Short-impending precipitation forecasting method and device based on high spatiotemporal resolution multi-source observation fusion

Abstract

This invention relates to a short-term precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion, comprising the following steps: acquiring multi-source observation data with different spatiotemporal resolutions; unifying the multi-source observation data onto a minute-level high-resolution grid; and calibrating the data using physical motion information and rain gauge observations to generate a high-quality precipitation field; and based on the high-quality precipitation field, performing short-term precipitation forecasting through a pre-established potential diffusion model. The potential diffusion model is constructed using deep learning, fusing radar echoes with the potential representation of the high-quality precipitation field, and generating precipitation in the potential space through a conditional diffusion mechanism. This invention ensures both the spatial detail and absolute accuracy of precipitation estimation, while maintaining temporal continuity and physical plausibility, making it suitable for operational scenarios such as convective weather monitoring, urban flooding early warning, and agricultural water resource management.

Description

Technical Field

[0001] This invention relates to the field of meteorological information technology, and in particular to a method and apparatus for short-term precipitation forecasting based on high spatiotemporal resolution multi-source observation fusion. Background Technology

[0002] Traditional short-term precipitation forecasts typically rely on operational radar mosaics, with a temporal resolution of approximately 6 minutes and a spatial resolution of approximately 1 km × 1 km. This makes it difficult to depict the rapid evolution of convective precipitation on a minute-scale basis, hindering the needs for refined operational scheduling and rapid early warning. In recent years, precipitation observation methods have diversified, including dual-polarization weather radar, X-band phased array radar, and high-density rain gauges. However, these different observation devices vary significantly in temporal and spatial resolution and observational physical mechanisms, resulting in heterogeneous and asynchronous observational data. Existing methods struggle to jointly utilize this observational information on a minute-scale basis, thus leading to poor accuracy in weather warnings, especially for severe convective weather. Summary of the Invention

[0003] The purpose of this invention is to at least address one of the shortcomings of the prior art and provide a method and apparatus for short-term precipitation forecasting based on high spatiotemporal resolution multi-source observation fusion.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: Specifically, a short-term precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion is proposed, including the following: Acquire multi-source observation data with different spatiotemporal resolutions, unify the multi-source observation data onto a minute-level high-resolution grid, and calibrate using physical motion information and rain gauge observations to generate a high-quality precipitation field; Based on the high-quality precipitation field, short-term precipitation forecasts are made using a pre-established potential diffusion model. The potential diffusion model is constructed using deep learning, which integrates radar echoes with the potential representation of the high-quality precipitation field, and generates precipitation in the potential space through a conditional diffusion mechanism.

[0005] Furthermore, specifically, the multi-source observation data includes the 6-minute mosaic reflectance of the operational large radar. X-band phased array radar 1-minute reflectivity and the 1-minute precipitation intensity of ground rain gauges ; Using the target generation time T as a reference, the multi-source observation data is aligned to time T through time interpolation. For operational large radar data, the echo motion vector field (u,v) is first calculated based on the pre-selected first few observations, and then the most recent time value is adsorbed to time T based on (u,v) using motion compensation time interpolation. For X-band phased array radar data and ground rain gauge data, the most recent time value is used to align to time T. The data is then resampled to a unified, preset high-resolution grid to ensure spatial consistency.

