Hectometer-level short temporary rainfall forecasting method based on generative adversarial network downscaling and physical constraint

Abstract

The invention provides a hectometer-level short temporary rainfall forecasting method based on generative adversarial network downscaling and physical constraint, which belongs to the technical field of weather forecast, and is characterized in that S-band radar jigsaw and high-resolution X-band radar observation data are fused through an adaptive rule to generate a fused radar echo combined reflectivity, and the combined reflectivity is used for forecasting the rainfall of the hectometer-level short temporary rainfall. A super-resolution downscaling model is constructed by using a generative adversarial network, a long-sequence high-resolution radar echo training data set is reconstructed, a deep learning forecasting model is constructed, a high-resolution radar echo sequence is used as input, space-time modeling training is performed by using a composite loss function, and a future radar echo forecasting field is output. The Z-R relation is converted into quantitative rainfall forecast, and finally the quantitative rainfall forecast with the spatial resolution reaching the hectometer level is generated. The interpretability of the model is enhanced through physical constraint, the bottleneck of insufficient hectometer-level resolution training data is effectively solved, and the spatial refinement degree of short temporary rainfall forecasting and the forecasting capacity for a severe convection system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a method for short-term precipitation forecasting at the 100-meter level based on generative adversarial networks for downscaling and physical constraints. Background Technology

[0002] With societal development, the demand for meteorological services is increasing across all sectors, with higher expectations for more accurate weather forecasts. Short-duration heavy rainfall is a key hazard causing urban flooding, flash floods, and geological disasters, exhibiting strong spatiotemporal heterogeneity in its formation and dissipation. Traditional operational forecasts, with their kilometer-level (1-5 km) resolution, struggle to effectively depict the fine structure of precipitation influenced by urban underlying surfaces and complex terrain, failing to meet the need for targeted, quantitative early warnings for key areas (such as transportation hubs and low-lying areas). Therefore, developing intelligent short-term forecasting technology with spatial resolution up to 100 meters and update frequencies up to minutes has become a core direction for improving the accuracy of meteorological disaster prevention and mitigation.

[0003] However, achieving operationally viable short-term precipitation forecasts at the 100-meter level still faces a series of interconnected technical bottlenecks: High-resolution training data is severely lacking: Data-driven deep learning models rely on massive amounts of high-quality samples. In existing operational radar networks, S-band radars, with their wide coverage, weak signal attenuation, and abundant historical data, have combined reflectivity spatial resolution (typically ≥1 km) that is insufficient to directly support 100-meter-level modeling. X-band radars, which possess 100-meter-level detection potential, suffer from sparse deployment, susceptibility to attenuation and obstruction interference, and a relatively short operational history, making it impossible to independently accumulate long-term, continuous, and complete high-resolution datasets. This constitutes the original data bottleneck for model training.

[0004] The challenge of fusing multi-source heterogeneous observation data: A single radar data source cannot simultaneously meet the dual requirements of coverage and resolution. S-band and X-band radars differ significantly in detection characteristics, spatial resolution, and error sources. Designing intelligent fusion algorithms to effectively combine low-resolution observations with high-resolution information to generate fused data that maintains large-scale integrity while containing detailed convective structures is the primary challenge for improving 100-meter-level forecasts.

[0005] The dilemma of balancing forecast model refinement and efficiency: Directly generating forecasts at the 100-meter grid level means a sharp increase in the dimensionality of the model output, posing a severe challenge to computational architecture and efficiency. Existing deep learning extrapolation models (such as optical flow and ConvLSTM) generally face problems of blurred details and long computation time when improving resolution. The model must reconstruct complex spatiotemporal evolution features (such as convection formation and merging) while meeting the hard constraints of operational systems on forecast lead time (usually requiring computation to be completed within a few minutes).

[0006] The risk of physical consistency in purely data-driven models: Deep learning models trained entirely on statistical patterns from historical data may generate results that violate fundamental physical principles of atmospheric motion. For example, they may produce false strong echoes in impossible areas or fail to reasonably reflect the forcing effect of topography on precipitation. This physical inconsistency significantly reduces forecasters' trust in the product and is a key obstacle to the operational application of artificial intelligence models.

