Beidou rainfall prediction method based on deep learning large model and related device

Abstract

This application relates to the field of meteorological forecasting technology, and provides a BeiDou rainfall forecasting method and related apparatus based on a deep learning large-scale model. The method acquires spatiotemporal reference data of the target area through the BeiDou satellite navigation system, including high-precision positioning data of BeiDou ground-based augmentation stations and meteorological element inversion data of BeiDou remote sensing payloads; collects real-time rainfall observation data from ground rainfall monitoring terminals, meteorological radars, and satellite remote sensing; performs spatiotemporal alignment processing on the spatiotemporal reference data and real-time rainfall observation data to construct a standardized dataset; constructs a deep learning large-scale model, trains the deep learning large-scale model using a spatiotemporal sequence prediction task, and inputs the standardized dataset into the trained deep learning large-scale model to output multi-scale rainfall forecast results for the target area for a preset duration in the future.

Description

Technical Field

[0001] This application relates to the field of meteorological forecasting technology, and in particular to a BeiDou rainfall forecasting method and related device based on a deep learning large model. Background Technology

[0002] In the field of meteorological forecasting, regional rainfall forecasting is of great significance for disaster prevention and mitigation, agricultural irrigation, and water resource management. Traditional rainfall forecasting methods mainly rely on numerical weather prediction models or statistical machine learning models, which have the following technical bottlenecks: First, insufficient spatiotemporal resolution makes it difficult to achieve fine-grained grid forecasting, and has limited ability to characterize small-scale meteorological element changes in complex terrain areas; Second, low efficiency of multi-source data fusion, especially lacking in-depth utilization of the unique spatiotemporal reference data of the BeiDou satellite navigation system (such as high-precision positioning data from ground-based augmentation stations and meteorological inversion data from remote sensing payloads), resulting in insufficient spatiotemporal alignment accuracy and feature correlation mining; Third, weak model adaptability, making it difficult to update the forecasting model in real time during extreme weather events, and lacking a dynamic calibration mechanism based on real-time feedback, leading to the accumulation of forecast errors over time. Summary of the Invention

[0003] This application provides a BeiDou rainfall prediction method and related device based on a deep learning large model. It aims to address the following problems that, despite the advantages of neural network-based prediction models in time-series data processing, existing methods generally suffer from the following issues as deep learning technology develops: ① They fail to fully integrate the spatiotemporal reference characteristics of the BeiDou system to construct dedicated feature representations, thus failing to effectively utilize the high-precision spatiotemporal coordinates and dynamic orbital parameters of BeiDou data; ② The multimodal feature fusion method is simple and lacks deep modeling of the spatiotemporal coupling characteristics of meteorological data (e.g., it does not introduce a dynamic attention mechanism to adapt to differences in terrain and meteorological conditions in different regions); ③ The calibration effect of satellite orbital parameters on spatiotemporal attention is not considered during training, and the prediction results lack spatial error calibration based on the accuracy of BeiDou differential positioning, leading to insufficient prediction accuracy in complex scenarios.

[0004] Firstly, this application provides a BeiDou rainfall prediction method based on a large deep learning model, including: Spatiotemporal reference data for the target area is obtained through the BeiDou satellite navigation system, including high-precision positioning data from BeiDou ground-based augmentation stations and meteorological element inversion data from BeiDou remote sensing payloads; real-time rainfall observation data from ground rainfall monitoring terminals, meteorological radars, and satellite remote sensing are collected; and the spatiotemporal reference data and real-time rainfall observation data are spatiotemporally aligned to construct a standardized dataset. A large-scale deep learning model is constructed, comprising a spatiotemporal encoding module, a multimodal feature fusion module, and a dynamic attention module. The spatiotemporal encoding module jointly encodes time-series features and spatial location features in standardized data to generate an embedding vector containing spatiotemporal location information. The multimodal feature fusion module fuses feature representations from BeiDou remote sensing data, ground monitoring data, and meteorological auxiliary data through a cross-modal attention mechanism. The dynamic attention module dynamically adjusts feature weights based on the terrain complexity and meteorological conditions of different regions. The deep learning model is trained using a spatiotemporal sequence prediction task. The standardized dataset is input into the trained deep learning model to output multi-scale rainfall prediction results for the target area for a predetermined duration in the future.

[0005] In some embodiments, training the deep learning large model using a spatiotemporal sequence prediction task includes: during the training process, introducing real-time orbital parameters of BeiDou satellites as a dynamic adjustment factor to adaptively calibrate the spatiotemporal attention weights of the deep learning large model; the loss function includes temporal prediction error, spatial distribution consistency error, and model complexity regularization term.

[0006] In some embodiments, the method further includes: when an extreme weather event occurs in the target area, automatically triggering an incremental learning mechanism to locally update the model parameters using newly added monitoring data in real time.

[0007] In some embodiments, the multi-scale rainfall prediction result includes a refined prediction grid; after outputting the multi-scale rainfall prediction result for the target area for a preset duration in the future, it further includes: using the wide-area differential positioning accuracy parameters of the BeiDou satellite to perform spatial error calibration on the prediction result, constructing a dynamic correction model in combination with real-time verification data from ground rain gauges, and correcting outliers in the refined prediction grid in real time.

[0008] In some embodiments, the method further includes: establishing a prediction result feedback mechanism based on BeiDou short message communication, receiving prediction effect feedback data from terminal devices in real time; evaluating the prediction confidence of the deep learning large model based on the feedback data, and automatically initiating the online fine-tuning process of the model when the prediction error exceeds a preset threshold, and pushing the final rainfall prediction result containing the accuracy level identifier to authorized users through the BeiDou system.

[0009] In some embodiments, the spatiotemporal alignment processing of the spatiotemporal reference data and real-time rainfall observation data to construct a standardized dataset includes: based on the high-precision spatiotemporal reference information provided by the BeiDou Navigation Satellite System, uniformly converting the timestamps of data from different sources into the BeiDou time standard time system; using a sliding window interpolation method to synchronize data sources with inconsistent time resolutions, generating time-series data sequences with equal time intervals; and uniformly transforming the coordinate system used by the positioning data of the BeiDou ground-based augmentation station with the coordinate systems of the ground rainfall monitoring terminal, meteorological radar, and satellite remote sensing data using coordinate transformation parameters provided by the BeiDou system. The data is then projected onto the Gauss-Krüger projection coordinate system of the target area. The target area is divided into grids according to a preset spatial grid resolution. Based on the sub-meter accuracy benchmark of BeiDou positioning data, each type of data is mapped to the corresponding grid node according to its spatial location. For cross-grid data, spatial interpolation is performed using the inverse distance weighting method. Multi-source data is normalized. For numerical data, a standardized scaling method is used, and for categorical meteorological element data, one-heat encoding is used. At the same time, outliers are removed and missing values ​​are filled by spatiotemporal neighborhood data, forming a structured dataset containing time dimension, spatial grid coordinates, and multimodal observation features.