[0006] Furthermore, specifically, calibration is performed using physical motion information and rain gauge observations to generate a high-quality precipitation field, including: For operational radar, the following reflectivity-rainfall intensity relationship is dynamically selected at each grid point to generate a preliminary precipitation estimate. ) ; When the reflectivity Z of the corresponding radar is less than 45, the R(A) type relationship is adopted; When reflectivity Z < 45 is not valid, based on the correlation coefficient of radar variables, if CC ≤ 0.95, use the R(KDP) type relation; if CC > 0.95, use the R(A) type relation. For X-band phased array radar, the above reflectivity-rain intensity relationship is dynamically selected at each grid point to generate a preliminary precipitation estimate. ); Obtain the confidence level of X-band phased array radar ( ) and the confidence level of the business radar ( ), calculate the weight W of the X-band phased array radar, The business radar weight is 1-W; The preliminary fusion field was then calculated. ; Based on the X-band phased array radar sequence within a preset time period, calculate the fine motion vector field (u(x,y),v(x,y)); The extrapolated field is obtained by translating the initial fusion field at time T-Δt along the fine motion vector field to time T. At the same time, the original observation at the current time T is reversed and transferred to the time T-Δt, where Δt is a preset value, and the consistency with the observation at that time is checked to evaluate the reliability of the current observation. Time-constrained optimization is performed using a pre-constructed cost function to obtain... The cost function is:

[0007] Where R is the target optimization field, which is obtained after optimization. Weight Dynamic adjustment Preset spatial smoothing constraints; Calculate calibration factor for each rain gauge location With discrete factor To constrain the process, a spatially continuous calibration factor field F(x,y) is generated through variational analysis, ultimately yielding a high-quality precipitation field. .

[0008] Furthermore, the method also includes obtaining a high-quality precipitation field. Then, Kalman filtering or spatiotemporal smoothing is applied to suppress unreasonable minute-by-minute fluctuations; the rationality of extreme values ​​is checked; and conservative extrapolation is performed in marginal areas based on climate background and recent trends.

[0009] Furthermore, specifically, the preset duration is 15 minutes.

[0010] Furthermore, specifically, the pre-established potential diffusion model employs two independent variational autoencoders (VAEs) to encode and decode radar echoes and precipitation products; among which the radar encoder... radar echo Mapping to latent representation Radar decoding Representation of radar echo potential Reconstructed as radar echo Precipitation encoder Rain products Mapping to latent representation Precipitation decoder Indicate the potential for precipitation Reconstruction into precipitation products ; To make radar latent variables comparable to precipitation latent variables, the two VAEs share a standard normal prior.

[0011] Furthermore, specifically, the data processing procedures for the pre-established potential diffusion model include: Input radar echo sequence and target precipitation products After encoding by the corresponding VAE, it is obtained and By piecing them together along the timeline, we obtain... and T is the input radar echo length, and Together they are the size parameters of the latent variables; Learning conditional distributions within the conditional flow matching framework. ; Let the target precipitation potential be represented as: ; noise: ; time: ; Constructing a bridging path based on conditions: ; The path The target velocity field is: ; This represents the intermediate state in the potential space from the "data endpoint" to the "noise endpoint"; It is a reference velocity obtained from bridging path resolution, used for supervised learning of the velocity field network; Conditional DiT velocity field network and flow matching loss: Conditional DiT is used as the velocity field network: ; in As a conditional input, DiT is injected via cross-attention; the velocity field is fitted using flow-matching loss. ; Inference: ODE inverse integration generates potential precipitation. During inference, in order to draw conclusions from the conditional distribution Sampling, first initialize: ; Under given conditions Solve for ODE: ; The resulting clear precipitation potential representation is obtained: ; Denoising the latent representation via precipitation decoder Decoding yields images of future precipitation. ).

[0012] Furthermore, specifically, the loss function of the pre-established potential diffusion model includes, The VAE loss, including reconstruction loss and KL divergence loss, respectively ensures that the difference between the input and reconstructed images is minimized and that the latent variables conform to a standard normal distribution. Potential alignment loss: Ensure that the mean and variance of the radar and precipitation potential spaces are aligned; Stream matching loss: by minimizing relative to target speed The mean square error, learning under conditions The following is a continuous transformation from noise to the potential representation of precipitation.

[0013] Furthermore, specifically, the training process of the pre-established potential diffusion model includes, Training VAEs: Training separately , and , Optimize the reconstruction loss and KL loss; Potential alignment: utilizing Force radar to align with potential precipitation space; Training the diffusion model: in radar latent representation Under these conditions, train the diffusion model Potential for predicting future precipitation.

