A method and system for intelligent extraction of short-term heavy precipitation near-term integrated information from multi-source data in a spatiotemporal manner.
By employing a multi-source data spatiotemporal collaboration and real-time element sliding correction method, combined with time-series deep learning decoding constrained by physical mechanisms such as water vapor flux, the data collaboration and mechanism constraint problems of short-term heavy precipitation forecasting were solved, achieving multi-dimensional refined output and improving the accuracy and reliability of the forecast.
Patent Information
- Application Number
- CN202611531739.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-30
- Publication Date
- 2026-11-10
AI Technical Summary
Existing short-term heavy precipitation nowcasting technologies suffer from a deep disconnect between data collaboration, mechanism constraints, and comprehensive analysis when dealing with high spatiotemporal resolution and severe, sudden convective weather. This makes it difficult to balance spatiotemporal accuracy, meteorological rationality, and decision-making reliability in forecast results. Furthermore, the lack of cross-validation mechanisms in single forecast models makes it easy to miss or false alarms, and the warning dimensions are too simplistic.
A method for intelligent extraction of near-term heavy precipitation information based on spatiotemporal collaboration of multi-source data is constructed. By sliding correction of multi-source spatiotemporal feature system and real-time elements, combined with temporal deep learning decoding constrained by physical mechanisms such as water vapor flux, a dual-chain cross-level mapping mechanism is established to achieve multi-dimensional and refined output of short-term heavy precipitation.
It significantly improves the meteorological rationality of forecast results and the accuracy and robustness of disaster prevention, mitigation and early warning. It can simultaneously output precipitation level, occurrence probability and comprehensive warning level, adapt to the real-time needs of meteorological operations, and reduce the rate of missed reports and false alarms.
Smart Images

Figure CN123069124A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of meteorological remote sensing monitoring technology, specifically to an intelligent extraction method and system for short-term heavy precipitation near-term integrated information based on spatiotemporal coordination of multi-source data. Background Technology
[0002] Existing short-term heavy precipitation nowcasting technologies generally face systemic technical bottlenecks when dealing with high spatiotemporal resolution and sudden convective weather, which are deeply separated from data coordination, mechanism constraints and comprehensive judgment. This makes it difficult for forecast results to take into account spatiotemporal accuracy, meteorological rationality and decision reliability. Specifically, traditional forecasting systems often employ static data overlay and simple extrapolation, lacking a dynamic spatiotemporal coordination mechanism based on sliding corrections of real-world elements. This results in significant lag and data blind spots in capturing the location and temporal abrupt changes of heavy precipitation. Simultaneously, AI forecasting models heavily reliant on pure data lack deep constraints on underlying meteorological and physical rules such as water vapor flux convergence and mass conservation, making them prone to generating distorted element fields that violate atmospheric advection laws when extrapolating complex evolution trends. Furthermore, existing single-prediction models lack a dual-chain cross-validation mechanism that integrates deep learning predictions with environmental physical parameter thresholds. This makes the system extremely intolerant to errors under complex weather conditions, highly susceptible to severe missed and false alarms, and the output warning dimensions are too singular, failing to simultaneously provide multi-dimensional, refined assessment information covering precipitation magnitude, probability of occurrence, and comprehensive warning level. Consequently, they struggle to substantially meet the stringent technical requirements of modern disaster prevention and mitigation operations for high-precision, highly robust short-term warnings.
[0003] Therefore, this invention provides a method and system for intelligent extraction of short-term heavy precipitation proximity comprehensive information based on spatiotemporal collaboration of multi-source data, in order to solve the above-mentioned problems. Summary of the Invention
[0004] To address the systemic shortcomings of existing short-term heavy precipitation nowcasting technologies, such as the lack of real-time dynamic correction in multi-source data fusion leading to spatiotemporal capture delays, the lack of deep constraints from meteorological and physical rules in deep learning models resulting in distorted forecast results, and the lack of cross-validation mechanisms in single forecast models leading to high false alarm rates and limited warning dimensions, this invention aims to provide an intelligent extraction method and system for comprehensive short-term heavy precipitation nowcasting information based on spatiotemporal collaboration of multi-source data. This is achieved by constructing a multi-source spatiotemporal collaborative feature system based on sliding correction of real-time elements, and coupling time-series deep learning decoding constrained by physical mechanisms such as water vapor flux with probability determination based on environmental physical thresholds. A dual-chain cross-level mapping mechanism between "forecast value and occurrence probability" is established, thereby enabling multi-dimensional and refined output of the probability of short-term heavy precipitation occurrence, precipitation magnitude, and comprehensive warning level. This fundamentally improves the meteorological rationality of forecast results and the accuracy and robustness of disaster prevention and mitigation warnings.