[0007] Currently, industry solutions often focus on individual aspects, such as using traditional interpolation methods for downscaling, extrapolating from a single data source, or directly running computationally expensive numerical models at the 100-meter level. These methods often suffer from significant loss of detail, limited forecasting capabilities, or inability to meet real-time business requirements, failing to form a systematic and implementable technical framework from data to products. Summary of the Invention

[0008] The purpose of this invention is to provide a method for short-term precipitation forecasting at the 100-meter level based on generative adversarial networks (GANs) for downscaling and physical constraints. This method addresses the shortcomings and technical problems of existing short-term nowcasting, such as the shortage of high-resolution training data and low accuracy in forecasting severe convective systems. It collaboratively overcomes the technical challenges across the entire chain, from high-quality data production to high-precision forecast model design and the generation of operationally usable products, thereby truly achieving stable, reliable, and efficient operational short-term precipitation forecasting capabilities at the 100-meter level.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for short-term precipitation forecasting at the 100-meter level based on generative adversarial networks and physical constraints, the method comprising the following steps: Step 1: Synchronously acquire S-band radar mosaic data and X-band radar observation data. Based on the preset adaptive fusion rules, perform quality control and spatiotemporal matching on the S-band and X-band data. Use the processed low-resolution S-band radar data as model input and the fused high-resolution radar echo data as supervision labels to construct a heterogeneous radar training sample set containing the correspondence between input and labels. Step 2: Construct a super-resolution downscaling model based on generative adversarial networks. Train the model using the heterogeneous radar training sample set to learn the nonlinear mapping relationship from low resolution to high resolution. Use the trained super-resolution downscaling model to process long sequences of historical S-band radar data in batches, reconstruct and generate long sequences of high-resolution radar echo reconstruction sequences with a spatial resolution of hundreds of meters, and establish a high-resolution reconstructed radar echo historical dataset. Step 3: Construct a radar echo sequence prediction model, which includes an encoder, an evolution network, and a generator; use the high-resolution reconstructed radar echo historical dataset as the sole training data source, and use a composite loss function with introduced physical constraints to perform spatiotemporal modeling training on the model, so that the model learns the characteristics of echo advection motion and intensity evolution at the hundred-meter scale.

[0010] Step 4: In the real-time forecasting stage, the latest S-band radar mosaic and X-band radar observation data are acquired simultaneously; the real-time S-band data is input into the trained super-resolution downscaling model to generate radar echoes with a resolution of 100 meters and use them as the background field; based on preset fusion rules, the radar echo background field with the 100-meter resolution is fused with the real-time X-band radar observation data to generate a real-time high-resolution fused echo field, wherein X-band observation data is used preferentially within the effective range of the X-band radar; the real-time high-resolution fused echo field is input into the trained radar echo sequence forecasting model to output the radar echo forecast field for the future preset time period, and after ZR relation inversion and intelligent correction, a quantitative precipitation forecast with a resolution of 100 meters is generated.

[0011] Furthermore, in step 1, the adaptive fusion rule is as follows: the S-band radar mosaic is used as the background field to ensure the continuity and integrity of the large-scale echo structure, and the X-band radar observation is used as the enhancement field; when the reflectivity factor detected by the X-band radar exceeds the first preset threshold and the difference between it and the background field exceeds the second preset threshold, the X-band observation value is used to replace the value of the corresponding grid point in the background field.

[0012] Furthermore, in step 2, the super-resolution downscaling model employs a generator containing residual connections or channel attention mechanisms, as well as a multi-scale discriminator. Its training adopts a two-stage strategy: the first stage jointly optimizes the generator and discriminator, and the second stage freezes the discriminator and adjusts the weights of the loss function to focus on improving the reconstruction accuracy of strong echo regions.

[0013] Furthermore, in step 2, the loss function of the super-resolution downscaling model is a composite form, which includes adversarial loss, pixel-level L1 loss, and a weighted loss term or perceptual loss term for the high reflectivity range.

[0014] Furthermore, in step 3, the prediction model is trained using a composite loss function, which includes: mean squared error loss for constraining the overall error of the sequence, bidirectional loss for improving spatiotemporal coherence, and adversarial loss for encouraging the generation of more realistic echo textures.