[0014] This invention also proposes a short-term precipitation forecasting device based on high spatiotemporal resolution multi-source observation fusion, comprising: The minute-level precipitation grid product generation module is used to acquire multi-source observation data with different spatiotemporal resolutions, unify the multi-source observation data onto a minute-level high-resolution grid, and use physical motion information and rain gauge observations for calibration to generate a high-quality precipitation field. The latent generation model extrapolation forecast module based on conditional flow matching is used to perform short-term precipitation forecasts based on the high-quality precipitation field and a pre-established latent diffusion model. The latent diffusion model is constructed through deep learning, which integrates radar echoes with the latent representation of the high-quality precipitation field, and completes precipitation generation in the latent space through a conditional diffusion mechanism.

[0015] The beneficial effects of this invention are as follows: This invention proposes a short-term precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion. By constructing a temporal-spatial dual-dimensional collaborative fusion framework and a flow-matching potential diffusion model, it enables the joint application of multi-source observation data in the generation of minute-level high-resolution precipitation products and short-term precipitation forecasts. This method comprehensively utilizes physical motion information, radar detection physical characteristics, and deep learning models, ensuring both spatial detail and absolute accuracy in precipitation estimation while maintaining temporal continuity and physical plausibility. It is suitable for operational scenarios such as convective weather monitoring, urban flooding early warning, and agricultural water resource management. Attached Figure Description

[0016] The above and other features of this disclosure will become more apparent from the detailed description of the embodiments illustrated in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same or similar elements. Obviously, the drawings described below are merely some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings: Figure 1The flowchart shown is a short-term precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion according to the present invention. Detailed Implementation

[0017] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The same reference numerals used throughout the accompanying drawings indicate the same or similar parts.

[0018] To provide minute-level resolution precipitation grid products to meet the needs of precision agricultural management, urban drainage early warning, and other applications for high spatiotemporal resolution precipitation information.

[0019] Based on the above discussion, and referring to... Figure 1 Example 1: This invention proposes a short-term precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion, including the following: Step 110: Acquire multi-source observation data with different spatiotemporal resolutions, unify the multi-source observation data onto a minute-level high-resolution grid, and calibrate using physical motion information and rain gauge observations to generate a high-quality precipitation field; Step 120: Based on the high-quality precipitation field, short-term precipitation forecasts are made using a pre-established potential diffusion model. The potential diffusion model is constructed using deep learning, which integrates radar echoes with the potential representation of the high-quality precipitation field, and completes precipitation generation in the potential space through a conditional diffusion mechanism.

[0020] In this embodiment 1, by constructing a temporal-spatial dual-dimensional collaborative fusion framework and a flow-matching potential diffusion model, the joint application of multi-source observation data in the generation of minute-level high-resolution precipitation products and short-term precipitation forecasts is realized. This method comprehensively utilizes physical motion information, radar detection physical characteristics, and deep learning models, ensuring both the spatial detail and absolute accuracy of precipitation estimation, while maintaining temporal continuity and physical rationality. It is suitable for operational scenarios such as convective weather monitoring, urban flooding early warning, and agricultural water resource management.

[0021] In a preferred embodiment of the present invention, the multi-source observation data specifically includes the 6-minute mosaic reflectance of the operational large radar. X-band phased array radar 1-minute reflectivity and the 1-minute precipitation intensity of ground rain gauges ; Using the target generation time T as a reference, the multi-source observation data is aligned to time T through time interpolation. For operational large radar data, the echo motion vector field (u,v) is first calculated based on the pre-selected first few observations, and then the most recent time value is adsorbed to time T based on (u,v) using motion compensation time interpolation. For X-band phased array radar data and ground rain gauge data, the most recent time value is used to align to time T. The data is then resampled to a unified, preset high-resolution grid to ensure spatial consistency.