[0010] In some embodiments, the construction of the deep learning large model includes: adopting a Transformer-based neural network architecture as the basic framework; the spatiotemporal coding module includes a time-dimensional location coding layer and a spatial-dimensional coordinate embedding layer, and maps timestamps, latitude and longitude, and altitude into high-dimensional spatiotemporal embedding vectors through a spatiotemporal joint coding function; the multimodal feature fusion module sets up a cross-modal interaction sublayer, and cross-fuses BeiDou remote sensing data features, ground monitoring data features, and meteorological auxiliary data features through a multi-head self-attention mechanism; the dynamic attention module constructs terrain complexity feature vectors and meteorological condition feature vectors as inputs for attention weight adjustment, and realizes dynamic adjustment of feature weights for different regions based on an adaptive normalization layer.

[0011] Secondly, this application provides a BeiDou rainfall prediction device based on a deep learning large model, the device comprising: The dataset construction unit is used to acquire spatiotemporal reference data of the target area through the BeiDou satellite navigation system, including high-precision positioning data of BeiDou ground-based augmentation stations and meteorological element inversion data of BeiDou remote sensing payloads; collect real-time rainfall observation data from ground rainfall monitoring terminals, meteorological radars and satellite remote sensing; and perform spatiotemporal alignment processing on the spatiotemporal reference data and real-time rainfall observation data to construct a standardized dataset. The model building unit is used to construct a large deep learning model, which includes a spatiotemporal encoding module, a multimodal feature fusion module, and a dynamic attention module. The spatiotemporal encoding module jointly encodes time-series features and spatial location features in standardized data to generate an embedding vector containing spatiotemporal location information. The multimodal feature fusion module fuses feature representations from BeiDou remote sensing data, ground monitoring data, and meteorological auxiliary data through a cross-modal attention mechanism. The dynamic attention module dynamically adjusts feature weights according to the terrain complexity and meteorological conditions of different regions. The result output unit is used to train the deep learning big model using a spatiotemporal sequence prediction task, input the standardized dataset into the trained deep learning big model, and output the multi-scale rainfall prediction results for the target area for a preset duration in the future.

[0012] Thirdly, embodiments of this application provide a computer device, including a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and, when executing the computer program, implement the BeiDou rainfall prediction method based on a deep learning large model as provided in any embodiment of this application.

[0013] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the BeiDou rainfall prediction method based on a deep learning large model as provided in any embodiment of this application.

[0014] This invention acquires high-precision spatiotemporal reference data (ground-based augmentation station positioning data and remote sensing payload inversion data) through the BeiDou system, and aligns it spatiotemporally with multi-source data such as ground monitoring and meteorological radar to construct a standardized dataset containing multi-dimensional spatiotemporal features. This significantly improves data consistency and feature richness, laying the foundation for accurate prediction. Based on the Transformer architecture, a spatiotemporal coding module, a multimodal feature fusion module, and a dynamic attention mechanism are designed to achieve joint encoding of time-series features and spatial location features. Through a cross-modal attention mechanism, multi-source data features are deeply fused, and feature weights are dynamically adjusted according to terrain complexity and meteorological conditions, effectively capturing the spatiotemporal coupling relationship under complex meteorological scenarios. Real-time orbital parameters of BeiDou satellites are introduced to calibrate spatiotemporal attention weights, combined with a composite loss function including temporal error and spatial consistency error, to improve model training accuracy. Through incremental learning mechanisms and real-time feedback fine-tuning, rapid adaptation to extreme weather conditions and dynamic correction of prediction errors are achieved, significantly enhancing model robustness. It outputs a refined prediction grid with high spatiotemporal resolution (minute-by-minute temporal resolution and 500-meter spatial resolution), uses BeiDou differential positioning accuracy parameters for spatial error calibration, and constructs a dynamic correction model based on ground verification data to achieve real-time correction of prediction results, breaking through the limitations of traditional methods in terms of spatiotemporal resolution and prediction accuracy in complex terrain.

[0015] In summary, this invention, through the deep integration of the BeiDou system and a large-scale deep learning model, forms a rainfall prediction technology solution with spatiotemporal reference perception, multimodal feature modeling, and dynamic adaptive optimization. This significantly improves the prediction accuracy and real-time performance in complex scenarios and has important engineering application value.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the steps of a BeiDou rainfall prediction method based on a large deep learning model provided in an embodiment of this application. Figure 2 This is a schematic diagram illustrating the principle of a BeiDou rainfall prediction method based on a large deep learning model, provided in one embodiment of this application. Figure 3 This is a schematic block diagram of the structure of a Beidou rainfall prediction device based on a deep learning large model provided in one embodiment of this application; Figure 4 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0022] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0023] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0024] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0025] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0026] In the field of meteorological forecasting, regional rainfall forecasting is of great significance for disaster prevention and mitigation, agricultural irrigation, and water resource management. Traditional rainfall forecasting methods mainly rely on numerical weather prediction models or statistical machine learning models, which have the following technical bottlenecks: First, insufficient spatiotemporal resolution makes it difficult to achieve fine-grained grid forecasting, and has limited ability to characterize small-scale meteorological element changes in complex terrain areas; Second, low efficiency of multi-source data fusion, especially lacking in-depth utilization of the unique spatiotemporal reference data of the BeiDou satellite navigation system (such as high-precision positioning data from ground-based augmentation stations and meteorological inversion data from remote sensing payloads), resulting in insufficient spatiotemporal alignment accuracy and feature correlation mining; Third, weak model adaptability, making it difficult to update the forecasting model in real time during extreme weather events, and lacking a dynamic calibration mechanism based on real-time feedback, leading to the accumulation of forecast errors over time.

[0027] This application provides a BeiDou rainfall prediction method and related device based on a deep learning large model. It aims to address the following problems that, despite the advantages of neural network-based prediction models in time-series data processing, existing methods generally suffer from the following issues as deep learning technology develops: ① They fail to fully integrate the spatiotemporal reference characteristics of the BeiDou system to construct dedicated feature representations, thus failing to effectively utilize the high-precision spatiotemporal coordinates and dynamic orbital parameters of BeiDou data; ② The multimodal feature fusion method is simple and lacks deep modeling of the spatiotemporal coupling characteristics of meteorological data (e.g., it does not introduce a dynamic attention mechanism to adapt to differences in terrain and meteorological conditions in different regions); ③ The calibration effect of satellite orbital parameters on spatiotemporal attention is not considered during training, and the prediction results lack spatial error calibration based on the accuracy of BeiDou differential positioning, leading to insufficient prediction accuracy in complex scenarios.

[0028] To solve the above problem, please refer to Figure 1 This application provides a BeiDou rainfall prediction method based on a large deep learning model. The method runs on a computer device, which can be deployed on a single server or a server cluster. It can also be deployed on handheld terminals, laptops, wearable devices, or robots, etc.

[0029] like Figure 1 As shown, the provided BeiDou rainfall prediction method based on a deep learning large model includes steps S101 to S103. Details are as follows: Step S101. Obtain spatiotemporal reference data of the target area through the BeiDou satellite navigation system, including high-precision positioning data of BeiDou ground-based augmentation stations and meteorological element inversion data of BeiDou remote sensing payloads; collect real-time rainfall observation data from ground rainfall monitoring terminals, meteorological radars and satellite remote sensing; perform spatiotemporal alignment processing on the spatiotemporal reference data and real-time rainfall observation data to construct a standardized dataset.