[0015] Furthermore, in step 4, the intelligent correction rule is as follows: when the precipitation intensity at a certain grid point in the 100-meter forecast product exceeds the third preset threshold, while the precipitation intensity in the corresponding area in the kilometer-level background field is lower than the fourth preset threshold, the forecast value of the background field is preferentially adopted or weighted; when both indicate precipitation, the 100-meter forecast product and the kilometer-level background field are dynamically weighted and fused.

[0016] Furthermore, in step 3, the forecast model is optimized using at least one of the following techniques: model pruning, quantization, or mixed-precision inference, to meet the timeliness requirements of operational real-time forecasting.

[0017] The present invention, by adopting the above-described technical solution, has the following beneficial effects: (1) This invention transforms widely existing low-resolution data into high-quality high-resolution data through innovative fusion and generation strategies, and breaks through the construction of long-sequence, hundred-meter-level meteorological training datasets, providing a solid data foundation for deep learning models.

[0018] (2) The accuracy of forecasts has been significantly improved: the end-to-end model trained on 100-meter-level data can directly output 100-meter-level grid forecasts, which effectively improves the forecast accuracy of short-term heavy precipitation and the spatial detail performance is better than that of traditional kilometer-level forecasts.

[0019] (3) Enhanced forecasting skills, especially for heavy precipitation: The weighted loss of extreme echoes in the downscaling model and the emphasis on spatiotemporal evolution and physical consistency in the forecasting model have jointly improved the hit rate (POD) and critical success index (CSI) for heavy precipitation events.

[0020] (4) Ensured the business availability and physical credibility of the product: By introducing evolutionary networks, physical transformation formulas and rule-based intelligent correction, the data-driven results were "constrained" on physical laws and business experience, which effectively avoided unreasonable forecasts and improved forecasters' trust in the product.

[0021] (5) A balance between efficiency and accuracy is achieved: Through lightweight network design, two-stage training strategy and model optimization, the entire system can complete the calculation within the time specified by the business (within a few minutes) to meet the timeliness requirements of real-time forecast. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the network structure of the adversarial network downscaling model and short-term precipitation forecasting model of this invention; Figure 3 This is a schematic diagram comparing the radar echo effects before and after processing by the downscaling method of this invention; Figure 4 This is a schematic diagram of the real-time forecast data generation and intelligent correction process of the present invention; Figure 5 This is a schematic diagram of the real-time 100-meter-level short-term precipitation forecast product of this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are merely to provide the reader with a thorough understanding of one or more aspects of the present invention, and these aspects of the invention can be implemented even without these specific details.

[0024] like Figure 1 As shown, the short-term precipitation forecasting method at the 100-meter level described in this invention mainly includes three core technical components: (1) multi-source data fusion and construction of high-resolution historical datasets; (2) construction and training of radar extrapolation models based on high-resolution datasets; and (3) generation of real-time forecast products and physical intelligent correction. The key steps are described in detail below with reference to the accompanying figures.

[0025] 1. Multi-source data fusion and construction of high-resolution historical datasets The goal of this step is to generate historical long-sequence training data (hundreds of meters in length) for model training; its core is... Figure 2 The physical information augmented generative adversarial network (GAN) downscaling model shown.

[0026] S1: Multi-source data acquisition and adaptive fusion First, S-band radar mosaic (1 km resolution) and X-band radar observation data (100 m resolution) within the forecast area are acquired simultaneously. Rigorous quality control measures, including attenuation correction and ground clutter suppression, are applied to the X-band data. The S-band mosaic and X-band data are then unified to a 100-meter target grid using bilinear interpolation to ensure spatial consistency.

[0027] Next, as Figure 1 The "Data Fusion" module demonstrates the implementation of an adaptive fusion algorithm: using S-band data as the background field to ensure continuous echoes over a wide area; and using X-band data as the enhancement field to provide fine structure. Fusion rules are designed; for example, when an X-band detection shows a reflectivity factor greater than 30 dBZ, and its value differs from the S-band background field by more than 10 dBZ, the X-band observation is adopted at that grid point. Finally, a fused reflectivity analysis field (denoted as fused radar reflectivity) with a unified spatial resolution of 100 meters is generated. This field will serve as the "high-resolution true value" for subsequent downscaling models.