[0005] This invention provides the following technical solution: a method for intelligent extraction of short-term heavy precipitation proximity comprehensive information based on spatiotemporal collaboration of multi-source data, the method comprising the following steps: S1: Acquire multi-source meteorological observation data and perform quality control processing and spatiotemporal collaborative registration to construct a standardized multi-source fusion spatiotemporal feature dataset; S2: Based on the real-world elements of the standardized multi-source fusion spatiotemporal feature dataset, the mesoscale numerical model forecast field is verified and biased to generate a three-dimensional element correction field. S3: Input the standardized multi-source fusion spatiotemporal feature dataset and the three-dimensional element correction field into the pre-trained intelligent fusion model for feature decoding, and output the grid-by-grid short-time heavy precipitation forecast value; the intelligent fusion model is a temporal deep learning model that integrates meteorological and physical consistency constraints and spatiotemporal attention mechanisms; S4: Based on the three-dimensional element correction field, extract environmental physical parameters point by point, and obtain the probability of short-term heavy precipitation by threshold judgment and weighted calculation. S5: Perform cross-level matching by combining the short-term heavy precipitation forecast value and the short-term heavy precipitation occurrence probability, and output short-term heavy precipitation near-term comprehensive early warning information. The early warning information includes precipitation level, occurrence probability and comprehensive early warning level.
[0006] Preferably, the step of acquiring multi-source meteorological observation data and performing quality control processing and spatiotemporal collaborative registration to construct a standardized multi-source fusion spatiotemporal feature dataset specifically includes: accessing heterogeneous meteorological observation data from multiple sources in the air, ground, and space, including radar observation data, ground rain gauge data, wind profiler detection data, meteorological satellite observation data, and atmospheric environmental field data; sequentially performing clutter filtering, outlier removal, and numerical standardization quality control processing on the multi-source meteorological observation data to obtain a uniform quality observation dataset; constructing a global standard spatial grid and setting a sliding temporal correction window; using a spatial interpolation algorithm to map the multi-source observation data to the unified spatial grid; and simultaneously completing the temporal alignment of the multi-source data within the sliding temporal correction window to generate a standardized multi-source fusion spatiotemporal feature dataset with a unified spatiotemporal benchmark.
[0007] Preferably, the intelligent fusion model includes a multi-source feature encoding layer, a spatiotemporal attention weighting layer, and a feature fusion output layer connected in sequence. The multi-source feature encoding layer adopts a recurrent neural network architecture, which encodes the input time series features through a hidden state recursion mechanism to extract the temporal evolution features of the precipitation system. The spatiotemporal attention weighting layer includes serially coupled temporal attention branches and spatial attention branches, which are used to dynamically allocate spatiotemporal weights to the encoded historical time series features to generate spatiotemporal context features for the forecast time, enabling the model to focus on key spatiotemporal regions associated with heavy precipitation. The feature fusion output layer is used to output grid-point short-term heavy precipitation forecast results.
[0008] Preferably, the intelligent fusion model completes the training and optimization of network parameters by minimizing the total loss function; the total loss function is composed of a weighted combination of a data-driven loss term, a physical constraint loss term, and a temporal consistency constraint loss term; the data-driven loss term is used to constrain the deviation between the forecast and actual precipitation values, and a temporal weighting mechanism is adopted to assign higher loss weights to the near forecast period; the physical constraint loss term is used to constrain the consistency between the physical quantities output by the model and meteorological physical laws; the temporal consistency constraint loss term is used to constrain the variation amplitude of physical quantities in adjacent time steps, ensuring the temporal smoothness of the precipitation evolution process.
