Short-term rainfall prediction system and method based on multi-source radar data fusion

Abstract

This invention discloses a short-term precipitation forecasting system and method based on multi-source radar data fusion, belonging to the field of meteorological forecasting technology. It simultaneously acquires multi-band radar data, preprocesses it to construct a standardized data system, employs a dynamic weighting algorithm to calculate reliability grid-by-grid, forming a weight distribution matrix, extracts multi-scale precipitation features in parallel, integrates these features using an attention fusion mechanism, and locates key precipitation areas. Finally, it generates high-resolution short-term precipitation forecasts, outputs visual products, and classifies warning levels. This invention overcomes the limitations of single-band detection through multi-source radar data integration and dynamic weighting, comprehensively improving the comprehensiveness and accuracy of precipitation feature extraction. Furthermore, it accurately locates key precipitation areas using an attention fusion mechanism, generates high-resolution forecasts, and scientifically classifies warning levels, thus improving the spatiotemporal accuracy and operational efficiency of short-term precipitation forecasts and providing reliable support for meteorological warnings and emergency responses.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, specifically to a short-term rainfall forecasting system and method based on multi-source radar data fusion. Background Technology

[0002] Short-term precipitation forecasting is a core component of the meteorological disaster prevention and mitigation system, directly determining the timeliness of disaster warnings and the effectiveness of emergency response. Radar observation is a key technology for acquiring real-time precipitation information and supporting short-term forecasting. With the continuous iteration of meteorological observation technology, radars of different bands, relying on their own detection advantages, have become core equipment for precipitation monitoring. Multi-source radar collaborative observation has become a development trend in the field of meteorological observation. Currently, meteorological operations and disaster prevention and mitigation work place higher demands on the spatiotemporal accuracy and early warning reliability of short-term precipitation forecasts, urgently requiring the integration of multi-source radar data to achieve comprehensive capture and in-depth analysis of precipitation information. However, in practical applications, existing multi-source radar data still has technical shortcomings in key aspects such as data standardization, adaptive weight allocation, and multi-scale feature fusion, failing to fully unleash the collaborative observation value of multi-source radar. The industry urgently needs a complete multi-source radar data fusion and forecasting technology solution to promote the upgrading of short-term precipitation forecasting towards precision and intelligence.

[0003] Traditional short-term precipitation forecasting techniques rely heavily on single-band radar for observation and analysis, making it difficult to simultaneously grasp both large-scale precipitation patterns and small-scale convective details. Observational data has inherent limitations, and traditional data processing lacks a unified spatiotemporal benchmark and standardized system. Issues such as ground clutter interference and inconsistent coordinates significantly reduce data quality. Multi-source radar data cannot be effectively coordinated and adapted, and weight allocation often adopts a fixed pattern, failing to dynamically adjust data reliability based on real-time detection conditions. Feature extraction methods are simplistic and do not incorporate parallel processing based on the characteristics of different radar bands. Precipitation feature capture is incomplete, feature fusion lacks a precise attention mechanism, the ability to locate key precipitation areas is weak, forecast resolution and accuracy are insufficient, warning level classification is relatively crude, forecast products cannot efficiently integrate with external operational systems, and forecast deviations and warning delays are prone to occur under extreme precipitation weather conditions, failing to meet the actual needs of modern meteorological operations and disaster prevention and mitigation. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a short-term rainfall prediction system and method based on multi-source radar data fusion. This system simultaneously acquires S, C, and X-band radar data, removes ground clutter, unifies coordinates, and aligns spatiotemporally to construct a standardized data system. A three-dimensional dynamic weighting algorithm is used to calculate weights grid-by-grid, generating a dynamic weight matrix. Multi-scale precipitation characteristics of each band are extracted in parallel. Combining the weight matrix with spatiotemporal attention factors, a unified fusion feature map is generated through a cross-band attention fusion algorithm. Finally, a high-resolution rainfall intensity forecast is generated, disaster warning levels are classified, and a visualization product is output.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On one hand, a short-term rainfall prediction system based on multi-source radar data fusion, the system comprising:

[0006] Multi-source radar preprocessing module: Simultaneously acquires real-time observation echo data from S-band, C-band and X-band phased array radars, performs ground clutter removal, coordinate unification transformation and spatiotemporal reference alignment, extracts estimated values ​​of detection range, signal-to-noise ratio, wavelength and precipitation intensity, and constructs a standardized radar data system;

[0007] Dynamic weight calculation module: Based on standardized radar data and related parameters, a three-dimensional linkage dynamic weight algorithm is used to calculate the weight of the entire area grid point by point. After normalization and integration, a three-dimensional dynamic weight distribution matrix that fully covers the detection area is formed.

[0008] Multi-scale feature extraction module: Based on the radar detection characteristics of each band, parallel feature processing is performed to extract multi-scale precipitation features corresponding to S-band, C-band and X-band radar data respectively.

[0009] Attention Feature Fusion Module: Combining a three-dimensional dynamic weight matrix with multi-scale precipitation features, the module integrates multi-dimensional features through a cross-band attention fusion algorithm. It locates key precipitation areas using spatiotemporal attention factors and outputs a unified fusion feature map containing full-scale precipitation information.

[0010] Rainfall forecast generation module: Based on the unified fusion feature map, the module extrapolates the spatiotemporal evolution of precipitation, generates high-resolution rainfall intensity forecasts for the next 0-2 hours, performs spatial smoothing and outlier verification, classifies disaster warning levels, generates visualized forecast products, and converts them into a common meteorological format to enable integration with external business systems.

[0011] Furthermore, in the multi-source radar preprocessing module, the real-time observed echo data includes basic data from S-band conventional weather radar, C-band Doppler weather radar, and X-band phased array weather radar, covering parameters such as reflectivity factor, radial velocity, spectral width, differential reflectivity, differential propagation phase, and correlation coefficient. The S-band and C-band radars adopt a volumetric scanning mode, sequentially completing planar position display scans at 14 elevation angles from 0.5° to 19.5°, completing a complete volumetric scan every 6 minutes. The X-band phased array radar adopts an electronic scanning mode, sequentially completing planar position display scans at 20 elevation angles from 0.5° to 30°, completing a complete volumetric scan every 3 minutes. The data format conforms to the meteorological industry standard radar basic data format specification.