[0028] S2: Downscaling model training, such as Figure 2As shown: To construct a long-term historical dataset at the 100-meter scale, we trained a specialized GAN model to upscale the long-term 1-kilometer S-band historical data to 100 meters.

[0029] Network architecture: such as Figure 2 As shown, the generator (G) employs an improved U-Net structure. Its encoder takes a low-resolution S-band image as input and extracts features through convolutional downsampling. The decoder gradually reconstructs high-resolution details through upsampling and skip connections, where a channel attention mechanism can be embedded to focus on important features. The discriminator (D) uses a multi-scale PatchGAN structure, which performs block discrimination on the generated image at multiple resolutions, forcing the generator to ensure texture realism at different scales simultaneously.

[0030] The specific process of DEM terrain processing is as follows: First, a high-resolution digital elevation model matching the forecast area is obtained. Then, key spatial structure features in the terrain are automatically extracted through a lightweight convolutional coding module. Finally, the terrain features are fused with meteorological data in the model feature layer, thereby significantly enhancing the model's ability to characterize the refined evolution process of precipitation in complex terrain.

[0031] Loss function and training: The model uses a composite loss function. .in, To combat the losses; L1 loss is used to constrain pixel values; To mitigate perceptual loss, image structure is constrained in the feature space. Specifically, this can be applied to high reflectivity regions (e.g., ≥40 dBZ). The loss function is given a higher weight to improve the prediction ability for strong echoes. The training employs a two-stage strategy: first... and A pre-trained generator is then introduced, followed by adversarial training with a discriminator, and the process can be stabilized using gradient penalty.

[0032] Results and Dataset Generation: A trained downscaling model is used to batch process long-term historical S-band radar data, reconstructing radar echo sequences covering a long period with a spatial resolution of 100 meters. Based on preset fusion rules, the 100-meter resolution radar echo background field is fused with real-time X-band radar observation data to generate a high-resolution fused echo field, thereby constructing a high-resolution reconstructed historical radar echo dataset. Figure 3As shown in the comparative diagram, this model can take low-resolution, blurry S-band radar echoes as input (b) and generate a 100-meter-level fused echo field (c) with clear convective cell boundaries and fine texture. Its reconstruction quality is significantly better than traditional interpolation methods, and it preserves the fine structure of the echoes from X-band radar observations (a). This dataset not only effectively restores the details of strong echo regions but also compensates for the observational gaps in attenuation regions from X-band radar, providing a high-quality training foundation for subsequent forecast models.

[0033] 2. Construction and Training of Radar Extrapolation Model Based on High-Resolution Dataset This section utilizes the aforementioned dataset to train a lightweight, end-to-end sequence prediction model, the structure of which is as follows: Figure 4 As shown.

[0034] S3: Forecast Model Construction and Training like Figure 4 As shown, the prediction model adopts an encoder-evolutionary network-generator architecture.

[0035] Encoder: The input is a fused radar reflectivity echo sequence over a past period (e.g., 1 hour, 10 frames). Spatiotemporal features are extracted using modules such as 3D convolution or ConvLSTM.

[0036] EvolveNet: This is a lightweight module used to simulate complex nonlinear evolution processes in echo fields, such as advection, generation, and cancellation. It receives encoded features and outputs the evolved state features.

[0037] Generator: Based on the characteristics of the output of the evolution network, it gradually decodes and generates radar echo fields for future moments (e.g., 3 hours in the future, a total of 30 frames), with the output spatial resolution maintained at 100 meters.

[0038] Training strategy: The model training adopts a composite objective function that includes mean squared error loss, bidirectional derivative loss (used to constrain spatiotemporal smoothness), and adversarial loss.