[0009] Preferably, the physical constraint loss term takes the physical relationship of water vapor flux convergence as the core constraint basis, calculates the predicted value of water vapor flux convergence based on the specific humidity field and wind vector field output by the model, and uses the deviation between the predicted value and the actual value of water vapor flux convergence as the loss constraint term, so that the model output conforms to the physical law of water vapor advection.
[0010] Preferably, the step of verifying and correcting the mesoscale numerical model forecast field based on the real-world elements of the standardized multi-source fusion spatiotemporal feature dataset specifically includes: acquiring historical hourly forecast products of the mesoscale numerical model; using a point-to-point moving average deviation algorithm to calculate the moving average deviation value between the forecast products and the real-world elements at each grid point during the historical period, and constructing a three-dimensional moving deviation field; and correcting the model forecast field for subsequent time periods based on the three-dimensional moving deviation field to generate the three-dimensional element correction field.
[0011] Preferably, the step of obtaining the probability of short-term heavy precipitation through threshold determination and weighted calculation specifically includes: selecting multiple types of water vapor, thermal, and dynamic environmental physical parameters that are strongly correlated with the occurrence of short-term heavy precipitation, and setting corresponding determination thresholds and weight coefficients for each type of physical parameter; determining whether each physical parameter meets the corresponding threshold conditions point by point, taking the valid determination value when it meets the conditions, and taking the invalid determination value when it does not meet the conditions; and weighting and summing the determination values of each physical parameter with the corresponding weight coefficients to obtain the probability of short-term heavy precipitation occurring point by point.
[0012] Preferably, the comprehensive short-term heavy precipitation forecast value and the short-term heavy precipitation occurrence probability are cross-graded and matched, using a dual-chain comprehensive judgment method of deep learning forecast and physical threshold discrimination. Specifically, this includes: dividing the short-term heavy precipitation forecast value output by deep learning into multiple intensity level intervals, and dividing the short-term heavy precipitation occurrence probability calculated by physical threshold into multiple probability level intervals; generating multiple warning levels through cross-mapping of intensity level intervals and probability level intervals, and outputting the corresponding comprehensive warning level for each grid point.
[0013] A multi-source data spatiotemporal collaborative short-term heavy precipitation near-term integrated information intelligent extraction system includes a multi-source data access and quality control module, which is used to acquire multi-source meteorological observation data and perform quality control processing and spatiotemporal collaborative registration to construct a standardized multi-source fusion spatiotemporal feature dataset; The model forecast correction module is used to perform real-world verification and bias correction on the mesoscale numerical model forecast field based on the real-world elements of the standardized multi-source fusion spatiotemporal feature dataset, and generate a three-dimensional element correction field. The deep learning inference module is used to input the standardized multi-source fusion spatiotemporal feature dataset and the three-dimensional feature correction field into the pre-trained intelligent fusion model for feature decoding, and output grid-by-grid short-time heavy precipitation forecast values; the intelligent fusion model is a temporal deep learning model that integrates meteorological and physical consistency constraints and spatiotemporal attention mechanisms; The physical probability calculation module is used to extract environmental physical parameters point by point based on the three-dimensional element correction field, and obtain the probability of short-term heavy precipitation through threshold judgment and weighted calculation. The dual-chain integrated analysis module is used to perform cross-level matching of the short-term heavy precipitation forecast value and the short-term heavy precipitation occurrence probability, and output short-term heavy precipitation imminent comprehensive early warning information. The early warning information includes precipitation level, occurrence probability and comprehensive early warning level.
[0014] A computer-readable storage medium storing computer program instructions, wherein when executed by a processor, the computer program instructions implement the steps of a method for intelligent extraction of short-term heavy precipitation proximity integrated information.
[0015] The beneficial effects of this invention are as follows: 1. This invention effectively compensates for the inherent defects of single radar observations, such as ground cover obstruction, blind spots in mountainous areas, and spatial discontinuities, by quality control processing, high-precision spatiotemporal registration, and collaborative fusion of multi-source heterogeneous meteorological data from air, ground, and space. It constructs a standardized dataset at the kilometer and minute levels with unified spatiotemporal reference and stable quality. At the same time, combined with real-time three-dimensional sliding bias correction technology, it corrects the systematic errors of mesoscale numerical models in real time, significantly improving the accuracy of the model initial field and providing a high-quality data foundation for the accurate identification and forecasting of short-term heavy precipitation.