[0012] Furthermore, in the multi-source radar preprocessing module, the standardized radar data system is built on a unified planar grid. The entire domain is divided into grid points with a fixed spatial resolution of 1km×1km. The grid points maintain a one-to-one spatial correspondence. The grid contains three basic data layers: multi-band radar echo intensity field, radial velocity field, and spectral width field. Each data layer is normalized according to a unified spatial scale. All data layers are synchronously matched to the same time node to form a time-series data sequence. The standardized radar data system synchronously embeds the radar detection range, signal-to-noise ratio, radar operating wavelength, and precipitation intensity estimation values ​​corresponding to each grid point. All values ​​are bound to the grid points and stored in an orderly manner, ultimately forming a multi-band radar standardized data system with unified spatial reference, unified time reference, and unified data dimension.

[0013] Furthermore, the mathematical expression of the three-dimensional linkage dynamic weight algorithm in the dynamic weight calculation module is as follows:

[0014] ;

[0015] in, This is the final dynamic weight of the single-grid multi-wavelength radar data, with a value range of 0-1, reflecting the reliability of the radar data at that grid point. This is an adaptive correction coefficient, with a value ranging from 0 to 1; The radar wavelength is represented by S-band, C-band, and X-band, which correspond to fixed reference values. I represents the estimated real-time precipitation intensity of a single grid point, reflecting the impact of precipitation intensity on the reliability of radar data. D represents the detection distance from a single grid point to the corresponding radar station, which is a quantized value of the spatial interval between the corresponding grid point and the radar station, characterizing the impact of detection distance on radar data quality. S represents the signal-to-noise ratio of the radar signal at a single grid point, with a value ranging from 0 to 1, characterizing the quality level of the radar signal.

[0016] Furthermore, in the dynamic weight calculation module, the three-dimensional dynamic weight distribution matrix is ​​a gridded weight dataset covering the entire detection area, corresponding to the Cartesian coordinate grid and vertical height layer of the standardized radar data system. The dynamic weight values ​​of each band radar data are stored with grid points as the basic unit. In the formation process, the dynamic weight of a single grid point of S-band, C-band, and X-band radar is first calculated grid by grid and height layer by height using a three-dimensional linkage dynamic weight algorithm. Then, the three types of weight values ​​of the same grid point are normalized so that the sum of the three types of weight values ​​is 1, eliminating the difference in numerical magnitude. Subsequently, a 3×3 mean filter kernel is used to complete the spatial smoothing of the weight field and remove isolated abrupt weight values. Finally, the data is integrated in the order of grid row and column number, height layer number, and band type to form a three-dimensional dynamic weight distribution matrix containing spatial location, height level, and band dimension. The parameters of each dimension of the matrix correspond one-to-one with the standardized radar data system, providing weight support for subsequent feature fusion.

[0017] Furthermore, in the multi-scale feature extraction module, parallel feature processing adopts a multi-threaded independent computing architecture. Dedicated processing threads are configured for the standardized data of S-band, C-band, and X-band radars. Each thread starts synchronously and runs independently without interfering with the processing flow. The S-band thread relies on the radar's wide-area detection characteristics to focus on the macroscopic distribution and overall movement trend of precipitation across the entire region, extracting large-scale background features. The C-band thread relies on the radar's balanced detection range and resolution to focus on the structure of precipitation bands and the aggregation morphology of convective cells, extracting medium-scale structural features. The X-band thread relies on the radar's high-resolution and fast scanning characteristics to focus on strong convective cores and small-scale vortex details, extracting refined convective features at the corresponding scale. After each thread completes processing, it synchronously outputs complementary feature results of the three scales, which are completely matched with the grid benchmark of the standardized radar data system, providing complete feature support for subsequent feature fusion.

[0018] Furthermore, in the attention feature fusion module, the mathematical expression for the cross-band attention fusion algorithm is:

[0019] ;

[0020] Wherein, F is the final output unified fusion feature map, which includes the global background, medium-scale structure and refined convection full-scale information of multi-band radar data; W is the dynamic weight of a single grid point calculated by the three-dimensional linkage dynamic weight algorithm, with a value range of 0-1, which is used to adjust the contribution ratio of different band features in the fusion process in real time. The large-scale background features extracted from S-band radar data reflect the macroscopic distribution and overall movement trend of precipitation across the entire region. The medium-scale structural features extracted from C-band radar data reflect the morphology of precipitation belts, the aggregation pattern of convective cells, and the changes in intensity gradient. This is a refined convective feature extracted from X-band radar data, capturing the location of strong convection cores, small-scale vortex structures, and details of abrupt changes in precipitation intensity; A is a spatiotemporal attention factor, ranging from 0 to 1, used to quantify the attention weight between precipitation abrupt change areas and strong convection core areas, automatically focusing on key precipitation areas.

[0021] Furthermore, in the attention feature fusion module, the process of locating key precipitation areas using spatiotemporal attention factors is as follows: the spatiotemporal attention factors correspond one-to-one with the grid points of the standardized radar data system, and each factor value corresponds to the degree of precipitation criticality in the grid area, covering all grid points in the entire detection area. For areas with drastic changes in precipitation intensity and active convection development, corresponding high-value factors are assigned, while for areas with stable precipitation and no obvious convection activity, corresponding low-value factors are assigned. During the cross-band attention fusion algorithm operation, the multi-scale features of key precipitation areas are integrated based on the factor values, and the features of non-key areas are processed accordingly. Finally, the feature location and presentation of key precipitation areas are completed in the unified fusion feature map.