[0030] Specifically, this step addresses the problems of insufficient spatiotemporal alignment accuracy and inadequate utilization of multi-source data in traditional methods by integrating spatiotemporal reference data from the BeiDou Navigation Satellite System with multi-source observation data. The core process includes four parts: BeiDou spatiotemporal reference data acquisition, multi-source observation data integration, spatiotemporal alignment processing, and standardized dataset construction.

[0031] The acquisition of BeiDou spatiotemporal reference data includes: High-precision positioning data from ground-based augmentation stations: Using BeiDou ground-based augmentation systems (such as CORS stations), the coordinates (latitude, longitude, and elevation), positioning accuracy parameters (such as differential positioning error covariance), and real-time timestamps of high-density distributed reference stations within the target area are acquired (within nanosecond accuracy). Meteorological inversion data from remote sensing payloads: Meteorological elements such as atmospheric humidity, temperature profiles, and cloud top height are inverted using remote sensing payloads carried by BeiDou (such as microwave radiometers and infrared detectors). Combined with satellite orbital parameters (ephemeris data), a gridded inversion result with spatiotemporal labels is constructed (e.g., a 500m × 500m resolution grid).

[0032] Ground rainfall monitoring terminal: Collects minute-by-minute rainfall data from equipment such as automatic rain gauges and tipping bucket rain gauges, along with the station's latitude and longitude and equipment calibration parameters.

[0033] Weather radar data: Acquire basic data such as radar reflectivity factor (Z value) and radial velocity, and generate gridded reflectivity products of the target area (e.g., 1km×1km resolution, updated every 5 minutes) through radar network mosaic technology.

[0034] Satellite remote sensing data: Integrating precipitation estimation products from other satellites (such as the Fengyun series) as supplementary observation data.

[0035] Spatiotemporal alignment processing includes: Spatial alignment: Projecting all data uniformly to a dedicated coordinate system for the target area (such as UTM projection), and correcting the coordinates of ground monitoring stations based on the high-precision coordinates of BeiDou ground-based augmentation stations (using differential positioning technology to eliminate ionospheric and tropospheric errors).

[0036] Kriging interpolation or inverse distance weighting is used to interpolate discrete point data (such as rain gauges and reference stations) to a unified grid (such as a 100m×100m fine grid), and the grid coordinates contain high-precision latitude and longitude information under the BeiDou spatiotemporal reference.

[0037] Time alignment: Based on the BeiDou time service system, the timestamps of all data are unified to UTC time, and the time resolution is aligned to the minute level (e.g., using a 5-minute time window, time gaps are filled by linear interpolation or the nearest neighbor method).

[0038] Feature engineering: Extract BeiDou spatiotemporal reference features (such as reference station coordinates, satellite orbital altitude, clock error parameters), meteorological inversion features (humidity, temperature), and observation features (radar reflectivity, measured rainfall) to form a multi-dimensional feature vector.

[0039] Data cleaning and normalization: Remove outliers (such as extreme values ​​exceeding physical thresholds), perform Z-score normalization on numerical features, and perform one-hot encoding on categorical features (such as terrain type).

[0040] Dataset partitioning: The dataset is divided into a training set (70%), a validation set (20%), and a test set (10%) in chronological order to ensure the continuity of the time series.

[0041] Step S102. Construct a large-scale deep learning model, which includes a spatiotemporal coding module, a multimodal feature fusion module, and a dynamic attention module. The spatiotemporal coding module jointly encodes the time-series features and spatial location features in the standardized data to generate an embedding vector containing spatiotemporal location information. The multimodal feature fusion module fuses the feature representations of BeiDou remote sensing data, ground monitoring data, and meteorological auxiliary data through a cross-modal attention mechanism. The dynamic attention module dynamically adjusts the feature weights according to the terrain complexity and meteorological conditions of different regions.

[0042] Specifically, this step constructs a deep learning model that includes a spatiotemporal coding module, a multimodal feature fusion module, and a dynamic attention module, addressing the problems of insufficient spatiotemporal feature modeling, simple multimodal fusion, and weak dynamic adaptability in traditional methods.

[0043] Time series encoding: An improved Transformer time encoder is adopted. The input includes time-related features such as timestamps, seasonal cycles, and historical rainfall sequences. The time dependencies are captured by location encoding (sine and cosine functions), and a time embedding vector T is generated by combining the high-precision timestamps of BeiDou time service.

[0044] Spatial location encoding: The grid coordinates (latitude, longitude, and elevation) and the positioning accuracy parameters of the BeiDou ground-based augmentation station (such as horizontal / vertical accuracy factors HDOP / VDOP) are used as spatial features. A spatial embedding vector S is generated by a graph neural network (GNN) or a two-dimensional convolutional neural network (CNN), where the size of the convolution kernel is dynamically adjusted according to the terrain complexity (such as using a smaller kernel to capture details in mountainous areas).

[0045] Joint encoding: The temporal embedding T and the spatial embedding S are concatenated by dimension, and a spatiotemporal joint embedding vector E=FC([T;S]) is generated through a fully connected layer.

[0046] Cross-modal attention mechanism: Input modalities include: BeiDou remote sensing inversion data (Modal A), ground monitoring data (Modal B), and meteorological auxiliary data (such as numerical weather prediction model output, Modal C). For each modality, features FA, FB, and FC are first extracted through an independent encoder (such as CNN or LSTM). Then, cross-attention is used to achieve inter-modal interaction: Ffusion = Attention(FA,FB,FB) + Attention(Ffusion,FC,FC). BeiDou orbital parameters (such as satellite elevation angle and azimuth angle) are specifically introduced as bias terms for attention query to enhance the modeling of the spatiotemporal reference characteristics of BeiDou data.

[0047] Topographic and meteorological condition inputs: Preprocessed topographic complexity indicators (such as slope and roughness calculated from DEM data) and real-time meteorological conditions (wind speed and pressure gradient) are used to generate a condition vector c through a multilayer perceptron (MLP). Adaptive weight generation: Using the spatiotemporal joint embedding E and the condition vector c as inputs, a dynamic weight network (such as a gated recurrent unit GRU) is used to generate attention weights αi for each grid point. A terrain mask is introduced during weight calculation to highlight the features of complex terrain areas: αi = Softmax(W * Tanh(WeEi + Wcc) ⊙ M terrain ) where Mterrain is a binary mask for terrain complexity (1 for complex areas, 0.5 otherwise).

[0048] Feature weighted fusion: The features after multimodal fusion are weighted according to the weight αi, and the dynamically adjusted feature representation Fdynamic is output.

[0049] Step S103. Train the deep learning big model using a spatiotemporal sequence prediction task, input the standardized dataset into the trained deep learning big model, and output the multi-scale rainfall prediction results for the target area for a preset duration in the future.

[0050] This step trains the model through a spatiotemporal sequence prediction task, which solves the problems of insufficient prediction accuracy and lack of dynamic calibration in traditional methods, and supports multi-scale (such as 100m, 500m, 1km resolution) rainfall prediction output.