[0039] Model Loss Function Design: In designing the model's loss function, a refined weighting mechanism based on echo intensity was established to address the need for accurate forecasting of strong echoes during short-term precipitation. Specifically, the radar echo sequence, divided into 6-minute intervals, is categorized into seven intensity levels and assigned progressive weights: 0–30 dBZ (weight 1), 30–40 dBZ (weight 2), 40–50 dBZ (weight 5), 50–55 dBZ (weight 10), 55–60 dBZ (weight 20), 60–65 dBZ (weight 50), and ≥65 dBZ (weight 90). This design significantly enhances the model's focus on strong echoes above 50 dBZ during training, effectively guiding it to improve its ability to capture and forecast strong precipitation.

[0040] A progressive training approach can be adopted: first, train all modules together, then fix the evolution network and fine-tune the generator to optimize details.

[0041] 3. Real-time forecast product generation and physical intelligent correction This section describes the process of applying the trained model to real-time business operations and generating the final product. The process is as follows: Figure 4 As shown, examples of real-time and forecast products are as follows: Figure 5 .

[0042] S4: Real-time echo sequence prediction In real-time operations, the system automatically acquires the latest S-band radar echoes, inputs them into a downscaling model to obtain a 100-meter-level echo background field (generated in step S2), and fuses them with real-time X-band observations to obtain a real-time fused radar reflectivity sequence, which is then input into a pre-trained forecast model (obtained in step S3). The model directly outputs a radar echo grid forecast sequence for the next 0-3 hours, with a time interval of 6 minutes and a spatial resolution of 100 meters.

[0043] ZR Relationship Conversion: Utilizing Physical Formulas (For example, convective precipitation parameters a=300, b=1.4), converting the reflectivity factor (Z) in the echo forecast sequence into the instantaneous precipitation rate (R).

[0044] Multi-source collaborative intelligent correction: External kilometer-level and minute-level quantitative precipitation forecasts (QPF) are introduced as the background field. Correction is performed using design rules. Suppressing "false" precipitation: If the precipitation rate at a certain grid point in the 100-meter level forecast is ≥10mm / 6min, while the corresponding area of ​​the background field in the kilometer level is ≤5mm / 6min, it is judged as a possible false signal, and the background field value is adopted first or significantly weighted.

[0045] Intensity rationalization and fusion: If both indicate significant precipitation, a dynamic weighted average is performed on the 100-meter level forecast and the kilometer level background field to make the intensity more reasonable while retaining the high-resolution structure.

[0046] Post-processing and product generation: The corrected precipitation rate field is time-accumulated (to obtain hourly precipitation), and trace precipitation (e.g., <0.1 mm / 6min) is removed, ultimately generating a standardized quantitative precipitation forecast product at the 100-meter level. The product is stored in NetCDF format and automatically plotted as shown in the image. Figure 5 The thematic forecast map shown clearly displays the fine distribution of precipitation at different time points in the future.

[0047] The entire process is integrated into an automated business system and is triggered periodically by a scheduler. The final generated short-term precipitation forecast product (e.g., 100-meter level) is... Figure 5 For example, data can be pushed in real time to the forecaster's work platform or the early warning release system through the data interface, providing high spatiotemporal resolution scientific and technological support for disaster prevention and mitigation decision-making.

[0048] The following example, using the forecast of precipitation for the next 0-3 hours starting at 19:00 (Beijing time) on June 4, 2024, illustrates the operational implementation process of this invention in detail: (1) Real-time data acquisition and preprocessing (triggered at 19:00). The system automatically triggers at the top of the hour to acquire and process the following multi-source heterogeneous data in real time: Radar data: Acquired S-band radar composite reflectivity mosaic from the national-level SWAN (6-minute temporal resolution, 1-kilometer spatial resolution), and simultaneously acquired the latest volume scan data from the Nanning X-band dual-polarization radar. Real-time quality control, attenuation correction, and ground clutter removal were performed on the X-band data, and the composite reflectivity at 100-meter resolution was calculated.

[0049] Background field data: Integrating quantitative precipitation forecast (QPF) products generated by a kilometer-level short-term precipitation forecast model as a collaborative background field for physical correction.

[0050] (2) Adaptive fusion of multi-source radar data. The system executes a rule-based intelligent fusion algorithm: First, the real-time acquired S-band radar mosaic is input into the trained super-resolution downscaling model to generate a radar echo background field with a resolution of 100 meters. Simultaneously, quality control and attenuation correction are performed on the real-time X-band radar observation data, and it is then unified to the same 100-meter target grid through bilinear interpolation to ensure spatial consistency.