[0022] Furthermore, in the rainfall forecast generation module, the rainfall intensity forecast result is a high-resolution gridded rainfall dataset for the next 0-2 hours generated based on the spatiotemporal evolution of precipitation using a unified fusion feature map. The spatial resolution is consistent with the 1km×1km grid of the standardized radar data system, and the temporal resolution is 6 minutes. R represents the rainfall intensity at a single grid point, with the unit being mm / h. The dataset covers the entire detection area and corresponds completely to the grid range and grid spacing of the fusion feature map. After spatial smoothing and outlier verification, the forecast results are classified into levels according to preset intensity thresholds: 0.1mm / h≤R<2.5mm / h is light rain, 2.5mm / h≤R<8.0mm / h is moderate rain, 8.0mm / h≤R<16.0mm / h is heavy rain, 16.0mm / h≤R<50.0mm / h is torrential rain, 50.0mm / h≤R<100.0mm / h is extremely heavy rain, and R≥100.0mm / h is exceptionally heavy rain.

[0023] Furthermore, in the rainfall forecast generation module, the specific steps for classifying disaster warning levels are as follows: using the rainfall intensity forecast results and the derived 1-hour and 2-hour cumulative rainfall as core judgment indicators, combined with the duration of regional rainfall and the rate of change of rainfall intensity, a four-level disaster warning level classification standard is set. A blue warning corresponds to a 1-hour cumulative rainfall ≥10mm and <30mm, or a 2-hour cumulative rainfall ≥15mm and <50mm, indicating that sporadic local rainfall has occurred and weather changes need to be monitored; a yellow warning corresponds to a 1-hour cumulative rainfall ≥30mm and <50mm, or a 2-hour cumulative rainfall ≥50mm and <80mm. A rainfall of 1 mm indicates increased regional rainfall intensity, which may easily lead to mild waterlogging. An orange alert corresponds to a cumulative rainfall of ≥50 mm and <100 mm in 1 hour, or ≥80 mm and <150 mm in 2 hours, indicating concentrated heavy rainfall and the need to guard against disasters such as flash floods and urban flooding. A red alert corresponds to a cumulative rainfall of ≥100 mm in 1 hour, or ≥150 mm in 2 hours, representing extreme heavy rainfall weather, requiring immediate activation of emergency response, preparation for personnel evacuation and rescue. The results of each level of alert are matched grid-based forecast data point by point, and the visualized regional distribution is synchronously linked to provide accurate hierarchical early warning support for external business systems.

[0024] On the other hand, the short-term rainfall prediction method based on multi-source radar data fusion has the following specific steps:

[0025] S100, multi-source radar preprocessing: synchronously acquires real-time echo data from S, C, and X band radars, completes ground clutter removal, coordinate unification and spatiotemporal benchmark alignment, extracts relevant parameters, and constructs a standardized radar data system;

[0026] S200, Dynamic Weight Calculation: Based on standardized radar data, a three-dimensional linkage dynamic weight algorithm is used to calculate the weights grid by grid. After normalization and integration, a three-dimensional dynamic weight distribution matrix covering the entire detection area is generated.

[0027] S300, multi-scale feature extraction: Based on the detection characteristics of radars in each band, feature processing is carried out in parallel to extract multi-scale precipitation features corresponding to the S, C, and X bands respectively.

[0028] S400, Attention Feature Fusion: Combining a three-dimensional dynamic weight matrix with multi-scale precipitation features, feature integration is achieved through a cross-band attention fusion algorithm. Key precipitation areas are located with the help of spatiotemporal attention factors, and a unified fused feature map is output.

[0029] S500, Rainfall Forecast Generation: Based on the unified fusion feature map, the spatiotemporal evolution of precipitation is inferred, and a high-resolution rainfall intensity forecast for the next 0-2 hours is generated. After spatial smoothing, outlier verification and disaster warning level classification, a visual forecast product is generated and converted into a general meteorological format.

[0030] Compared with existing technologies, this short-term rainfall prediction system and method based on multi-source radar data fusion has the following advantages:

[0031] I. This invention integrates multi-band radar observation data synchronously and performs standardized preprocessing, unifying the spatiotemporal benchmark and dimensional system of the data. This effectively eliminates interference information and ensures the collaborative adaptation of multi-source data. Relying on a three-dimensional linkage dynamic weighting algorithm, it calculates and generates a global three-dimensional weight matrix point by point. Based on different detection conditions, it adaptively adjusts the data reliability weights, avoiding the inherent limitations of single-band data. Combining the radar detection characteristics of each band, it conducts parallel multi-scale feature extraction, capturing precipitation structure information at different scales to form a complementary and complete precipitation feature system. This comprehensively improves the comprehensiveness and accuracy of precipitation feature extraction, solving the problem that traditional single-source radar detection is difficult to balance in terms of coverage, resolution, and timeliness. It lays a high-quality data foundation for subsequent feature fusion and forecasting, optimizes the efficiency of radar data utilization and the ability to represent precipitation information, and fully releases the collaborative value of multi-source radar data.

[0032] II. This invention integrates cross-band attention fusion algorithms with a three-dimensional dynamic weight matrix and multi-scale precipitation features to achieve unified integration. It uses spatiotemporal attention factors to accurately locate key precipitation areas, strengthens the fusion weight of core precipitation information, and outputs a unified fusion feature map covering all scales of precipitation information. Based on the fusion feature map, it performs spatiotemporal evolution simulations of precipitation, generating high-resolution short-term precipitation forecasts. After smoothing verification and disaster warning level classification, it forms a standardized visualization product, enabling efficient integration with external business systems. This significantly improves the spatial and temporal accuracy of short-term precipitation forecasts, accurately identifies key precipitation areas such as severe convection, and scientifically classifies disaster warning levels, providing timely and reliable support for meteorological warnings and emergency response. It effectively enhances the operational practicality and scenario adaptability of short-term precipitation forecasts, facilitating the precise and efficient implementation of meteorological forecast services.

[0033] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0035] Figure 1A flowchart of a short-term rainfall prediction system based on multi-source radar data fusion;

[0036] Figure 2 This is a schematic diagram of data transmission in a short-term rainfall forecasting system based on multi-source radar data fusion.