[0051] Input and output format: The input is multimodal features within a spatiotemporal sliding window (e.g., the past 6 hours, with a time step of 5 minutes, for a total of 72 time steps), and the output is a multi-scale rainfall grid for the next 12 hours, with each hour as a base (e.g., using 100m resolution as a base, and generating 500m and 1km resolution results through upsampling).

[0052] Loss function: A multi-task loss function is adopted, combining mean squared error (MSE) and quantile loss to improve the prediction capability of extreme precipitation scenarios. .

[0053] Where λ1 and λ2 are weight coefficients, which are optimized through the validation set.

[0054] Orbit parameter calibration reduces the impact of orbit errors on prediction by using the dynamic orbit parameters of BeiDou satellites (such as the position error covariance in ephemeris data) as a regularization term during training to constrain the weights of the spatiotemporal attention module.

[0055] Differential positioning error calibration involves correcting the spatial error of the prediction results for each grid during the prediction phase, based on the differential positioning accuracy of the BeiDou ground-based augmentation stations (such as the real-time published standard deviation of the positioning error). .

[0056] Where σpos is the positioning error corresponding to the grid point, and Dref is the reference distance (e.g., 1km).

[0057] The basic grid prediction directly outputs a finely detailed 100m resolution grid of rainfall as the base prediction result. Upsampling and fusion generate 500m and 1km resolution grids from the basic grid using bilinear interpolation. Combining historical error statistics at different resolutions (e.g., high-resolution grids have smaller errors in mountainous areas), a weighted average method is used to fuse multi-scale results, improving the prediction robustness under different scenarios.

[0058] In some embodiments, training the deep learning large model using a spatiotemporal sequence prediction task includes: during the training process, introducing real-time orbital parameters of BeiDou satellites as a dynamic adjustment factor to adaptively calibrate the spatiotemporal attention weights of the deep learning large model; the loss function includes temporal prediction error, spatial distribution consistency error, and model complexity regularization term.

[0059] During the model training phase, real-time orbital parameters of BeiDou satellites are introduced as dynamic adjustment factors to calibrate the spatiotemporal attention weights. At the same time, a multi-dimensional loss function is designed, which includes temporal prediction error, spatial distribution consistency error, and model complexity regularization term, to improve the model's ability to accurately model spatiotemporal dependencies.

[0060] Parameter acquisition is achieved by acquiring BeiDou satellite ephemeris data (including orbital position, velocity, clock error, and orbital error covariance matrix) in real time, through decoding BeiDou navigation messages or accessing official data interfaces, at a frequency consistent with the satellite data update rate (e.g., once per second).

[0061] Attention weight adjustment incorporates orbital altitude and the geometric distance between the satellite and the target grid point (calculated inversely using latitude and longitude) as bias terms when calculating the attention score in the spatiotemporal attention module. The formula is as follows: .

[0062] Where β is a hyperparameter, and Orbit_Bias contains the normalized orbital altitude value and the reciprocal of the geometric distance. When a satellite enters a high-orbit or low-orbit region, the attenuation coefficient of the attention weight is automatically adjusted to enhance the focus on the features of the inverted data from low-Earth orbit satellites.

[0063] The mean square error of time-series forecasts (MSE-T) is calculated by comparing the predicted rainfall series with the actual series, focusing on the continuity of the time dimension. (T is the time step, N is the number of grid points).

[0064] Spatial Distribution Consistency Error (SSIM-S): This index assesses the consistency between predicted and actual rainfall spatial distribution patterns, suppressing unreasonable spatial outliers. ; Model complexity regularization (L2+Dropout): This applies L2 norm constraints to the model weights and introduces Dropout layers in the spatiotemporal encoding module and attention layer to prevent overfitting. ; Total loss function: (α is the spatial error weight, which is tuned using the validation set).

[0065] In some embodiments, the method further includes: when an extreme weather event occurs in the target area, automatically triggering an incremental learning mechanism to locally update the model parameters using newly added monitoring data in real time.

[0066] An incremental learning mechanism is designed for extreme weather events (such as typhoons and torrential rains). By triggering local updates of model parameters through real-time monitoring data, the mechanism solves the problem that traditional models cannot adapt quickly in sudden scenarios and achieves dynamic calibration of the model.

[0067] Extreme weather event detection includes: Data triggering conditions: When ≥30% of ground rain gauges in the target area detect rainfall ≥50mm within 1 hour (extreme precipitation threshold), or when meteorological radar detects strong echoes with a combined reflectivity ≥50dBz lasting for more than 30 minutes, incremental learning is automatically triggered. Multi-source data fusion judgment: Combining abrupt changes in atmospheric precipitable water (PWV) retrieved from BeiDou remote sensing (increase ≥20% within 3 hours) with extreme weather early warning products from numerical models improves detection accuracy.

[0068] The incremental learning execution process includes: Local parameter selection: Only the weights of the dynamic attention module and the multimodal feature fusion module are updated, while the underlying parameters of the spatiotemporal coding module are fixed (to avoid damaging the basic spatiotemporal feature extraction capability). The parameter update range is controlled through gradient masking technology. Real-time data processing: Newly added extreme weather period data (such as minute-by-minute monitoring data from the past 2 hours) are spatiotemporally aligned to construct a temporary incremental training set (sample size ≥ 100 time steps). Lightweight training strategy: The AdamW optimizer is used, with the learning rate set to 10 times the base training rate (e.g., 1e-3). The training epochs are 5-10, and each epoch only calculates the gradient of the incremental data and updates the local parameters.

[0069] Model state saving and rollback are achieved by saving a snapshot of the current model parameters before triggering. If the validation set error increases by more than 5% after incremental learning, the model will automatically roll back to the historical optimal parameters to avoid model performance degradation caused by noisy data.

[0070] In some embodiments, the multi-scale rainfall prediction result includes a refined prediction grid; after outputting the multi-scale rainfall prediction result for the target area for a preset duration in the future, it further includes: using the wide-area differential positioning accuracy parameters of the BeiDou satellite to perform spatial error calibration on the prediction result, constructing a dynamic correction model in combination with real-time verification data from ground rain gauges, and correcting outliers in the refined prediction grid in real time.

[0071] After refining the grid prediction, spatial error calibration is performed using the BeiDou wide-area differential positioning accuracy parameters. A dynamic correction model is then constructed by combining real-time data from ground stations to correct abnormal prediction values ​​in real time, thereby improving the prediction accuracy in complex terrain areas.

[0072] The application of BeiDou differential positioning accuracy parameters is achieved through error model construction: based on the real-time differential positioning accuracy (such as horizontal accuracy σ_h, vertical accuracy σ_v) released by BeiDou ground-based augmentation stations, a linear relationship between grid point prediction error and positioning accuracy is established. Spatial error calibration is performed by adjusting the predicted value for each grid point.

[0073] The construction of the dynamic correction model includes: real-time verification data acquisition: acquiring measured data from ground rain gauges every 10 minutes, and generating a verification grid Yverify with the same resolution through inverse distance weighting interpolation.

