the performance of the present invention in heavy precipitation forecasting is better. Through a two-step forecasting method, the preliminary forecast results are first generated by cGAN, and then the spatial information transmission between GNN learning nodes is used to correct the preliminary forecast results in combination with high-precision meteorological observation data, thereby significantly improving the forecast accuracy and reliability of extreme precipitation events.

[0013] Compared with the prior art, the main advantages of the present invention are: 1. Improve data diversity and physical law modeling: Traditional GAN ​​models do not explicitly consider the learning differences between long-term and short-term rainfall patterns, lack physical meaning, and only capture structure and intensity information by optimizing the loss function, without fully considering the impact of synergy factors on rainfall forecasts. This invention innovatively designs a dual-branch learning structure for long-term periodic laws and short-term fluctuations, capturing long-term trends and short-term changes respectively, which is more in line with meteorological physical laws. At the same time, the conditional GAN ​​(cGAN) framework is used to input rainfall synergy factors (such as temperature, humidity, wind speed, etc.) as conditions into the generator, while also considering the constraints of synergy factors in the discriminator, forcing the model to pay attention to and learn the potential nonlinear relationship between rainfall and synergy factors, thereby generating more accurate precipitation forecasts that are more in line with physical laws.

[0014] 2. Improvement of the prediction accuracy of heavy precipitation: Traditional GAN ​​models often have over-smoothing in heavy precipitation forecasting, resulting in insufficient prediction ability of extreme precipitation events. This paper proposes a two-step forecasting method: first, the preliminary forecast results are generated by cGAN, and then the spatial information transmission between learning nodes is used by graph neural network (GNN), and the preliminary forecast results are corrected in combination with high-precision meteorological observation data. This method effectively alleviates the over-smoothing effect of cGAN on heavy precipitation and significantly improves the prediction accuracy and reliability of extreme precipitation events.

[0015] Other beneficial effects of the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Flow chart of a regional extreme precipitation forecast and warning method according to an embodiment of the present invention.

[0017] Figure 2 The network architecture of cGAN and GNN in the embodiments of the present invention.

[0018] Figure 3 A multi-scale feature extraction module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following is a detailed description of the embodiments of the present invention. It should be emphasized that the following description is only exemplary and is not intended to limit the scope and application of the present invention.

[0020] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0021] In view of the technical problems of insufficient spatiotemporal feature recognition capability, oversmoothing phenomenon and limited forecast accuracy in the rainfall forecast model in the prior art, this paper proposes an extreme precipitation intelligent forecast and warning method and system based on conditional generative adversarial network (cGAN). This invention significantly improves the spatiotemporal forecast accuracy and warning reliability of extreme precipitation events through an innovative two-step forecast framework.

[0022] The precipitation forecast and correction method based on cGANs focuses on multi-source data fusion and stronger spatiotemporal feature extraction, as well as strengthening the physical constraints of the model to improve the accuracy of heavy rainfall forecasts. Specifically, multi-source meteorological data such as satellites, radars, and reanalysis data can be combined to capture the complex spatiotemporal evolution characteristics of the precipitation system through the generator and discriminator architecture of cGANs. At the same time, the meteorological observation station data is introduced to further correct the model prediction results to enhance the consistency between the forecast data and the actual precipitation, and reduce the over-smoothing effect of the model on extreme precipitation events. Through this method, it aims to improve the reliability of precipitation forecasts and provide strong support for accurate meteorological services, especially playing an important role in early warning and disaster prevention and mitigation of heavy precipitation events.

[0023] See also Figures 1 to 3 The embodiment of the present invention provides a regional extreme precipitation forecasting and warning method based on a conditional generative adversarial network (cGAN), comprising the following steps: S1. Preprocessing of multi-source meteorological data: standardizing multi-source meteorological data and aligning the spatial consistency of data with different resolutions. The multi-source meteorological data includes rainfall and its synergistic factors.