[0037] Figure 3 This is a schematic diagram of the data transmission for attention feature fusion according to the present invention. Detailed Implementation

[0038] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0039] Example 1:

[0040] This embodiment is applied to a scenario in an urban meteorological center where severe convective weather is widespread and frequent during the flood season, with concurrent monitoring of precipitation in multiple regions. In this situation, the system needs to simultaneously access radar observation data from multiple stations and respond to high-frequency requests from external business systems, operating under a high-load and busy state with extremely large data access volumes, intensive computational tasks, and concurrent service requests. Relying on the complete module of the short-term precipitation forecasting system based on multi-source radar data fusion of this invention, the system can stably complete the entire process of multi-source radar data processing, feature fusion, and precipitation forecasting, ensuring the accuracy of forecast results, timeliness of output, and service stability under busy scenarios. Figure 1 As shown, the specific implementation process is as follows:

[0041] The multi-source radar preprocessing module synchronously receives real-time observation echo data from S-band conventional weather radar, C-band Doppler weather radar, and X-band phased array weather radar. The data covers core parameters such as reflectivity factor, radial velocity, spectral width, differential reflectivity, differential propagation phase, and correlation coefficient. The S-band and C-band radars output observation data in volumetric scanning mode, sequentially completing planar position display scans at 14 elevation angles from 0.5° to 19.5°, completing a full volumetric scan every 6 minutes. The X-band phased array radar outputs observation data in electronic scanning mode, sequentially completing 20 elevation angle scans from 0.5° to 30°. The module displays the planar position of the angle during scanning, completing a full volume scan every 3 minutes. All data conforms to the meteorological industry standard radar basic data format specifications. The module first performs ground clutter removal processing on various echo data, thoroughly eliminating interference signals generated by non-precipitation echoes such as ground buildings, mountains, and vegetation. This ensures that subsequent analysis only retains true and valid precipitation-related data, avoiding prediction deviations caused by clutter interference at the source. Then, a coordinate transformation operation is performed on the multi-source data, standardizing the observation coordinates of different radar devices to the same spatial coordinate system, eliminating data spatial misalignment caused by differences in observation coordinates between different devices. Simultaneously, spatiotemporal benchmark alignment was performed to match all radar data to a unified time node and spatial benchmark, avoiding data logic errors caused by temporal discrepancies. Subsequently, the detection range, signal-to-noise ratio, radar operating wavelength, and precipitation intensity estimates corresponding to each grid point were extracted. A standardized radar data system was built using a unified planar grid as the carrier, and the entire area was divided into grid points according to a fixed spatial resolution of 1km×1km to ensure a one-to-one correspondence between the spatial locations of all grid points. Within the grid, three basic data layers were constructed: a multi-band radar echo intensity field, a radial velocity field, and a spectral width field. Each data layer was then processed according to a unified spatial scale. Numerical normalization processing synchronizes all data layers to the same time node to form a time-series data sequence. The extracted parameters are bound to grid points and stored in an orderly manner. Finally, a multi-band radar standardized data system with unified spatial reference, unified time reference, and unified data dimension is constructed. This provides regular, unified, and interference-free basic data support for subsequent data processing in busy system scenarios, ensuring the smoothness of data reading, calling, and calculation under high concurrency, avoiding calculation lag or result deviation caused by data disorder and inconsistent references, and enabling the system to quickly retrieve standardized data for subsequent processing even under busy conditions.

[0042] The dynamic weight calculation module retrieves the preprocessed standardized radar data system and its bound parameters, and initiates the three-dimensional linkage dynamic weight algorithm to perform point-by-point and height-by-height layer calculations on the entire grid within the detection area. The mathematical expression of the three-dimensional linkage dynamic weight algorithm is as follows:

[0043] ;

[0044] in, The final dynamic weights for single-grid multi-wavelength radar data; For adaptive correction coefficients; Where is the radar wavelength; I represents the estimated real-time precipitation intensity at a single grid point; D represents the detection distance from the single grid point to the corresponding radar station; and S represents the signal-to-noise ratio of the radar signal at a single grid point. Dynamic weights for single grid points are calculated separately for S-band, C-band, and X-band radar data to accurately reflect the reliability level of each band's data at the corresponding grid point. After calculating the single grid point weights, normalization and integration processing is performed on the weight values ​​of the three bands at the same grid point to eliminate the magnitude differences in weight values ​​between different bands, ensuring the sum of the three weight values ​​is 1. Then, a 3×3 mean filter is used to spatially smooth the weight field, removing isolated abrupt changes in the weight data, making the weight distribution more consistent with the actual detection environment and data quality status. Finally, weights are calculated according to grid row and column number, height layer number, and band type. The data integration is completed sequentially, generating a three-dimensional dynamic weight distribution matrix that fully covers the detection area and precisely matches the dimensions of the standardized radar data system. This matrix stores the dynamic weight values ​​of radar data in each band using grid points as basic units. It can adapt to changes in data quality under different detection conditions in real time, providing a precise and adaptive basis for weight adjustment in the subsequent feature fusion stage. Under heavy system load and high computational pressure, the standardized weight matrix simplifies the computational logic of feature fusion, reduces the time spent on invalid computation, improves data fusion efficiency, and ensures the accuracy of weight adjustment. This allows radar data of different qualities to be reasonably adapted in subsequent processing, avoids low-quality data from interfering with the overall prediction results, and ensures the scientific and reliable nature of data fusion in busy scenarios.