[0051] A "background field-enhanced field" fusion strategy was adopted: the downscaled 100-meter-level echo was used as the background field to maintain the continuity and integrity of the large-scale echo structure; X-band observations were used as the enhanced field to embed fine-scale echo information. In the analysis at 19:00 in this example, the system identified a convective echo band generated in the northern part of Nanning city with a reflectivity exceeding 45 dBZ in the background field, while X-band observations in this area only showed about 30 dBZ due to attenuation, especially with significant signal loss behind strong echoes. Based on preset fusion rules (such as adopting background field data when X-band reflectivity exceeds a threshold and differs significantly from the background field), the system generated a fused radar reflectivity analysis field with physically corrected intensity and a spatial resolution of 100 meters. This fused field effectively compensated for the signal in the attenuated region while preserving high-resolution details of the X-band, significantly improving the intensity characterization and structural fineness of the convective echo.

[0052] (3) Input sequence construction and deep learning extrapolation (19:01–19:03). The system extracts the fused radar reflectivity sequence (10, 600, 600) from the past hour (18:06–19:00, 10 frames) and inputs it into the pre-trained encoder-evolutionary network-generator end-to-end forecast model. Based on the learned complex spatiotemporal evolution laws, the model outputs the radar echo grid forecast field (30, 600, 600) with a resolution of 100 meters for the next 3 hours (19:06–22:00, 30 frames). The entire inference process takes about 100 seconds on a conventional CPU server, meeting the real-time requirements of the business.

[0053] (4) Physical transformation and intelligent correction (19:03–19:05). The following refinement processing is performed on the predicted echo field: ZR physical conversion: using the ZR relationship applicable to convective precipitation ( This converts each frame of echo field into an instantaneous precipitation rate field.

[0054] Multi-source collaborative intelligent correction: A kilometer-level QPF background field is introduced for physical constraints. According to the "extreme precipitation and false precipitation suppression rules", precipitation forecasts from the background field are prioritized, while dynamic weighted fusion is performed to optimize the intensity distribution while preserving the fine morphology of the hundred-meter-level forecasts.

[0055] Post-processing integration: The instantaneous precipitation rate is accumulated to obtain the cumulative precipitation for the next 1, 2, and 3 hours, and trace precipitation of <0.1 mm is removed, and a reasonableness check is performed.

[0056] (5) Product generation and business distribution (to be completed before 19:06). The system automatically generates and publishes the final forecast product: Archived in standard NetCDF format with 100-meter gridded forecast data, adhering to the CF convention, and including complete coordinates, attributes, and forecast metadata. Thematic forecast maps are automatically generated, using commonly used operational precipitation color scales to visually display the fine spatial structure of forecast precipitation (e.g., ...). Figure 5 (As shown).

[0057] Compared with traditional methods: In this case, traditional extrapolation methods, limited by the low resolution and smoothing effect of a single SWAN mosaic, failed to effectively capture convective signals in the northern part of Nanning city, and lacked the ability to characterize the fine structure of strong echoes. This invention, through a full-chain collaboration of multi-source fusion, generative adversarial network downscaling, end-to-end deep learning extrapolation, and physical rule correction, achieves high-resolution, high-precision, and physically constrained forecasts of the generation and dissipation evolution of small- and medium-scale convective systems. This significantly improves the lead time and spatial precision of short-term heavy rainfall warnings, demonstrating its operational value in the context of heavy rainfall events during the flood season.

[0058] Matters not covered in this invention are common knowledge.