[0024] In some embodiments, step S1 specifically includes: using an adaptive standardization method to uniformly process multi-source meteorological data (such as satellite, radar, reanalysis data, etc.), including the spatiotemporal alignment of historical precipitation data; using a differentiable interpolation algorithm to perform spatial consistency processing on the standardized multi-source meteorological data, and interpolating data of different resolutions to a unified spatial resolution; adding random noise disturbances to the standardized multi-source meteorological data to enhance the robustness of the model to data uncertainty and noise. Further, in step S1, rainfall and its synergistic factor data of multiple time steps (t1, t2, ..., tn) are used as model inputs, and the synergistic factors include variables such as wind field, temperature, humidity, and air pressure, and ERA5 data is used as the true value source to construct a time series data set for model training and prediction.

[0025] S2. Feature extraction and fusion: The preprocessed data is input into the encoding module of the conditional generative adversarial network (cGAN). The interannual daily average features and anomaly features are extracted through a multi-branch structure. Multi-source features are fused based on the attention mechanism, and a multi-scale feature pyramid is constructed to enhance the representation ability of rainfall spatial modality.

[0026] In some embodiments, step S2 specifically includes: inputting the preprocessed multi-source meteorological data into the encoder of cGAN, mapping the data to a spatial latent representation, and using KL divergence to regularize the latent space; uniformly processing rainfall and its synergistic factor data of different spatial resolutions through a differentiable interpolation layer, the interpolation layer is composed of a convolution layer and a pooling layer to ensure the spatial consistency of the data; decomposing the data into interannual average daily data and anomaly data, and using a spatiotemporal feature extraction module of a Conv-Transformer combination to extract features of the two types of data respectively; splicing the interannual average daily features and anomaly features of multiple data through a concat layer, using an attention mechanism to calculate the attention of each feature, and fusing the features based on the attention score; performing spatial multi-scale feature extraction on the fused features to enhance the representation capability of semantic information and fine-grained information.

[0027] S3. Generate forecast data: Input the fused features into the generator of the conditional generative adversarial network (cGAN), and generate forecast data through the decoding module; use the conditional discriminator to supervise the generated data, and the discriminator explicitly introduces the rainfall synergy factor as a physical constraint condition, and optimizes the network parameters through the adversarial loss function and the physical consistency loss, so that the generated data conforms to the laws of meteorological physics.

[0028] In some embodiments, step S3 specifically includes: inputting the enhanced multi-scale spatial features into the decoding module of the generator of the cGAN, and mapping the potential features into the rainfall and its synergistic factor data at the next moment through layer-by-layer convolution operations; in the decoding module, using the attention mechanism and feature concatenation (concat) to further fuse the multi-scale spatial feature information to predict the rainfall and its synergistic factor data at the next moment; inputting the forecast data generated by the generator into the discriminator, and the discriminator determines whether the generated data comes from the real data distribution by calculating the discriminant loss Loss, further constraining and optimizing the output of the generator, and improving its modeling ability for real rainfall data; and inputting the generated forecast data as a new input into the trained cGAN model to achieve multi-step prediction.

[0029] S4. Graph neural network post-processing: A topological map is constructed based on the coordinates of meteorological stations and grid points. The meteorological station observations are taken as true values. The graph neural network (GNN) is used to perform spatial dependency modeling and optimization correction on the generated forecast data to improve the forecast accuracy of extreme precipitation events.

[0030] In some embodiments, in step S4, a topological map is constructed based on the coordinates of the meteorological station and the ERA5 data grid points to characterize the spatial relationship; the ERA5 rainfall data and the data interpolated to the meteorological station are used as the input of the graph neural network (GNN); the observation value of the meteorological station is used as the true value, and the GNN network is trained as a post-processing module to perform spatial dependency modeling and optimization correction on the generated forecast data to improve the forecast accuracy.