[0045] The multi-scale feature extraction module employs a multi-threaded independent computing architecture for parallel feature processing. Dedicated processing threads are configured for standardized data from S-band, C-band, and X-band radars. Each thread starts synchronously and runs independently, without interfering with its processing flow, effectively distributing system computational pressure and perfectly adapting to high-load operation under busy system scenarios. The S-band data processing thread leverages the wide-area detection capabilities of this radar band to focus on the macroscopic distribution and overall movement trend of precipitation across the entire region, accurately extracting large-scale background features and fully presenting the overall trend and spatial coverage of precipitation, providing macroscopic evidence for precipitation trend analysis. The C-band data processing thread, utilizing the balanced detection range and resolution of this radar band, focuses on the structure of precipitation bands and the aggregation morphology of convective cells, extracting mesoscale structural features, clearly demonstrating the internal structural distribution and intensity gradient changes of precipitation, filling in the gaps between macroscopic and microscopic features. To fill the information gaps, the X-band data processing thread leverages the high-resolution, rapid scanning capabilities of radar in this band to focus on the core of strong convection and the details of small-scale vortices, extracting refined convective features. This accurately captures the core regions and subtle structural changes within precipitation, leaving no crucial information about heavy precipitation unexplored. Each thread outputs results synchronously after feature extraction, ensuring all feature results perfectly match the standardized radar data system's 1km×1km grid benchmark. This forms a complete precipitation feature system that complements large-scale, medium-scale, and fine-scale features, comprehensively covering macroscopic, mesoscopic, and microscopic information about precipitation. This compensates for the limitations of single-band feature extraction, providing comprehensive, detailed, and complete feature support for subsequent feature fusion. Simultaneously, the multi-threaded parallel processing mode significantly improves feature extraction speed, ensuring rapid feature extraction even when the system is busy, without delaying the overall forecast process and maintaining the timeliness of forecast services. Figure 3 As shown.

[0046] The attention feature fusion module retrieves the three-dimensional dynamic weight distribution matrix and the three complementary precipitation features output by the multi-scale feature extraction module, and initiates the cross-band attention fusion algorithm to integrate the multi-dimensional features. The weight values ​​in the three-dimensional dynamic weight matrix are incorporated into the fusion operation, and the contribution ratio of different band features in the fusion process is adjusted in real time. The mathematical expression of the cross-band attention fusion algorithm is as follows:

[0047] ;

[0048] Where F is the final output unified fusion feature map; W is the dynamic weight of a single grid point calculated by the three-dimensional linkage dynamic weight algorithm; Large-scale background features extracted from S-band radar data; Medium-scale structural features extracted from C-band radar data; This section describes the refined convective features extracted from X-band radar data. A represents the spatiotemporal attention factor, which allows features from high-confidence data to play a greater role, while features from low-confidence data are appropriately adapted, fully leveraging the core value of high-quality data. Simultaneously, the spatiotemporal attention factor is used to locate key precipitation areas. This factor corresponds one-to-one with 1km×1km grid points in the standardized radar data system. High-value factors are automatically assigned to areas with drastic changes in precipitation intensity and active convection, while low-value factors are assigned to areas with stable precipitation and no significant convective activity. During the fusion calculation, the characteristic information of key precipitation areas is emphasized, while non-key areas are neglected. Features are adapted to accurately highlight core precipitation information and eliminate redundant and invalid features. The final output is a unified fused feature map that includes global background, mesoscale structure, and refined convective full-scale precipitation information. This feature map integrates all the advantages of multi-band radar, fully restores the true spatiotemporal state of precipitation, and provides a high-quality feature foundation for subsequent spatiotemporal evolution inference of precipitation. In busy system scenarios, accurate feature fusion reduces invalid calculations and improves the processing efficiency of the inference process, while ensuring the integrity and accuracy of feature information, ensuring the reliability of subsequent forecast results, and ensuring that the prediction results still have high accuracy even in busy scenarios.

[0049] The rainfall forecast generation module uses a unified fusion feature map to extrapolate the spatiotemporal evolution of precipitation, rapidly generating high-resolution rainfall intensity forecasts for the next 0-2 hours. The forecast data employs a 1km×1km spatial resolution and a 6-minute temporal resolution, covering the entire detection area and perfectly corresponding to the grid range and spacing of the fusion feature map. After generating the forecast results, spatial smoothing is first performed to make the rainfall intensity distribution more closely match actual precipitation patterns, optimizing the presentation of the forecast data. Then, outlier verification is conducted to remove unreasonable data generated during the extrapolation process, ensuring the authenticity and validity of the forecast data. Subsequently, rainfall intensity levels are classified according to preset intensity thresholds, based on the rainfall intensity at a single grid point, distinguishing between light rain, moderate rain, heavy rain, torrential rain, extremely heavy rain, and exceptionally heavy rain. Finally, rainfall intensity and 1-hour cumulative rainfall are used as the basis for further classification. Using 2-hour cumulative rainfall as the core indicator, and combining the duration and rate of change of rainfall intensity, four levels of disaster warning are defined. The blue, yellow, orange, and red warning results are matched with gridded forecast data grid by grid point to generate visualized forecast products. These products are then converted into a common meteorological format to achieve stable integration with external business systems such as meteorological operational platforms, emergency management systems, and urban management platforms. This ensures the continuous output of accurate forecast and warning information. Even under heavy system loads and constant external requests, the system can reliably output standardized and directly applicable forecast products, guaranteeing the timely implementation of meteorological warnings, emergency response, and urban disaster prevention. This avoids forecast interruptions, information delays, or result distortions caused by system overload, comprehensively improving the stability and practicality of short-term rainfall forecast services and providing solid technical support for urban flood season disaster prevention and mitigation.

[0050] This embodiment demonstrates the complete operation of a short-term precipitation forecasting system under high-concurrency and busy scenarios. It incorporates parameters such as 1km×1km spatial resolution, 6min / 3min scan cycles, 14 / 20 elevation angles, a 3×3 mean filter kernel, and a 0-2 hour forecast duration. Through standardized data preprocessing, a three-dimensional dynamic weighting algorithm, multi-threaded parallel feature extraction, and a cross-band attention fusion algorithm, it stably outputs high-resolution forecasts and level-four early warnings, which are then integrated with operational systems. The system maintains smooth operation under high load throughout, effectively addressing computational bottlenecks in busy scenarios, fully leveraging the value of multi-source radar fusion, improving forecast accuracy and timeliness, and providing stable and reliable support for meteorological services and disaster prevention and mitigation during the flood season.