[0022] In this preferred embodiment, the multi-source data preprocessing and spatiotemporal resampling process includes: Input data: Business radar 6-minute mosaic reflectivity ( ); X-band phased array radar 1-minute reflectivity ( ); 1-minute precipitation intensity at ground rain gauges ( ).

[0023] Time alignment: Based on the target generation time (T), various data types are aligned to time (T) through time interpolation. For 6-minute radar data, the echo motion vector field ((u,v)) is first calculated based on the previous observations, and then motion-compensated time interpolation is used to align the most recent time value to time (T). For 1-minute data, the most recent time value is used.

[0024] Spatial resampling: Resample all data to a uniform high-resolution grid (such as a 1-kilometer grid) to ensure spatial consistency.

[0025] In a preferred embodiment of the present invention, specifically, calibration is performed using physical motion information and rain gauge observations to generate a high-quality precipitation field, including... 1) Selection of dynamic precipitation relationships and preliminary precipitation estimation based on physical processes For operational radar, the following reflectivity-rainfall intensity relationship is dynamically selected at each grid point to generate a preliminary precipitation estimate. ) ; When the reflectivity Z of the corresponding radar is less than 45, the R(A) type relationship is adopted; When reflectivity Z < 45 is not valid, based on the correlation coefficient of radar variables, if CC ≤ 0.95, use the R(KDP) type relationship; if CC > 0.95, use the R(A) type relationship. The correlation coefficient of radar variables is a parameter commonly used in this field. For X-band phased array radar, the above reflectivity-rain intensity relationship is dynamically selected at each grid point to generate a preliminary precipitation estimate. ); 2) Adaptive fusion of multi-scale radar data (spatial dimensional fusion) Obtain the confidence level of X-band phased array radar ( ) and the confidence level of the business radar ( ), calculate the weight W of the X-band phased array radar, The business radar weight is 1-W; The preliminary fusion field was then calculated. ; Among them, the confidence level of X-band radar ( The confidence level is determined based on factors such as signal-to-noise ratio, beam obstruction, and distance. It is high within the effective detection range and decreases with increasing distance and obstruction. Business radar confidence level ( Based on the puzzle quality indicators and the consistency assessment with the X-band estimate, the confidence level is high if the two estimates are close within the X-band coverage area, and lower otherwise; outside the coverage area, the confidence level is determined by historical statistics.

[0026] 3) Physically guided time-dimensional extrapolation and current-moment optimization (time-dimensional fusion) This is one of the key innovations of the present invention: introducing physical motion information into the time dimension, and optimizing the precipitation field at the current moment through forward extrapolation and backward tracing.

[0027] Short-time motion vector field estimation: Based on the X-band phased array radar sequence within a preset time period, calculate the fine motion vector field (u(x,y),v(x,y)). The calculation method can be optical flow method or cross-correlation method. Forward extrapolation and backward tracing: The preliminary fusion field at time T-Δt is translated along the fine motion vector field to time T to obtain the extrapolated field. At the same time, the original observation at the current time T is reversed and transferred to the time T-Δt, where Δt is a preset value, and the consistency with the observation at that time is checked to evaluate the reliability of the current observation. Time-constrained optimization is performed using a pre-constructed cost function to obtain... The cost function is: ; Where R is the target optimization field, which is obtained after optimization. Weight Dynamic adjustment; The first item is to maintain the field of spatial integration. The approach, the second encouragement and physical motion extrapolation field Consistent, the third term introduces appropriate spatial smoothness constraints, weights Dynamic adjustment: Increase during the stable movement phase of precipitation. Reduce during the rapid development / dissipation phase Finally obtained .

[0028] Progressive point-area calibration (absolute accuracy calibration) of rain gauge observations. To ensure the absolute accuracy of the product, this step uses rain gauge observations as anchor points to calibrate the optimized precipitation field: Calculate calibration factor for each rain gauge location With discrete factor To constrain the process, a spatially continuous calibration factor field F(x,y) is generated through variational analysis. Considering distance attenuation, topographic consistency, and precipitation type consistency, the calibration factors are smoothly transitioned spatially, ultimately yielding a high-quality precipitation field. .