[0059] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for short-term precipitation forecasting at the 100-meter level based on generative adversarial networks for downscaling and physical constraints, characterized in that... The method includes the following steps: Step 1: Synchronously acquire S-band radar mosaic data and X-band radar observation data. Based on the preset adaptive fusion rules, perform quality control and spatiotemporal matching on the S-band and X-band data. Use the processed low-resolution S-band radar data as model input and the fused high-resolution radar echo data as supervision labels to construct a heterogeneous radar training sample set containing the correspondence between input and labels. Step 2: Construct a super-resolution downscaling model based on generative adversarial networks. Train the model using the heterogeneous radar training sample set to learn the nonlinear mapping relationship from low resolution to high resolution. Use the trained super-resolution downscaling model to process long sequences of historical S-band radar data in batches, reconstruct and generate long sequences of high-resolution radar echo reconstruction sequences with a spatial resolution of hundreds of meters, and establish a high-resolution reconstructed radar echo historical dataset. Step 3: Construct a radar echo sequence prediction model, which includes an encoder, an evolution network, and a generator; use the high-resolution reconstructed radar echo historical dataset as the sole training data source, and use a composite loss function with introduced physical constraints to perform spatiotemporal modeling training on the model, so that the model learns the characteristics of echo advection motion and intensity evolution at the hundred-meter scale. Step 4: In the real-time forecasting stage, the latest S-band radar mosaic and X-band radar observation data are acquired simultaneously; the real-time S-band data is input into the trained super-resolution downscaling model to generate radar echoes with a resolution of 100 meters and use them as the background field; based on preset fusion rules, the radar echo background field with the 100-meter resolution is fused with the real-time X-band radar observation data to generate a real-time high-resolution fused echo field, wherein X-band observation data is used preferentially within the effective range of the X-band radar; the real-time high-resolution fused echo field is input into the trained radar echo sequence forecasting model to output the radar echo forecast field for the future preset time period, and after ZR relation inversion and intelligent correction, a quantitative precipitation forecast with a resolution of 100 meters is generated.

2. The method for short-term precipitation forecasting at the 100-meter level based on generative adversarial networks and physical constraints according to claim 1, characterized in that: In step 1, the adaptive fusion rule is as follows: the S-band radar mosaic is used as the background field to ensure the integrity of the large-scale echo structure and to compensate for the reflectivity attenuation problem of the X-band radar combination; the X-band radar observation is used as the enhancement field. When the reflectivity factor detected by the X-band radar exceeds the first preset threshold and the difference between it and the background field exceeds the second preset threshold, the X-band observation value is used to replace the value of the corresponding grid point in the background field.

3. The method for short-term precipitation forecasting at the 100-meter level based on generative adversarial networks and physical constraints according to claim 1, characterized in that: In step 2, the super-resolution downscaling model employs a generator that includes residual connections or channel attention mechanisms, as well as a multi-scale discriminator. Its training adopts a two-stage strategy. The first stage jointly optimizes the generator and discriminator, and the second stage freezes the discriminator and adjusts the weights of the loss function to focus on improving the reconstruction accuracy of strong echo regions.

4. The method for short-term precipitation forecasting at the 100-meter level based on generative adversarial networks and physical constraints according to claim 1, characterized in that: In step 2, the loss function of the super-resolution downscaling model is a composite form, which includes adversarial loss, pixel-level L1 loss, and a weighted loss term or perceptual loss term for the high reflectivity range.

5. The method for short-term precipitation forecasting at the 100-meter level based on generative adversarial networks and physical constraints according to claim 1, characterized in that: In step 3, the training of the prediction model is optimized using a composite loss function, which includes: mean squared error loss to constrain the overall error of the sequence, bidirectional loss to improve spatiotemporal coherence, and adversarial loss to encourage the generation of more realistic echo textures.

6. The method for short-term precipitation forecasting at the 100-meter level based on generative adversarial networks and physical constraints according to claim 1, characterized in that: In step 4, the intelligent correction rule is as follows: when the precipitation intensity at a certain grid point in the 100-meter forecast product exceeds the third preset threshold, while the precipitation intensity in the corresponding area in the kilometer-level background field is lower than the fourth preset threshold, the forecast value of the background field is adopted or weighted more towards it; when both indicate precipitation, the 100-meter forecast product and the kilometer-level background field are dynamically weighted and fused.

7. The method for short-term precipitation forecasting at the 100-meter level based on generative adversarial networks and physical constraints according to claim 1, characterized in that: In step 3, the forecast model is optimized using at least one of the following techniques: model pruning, quantization, or mixed-precision inference, to meet the timeliness requirements of operational real-time forecasts.

Images