[0051] Example 2:

[0052] This embodiment strictly follows the short-term precipitation forecasting method based on multi-source radar data fusion of the present invention, fully executing all steps from data preprocessing to forecast output. It achieves standardized implementation of multi-source radar data fusion and short-term precipitation forecasting, and the method is replicable and scalable, applicable to short-term precipitation forecasting in various meteorological observation scenarios such as urban meteorological monitoring, mountain flood warning, and traffic meteorological support. Figure 2 As shown, the specific implementation steps are as follows:

[0053] S100 Multi-Source Radar Preprocessing: Simultaneously acquires real-time observation echo base data from S-band, C-band, and X-band radars. The data includes core parameters such as reflectivity factor, radial velocity, spectral width, differential reflectivity, differential propagation phase, and correlation coefficient. S-band and C-band radars use volumetric scanning mode, sequentially completing planar position display scans at 14 elevation angles from 0.5° to 19.5°, completing a full volumetric scan every 6 minutes. X-band phased array radar uses electronic scanning mode, sequentially completing planar position display scans at 20 elevation angles from 0.5° to 30°, completing a full volumetric scan every 3 minutes. All data adheres to meteorological industry standard radar base data format specifications. First, ground clutter is removed from the echo data to eliminate invalid interference signals generated by ground object reflections, retaining pure precipitation echo information to ensure data validity from the source. Then, coordinate unification is performed to standardize the observation spatial coordinates of different radars, eliminating spatial observation biases between equipment and providing a unified spatial reference for multi-source data. Simultaneously, spatiotemporal reference alignment is completed. To ensure consistency, multi-source data is matched to the same temporal and spatial reference, forming a temporally coherent and spatially unified data foundation. This avoids analytical errors caused by temporal and spatial misalignment. Subsequently, the detection range, signal-to-noise ratio, radar operating wavelength, and precipitation intensity estimates for each grid point are extracted. A standardized radar data system is built using a unified planar grid. The entire domain is divided into grid points with a fixed spatial resolution of 1km×1km, ensuring a one-to-one spatial correspondence between grid points. Within the grid, three basic data layers—echo intensity field, radial velocity field, and spectral width field—are constructed. Numerical normalization is performed on each data layer at a uniform scale. All data layers are synchronized to the same time node to form a time-series data sequence. The extracted parameters are bound and stored with the grid points. Finally, a multi-band radar standardized data system with fully unified spatial, temporal, and data dimensions is constructed. This provides regular, standardized, and high-quality basic data for subsequent method execution, ensuring that subsequent steps such as weight calculation and feature extraction have a stable and reliable data support. This avoids process bottlenecks or result errors caused by data chaos, giving the entire prediction method a solid data foundation.

[0054] S200, Dynamic Weight Calculation: Based on pre-processed standardized radar data and built-in parameters, a three-dimensional linkage dynamic weight algorithm is used to perform calculations on the entire detection area grid grid by grid and by height layer. The dynamic weights of S-band, C-band, and X-band radar data at individual grid points are calculated separately, accurately reflecting the reliability level of each band data at the corresponding point. This adapts to data states under different detection distances, signal quality, and precipitation intensities. After completing the weight calculation for each grid point, the weight values ​​of the three bands at the same grid point are normalized and integrated, ensuring the sum of the three weight values ​​is 1. This balances the numerical differences in weights across different bands, placing the weight data within a uniform range for easier subsequent fusion calculations. Finally, a 3×3 averaging... The value filtering kernel performs spatial smoothing of the weight field, removing isolated abrupt weight data and making the weight distribution more consistent with the actual detection environment and data quality status, thus improving the rationality of the weight data. Finally, the data is integrated according to grid row and column numbers, height layer numbers, and band types to generate a three-dimensional dynamic weight distribution matrix that covers the entire detection area and is fully compatible with the 1km×1km standardized radar data system. This matrix provides a precise basis for weight adjustment for feature fusion, allowing radar data of different qualities and bands to be reasonably allocated during the fusion process, improving the scientificity and accuracy of feature fusion, laying a core foundation for subsequent integrated feature integration, and ensuring that the feature fusion stage can adaptively adjust the weights according to data quality.

[0055] S300, Multi-scale Feature Extraction: Based on the detection characteristics of S-band, C-band, and X-band radars, multi-scale precipitation feature extraction is carried out in parallel processing. Independent processing workflows are configured for each band of data, and feature extraction operations are performed simultaneously to improve overall processing efficiency. S-band data, leveraging its wide-area detection advantage, extracts large-scale background features of the macroscopic distribution and overall movement trend of precipitation across the entire region, grasping the overall development direction of precipitation and providing macroscopic guidance for regional precipitation analysis. C-band data, utilizing its balanced detection range and resolution, extracts mesoscale structural features of precipitation band structure, convective cell aggregation morphology, and intensity gradient changes, revealing the internal structural morphology of precipitation and connecting... Macroscopic and microscopic features: X-band data, leveraging its high-resolution and rapid scanning advantages, extracts refined convective features of strong convection cores, small-scale vortices, and details of abrupt changes in precipitation intensity. It captures the core subtle changes in precipitation, identifies key areas of heavy precipitation, and outputs these three types of features simultaneously. All features are matched with the grid benchmark of a 1km×1km standardized radar data system, forming a complementary and complete set of precipitation features that comprehensively covers different scales of precipitation information. This solves the problem of incomplete feature extraction from a single band and provides rich, complete, and multi-layered feature support for subsequent feature fusion. The fused features can fully restore the true state of precipitation and provide complete information for accurate projection.

[0056] S400 Attention Feature Fusion: This system combines a three-dimensional dynamic weight matrix with multi-scale precipitation features, employing a cross-band attention fusion algorithm to integrate multi-dimensional features. The three-dimensional dynamic weight matrix adjusts the fusion contribution ratio of different band features in real time, ensuring that features from high-reliability data dominate, maximizing the value of high-quality data. Simultaneously, it uses a spatiotemporal attention factor to identify key precipitation areas. This factor corresponds one-to-one with 1km×1km grid points, strengthening feature weights for areas with active convection and abrupt intensity changes, while weakening unnecessary features in areas with stable precipitation. This accurately locates and highlights core precipitation information, avoiding redundant information from interfering with prediction results. After optimizing and integrating all-dimensional features, it outputs a unified fused feature map containing global background, mesoscale structure, and refined full-scale convection information. This feature map integrates all the advantages of multi-source radar, eliminating redundant interference information and fully presenting the spatiotemporal distribution and evolution characteristics of precipitation. It provides the most core and accurate feature basis for subsequent spatiotemporal evolution extrapolation of precipitation, ensuring that the extrapolation results conform to the actual development law of precipitation, and providing accurate feature support for subsequent forecast generation.