[0029] In a preferred embodiment of the present invention, the method further includes obtaining a high-quality precipitation field. Then, Kalman filtering or spatiotemporal smoothing is applied to suppress unreasonable minute-by-minute fluctuations; the rationality of extreme values ​​is checked; and conservative extrapolation is performed in marginal areas based on climate background and recent trends.

[0030] In this preferred embodiment, product quality control and continuity assurance are achieved through the above-described methods.

[0031] Regarding the process of generating high-quality precipitation fields in the upper part, this invention has the following innovations in the generation of minute-level precipitation grid products: The time-space dual-dimensional collaborative fusion framework not only integrates radar data with different spatial resolutions, but also uses motion advection extrapolation as a physical constraint to optimize the estimation of the current moment, solving the problems of radar scanning asynchrony and missing information during the intermittent period.

[0032] Physically guided dynamic data weight allocation: Confidence is calculated in real time based on radar detection physics (signal-to-noise ratio, beam obstruction, attenuation characteristics) to achieve true adaptive fusion.

[0033] Progressive point-area calibration technology: First, a physically consistent precipitation field is obtained through multi-radar fusion and time extrapolation. Then, progressive calibration is performed using rain gauges as anchor points, taking into account both spatial details and absolute accuracy.

[0034] Dynamic precipitation relationship selection engine: Real-time determination of precipitation type and selection of appropriate reflectance-rain intensity relationship, significantly improving the inversion accuracy of different precipitation types.

[0035] Temporal consistency constraint of motion perception: Introducing extrapolation field constraints based on physical motion into variational optimization, so that the product maintains physical continuity in the time dimension, which is suitable for fast-moving convection systems.

[0036] In a preferred embodiment of the present invention, the preset duration is 15 minutes.

[0037] After obtaining high-quality minute-level precipitation grid products, this invention further utilizes deep learning to construct a latent diffusion model for short-term precipitation extrapolation. This model integrates the latent representations of radar echoes and precipitation products, and completes precipitation generation in the latent space through a conditional diffusion mechanism.

[0038] 1. Problem Definition Input the radar echo sequence from the past hour ( Predicting precipitation products at each time step over the next 3 hours. Radar echoes and precipitation products are both represented as (). ) grid image.

[0039] 2. Latent Representation and Network Modules Radar VAE encoder / decoder: ; Precipitation VAE encoder / decoder:

[0040] Radar potential: (Conditional information); Precipitation potential: (Generate target); DiT velocity field network: ; Time variable: ; noise: ; Clear target potential: ; Potential results generated: ; In a preferred embodiment of the present invention, specifically, the pre-established potential diffusion model employs two independent variational autoencoders (VAEs) to encode and decode radar echoes and precipitation products; wherein the radar encoder radar echo Mapping to latent representation Radar decoding Representation of radar echo potential Reconstructed as radar echo Precipitation encoder Rain products Mapping to latent representation Precipitation decoder Indicate the potential for precipitation Reconstruction into precipitation products ; Among them, radar VAE: ; ; Precipitation VAE: ; ; , It is random noise that follows a normal distribution.

[0041] To make radar latent variables comparable to precipitation latent variables, the two VAEs share a standard normal prior.

[0042]

[0043] And introduce a potential alignment loss (mean and variance alignment): And the basic losses of the two VAEs: ; .