[0016] 2. This invention constructs a temporal deep learning model embedded with a spatiotemporal dual attention mechanism. It captures the continuous evolution of precipitation systems through temporal attention and dynamically focuses on the core areas of heavy precipitation through spatial attention, thus achieving refined joint extraction of spatiotemporal features. At the same time, it introduces physical constraints on water vapor flux convergence and temporal smoothing constraints in model training, which fundamentally suppresses the distorted results that violate meteorological laws, such as false precipitation and abrupt changes in physical quantities, which are prone to occur in pure data-driven models. This ensures that the forecast results have both the precision of deep learning and conform to the atmospheric physical evolution mechanism, significantly improving the operational usability.
[0017] 3. This invention employs a dual-chain comprehensive judgment method combining "deep learning quantitative forecasting + physical threshold probability discrimination." This method leverages the advantages of deep learning models in providing refined quantitative forecasts of precipitation intensity and location, while also verifying the probability of heavy precipitation from a causal perspective through threshold checks of 13 environmental physical parameters. This two-way cross-validation effectively avoids the technical shortcomings of single methods and simultaneously reduces the false alarm and missed alarm rates. Ultimately, it can simultaneously output three core types of information: precipitation level, probability of occurrence, and comprehensive warning level, providing a more comprehensive warning dimension and offering all-round data support for disaster prevention and mitigation decision-making.
[0018] 4. The overall algorithm architecture of this invention is lightweight and computationally efficient, enabling minute-level rolling updates and fully adapting to the real-time requirements of 5-minute rolling early warnings in meteorological operations. The solution is compatible with various disaster-causing precipitation scenarios such as small- and medium-scale convective rainstorms, Meiyu front rainstorms, and typhoon rainstorms. It can be widely applied to multiple business areas such as short-term meteorological early warning, urban flood risk prediction, prevention and control of mountain torrents and geological disasters, watershed hydrological flood control scheduling, and transportation meteorological support. It has low barriers to engineering implementation and strong business adaptability. Attached Figure Description
[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0020] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system flow of the present invention; Figure 3 This is a map showing the actual precipitation from 2:00 AM to 3:00 AM on June 7, 2025. Figure 4 Precipitation forecast map for June 7, 2025, from 2:00 AM to 3:00 AM, provided by the State. Figure 5This is a provincial-level guidance map for precipitation forecast from 2:00 AM to 3:00 AM on June 7, 2025. Figure 6 This is the precipitation forecast map for 2:00-3:00 on June 7, 2025, as presented in this invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific application scenarios and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0022] Example 1: As Figures 1 to 2 As shown in the figure, this embodiment provides an intelligent extraction method for short-term heavy precipitation proximity comprehensive information based on spatiotemporal collaboration of multi-source data. The method includes the following detailed steps: Step S1: Acquire multi-source meteorological observation data and perform quality control processing and spatiotemporal collaborative registration to construct a standardized multi-source fusion spatiotemporal feature dataset; 1. Multi-source data access and quality control; Real-time synchronous acquisition of multi-source heterogeneous meteorological observation data from air, space, and ground, specifically including: (1) Dual polarization radar observation data: including horizontal reflectivity, differential reflectivity, differential propagation phase shift rate, correlation coefficient and hydrogel classification products; (2) Ground rain gauge data: including minute-by-minute actual precipitation intensity observation records; (3) Meteorological satellite observation data: including infrared brightness temperature, cloud top height and microwave liquid water content; (4) Atmospheric environmental field detection data: including three-dimensional wind field detected by wind profile detection, GPS water vapor inversion data and ground temperature, pressure and humidity observation elements.
[0023] For subsequent water vapor flux calculations, the environmental wind field extracted in this step specifically refers to the horizontal wind field. in The wind speed is in the zonal direction. This refers to the meridional wind speed.