[0057] S500 Rainfall Forecast Generation: Based on a unified fusion feature map, the spatiotemporal evolution of precipitation is extrapolated, generating high-resolution gridded rainfall intensity forecast data for the next 0-2 hours. The forecast data uses a 1km×1km spatial resolution and a 6-minute temporal resolution, fully covering the detection area to ensure comprehensiveness and precision. After generating the forecast data, spatial smoothing is first performed to optimize the data distribution, making the forecast results more closely match actual precipitation changes. Then, outlier verification is conducted to remove invalid data, ensuring the accuracy and effectiveness of the forecast data. Finally, rainfall intensity levels are classified according to preset thresholds, using single-grid rainfall intensity as an indicator to distinguish different rainfall levels. Rainfall levels are determined by combining 1-hour cumulative rainfall, 2-hour cumulative rainfall, rainfall duration, and the rate of change in intensity to classify disaster warning levels into four levels: blue, yellow, orange, and red. Warning information is then matched with gridded forecast data to create visualized forecast products, which are then converted into a common meteorological format. This ultimately enables integration with external operational systems, providing standardized and directly usable forecast results for weather forecasting, disaster warning, and emergency response. This allows for the rapid implementation of short-term rainfall prediction results, improving the response speed and handling capabilities of meteorological disaster prevention and mitigation, and truly transforming the forecasting method of this invention into an effective technical means that serves practical operations.

[0058] This embodiment strictly follows the methodology to complete the entire short-term precipitation forecasting process, incorporating key parameters such as a 1km×1km grid, multi-band scanning parameters, weight normalization, 3×3 smoothing filtering, 0-2 hour forecasts, 6-minute temporal resolution, and four-level warnings. The entire process, from data standardization to forecast product output, is standardized and controllable. This method fully integrates the advantages of multi-band radar detection, relying on adaptive weighting and spatiotemporal attention mechanisms to optimize processing logic. It can stably output accurate forecasts and tiered warnings, possessing high replicability and practicality, and providing a standardized technical solution for various types of short-term meteorological forecasting.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A nowcasting rainfall prediction system based on multi-source radar data fusion, characterized in that, The system includes: Multi-source radar preprocessing module: Simultaneously acquires real-time observation echo data from S-band, C-band and X-band phased array radars, performs ground clutter removal, coordinate unification transformation and spatiotemporal reference alignment, extracts estimated values ​​of detection range, signal-to-noise ratio, wavelength and precipitation intensity, and constructs a standardized radar data system; Dynamic weight calculation module: Based on standardized radar data and related parameters, a three-dimensional linkage dynamic weight algorithm is used to calculate the weight of the entire area grid point by point. After normalization and integration, a three-dimensional dynamic weight distribution matrix that fully covers the detection area is formed. Multi-scale feature extraction module: Based on the radar detection characteristics of each band, parallel feature processing is performed to extract multi-scale precipitation features corresponding to S-band, C-band and X-band radar data respectively. Attention Feature Fusion Module: Combining a three-dimensional dynamic weight matrix with multi-scale precipitation features, the module integrates multi-dimensional features through a cross-band attention fusion algorithm. It locates key precipitation areas using spatiotemporal attention factors and outputs a unified fusion feature map containing full-scale precipitation information. Rainfall forecast generation module: Based on the unified fusion feature map, the module performs spatiotemporal evolution simulation of precipitation, generates high-resolution rainfall intensity forecast results, performs spatial smoothing and outlier verification, classifies disaster warning levels, generates visual forecast products, and converts them into a common meteorological format.

2. The nowcasting rainfall prediction system based on multi-source radar data fusion of claim 1, wherein, In the multi-source radar preprocessing module, the real-time observation echo data includes basic data from S-band conventional weather radar, C-band Doppler weather radar, and X-band phased array weather radar, covering parameters such as reflectivity factor, radial velocity, spectral width, differential reflectivity, differential propagation phase, and correlation coefficient. The S-band and C-band radars use volumetric scanning mode to sequentially complete planar position display scans at 14 elevation angles from 0.5° to 19.5°, completing a full volumetric scan every 6 minutes. The X-band phased array radar uses electronic scanning mode to sequentially complete planar position display scans at 20 elevation angles from 0.5° to 30°, completing a full volumetric scan every 3 minutes. The data format conforms to the meteorological industry standard radar basic data format specification.

3. The nowcasting rainfall prediction system based on multi-source radar data fusion of claim 1, wherein, In the multi-source radar preprocessing module, the standardized radar data system is built on a unified planar grid. The entire domain is divided into grid points with a fixed spatial resolution of 1km×1km. The grid points maintain a one-to-one spatial correspondence. The grid contains three basic data layers: multi-band radar echo intensity field, radial velocity field, and spectral width field. Each data layer is normalized according to a unified spatial scale. All data layers are synchronously matched to the same time node to form a time-series data sequence. The standardized radar data system synchronously embeds the radar detection range, signal-to-noise ratio, radar operating wavelength, and precipitation intensity estimation values ​​corresponding to each grid point. All values ​​are bound to the grid points and stored in an orderly manner.

4. The short-term rainfall prediction system based on multi-source radar data fusion according to claim 1, characterized in that, The mathematical expression for the three-dimensional linkage dynamic weight algorithm in the dynamic weight calculation module is as follows: ； in, The final dynamic weights for single-grid multi-wavelength radar data; For adaptive correction coefficients; denoted as the radar wavelength; I represents the estimated real-time precipitation intensity at a single grid point; D represents the detection distance from the single grid point to the corresponding radar station; and S represents the signal-to-noise ratio of the radar signal at the single grid point.