[0031] In some embodiments, in step S4, a forecast correction module is used to map the forecast data generated by cGAN to the location of the meteorological station through interpolation, and the difference with the real station data is optimized. Combined with the GNN spatial information transmission mechanism, the rainfall data between adjacent grid points is adjusted to achieve spatial detail enhancement of the generated regular forecast data, so as to solve the over-smoothing problem of extreme rainfall forecasts and make the generated data more in line with the actual surface observation rainfall data.

[0032] In some embodiments, step S4 specifically includes the following steps: interpolating the forecast data generated by cGAN to the spatial position of the meteorological station, and mapping the forecast data to each meteorological station using an interpolation function; merging the interpolated forecast data with the original forecast data as the input of the graph neural network (GNN); constructing a graph structure of the graph neural network, in which the nodes include meteorological stations and cGAN forecast data grid points, and the edges are constructed according to the distances between the stations; inputting the node feature matrix and the adjacency matrix into the GNN, capturing spatial dependencies and generating optimized forecast data through message passing, message aggregation, node update, and multi-layer propagation mechanisms; using the meteorological station observation value as the true value, calculating the mean square error (MSE) loss, and updating the GNN network parameters through back propagation to achieve optimized correction of the forecast data.

[0033] In the present invention, multi-source meteorological data preprocessing provides standardized input for feature extraction and fusion, feature extraction and fusion provide enhanced features for generating forecast data, the generated forecast data generates preliminary results through cGAN, and graph neural network post-processing optimizes and corrects the preliminary results, ultimately achieving high-precision extreme precipitation forecast and warning.

[0034] The following further describes specific embodiments of the present invention and examples of algorithm implementation thereof.

[0035] The present invention proposes a regional extreme precipitation forecasting and warning method based on conditional generative adversarial network (cGAN). A multi-source meteorological data feature fusion framework is constructed based on the attention mechanism, a rainfall synergy factor and a physical consistency discriminator of rainfall are added, and a graph neural network (GNN) optimization module is added, which significantly improves the forecasting accuracy and reliability of extreme precipitation events.

[0036] In one embodiment of the present invention, an adaptive normalization method is first used to uniformly process multi-source meteorological data, including the spatiotemporal alignment of historical precipitation data, and a differentiable interpolation algorithm is used to achieve spatial consistency of data with different resolutions.

[0037] In terms of model architecture, an embodiment of the present invention proposes a feature fusion cGAN network based on an attention mechanism, which adopts a multi-branch structure to capture the interannual average daily characteristics and anomaly characteristics, and constructs a multi-scale feature pyramid to enhance the representation ability of rainfall spatial modalities. At the same time, a rainfall synergy factor and a physical consistency discriminator of rainfall are constructed to enhance physical constraints. In addition, a two-step mechanism is innovatively introduced to further improve the forecast accuracy. Specifically, in order to make the forecast results more consistent with the actual meteorological observation results and solve the problem of over-smoothing of extreme rainfall events, the present invention constructs a topological map based on the coordinates of meteorological stations and grid points, takes the reanalysis data and its interpolation results at the observation station as input, and takes the meteorological station observation value as the true value. The forecast results are further optimized using a graph neural network (GNN), thereby improving the accuracy of the forecast.

[0038] The present invention adopts the following main steps (see Figure 1 ): 1) Preprocessing of multi-source meteorological data: random noise disturbance is added to all the input standardized multi-source meteorological data, and the data of different resolutions are aligned through a differentiable interpolation algorithm to ensure spatial consistency and reduce spatial bias.

[0039] 2) Feature extraction and fusion: The preprocessed data is input into the encoding module of cGAN. Specifically, the annual daily average features and anomaly features are extracted through a multi-branch structure, and multi-source features are fused based on the attention mechanism to improve the multi-source data representation capability. A multi-scale feature extraction module is constructed to enhance the spatial modal representation capability of rainfall.

[0040] 3) Generate forecast data: Input the encoded features into the decoder module of the cGAN network, and enhance the decoder’s generation capability through the conditional discriminator.