[0024] All access data undergoes pre-processing such as clutter filtering, outlier removal, and outlier removal to eliminate invalid and abnormal samples. Subsequently, the Z-score normalization algorithm is used to eliminate the inherent dimensional differences of various element data and map them to a dimensionless unified numerical range to obtain a uniform observation dataset.
[0025] 2. Temporal gap completion and spatiotemporal co-registration; A global standard Cartesian spatial grid was established with a spatial resolution of 1 km × 1 km. This resolution is suitable for the fine-grained detection accuracy of conventional X-band dual-polarization radar and is sufficient to characterize the local features of small- to medium-scale convective cloud clusters. To address the spatial heterogeneity of multi-source data, a classification interpolation strategy was adopted: nearest-neighbor interpolation was performed on discrete station data, while bilinear interpolation was performed on surface-source grid data, uniformly mapping all observation data to the standard spatial grid.
[0026] The sliding time series correction window is set to the past hour, with a time step of 6 minutes. This step size is consistent with the conventional volume scan cycle of the radar, ensuring the continuity of the evolution of precipitation system characteristics. To address the issue of missing time series data due to inconsistent time resolutions from different data sources, the following completion rules are adopted: for minor discontinuities with a missing duration of less than or equal to one step, linear interpolation is used for completion; for data with a missing duration greater than one step, a weighted moving average of the preceding and following valid time periods is used for completion.
[0027] The above processing generates a standardized multi-source fusion spatiotemporal feature dataset that is spatially regular and temporally aligned.
[0028] Step S2: Based on the real-world elements of the standardized multi-source fusion spatiotemporal feature dataset, perform real-world verification and bias correction on the mesoscale numerical model forecast field to generate a three-dimensional element correction field. 1. Construction of three-dimensional deviation field; Historical hourly forecast products from mesoscale numerical models are obtained, and the forecast bias of the past 3 hours is calculated using the standardized multi-source fusion spatiotemporal feature dataset output in step S1 as the ground truth. This window length can reflect the recent systematic error trend of the model and avoid historical noise interference from earlier weather systems.
[0029] For ground features and atmospheric features at different altitudes (such as temperature, humidity, and wind fields at isobaric surfaces of 850 hPa, 700 hPa, and 500 hPa), a three-dimensional point-to-point moving average deviation algorithm is employed. The actual features are assumed to be located at three-dimensional spatial grid points. The value at time t (where z is the vertical height layer or isobaric surface layer) is The model forecast value is The three-dimensional sliding deviation within the time window The calculation model is as follows: ; In the formula, K represents the total number of valid time samples within the past 3-hour time window. The first within the time window Each observation time.
[0030] 2. Layered correction and rolling updates; Deviation field Instead of being a static value, it is updated every 6 minutes as real-time data is accessed, with each calculation using the latest sample from the past 3 hours to dynamically correct systematic biases in the model. The calculated three-dimensional slip bias field is used as a dynamic compensation to perform layer-by-layer real-time corrections on the model forecast field for the next 0-2 hours, generating a three-dimensional element correction field that eliminates systematic errors, effectively reducing forecast biases introduced by underlying surface and local microclimate differences.
[0031] Step S3: Input the three-dimensional feature correction field into the pre-trained intelligent fusion model for feature decoding, and output the grid-by-grid short-time heavy precipitation forecast value. The intelligent fusion model is a temporal deep learning model that integrates meteorological and physical consistency constraints with spatiotemporal attention mechanisms. Its network structure includes a multi-source feature encoding layer, a spatiotemporal dual attention weighting layer, and a feature fusion output layer.
[0032] 1. Model input / output tensor logic; Input Construction: The model uses the multi-source spatiotemporal features from the past hour (10 time steps) as the initial sequence input to the encoder. Simultaneously, the 3D feature correction field generated in step S2 for the next 0-2 hours is used as the boundary condition, embedded along the channel dimension at the corresponding time step of the decoder. The two are then concatenated to form the model's input tensor. The input features consist of 15 meteorological channels, covering core meteorological elements such as 3D wind field, specific humidity, temperature, and radar reflectivity. Output format: The model outputs the grid-by-grid precipitation forecast for the next 2 hours with a time step of 6 minutes (a total of 20 future time steps) in a single forward inference, abbreviated as DL-SHRN, and the unit is mm / h.