5. The short-term rainfall prediction system based on multi-source radar data fusion according to claim 1, characterized in that, In the dynamic weight calculation module, the three-dimensional dynamic weight distribution matrix corresponds to the Cartesian coordinate grid and vertical height layer of the standardized radar data system. The dynamic weight values ​​of each band radar data are stored with grid points as the basic unit. In the formation process, the dynamic weight of a single grid point of S-band, C-band, and X-band radar is first calculated grid by grid and height layer by height using a three-dimensional linkage dynamic weight algorithm. Then, the three types of weight values ​​of the same grid point are normalized so that the sum of the three types of weight values ​​is 1, eliminating the difference in numerical magnitude. Subsequently, a 3×3 mean filter kernel is used to complete the spatial smoothing of the weight field and remove isolated abrupt weight values. Finally, the data is integrated according to the grid row and column order to form a three-dimensional dynamic weight distribution matrix containing spatial location, height level, and band dimension. The parameters of each dimension of the matrix correspond one-to-one with the standardized radar data system.

6. The short-term rainfall prediction system based on multi-source radar data fusion according to claim 1, characterized in that, In the multi-scale feature extraction module, the parallel feature processing adopts a multi-threaded independent computing architecture. Dedicated processing threads are configured for the standardized data of S-band, C-band, and X-band radars. Each thread starts synchronously and runs independently without interfering with the processing flow. The S-band thread relies on the radar's wide-range detection characteristics to focus on the macroscopic distribution and overall movement trend of precipitation across the entire region and extract wide-range background features. The C-band thread leverages the balanced characteristics of radar detection range and resolution to focus on the structure of precipitation belts and the aggregation morphology of convective cells, extracting medium-scale structural features; the X-band thread leverages the high-resolution and rapid scanning characteristics of radar to focus on strong convective cores and small-scale vortex details, extracting refined convective features at the corresponding scales.

7. The short-term rainfall prediction system based on multi-source radar data fusion according to claim 1, characterized in that, In the attention feature fusion module, the mathematical expression of the cross-band attention fusion algorithm is: ； Where F is the final output unified fusion feature map; W is the dynamic weight of a single grid point calculated by the three-dimensional linkage dynamic weight algorithm; Large-scale background features extracted from S-band radar data; Medium-scale structural features extracted from C-band radar data; A represents the refined convection features extracted from X-band radar data; A is the spatiotemporal attention factor.

8. The short-term rainfall prediction system based on multi-source radar data fusion according to claim 1, characterized in that, In the rainfall forecast generation module, the rainfall intensity forecast result is a high-resolution gridded rainfall dataset for the next 0-2 hours, generated based on the spatiotemporal evolution of precipitation using a unified fusion feature map. The spatial resolution is consistent with the 1km×1km grid of the standardized radar data system, and the temporal resolution is 6 minutes. R represents the rainfall intensity at a single grid point, with the unit being mm / h. The dataset covers the entire detection area and corresponds completely to the grid range and grid spacing of the fusion feature map. After spatial smoothing and outlier verification, the forecast results are classified into levels according to preset intensity thresholds: 0.1mm / h≤R<2.5mm / h is light rain, 2.5mm / h≤R<8.0mm / h is moderate rain, 8.0mm / h≤R<16.0mm / h is heavy rain, 16.0mm / h≤R<50.0mm / h is torrential rain, 50.0mm / h≤R<100.0mm / h is extremely heavy rain, and R≥100.0mm / h is exceptionally heavy rain.

9. The short-term rainfall prediction system based on multi-source radar data fusion according to claim 1, characterized in that, The specific steps for classifying disaster warning levels in the rainfall forecast generation module are as follows: Using the rainfall intensity forecast results and the derived 1-hour cumulative rainfall as the core judgment indicators, and combining the duration of regional rainfall and the rate of change in rainfall intensity, a four-level disaster warning level classification standard is set. A blue warning corresponds to a 1-hour cumulative rainfall ≥10mm and <30mm, indicating sporadic local rainfall, requiring attention to weather changes; a yellow warning corresponds to a 1-hour cumulative rainfall ≥30mm and <50mm, indicating increased regional rainfall intensity, which may easily cause mild waterlogging; an orange warning corresponds to a 1-hour cumulative rainfall ≥50mm and <100mm, indicating concentrated heavy rainfall, requiring precautions against flash floods and urban flooding; a red warning corresponds to a 1-hour cumulative rainfall ≥100mm, representing extreme heavy rainfall weather, requiring immediate activation of emergency response and preparation for personnel evacuation and rescue operations.

10. A method for short-term rainfall prediction based on multi-source radar data fusion, applicable to the short-term rainfall prediction system based on multi-source radar data fusion as described in any one of claims 1-9, characterized in that, The specific steps of this method are as follows: S100, multi-source radar preprocessing: synchronously acquires real-time echo data from S, C, and X band radars, completes ground clutter removal, coordinate unification and spatiotemporal benchmark alignment, extracts relevant parameters, and constructs a standardized radar data system; S200, Dynamic Weight Calculation: Based on standardized radar data, a three-dimensional linkage dynamic weight algorithm is used to calculate the weights grid by grid. After normalization and integration, a three-dimensional dynamic weight distribution matrix covering the entire detection area is generated. S300, multi-scale feature extraction: Based on the detection characteristics of radars in each band, feature processing is carried out in parallel to extract multi-scale precipitation features corresponding to the S, C, and X bands respectively. S400, Attention Feature Fusion: Combining a three-dimensional dynamic weight matrix with multi-scale precipitation features, feature integration is achieved through a cross-band attention fusion algorithm. Key precipitation areas are located with the help of spatiotemporal attention factors, and a unified fused feature map is output. S500, Rainfall Forecast Generation: Based on the unified fusion feature map, the spatiotemporal evolution of precipitation is inferred, and a high-resolution rainfall intensity forecast for the next 0-2 hours is generated. After spatial smoothing, outlier verification and disaster warning level classification, a visual forecast product is generated and converted into a general meteorological format.

Images