[0041] 4) cGNN post-processing: A topological map is constructed based on the coordinates of the meteorological stations and ERA5 data grid points. The ERA5 rainfall data and the data interpolated to the meteorological stations based on ERA5 are used as input. The meteorological station observations are taken as the true values, and the GNN network is trained as a post-processing module to optimize the cGAN forecast results and improve their accuracy and reliability.

[0042] In steps 1) and 2), the present invention first interpolates the standardized and random noise-perturbed multi-source meteorological data to a uniform spatial resolution through a differentiable interpolation algorithm to ensure spatial consistency. Subsequently, preliminary feature extraction is performed on the multi-source data, including interannual daily average features and anomaly features, and features from different data sources are fused through an attention mechanism method. The fused features are then subjected to spatial scale feature extraction to enhance feature representation capabilities.

[0043] In step 3), the present invention generates forecast data by decoding the coded features. Then, in order to detect whether the generated forecast data is more realistic, the present invention calculates the loss of the forecast data and the true value, and calculates the discriminant loss through the cGAN discriminator, and then reversely optimizes the network parameters. At the same time, the conditional discriminator supervises the output of the generator to ensure the accuracy of the generated data while making it more in line with physical laws.

[0044] In step 4), in order to make the generated data more consistent with the actual surface observed rainfall data and solve the over-smoothing problem of the forecast model for extreme rainfall conditions, the present invention adds a forecast correction module after the cGAN forecast model to increase the spatial description of the actual rainfall data.

[0045] The specific method flow of the present invention is as follows Figure 1 As shown, the network structures of cGAN and GNN are as follows Figure 2 shown.

[0046] The present invention can be divided into two stages.

[0047] The first stage is forecast data generation, the specific steps are as follows: In the forward forecasting work based on cGAN, the present invention takes the rainfall and its synergistic factor data at time t1, t2, ..., tn as the input of the model, and takes the rainfall and its synergistic factor (such as wind field, temperature, humidity, air pressure and other variables, using ERA5 data) at time t as the true value, that is, .

[0048] cGAN consists of a generator G and a discriminator D. The input data first passes through the trained encoder ɛ of G to map the multi-source data to the spatial potential representation Z, that is, , KL divergence will be used to regularize the latent space. The multi-source data input layer will first pass through a differentiable interpolation layer, which is mainly composed of a convolution (Conv) and a pooling layer (Pool), to unify the rainfall and its synergy factor data of different spatial resolutions to the same spatial scale. Secondly, each layer of data will be decomposed into inter-annual average daily data and anomaly data. The corresponding features are preliminarily extracted through the spatiotemporal feature extraction module of the Conv-transformer combination. The two features of multiple data are preliminarily spliced ​​through the concat layer, and the attention mechanism method is used to calculate the attention to each feature, and these features are fused based on the attention score. Then, the spatial multi-scale feature extraction operation is performed on the preliminarily fused features (see Figure 3 ) to obtain features that have stronger representation capabilities for semantic information and fine-grained information.

[0049] Next, the enhanced multi-scale spatial features are input into the decoding module γ of the subsequent cGAN G. Through layer-by-layer convolution operations, the potential features are mapped to the rainfall and its synergistic factor data at the next moment, that is, Specifically, in the decoding module, the extracted multi-scale spatial feature information is further spliced ​​and fused through the attention mechanism and concat, and the data of rainfall and its synergistic factors at time t+1 are predicted through layer-by-layer convolution operations. The discriminator D needs to judge the (X t+1 ,Y t+1 )_predict, whether it comes from the real data distribution (X t+1 ,Y t+1 )_true, so the decoder should generate (X t+1 ,Y t+1 )_predict is input into D to calculate the discriminative loss (Loss_D = D((X t+1 ,Y t+1 )_predict,(X t+1 ,Y t+1 )_true)), further constraining and optimizing the output of the generator to improve its modeling ability for real rainfall data. In addition, the forecast data at time t+1 will be used as input to the trained cGAN model to further generate forecast data at time t+2 to achieve multi-step prediction.