[0044] In a preferred embodiment of the present invention, specifically, the data processing procedure for the pre-established potential diffusion model includes: (3.1) Conditional bridging path and target velocity field Input radar echo sequence and target precipitation products After encoding by the corresponding VAE, it is obtained and By piecing them together along the timeline, we obtain... and ; Learning conditional distributions within the conditional flow matching framework. ; Let the target precipitation potential be represented as: ; noise: ; time: ; Constructing a bridging path based on conditions: ; The path The target velocity field is: ; This represents the intermediate state in the potential space from the "data endpoint" to the "noise endpoint"; It is a reference velocity obtained from bridging path resolution, used for supervised learning of the velocity field network; (3.2) Conditional DiT velocity field network and flow matching loss: Conditional DiT is used as the velocity field network: ; in As a conditional input, DiT is injected via cross-attention; the velocity field is fitted using flow-matching loss. ; (3.3) Inference: ODE inverse integration generates precipitation potential During inference, in order to draw conclusions from the conditional distribution Sampling, first initialize: ; Under given conditions Solve for ODE: ; The resulting clear precipitation potential representation is obtained: ; (4) Decoder (VAE Decoder) Denoising the latent representation via precipitation decoder Decoding yields images of future precipitation. ).

[0045] In a preferred embodiment of the present invention, specifically, the loss function of the pre-established potential diffusion model includes, The VAE loss, including reconstruction loss and KL divergence loss, respectively ensures that the difference between the input and reconstructed images is minimized and that the latent variables conform to a standard normal distribution. Potential alignment loss: Ensure that the mean and variance of the radar and precipitation potential spaces are aligned; Stream matching loss: by minimizing relative to target speed The mean square error, learning under conditions The following is a continuous transformation from noise to the potential representation of precipitation.

[0046] In a preferred embodiment of the present invention, the training process of the pre-established potential diffusion model specifically includes: Training VAEs: Training separately , and , Optimize the reconstruction loss and KL loss; Potential alignment: utilizing Force radar to align with potential precipitation space; Training the diffusion model: in radar latent representation Under these conditions, train the diffusion model Potential for predicting future precipitation.

[0047] The algorithmic advantages of this pre-established potential diffusion model: Deep modeling of radar-precipitation relationship: By sharing potential space and potential alignment loss, the statistical relationship between radar echo and precipitation products is explicitly modeled to address the potential distributional differences between the two.

[0048] Conditional diffusion generation: The conditional diffusion model is used for generation in the latent space. Compared with direct modeling in the pixel space, it has lower noise, more efficient inference, and can better capture the spatiotemporal structure of precipitation.

[0049] High-precision short-term forecast: Combining minute-level high-quality precipitation products and potential diffusion models, it can achieve high-precision extrapolation forecasts of precipitation intensity for the next three hours, which is especially suitable for severe convective weather warnings.

[0050] Example 2: This invention also proposes a short-term precipitation forecasting device based on high spatiotemporal resolution multi-source observation fusion, comprising: The minute-level precipitation grid product generation module is used to acquire multi-source observation data with different spatiotemporal resolutions, unify the multi-source observation data onto a minute-level high-resolution grid, and use physical motion information and rain gauge observations for calibration to generate a high-quality precipitation field. The latent generation model extrapolation forecast module based on conditional flow matching is used to perform short-term precipitation forecasts based on the high-quality precipitation field and a pre-established latent diffusion model. The latent diffusion model is constructed through deep learning, which integrates radar echoes with the latent representation of the high-quality precipitation field, and completes precipitation generation in the latent space through a conditional diffusion mechanism.

[0051] In this embodiment 2, the minute-level precipitation grid product generation module generates a high-resolution minute-level precipitation grid product with physical consistency and temporal continuity through spatiotemporal fusion of multi-scale radar and rain gauge observations. This module includes steps such as data preprocessing, dynamic precipitation relationship selection, adaptive fusion, physically guided temporal optimization, rain gauge calibration, and product quality control.

[0052] The Latent Diffusion Model (LDM) extrapolation forecast module based on conditional flow matching uses minute-level fused precipitation products as high-quality training samples and combines radar reflectivity information to construct a Latent Diffusion Model (LDM) for high-precision short-term extrapolation forecasts of precipitation intensity for the next three hours.

[0053] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment, depending on actual needs.