[0074] Outlier detection: Calculate the difference between the calibrated predicted value and the validation grid. When the absolute value of the difference exceeds 3 times the historical standard deviation of the grid points, it is marked as an outlier.

[0075] The correction strategy corrects outlier grids by using predicted and validated values ​​from normal grids within the spatiotemporal neighborhood, employing a weighted average method (weights are inversely proportional to distance). The correction effect is evaluated in real-time by calculating the root mean square error (RMSE) of the grid points after each correction. If the error decreases by ≥15% after correction, the result is retained; otherwise, the correction is skipped.

[0076] In some embodiments, the method further includes: establishing a prediction result feedback mechanism based on BeiDou short message communication, receiving prediction effect feedback data from terminal devices in real time; evaluating the prediction confidence of the deep learning large model based on the feedback data, and automatically initiating the online fine-tuning process of the model when the prediction error exceeds a preset threshold, and pushing the final rainfall prediction result containing the accuracy level identifier to authorized users through the BeiDou system.

[0077] Establish a prediction result feedback mechanism based on BeiDou short message service to collect feedback data from terminal devices (such as disaster prevention and mitigation terminals and agricultural sensors) in real time, dynamically evaluate the model confidence level, trigger online fine-tuning, and push results with accuracy levels to form a "prediction-feedback-optimization" closed loop.

[0078] BeiDou short message communication: Terminal devices send feedback data (including forecast period, grid point ID, measured rainfall, and device location) through the BeiDou short message module. The data encoding adopts a custom protocol to ensure low power consumption and anti-interference.

[0079] Feedback data processing: Perform spatiotemporal alignment verification on the received data (such as matching device location with grid points), filter invalid feedback (such as timeout data, incorrect format), and store it in the feedback database.

[0080] Confidence Calculation: For each grid point, the conditional probability distribution P(Error|Feature) of the prediction error is calculated based on historical feedback data. When the mean historical error corresponding to the current predicted feature combination (such as terrain complexity, meteorological elements) exceeds a preset threshold (such as 20% relative error), it is marked as a low-confidence area. Online Fine-tuning Trigger: When the proportion of low-confidence grids is ≥15%, online fine-tuning is automatically initiated. The fine-tuning data uses a mixture of feedback data from the past 7 days and the original training data (ratio 1:3) to prevent data shift.

[0081] Accuracy level classification: Based on the historical correction error of grid points, it is divided into three levels: Level 1 (error ≤ 10%): green indicator, high reliability; Level 2 (10% < error ≤ 20%): yellow indicator, manual verification required; Level 3 (error > 20%): red indicator, trend reference only. Push strategy: Prediction results are sent to authorized users via BeiDou short message service or API interface, along with the accuracy level of each grid point. For areas marked with red, ground station measured data are simultaneously pushed as a supplement.

[0082] In some embodiments, the spatiotemporal alignment processing of the spatiotemporal reference data and real-time rainfall observation data to construct a standardized dataset includes: based on the high-precision spatiotemporal reference information provided by the BeiDou Navigation Satellite System, uniformly converting the timestamps of data from different sources into the BeiDou time standard time system; using a sliding window interpolation method to synchronize data sources with inconsistent time resolutions, generating time-series data sequences with equal time intervals; and uniformly transforming the coordinate system used by the positioning data of the BeiDou ground-based augmentation station with the coordinate systems of the ground rainfall monitoring terminal, meteorological radar, and satellite remote sensing data using coordinate transformation parameters provided by the BeiDou system. The data is then projected onto the Gauss-Krüger projection coordinate system of the target area. The target area is divided into grids according to a preset spatial grid resolution. Based on the sub-meter accuracy benchmark of BeiDou positioning data, each type of data is mapped to the corresponding grid node according to its spatial location. For cross-grid data, spatial interpolation is performed using the inverse distance weighting method. Multi-source data is normalized. For numerical data, a standardized scaling method is used, and for categorical meteorological element data, one-heat encoding is used. At the same time, outliers are removed and missing values ​​are filled by spatiotemporal neighborhood data, forming a structured dataset containing time dimension, spatial grid coordinates, and multimodal observation features.

[0083] Based on the BeiDou high-precision spatiotemporal reference, spatiotemporal alignment of multi-source data is achieved, including time synchronization (BeiDou time standard), coordinate unification (Gauss-Krüger projection), gridded interpolation, and data standardization, to construct a structured dataset containing spatiotemporal dimensions.

[0084] The time system is unified by converting the timestamps of all data sources to BeiDou Time (BDT), obtaining real-time BDT time through the BeiDou time synchronization module, and calibrating other time system data such as GPS time and UTC time using publicly available time conversion parameters (such as the deviation file between BDT and UTC).

[0085] Time synchronization processing generates equally spaced (1-minute) time series by using sliding window interpolation (window size 10 minutes, step size 1 minute) for data with inconsistent time resolution (such as radar data 5 minutes / time, rain gauge data 1 minute / time). Missing values ​​are filled by linear interpolation of the preceding and following times.

[0086] Coordinate System 1: The ITRF coordinate system data of the BeiDou ground-based augmentation station is converted into the Gauss-Krüger projection coordinate system of the target area (e.g., 3-degree zone, central meridian 117°E) using the coordinate transformation parameters (seven-parameter method) provided by the BeiDou system. The ground station, radar, and remote sensing data are converted in the same way.

[0087] Grid division: The target area is divided into grids according to a preset resolution (e.g., 100m×100m). The coordinates of the center point of each grid are used as a reference. The distance from each data point to the grid center is calculated for subsequent interpolation.

[0088] Inverse distance weighted interpolation: By interpolating discrete point data (such as rain gauges) to the grid according to the formula zi=∑wjzj, the weight wj=1 / dj^2 (dj is the distance from the data point to the grid center), and directly mapping the cross-grid radar remote sensing gridded data according to coordinates.

[0089] Missing value imputation: For grids with ≥3 valid data points in the spatiotemporal neighborhood, the spatiotemporal KNN algorithm (considering the 3 most recent time moments and the 5 most recent spatial grids) is used to imput missing values; otherwise, the grids are marked as invalid.

[0090] Numerical data: Z-score normalization is used (x′=(x μ) / σ), including rainfall, temperature, humidity, etc.; Categorical data: terrain types (mountains, plains, water) are encoded using one-hot encoding to generate 3D feature vectors; Structured dataset: each sample contains a timestamp (BDT format), grid coordinates (x, y, z), multimodal features (BeiDou positioning parameters, remote sensing inversion values, monitoring data), stored in HDF5 format for easy batch processing.

[0091] In some embodiments, the construction of the deep learning large model includes: adopting a Transformer-based neural network architecture as the basic framework; the spatiotemporal coding module includes a time-dimensional location coding layer and a spatial-dimensional coordinate embedding layer, and maps timestamps, latitude and longitude, and altitude into high-dimensional spatiotemporal embedding vectors through a spatiotemporal joint coding function; the multimodal feature fusion module sets up a cross-modal interaction sublayer, and cross-fuses BeiDou remote sensing data features, ground monitoring data features, and meteorological auxiliary data features through a multi-head self-attention mechanism; the dynamic attention module constructs terrain complexity feature vectors and meteorological condition feature vectors as inputs for attention weight adjustment, and realizes dynamic adjustment of feature weights for different regions based on an adaptive normalization layer.