[0050] Among them, the noise z is a random vector sampled from a prior distribution (such as a standard normal distribution); the conditional information y is the input data y= (x r t , x c t );G r(z|y) is the rainfall data generated at time t+1; G c (z|y) is the generated meteorological synergy factor at time t+1.

[0051] The second stage is forecast post-processing. The specific steps are as follows: Specifically, the trained cGAN is used to generate the rainfall forecast data G at time t+1 r (z|y). Then the generated forecast data G r (z|y) is interpolated to the spatial position of the meteorological station. Assume that the position coordinates of the meteorological station are (S i ) i=1:N , where S i is the coordinate of the ith station, using the Kriging interpolation function Interpolate forecast data to station locations: The interpolated forecast data is combined with the original forecast data as input data.

[0052] The merged data is input into the trained graph neural network (GNN) to generate post-processed forecast data. The graph structure of the GNN network is as follows: 1) Node: Each meteorological station S i The node corresponding to the cGAN prediction data grid point, the node feature is X i t+1 ; 2) Edge: construct edges based on the distance between sites, and the adjacency matrix is ​​A. The input of GNN is the node feature matrix X={X i} i=1:N And the adjacency matrix A. The output of GNN is the post-processed forecast data. The present invention takes the meteorological station observation value as the true value, calculates the MSE loss, and gradually updates the GNN network parameters: In GNN, neighbor information propagation is one of the core mechanisms, which transmits and aggregates information between nodes through the graph structure (nodes and edges). Specifically: (1) Message passing: For each node v i , from its neighbor nodes Collect information. The message function M defines how to get information from neighbor nodes v j Pass information to node v i (2)Message aggregation: for node v i Aggregate all neighbor information. (3) Node update: Use the update function U to aggregate the aggregated message mi with the node’s current feature Combined, the feature representation of the node in the L+1 layer is generated. (4) Multi-layer propagation: Through multi-layer information transmission, the node features gradually merge the global information. (4) Output: The node features of the last layer It can be used as post-processed forecast data. Through multi-layer propagation, GNN can capture complex spatial dependencies and generate more reliable forecast results.

[0053] In summary, the present invention innovatively constructs a multi-source meteorological data feature fusion framework based on the attention mechanism, adds a rainfall synergy factor and a physical consistency discriminator of rainfall, and adds a graph neural network (GNN) optimization module, which significantly improves the forecast accuracy and reliability of extreme precipitation events.

[0054] The important features and innovative contributions of the present invention are: 1. Multi-source data enhancement and robustness improvement: By introducing Gaussian noise into multi-source input data, the model’s robustness to data uncertainty and noise is enhanced, and the stability of extreme rainfall prediction is improved. At the same time, the fusion of multi-source data will bring more information gain.

[0055] 2. Conditional GAN ​​modeling of nonlinear relationships: Using the conditional GAN ​​model, by taking meteorological synergistic factors (such as temperature, humidity, wind speed, etc.) as conditional information, the model is forced to learn the potential nonlinear relationship between rainfall and synergistic factors, thereby improving the accuracy of the prediction and enhancing the physical consistency and reliability of the forecast results.

[0056] 3. Post-processing correction driven by observation data: Based on the powerful modeling ability of graph neural networks (GNN) for complex spatial dependencies, this paper combines the topological structure between meteorological stations and grid points, and uses the neighboring information propagation mechanism to dynamically update node features, thereby effectively capturing local details and global spatial change patterns. This method performs post-processing correction on the initial prediction results of cGAN, which can significantly reduce system errors, enhance the model's ability to characterize extreme precipitation events, effectively alleviate the over-smoothing phenomenon in the prediction of heavy precipitation events, and improve the precision and reliability of forecast results.