[0054] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0055] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0056] Although the description of the invention has been quite detailed and particularly of several described embodiments, it is not intended to limit it to any of these details or embodiments or any particular embodiment, but should be considered as providing a broad possible interpretation of the claims by referring to the appended claims and taking into account the prior art, thereby effectively covering the intended scope of the invention. Furthermore, the invention has been described above with respect to embodiments foreseeable by the inventors in order to provide a useful description, and non-substantial modifications to the invention that have not yet been foreseen may still represent equivalent modifications.

[0057] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. Any embodiment that achieves the technical effects of the present invention using the same means should fall within the protection scope of the present invention. Within the protection scope of the present invention, various modifications and variations can be made to the technical solutions and / or implementation methods.

Claims

1. A short-impending precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion, characterized in that, Including the following: Acquire multi-source observation data with different spatiotemporal resolutions, unify the multi-source observation data onto a minute-level high-resolution grid, and calibrate using physical motion information and rain gauge observations to generate a high-quality precipitation field; Based on the high-quality precipitation field, short-term precipitation forecasts are made using a pre-established potential diffusion model. The potential diffusion model is constructed using deep learning, which integrates radar echoes with the potential representation of the high-quality precipitation field, and generates precipitation in the potential space through a conditional diffusion mechanism.

2. The short-term precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion as described in claim 1, characterized in that, In particular, the multi-source observation data includes a business large radar 6-minute mosaic reflectivity , an X-band phased array radar 1-minute reflectivity , and a 1-minute precipitation intensity of a ground rain gauge wherein denotes a longitude coordinate, denotes a latitude coordinate, is an observation time of the 6-minute mosaic reflectivity, is an observation time of the 1-minute mosaic reflectivity, is an observation time of the rain gauge. Using the target generation time T as a reference, the multi-source observation data is aligned to time T through time interpolation. For operational large radar data, the echo motion vector field (u,v) is first calculated based on the pre-selected first few observations, and then the most recent time value is adsorbed to time T based on (u,v) using motion compensation time interpolation. For X-band phased array radar data and ground rain gauge data, the most recent time value is used to align to time T. The data is then resampled to a unified, preset high-resolution grid to ensure spatial consistency.

3. The short-term precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion as described in claim 1, characterized in that, Specifically, calibration is performed using physical motion information and rain gauge observations to generate a high-quality precipitation field, including: For operational radar systems, the following reflectivity-rainfall intensity relationship is dynamically selected for each grid point to generate a preliminary precipitation estimate. ; When the reflectivity Z of the corresponding radar is less than 45, the R(A) type relationship is adopted; When reflectivity Z < 45 is not valid, based on the correlation coefficient of radar variables, if CC ≤ 0.95, use the R(KDP) type relation; if CC > 0.95, use the R(A) type relation. For X-band phased array radar, the above reflectivity-rain intensity relationship is dynamically selected at each grid point to generate a preliminary precipitation estimate. ; Obtain the confidence level of X-band phased array radar ( ) and the confidence level of the business radar ( ), calculate the weight W of the X-band phased array radar, The business radar weight is 1-W; The preliminary fusion field was then calculated. ; Based on the X-band phased array radar sequence within a preset time period, calculate the fine motion vector field (u(x,y),v(x,y)); The extrapolated field is obtained by translating the initial fusion field at time T-Δt along the fine motion vector field to time T. At the same time, the original observation at the current time T is reversed and transferred to the time T-Δt, where Δt is a preset value, and the consistency with the observation at that time is checked to evaluate the reliability of the current observation. Time-constrained optimization is performed using a pre-constructed cost function to obtain... The cost function is: ； Where R is the target optimization field, which is obtained after optimization. Weight Dynamic adjustment Preset spatial smoothing constraints; Calculate calibration factor for each rain gauge location , Let i be the precipitation at the i-th rain gauge station. Let i be the longitude of the i-th rain gauge station. Let be the latitude of the i-th rain gauge, with a discrete factor of . To constrain the process, a spatially continuous calibration factor field F(x,y) is generated through variational analysis, ultimately yielding a high-quality precipitation field. .