[0092] We use the Transformer architecture to build a large deep learning model and design a spatiotemporal joint coding module, a cross-modal multi-head self-attention fusion module, and a terrain-meteorological driven dynamic attention module to enhance the modeling capability of the spatiotemporal coupling characteristics of meteorological data.

[0093] The spatiotemporal coding module design includes: Time position coding: Adopting the Transformer sine / cosine position coding, combined with minute-level precision from BeiDou, the coding function is: (dmodel is the embedded dimension, d is the current dimension).

[0094] Spatial coordinate embedding: Gauss-Kruger coordinates (x, y) and altitude z are mapped to a spatial embedding vector S=Linear([x,y,z]) through a linear layer, and a localization accuracy factor HDOP is added as an auxiliary feature (concatenated after normalization). Spatiotemporal joint encoding: The temporal location code T is added to the elements of the spatial embedding S to obtain the spatiotemporal embedding vector Espatiotemporal = T + S.

[0095] Cross-modal inputs utilize BeiDou remote sensing features (such as humidity profiles), ground monitoring features (rainfall, air pressure), and meteorological auxiliary features (wind speed forecasts from numerical models) to extract local features through independent CNNs, and then unify the dimensions through linear layers. Multi-head self-attention fusion employs an 8-head self-attention mechanism to calculate the interactions between different modalities. ; Q, K, and V are features from different modalities, and cross-modal queries enhance feature complementarity.

[0096] Conditional vector generation uses terrain complexity (slope and roughness calculated based on DEM) and meteorological conditions (current wind speed and air pressure gradient) to generate conditional vectors cterrain and cmet through a fully connected layer. After concatenation, the vectors are input into an adaptive normalization layer (AdaNorm).

[0097] Dynamic weight adjustment: AdaNorm adjusts the normalization parameters (mean and variance) of the attention layer based on the conditional vector, using the following formula: ; (γ,β are generated from the conditional vectors via MLP).

[0098] Regional adaptive weighting: For regions with complex terrain (such as mountainous areas), the weight of radar reflectivity features is automatically increased; for plains, the weight of ground monitoring data is enhanced, and regional differential weighting is achieved through terrain masking matrix.

[0099] In some embodiments, to address the issues of cross-regional data privacy protection and computing power balance, a federated learning framework supported by BeiDou short message communication is designed to enable distributed meteorological data collaborative training in different regions (such as provinces / cities). Without sharing the original data, multi-regional model parameters are aggregated to improve the generalization ability of the global model to heterogeneous geographical environments.

[0100] The federated learning architecture includes: Node composition: a central server (National Meteorological Administration), regional nodes (provincial meteorological bureaus), and edge devices (BeiDou ground-based augmentation stations). Parameters are transmitted with encryption via the BeiDou short message channel (supporting asynchronous aggregation in offline environments). Data partitioning: Each regional node partitions data according to "spatiotemporal domain + terrain type" (e.g., mountainous / plain subsets) to ensure that local data distribution has regional specificity. The federated task is to optimize the global rainfall prediction model.

[0101] The collaborative training process includes: Local training: Regional nodes train sub-models based on local data (including grid features after BeiDou coordinate transformation), focusing on updating the parameters of the dynamic attention module and multimodal fusion layer, while fixing the underlying spatiotemporal coding parameters (reducing communication overhead). Parameter aggregation: The central server adopts a weighted aggregation strategy (weight is region area / number of grids), incorporating differential privacy protection (ε=0.5, Laplace noise), with an aggregation interval of 6 hours (matching the numerical pattern update frequency). Cross-domain adaptation mechanism: A domain adaptation layer is added to the central server, aligning the feature distribution of different regions through adversarial training (Domain Adversarial Neural Network) to alleviate the data heterogeneity between the "northern plains and southern hills".

[0102] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of a BeiDou rainfall prediction device 200 based on a deep learning large-scale model provided in an embodiment of this application. The BeiDou rainfall prediction device 200 based on a deep learning large-scale model is used to execute the steps of the BeiDou rainfall prediction method based on a deep learning large-scale model shown in any embodiment of this application. The BeiDou rainfall prediction device 200 based on a deep learning large-scale model can be a single server or a server cluster, or it can be a terminal, such as a handheld terminal, a laptop computer, a wearable device, or a robot.

[0103] like Figure 3 As shown, the BeiDou rainfall prediction device 200 based on a deep learning large model includes: The dataset construction unit 201 is used to acquire spatiotemporal reference data of the target area through the BeiDou satellite navigation system, including high-precision positioning data of BeiDou ground-based augmentation stations and meteorological element inversion data of BeiDou remote sensing payloads; collect real-time rainfall observation data from ground rainfall monitoring terminals, meteorological radars and satellite remote sensing; and perform spatiotemporal alignment processing on the spatiotemporal reference data and real-time rainfall observation data to construct a standardized dataset. Model building unit 202 is used to build a large deep learning model, which includes a spatiotemporal encoding module, a multimodal feature fusion module, and a dynamic attention module. The spatiotemporal encoding module jointly encodes the time series features and spatial location features in the standardized data to generate an embedding vector containing spatiotemporal location information. The multimodal feature fusion module fuses feature representations from BeiDou remote sensing data, ground monitoring data, and meteorological auxiliary data through a cross-modal attention mechanism. The dynamic attention module dynamically adjusts feature weights according to the terrain complexity and meteorological conditions of different regions. The result output unit 203 is used to train the deep learning big model using a spatiotemporal sequence prediction task, input the standardized dataset into the trained deep learning big model, and output the multi-scale rainfall prediction result for the target area for a preset duration in the future.

[0104] In some embodiments, training the deep learning large model using a spatiotemporal sequence prediction task includes: during the training process, introducing real-time orbital parameters of BeiDou satellites as a dynamic adjustment factor to adaptively calibrate the spatiotemporal attention weights of the deep learning large model; the loss function includes temporal prediction error, spatial distribution consistency error, and model complexity regularization term.

[0105] In some embodiments, the method further includes: when an extreme weather event occurs in the target area, automatically triggering an incremental learning mechanism to locally update the model parameters using newly added monitoring data in real time.

[0106] In some embodiments, the multi-scale rainfall prediction result includes a refined prediction grid; after outputting the multi-scale rainfall prediction result for the target area for a preset duration in the future, it further includes: using the wide-area differential positioning accuracy parameters of the BeiDou satellite to perform spatial error calibration on the prediction result, constructing a dynamic correction model in combination with real-time verification data from ground rain gauges, and correcting outliers in the refined prediction grid in real time.

[0107] In some embodiments, the method further includes: establishing a prediction result feedback mechanism based on BeiDou short message communication, receiving prediction effect feedback data from terminal devices in real time; evaluating the prediction confidence of the deep learning large model based on the feedback data, and automatically initiating the online fine-tuning process of the model when the prediction error exceeds a preset threshold, and pushing the final rainfall prediction result containing the accuracy level identifier to authorized users through the BeiDou system.