[0057] Compared with the prior art, the present invention has the following significant advantages: 1. Improve data diversity and physical law modeling: Traditional GAN ​​models do not explicitly consider the learning differences between long-term and short-term rainfall patterns, lack physical meaning, and only capture structure and intensity information by optimizing the loss function, without fully considering the impact of synergy factors on rainfall forecasts. This invention innovatively designs a dual-branch learning structure for long-term periodic laws and short-term fluctuations, capturing long-term trends and short-term changes respectively, which is more in line with meteorological physical laws. At the same time, the conditional GAN ​​(cGAN) framework is used to input rainfall synergy factors (such as temperature, humidity, wind speed, etc.) as conditions into the generator, while also considering the constraints of synergy factors in the discriminator, forcing the model to pay attention to and learn the potential nonlinear relationship between rainfall and synergy factors, thereby generating more accurate precipitation forecasts that are more in line with physical laws.

[0058] 2. Improvement of the prediction accuracy of heavy precipitation: Traditional GAN ​​models often have over-smoothing in heavy precipitation forecasting, resulting in insufficient prediction ability of extreme precipitation events. This paper proposes a two-step forecasting method: first, the preliminary forecast results are generated by cGAN, and then the spatial information transmission between learning nodes is used by graph neural network (GNN), and the preliminary forecast results are corrected in combination with high-precision meteorological observation data. This method effectively alleviates the over-smoothing effect of cGAN on heavy precipitation and significantly improves the prediction accuracy and reliability of extreme precipitation events.

[0059] An embodiment of the present invention further provides a storage medium for storing a computer program, which at least performs the above method when executed.

[0060] An embodiment of the present invention further provides a control device, comprising a processor and a storage medium for storing a computer program; wherein the processor is configured to execute at least the method described above when executing the computer program.

[0061] An embodiment of the present invention further provides a processor, wherein the processor executes a computer program and at least executes the method described above.

[0062] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a ferromagnetic random access memory (Flash Memory), a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory can be a disk memory or a tape memory. The storage medium described in the embodiments of the present invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0063] In the several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0064] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0065] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0066] A person skilled in the art can understand that: all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, etc. Various media that can store program codes.

[0067] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0068] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0069] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0070] The features disclosed in several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0071] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, several equivalent substitutions or obvious variations can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A regional extreme precipitation forecasting and warning method based on cGAN, characterized in that: The following steps are involved: S1. Preprocessing of multi-source meteorological data: standardizing multi-source meteorological data and aligning spatial consistency of data with different resolutions, wherein the multi-source meteorological data includes rainfall and its synergistic factors; S2. Feature extraction and fusion: The preprocessed data is input into the encoding module of the conditional generative adversarial network (cGAN), and the interannual daily average features and anomaly features are extracted through a multi-branch structure. Multi-source features are fused based on the attention mechanism, and a multi-scale feature pyramid is constructed to enhance the representation ability of rainfall spatial modality. S3. Generate forecast data: Input the fused features into the generator of the conditional generative adversarial network (cGAN), and generate forecast data through the decoding module; Use the conditional discriminator to supervise the generated data, and the discriminator explicitly introduces the rainfall coordination factor as a physical constraint condition, and optimizes the network parameters through the adversarial loss function and the physical consistency loss, so that the generated data conforms to the laws of meteorological physics; S4. Graph neural network post-processing: A topological map is constructed based on the coordinates of meteorological stations and grid points. The meteorological station observations are taken as true values. The graph neural network (GNN) is used to perform spatial dependency modeling and optimization correction on the generated forecast data, thereby improving the forecast accuracy of extreme precipitation events.

2. The regional extreme precipitation forecasting and early warning method based on cGAN according to claim 1 is characterized in that: Step S1 specifically includes: Adaptive normalization methods are used to uniformly process multi-source meteorological data, including the spatiotemporal alignment of historical precipitation data; The standardized multi-source meteorological data are processed for spatial consistency using a differentiable interpolation algorithm, and data with different resolutions are interpolated to a uniform spatial resolution. Random noise perturbations are added to the standardized multi-source meteorological data to enhance the robustness of the model to data uncertainty and noise.