[0033] 2. Intelligent fusion model architecture coupled with dual attention; This embodiment adopts a serially coupled spatiotemporal dual attention architecture, first performing temporal attention weighting, and then performing spatial attention weighting.
[0034] Multi-source feature encoding layer: Employing a Long Short-Term Memory (LSTM) network architecture, temporal encoding is performed independently for each spatial grid point. For the input single-grid time series features... Encoding is performed using a hidden state recursion mechanism: ; The hidden state of the coding layer simultaneously generates predicted values for three types of physical quantities through a nonlinear mapping layer: ; In the formula, For the first The hidden state at any given moment. For the nonlinear transformation function of the LSTM unit; For the model in the first The single-grid input feature vector at time t; For the first The hidden state at any given moment. For the predicted precipitation, For the predicted specific humidity, For the predicted wind vector; This is the nonlinear mapping function for the fully connected output layer. The temporal evolution and advection characteristics of the precipitation system are extracted through temporal recursion of the hidden states. The encoded hidden states of all grid points together constitute a temporal feature map with spatial dimension.
[0035] Spatiotemporal dual attention weighted layer: Temporal attention weighting: In the encoder-decoder architecture of an RNN, when the decoder predicts future time steps, it first processes all the historical hidden states of the encoder. Attention weighting is applied to generate a temporal context vector. : ; Among them, weight Indicating the prediction of the future When taking the first step, regarding the first step in history The level of attention during the stealth state is calculated as follows: ; In the formula, This represents the hidden state of the decoder at the previous moment; It is a trainable feedforward alignment network used to compute feature matching scores between historical hidden states and the current prediction step; For feedforward aligned network The calculated historical number Alignment score (matching score) between the hidden step state and the current forecast step; The summation index for traversing all historical time steps.
[0036] The temporal context vectors of all spatial grid points together constitute the temporal context feature map. .
[0037] Spatial attention weighting: After acquiring the temporally weighted features H, spatial attention calculations are performed sequentially. For each time step's input features (such as radar or satellite grid features), the spatial attention mechanism calculates the weights for different spatial locations, focusing on the core area of heavy precipitation. Definition A trainable 1×1 convolutional kernel is used to compress and map C-channel features into a single-channel spatial weight map; the spatial weights are obtained through Softmax normalization. Weighted summation yields spatially augmented features: ; In the formula, For the first Spatial enhancement features after spatial attention weighting at each moment; For the first Time of the first Attention weights for each spatial location; For the first Time of the first The input feature vector of each spatial location; H and W are the number of height grid points and width grid points of the spatial grid, respectively, and H×W is the total number of grid points of the spatial grid; i is the summation index for traversing each spatial grid point.
[0038] The formula for calculating spatial weight is: ; In the formula, H is the spatiotemporal feature map obtained after the aforementioned time attention weighting; Let (i,j) be the feature vector of feature map H at spatial location (i,j), where (i,j) is the row-major sequence in the previous formula. A spatial grid point; It is a trainable 1×1 convolutional kernel used to linearly map feature vectors to attention scores; (·) represents the exponential function; (x,y) in the denominator is the summation index for all positions across the entire spatial grid. For the first Time of the first The attention weights of each spatial grid point after Softmax normalization are between 0 and 1, and the sum of the weights of each spatial grid point is 1.
[0039] Finally, the spatially weighted features are fused with the original features pixel by pixel to guide the model to accurately focus on the core physical region of heavy precipitation and avoid computational interference from background areas without precipitation.
[0040] Feature fusion output layer: Nonlinear inverse mapping is performed on the features that have undergone spatiotemporal weighting to output short-term heavy precipitation forecast values for each grid point.
[0041] 3. The entire model training process and loss function; Gridded label definition: Based on real-time observations from ground rain gauges, spatial objective analysis and interpolation are performed using radar quantitative precipitation estimation (QPE) products to generate a gridded matrix of future 0-2 hour precipitation levels, which serves as a supervisory label for model training.