[0108] In some embodiments, the spatiotemporal alignment processing of the spatiotemporal reference data and real-time rainfall observation data to construct a standardized dataset includes: based on the high-precision spatiotemporal reference information provided by the BeiDou Navigation Satellite System, uniformly converting the timestamps of data from different sources into the BeiDou time standard time system; using a sliding window interpolation method to synchronize data sources with inconsistent time resolutions, generating time-series data sequences with equal time intervals; and uniformly transforming the coordinate system used by the positioning data of the BeiDou ground-based augmentation station with the coordinate systems of the ground rainfall monitoring terminal, meteorological radar, and satellite remote sensing data using coordinate transformation parameters provided by the BeiDou system. The data is then projected onto the Gauss-Krüger projection coordinate system of the target area. The target area is divided into grids according to a preset spatial grid resolution. Based on the sub-meter accuracy benchmark of BeiDou positioning data, each type of data is mapped to the corresponding grid node according to its spatial location. For cross-grid data, spatial interpolation is performed using the inverse distance weighting method. Multi-source data is normalized. For numerical data, a standardized scaling method is used, and for categorical meteorological element data, one-heat encoding is used. At the same time, outliers are removed and missing values ​​are filled by spatiotemporal neighborhood data, forming a structured dataset containing time dimension, spatial grid coordinates, and multimodal observation features.

[0109] In some embodiments, the construction of the deep learning large model includes: adopting a Transformer-based neural network architecture as the basic framework; the spatiotemporal coding module includes a time-dimensional location coding layer and a spatial-dimensional coordinate embedding layer, and maps timestamps, latitude and longitude, and altitude into high-dimensional spatiotemporal embedding vectors through a spatiotemporal joint coding function; the multimodal feature fusion module sets up a cross-modal interaction sublayer, and cross-fuses BeiDou remote sensing data features, ground monitoring data features, and meteorological auxiliary data features through a multi-head self-attention mechanism; the dynamic attention module constructs terrain complexity feature vectors and meteorological condition feature vectors as inputs for attention weight adjustment, and realizes dynamic adjustment of feature weights for different regions based on an adaptive normalization layer.

[0110] It should be noted that, for the sake of convenience and brevity, the specific working process of the Beidou rainfall prediction device 200 based on a deep learning large model and its various modules, as described above, can be found by referring to... Figure 1 The corresponding processes in the embodiments of the BeiDou rainfall prediction method based on deep learning large models described in the corresponding implementation examples will not be repeated here.

[0111] Figure 1 The corresponding BeiDou rainfall prediction method based on a large deep learning model can be implemented as a computer program, which can be used in various ways, such as... Figure 3 It runs on the device shown.

[0112] Please see Figure 4 , Figure 4 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application. The computer device includes a processor, a memory, and a network interface connected via a device bus, wherein the memory may include a storage medium and internal memory.

[0113] The storage medium can store operating devices and computer programs. The computer program includes program instructions that, when executed, cause the processor to perform... Figure 1 This corresponds to any BeiDou rainfall prediction method based on a large deep learning model.

[0114] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0115] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to execute any BeiDou rainfall prediction method based on a large deep learning model.

[0116] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the terminal to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0117] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0118] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: Spatiotemporal reference data for the target area is obtained through the BeiDou satellite navigation system, including high-precision positioning data from BeiDou ground-based augmentation stations and meteorological element inversion data from BeiDou remote sensing payloads; real-time rainfall observation data from ground rainfall monitoring terminals, meteorological radars, and satellite remote sensing are collected; and the spatiotemporal reference data and real-time rainfall observation data are spatiotemporally aligned to construct a standardized dataset. A large-scale deep learning model is constructed, comprising a spatiotemporal encoding module, a multimodal feature fusion module, and a dynamic attention module. The spatiotemporal encoding module jointly encodes time-series features and spatial location features in standardized data to generate an embedding vector containing spatiotemporal location information. The multimodal feature fusion module fuses feature representations from BeiDou remote sensing data, ground monitoring data, and meteorological auxiliary data through a cross-modal attention mechanism. The dynamic attention module dynamically adjusts feature weights based on the terrain complexity and meteorological conditions of different regions. The deep learning model is trained using a spatiotemporal sequence prediction task. The standardized dataset is input into the trained deep learning model to output multi-scale rainfall prediction results for the target area for a predetermined duration in the future.

[0119] In some embodiments, training the deep learning large model using a spatiotemporal sequence prediction task includes: during the training process, introducing real-time orbital parameters of BeiDou satellites as a dynamic adjustment factor to adaptively calibrate the spatiotemporal attention weights of the deep learning large model; the loss function includes temporal prediction error, spatial distribution consistency error, and model complexity regularization term.

[0120] In some embodiments, the method further includes: when an extreme weather event occurs in the target area, automatically triggering an incremental learning mechanism to locally update the model parameters using newly added monitoring data in real time.

[0121] In some embodiments, the multi-scale rainfall prediction result includes a refined prediction grid; after outputting the multi-scale rainfall prediction result for the target area for a preset duration in the future, it further includes: using the wide-area differential positioning accuracy parameters of the BeiDou satellite to perform spatial error calibration on the prediction result, constructing a dynamic correction model in combination with real-time verification data from ground rain gauges, and correcting outliers in the refined prediction grid in real time.

[0122] In some embodiments, the method further includes: establishing a prediction result feedback mechanism based on BeiDou short message communication, receiving prediction effect feedback data from terminal devices in real time; evaluating the prediction confidence of the deep learning large model based on the feedback data, and automatically initiating the online fine-tuning process of the model when the prediction error exceeds a preset threshold, and pushing the final rainfall prediction result containing the accuracy level identifier to authorized users through the BeiDou system.

[0123] In some embodiments, the spatiotemporal alignment processing of the spatiotemporal reference data and real-time rainfall observation data to construct a standardized dataset includes: based on the high-precision spatiotemporal reference information provided by the BeiDou Navigation Satellite System, uniformly converting the timestamps of data from different sources into the BeiDou time standard time system; using a sliding window interpolation method to synchronize data sources with inconsistent time resolutions, generating time-series data sequences with equal time intervals; and uniformly transforming the coordinate system used by the positioning data of the BeiDou ground-based augmentation station with the coordinate systems of the ground rainfall monitoring terminal, meteorological radar, and satellite remote sensing data using coordinate transformation parameters provided by the BeiDou system. The data is then projected onto the Gauss-Krüger projection coordinate system of the target area. The target area is divided into grids according to a preset spatial grid resolution. Based on the sub-meter accuracy benchmark of BeiDou positioning data, each type of data is mapped to the corresponding grid node according to its spatial location. For cross-grid data, spatial interpolation is performed using the inverse distance weighting method. Multi-source data is normalized. For numerical data, a standardized scaling method is used, and for categorical meteorological element data, one-heat encoding is used. At the same time, outliers are removed and missing values ​​are filled by spatiotemporal neighborhood data, forming a structured dataset containing time dimension, spatial grid coordinates, and multimodal observation features.