3. The regional extreme precipitation forecasting and early warning method based on cGAN according to claim 1 is characterized in that: In step S1, rainfall data of multiple time steps and its synergistic factor data are used as model inputs. The synergistic factors include wind field, temperature, humidity, and air pressure variables, and ERA5 data is used as the true value source to construct a time series data set for model training and prediction.

4. The regional extreme precipitation forecasting and early warning method based on cGAN according to claim 1 is characterized in that: Step S2 specifically includes: The preprocessed multi-source meteorological data is input into the encoder of cGAN, the data is mapped to the spatial latent representation, and the KL divergence is used to regularize the latent space; The rainfall and its synergistic factor data of different spatial resolutions are uniformly processed through a differentiable interpolation layer, which consists of a convolutional layer and a pooling layer; The data is decomposed into interannual daily average data and anomaly data, and the spatiotemporal feature extraction module of Conv-Transformer combination is used to extract features of the two types of data respectively; The annual daily average features and anomaly features of multiple data are concatenated through the concat layer, the attention mechanism is used to calculate the attention of each feature, and the features are fused based on the attention scores; The fused features are subjected to spatial multi-scale feature extraction to enhance the representation capability of semantic information and fine-grained information.

5. The regional extreme precipitation forecasting and early warning method based on cGAN according to claim 1 is characterized in that: Step S3 specifically includes: The enhanced multi-scale spatial features are input into the decoding module of the cGAN generator, and the potential features are mapped into the rainfall and its synergistic factor data at the next moment through layer-by-layer convolution operations; In the decoding module, the attention mechanism and feature concatenation are used to further integrate multi-scale spatial feature information to predict the rainfall and its synergistic factor data at the next moment; The forecast data generated by the generator is input into the discriminator. The discriminator determines whether the generated data comes from the real data distribution by calculating the discriminant loss, further constraining and optimizing the output of the generator and improving its modeling ability for real rainfall data. The generated forecast data is used as new input into the trained cGAN model to achieve multi-step prediction.

6. The regional extreme precipitation forecasting and early warning method based on cGAN according to claim 1 is characterized in that: In step S4, a topological map is constructed based on the coordinates of the meteorological stations and the ERA5 data grid points to characterize the spatial relationship; The ERA5 rainfall data and its interpolation to the meteorological station data are used as the input of the graph neural network (GNN); The observation values ​​of meteorological stations are taken as the true values, and the GNN network is trained as the post-processing module to perform spatial dependency modeling and optimization correction on the generated forecast data to improve the forecast accuracy.

7. The regional extreme precipitation forecasting and early warning method based on cGAN according to claim 1 is characterized in that: In step S4, the forecast correction module is used to map the forecast data generated by cGAN to the location of the meteorological station through interpolation, and the difference with the real station data is optimized. Combined with the GNN spatial information transmission mechanism, the rainfall data between adjacent grid points is adjusted to enhance the spatial details of the generated regular forecast data, so as to solve the over-smoothing problem of extreme rainfall forecasts and make the generated data more consistent with the actual surface observation rainfall data.

8. The regional extreme precipitation forecasting and early warning method based on cGAN according to claim 1 is characterized in that: Step S4 specifically includes the following steps: Interpolate the forecast data generated by cGAN to the spatial location of the meteorological station, and use the interpolation function to map the forecast data to each meteorological station; The interpolated forecast data is combined with the original forecast data as the input of the graph neural network (GNN); Construct the graph structure of the graph neural network, where the nodes include meteorological stations and cGAN forecast data grid points, and the edges are constructed based on the distance between stations; The node feature matrix and adjacency matrix are input into GNN to capture spatial dependencies and generate optimized forecast data through message passing, message aggregation, node update and multi-layer propagation mechanisms; The meteorological station observations are taken as the true values, the mean square error (MSE) loss is calculated, and the GNN network parameters are updated through back propagation to achieve optimal correction of the forecast data.

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

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images