[0042] Hyperparameter configuration: The Adaptive Moment Estimator (Adam) optimizer is used, with the initial learning rate set to [value missing]. Batch size is set to 16; maximum number of iterations is set to 200. An early stopping strategy is enabled: training terminates when the validation set loss does not decrease for 15 consecutive iterations to prevent overfitting.
[0043] Total Loss Function: The model optimizes its parameters by minimizing the total loss function. The total loss is a weighted combination of data-driven loss, physical constraint loss, and temporal consistency loss. ; in, The physical constraint equilibrium coefficient, This is the timing smoothness constraint coefficient; in this embodiment Take 0.3, We set it to 0.1 to ensure that while pursuing regression accuracy, we also take into account the laws governing meteorological and physical evolution.
[0044] Data-driven loss (time-weighted MSE): ; In the formula, T is the number of forecast time steps. In this embodiment, the forecast is for the next 2 hours, with a step size of 6 minutes, so T=20; t is the summation index of the forecast time steps; N is the total number of spatial grid points involved in the calculation; i is the summation index of each spatial grid point. For the model in the first The forecast time, the first Precipitation forecast values output by each spatial grid point; The actual precipitation observation values (i.e., training labels) are for the corresponding time and grid point. As a time decay weight, a higher loss weight is assigned to the forecast time that is closer to the forecast time.
[0045] Physical constraint loss (temporal MFC consistency): Water vapor flux convergence is calculated using the physical quantities output by the model and compared with its corresponding actual values. The model output is constrained to conform to the physical laws of water vapor advection. Horizontal wind field. The partial derivatives of both the specific humidity q and the partial derivatives of the water vapor flux convergence forecast (MFC) at time t are calculated using the central difference method with a discrete grid. ; In the formula: These are the specific humidity gradients in the latitudinal and meridional directions, respectively. These are the divergence components of zonal and meridional winds, respectively; all partial derivatives are calculated using the central difference method with discrete grid points. The first At any given time, the zonal and meridional wind speeds; For the first The relative humidity at any given time; These are the horizontal latitudinal coordinates and the longitudinal coordinates, respectively.
[0046] The formula for calculating physical constraint loss is: ; In the formula, The meaning is the same as that in the aforementioned data-driven loss; The first is calculated based on the predicted wet field and wind field output by the model. The forecast time, the first Predicted values of water vapor flux convergence at each spatial grid point; The actual value of water vapor flux convergence at the corresponding time and grid point is calculated based on real-time observation data; ||·||2 represents the L2 norm (Euclidean norm) of the vector, which is used to measure the deviation between the predicted and actual values of water vapor flux convergence at the same grid point.
[0047] This constraint mechanism effectively suppresses distorted forecast results from deep learning models that violate atmospheric water vapor transport patterns.
[0048] Temporal consistency loss: By penalizing drastic changes in physical quantities at adjacent time steps, temporal consistency is enhanced, ensuring the continuity and smoothness of precipitation system evolution. ; In the formula, the constraint object This is a vector of physical quantities consisting of precipitation and water vapor flux convergence at time t, specifically including precipitation. Water vapor flux convergence ;‖·‖² represents the square of the L2 norm of the vector; T is the forecast time step, and T-1 is the logarithm of adjacent time steps; by penalizing drastic abrupt changes in physical quantities in adjacent time steps, the temporal continuity of the precipitation system evolution is enhanced, which conforms to the laws of meteorological physical evolution.