[0124] In some embodiments, the construction of the deep learning large model includes: adopting a Transformer-based neural network architecture as the basic framework; the spatiotemporal coding module includes a time-dimensional location coding layer and a spatial-dimensional coordinate embedding layer, and maps timestamps, latitude and longitude, and altitude into high-dimensional spatiotemporal embedding vectors through a spatiotemporal joint coding function; the multimodal feature fusion module sets up a cross-modal interaction sublayer, and cross-fuses BeiDou remote sensing data features, ground monitoring data features, and meteorological auxiliary data features through a multi-head self-attention mechanism; the dynamic attention module constructs terrain complexity feature vectors and meteorological condition feature vectors as inputs for attention weight adjustment, and realizes dynamic adjustment of feature weights for different regions based on an adaptive normalization layer.

[0125] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform as described in this application. Figure 1 The steps of the BeiDou rainfall prediction method based on a deep learning large model provided in any corresponding embodiment.

[0126] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0127] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for Beidou rainfall prediction based on a deep learning large model, characterized in that, The method includes: Spatiotemporal reference data of the target area is obtained through the BeiDou satellite navigation device, including high-precision positioning data of BeiDou ground-based augmentation stations and meteorological element inversion data of BeiDou remote sensing payloads; real-time rainfall observation data are collected from ground rainfall monitoring terminals, meteorological radars and satellite remote sensing; the spatiotemporal reference data and real-time rainfall observation data are spatiotemporally aligned to construct a standardized dataset. A large-scale deep learning model is constructed, comprising a spatiotemporal encoding module, a multimodal feature fusion module, and a dynamic attention module. The spatiotemporal encoding module jointly encodes time-series features and spatial location features in standardized data to generate embedding vectors containing spatiotemporal location information. The multimodal feature fusion module fuses feature representations from BeiDou remote sensing data, ground monitoring data, and meteorological auxiliary data through a cross-modal attention mechanism. The dynamic attention module dynamically adjusts feature weights based on the terrain complexity and meteorological conditions of different regions. The construction of the large-scale deep learning model includes: employing a Transformer-based approach... The neural network architecture serves as the basic framework. The spatiotemporal coding module includes a location coding layer in the time dimension and a coordinate embedding layer in the spatial dimension. It maps timestamps, latitude and longitude, and altitude into high-dimensional spatiotemporal embedding vectors through a spatiotemporal joint coding function. The multimodal feature fusion module sets up a cross-modal interaction sublayer, which cross-fuses BeiDou remote sensing data features, ground monitoring data features, and meteorological auxiliary data features through a multi-head self-attention mechanism. The dynamic attention module constructs terrain complexity feature vectors and meteorological condition feature vectors as inputs for attention weight adjustment, and realizes dynamic adjustment of feature weights for different regions based on an adaptive normalization layer. The deep learning model is trained using a spatiotemporal sequence prediction task, including: during training, real-time orbital parameters of BeiDou satellites are introduced as a dynamic adjustment factor to adaptively calibrate the spatiotemporal attention weights of the deep learning model; the loss function of the deep learning model includes temporal prediction error, spatial distribution consistency error, and model complexity regularization term; the standardized dataset is input into the trained deep learning model to output multi-scale rainfall prediction results for the target area for a preset duration; the multi-scale rainfall prediction results include a refined prediction grid. Among them, when extreme weather events occur in the target area, an incremental learning mechanism is automatically triggered to locally update the model parameters using real-time newly added monitoring data; After outputting the multi-scale rainfall prediction results for the target area for a preset time in the future, the method further includes: using the wide-area differential positioning accuracy parameters of the Beidou satellite to perform spatial error calibration on the prediction results, constructing a dynamic correction model by combining real-time verification data from ground rain gauge stations, and correcting outliers in the refined prediction grid in real time. The method further includes: establishing a prediction result feedback mechanism based on BeiDou short message communication to receive prediction effect feedback data from terminal devices in real time; evaluating the prediction confidence of the deep learning large model based on the feedback data; automatically initiating the online fine-tuning process of the model when the prediction error exceeds a preset threshold; and pushing the final rainfall prediction result containing accuracy level identifier to authorized users through the BeiDou device.

2. The method of claim 1, wherein, The process of performing spatiotemporal alignment processing on the spatiotemporal reference data and real-time rainfall observation data to construct a standardized dataset includes: Based on the high-precision spatiotemporal reference information provided by the BeiDou satellite navigation device, the timestamps of data from different sources are uniformly converted into the BeiDou time standard time system. For data sources with inconsistent time resolutions, the sliding window interpolation method is used for time synchronization processing to generate time-series data sequences with equal time intervals. The coordinate system used for positioning data from BeiDou ground-based augmentation stations is uniformly transformed with the coordinate systems of ground rainfall monitoring terminals, meteorological radar, and satellite remote sensing data through coordinate transformation parameters provided by the BeiDou device, and then projected onto the Gauss-Krüger projection coordinate system of the target area. The target area is divided into grids according to the preset spatial grid resolution. Based on the sub-meter accuracy benchmark of BeiDou positioning data, each type of data is mapped to the corresponding grid node according to its spatial location. For cross-grid data, the inverse distance weighting method is used for spatial interpolation filling. The multi-source data is normalized, and a standardization scaling method is used for numerical data and a unique thermal encoding method is used for categorical meteorological element data. At the same time, outliers are removed and missing values ​​are filled in by spatiotemporal neighborhood data, forming a structured dataset that includes time dimension, spatial grid coordinates and multimodal observation features.

3. A BeiDou rainfall prediction device based on a deep learning large model, used to implement the method according to any one of claims 1-2, characterized in that, The device includes: The dataset construction unit is used to acquire spatiotemporal reference data of the target area through the BeiDou satellite navigation system, including high-precision positioning data of BeiDou ground-based augmentation stations and meteorological element inversion data of BeiDou remote sensing payloads; collect real-time rainfall observation data from ground rainfall monitoring terminals, meteorological radars and satellite remote sensing; and perform spatiotemporal alignment processing on the spatiotemporal reference data and real-time rainfall observation data to construct a standardized dataset. The model building unit is used to construct a large deep learning model, which includes a spatiotemporal encoding module, a multimodal feature fusion module, and a dynamic attention module. The spatiotemporal encoding module jointly encodes time-series features and spatial location features in standardized data to generate an embedding vector containing spatiotemporal location information. The multimodal feature fusion module fuses feature representations from BeiDou remote sensing data, ground monitoring data, and meteorological auxiliary data through a cross-modal attention mechanism. The dynamic attention module dynamically adjusts feature weights according to the terrain complexity and meteorological conditions of different regions. The result output unit is used to train the deep learning big model using a spatiotemporal sequence prediction task, input the standardized dataset into the trained deep learning big model, and output the multi-scale rainfall prediction results for the target area for a preset duration in the future.

4. A computer device, characterized in that, The computer device includes a memory and a processor; The memory is used to store computer programs; The processor is configured to execute the computer program and, in executing the computer program, implement the method as described in any one of claims 1 to 2.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to implement the method as described in any one of claims 1 to 2.

Images