4. The short-term precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion according to claim 3, characterized in that, The method also includes obtaining a high-quality precipitation field. Then, Kalman filtering or spatiotemporal smoothing is applied to suppress unreasonable minute-by-minute fluctuations; the rationality of extreme values ​​is checked; and conservative extrapolation is performed in marginal areas based on climate background and recent trends.

5. The short-term precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion according to claim 1, characterized in that, Specifically, the preset duration is 15 minutes.

6. The short-term precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion according to claim 1, characterized in that, Specifically, the pre-established potential diffusion model uses two independent variational autoencoders (VAEs) to encode and decode radar echoes and precipitation products; among which the radar encoder... radar echo Mapping to latent representation Radar decoding Representation of radar echo potential Reconstructed as radar echo Precipitation encoder Rain products Mapping to latent representation Precipitation decoder Indicate the potential for precipitation Reconstruction into precipitation products ; To make radar latent variables comparable to precipitation latent variables, the two VAEs share a standard normal prior.

7. The short-term precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion according to claim 6, characterized in that, Specifically, the data processing procedure for the pre-established potential diffusion model includes: Input radar echo sequence and target precipitation products After encoding by the corresponding VAE, it is obtained. and By piecing them together along the timeline, we obtain... and ; Learning conditional distributions within the conditional flow matching framework. ; Let the target precipitation potential be represented as: ； noise: ； time: ； Constructing a bridging path based on conditions: ； The path The target velocity field is: ； This represents the intermediate state between the "data endpoint" and the "noise endpoint" in the potential space. It is a reference velocity obtained from bridging path resolution, used for supervised learning of the velocity field network; Conditional DiT velocity field network and flow matching loss: Conditional DiT is used as the velocity field network: ； in As a conditional input, DiT is injected via cross-attention; the velocity field is fitted using flow-matching loss. ； Inference: ODE inverse integration generates potential precipitation. During inference, in order to draw conclusions from the conditional distribution Sampling, first initialize: ； Under given conditions Solve for ODE: ； The resulting clear precipitation potential representation is obtained: ； Denoising the latent representation via precipitation decoder Decoding yields images of future precipitation. ).

8. The short-term precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion according to claim 7, characterized in that, Specifically, the loss function of the pre-established potential diffusion model includes, The VAE loss, including reconstruction loss and KL divergence loss, respectively ensures that the difference between the input and reconstructed images is minimized and that the latent variables conform to a standard normal distribution. Potential alignment loss: Ensure that the mean and variance of the radar and precipitation potential spaces are aligned; Stream matching loss: by minimizing relative to target speed The mean square error, learning under conditions The following is a continuous transformation from noise to the potential representation of precipitation.

9. The short-term precipitation forecasting method based on high spatiotemporal resolution multi-source observation fusion according to claim 8, characterized in that, Specifically, the training process of the pre-established potential diffusion model includes, Training VAEs: Training separately , and , Optimize the reconstruction loss and KL loss; Potential alignment: utilizing Force radar to align with potential precipitation space; Training the diffusion model: in radar latent representation Under these conditions, train the diffusion model Potential for predicting future precipitation.

10. A short-term precipitation forecasting device based on high spatiotemporal resolution multi-source observation fusion, characterized in that, include: The minute-level precipitation grid product generation module is used to acquire multi-source observation data with different spatiotemporal resolutions, unify the multi-source observation data onto a minute-level high-resolution grid, and use physical motion information and rain gauge observations for calibration to generate a high-quality precipitation field. The latent generation model extrapolation forecast module based on conditional flow matching is used to perform short-term precipitation forecasts based on the high-quality precipitation field and a pre-established latent diffusion model. The latent diffusion model is constructed through deep learning, which integrates radar echoes with the latent representation of the high-quality precipitation field, and completes precipitation generation in the latent space through a conditional diffusion mechanism.

Images