[0049] Step S4: Extract environmental physical parameters grid-by-grid based on the three-dimensional element correction field, and obtain the probability of short-term heavy precipitation by threshold determination and weighted calculation; The physical parameter thresholds and weighting coefficients in this embodiment are calibrated for the characteristics of short-duration heavy precipitation in the Jianghuai region. In practical applications, other regions can be adaptively adjusted based on local climate statistics. Thirteen types of water vapor, thermal, and dynamic environmental physical parameters strongly correlated with the occurrence of short-duration heavy precipitation are selected, and corresponding judgment thresholds and weighting coefficients are set for each type of parameter. Each physical parameter is judged at each grid point to determine whether it meets the corresponding threshold condition. If it meets the condition, the judgment value is set to 1; otherwise, it is set to 0. The judgment values of each parameter are weighted and summed with the corresponding weighting coefficients to obtain the short-duration heavy precipitation occurrence probability (SHRP) at each grid point. The 13 types of physical parameters and their corresponding thresholds and weights are as follows: 1.850 hPa specific humidity (q850): threshold ≥14 g / kg, weighting coefficient 0.3; 2. Total atmospheric water vapor content (PWV): Threshold ≥ 50 mm, weighting coefficient 0.23; 3. K-index (K): Threshold ≥ 32℃, weighting coefficient 0.13; 4.850 hPa pseudo-equivalent potential temperature ( Threshold ≥ 345K, weighting coefficient 0.05; 5. Convective Available Potential Energy (BCAPE): Threshold ≥ 240 J / kg, weighting coefficient 0.05; 6.925 hPa divergence (D925): threshold ≤ 1.0 × 10 -5 / s, weighting coefficient 0.02; 7. Ground divergence (Ds): Threshold ≤ -1.0 × 10 -5 / s, weighting factor 0.04; 8. Temperature difference (ΔT) between T850 and T500: Threshold ≥ 22℃, weighting coefficient 0.03; 9.850 hPa temperature (T850): threshold ≥17℃, weighting coefficient 0.03; 10.700 hPa meridional wind (F700): threshold ≥ 1.5 m / s, weighting coefficient 0.03; Vertical velocity at 11.700 hPa (W700): threshold ≤ -0.1 Pa / s, weighting coefficient 0.04; Vertical wind shear of 12.0-3 km (SHR3): threshold ≥7 m / s, weighting coefficient 0.03; 13. Optimal Lift Index (BLI): Threshold ≤ -1℃, weighting coefficient 0.02.
[0050] The weighted calculation model for the probability of short-duration heavy rainfall is as follows: ; In the formula, For the first The weighting coefficients of the physical parameters are calculated, and the sum of the weighting coefficients of the 13 parameters is 1. The threshold value for the corresponding parameter is 1 if the threshold is met, and 0 if the threshold is not met. The output probability value ranges from [0,1].
[0051] Step S5: Perform cross-level matching by combining the short-term heavy precipitation forecast values and the occurrence probability to output short-term heavy precipitation near-term comprehensive early warning information; To compensate for the potential false alarms that may occur with simple deep learning extrapolation and the inability to quantify physical threshold determination, this step adopts a dual-chain comprehensive judgment method combining deep learning prediction and physical threshold discrimination, and executes bidirectional cross-validation logic.
[0052] A unified interval representation rule is used to discretize the precipitation forecast values (DL-SHRN) output by deep learning into four intensity intervals: , The above intensity grading thresholds are based on the Jiangsu Provincial meteorological operational standard of ≥20mm of 1-hour precipitation for short-duration heavy rainfall, combined with the statistical results of the probability distribution of historical short-duration heavy rainfall samples in the Jianghuai region, and matched with the five-level meteorological disaster warning grading system. 20mm is the entry-level threshold for short-duration heavy rainfall, 50mm corresponds to the boundary of rainfall intensity with relatively strong disaster risk, and 80mm corresponds to the boundary of rainfall intensity for extreme short-duration heavy rainfall. The probability grading thresholds are based on the statistical law of the probability of combination hits of 13 environmental physical parameters, forming a cross-judgment matrix that matches the dimensionality of the intensity grading. These grading thresholds are the preferred implementation parameters for the Jianghuai region, and can be adjusted adaptively according to the climate characteristics and operational warning standards of different regions in practical applications. The occurrence probability (SHRP) obtained from physical discrimination is divided into four probability intervals: All critical values are assigned to the next higher level interval.
[0053] Based on the following 16 cross-mapping judgment rules, the corresponding comprehensive warning level is output for each grid point. The levels are divided into five levels from high to low: extremely strong (red), strong (orange), relatively strong (yellow), moderate (blue), and weak (green). when Time: If Determined to be exceptionally strong (red); if Determined as strong (orange); if Determined as relatively strong (yellow); if Classified as normal (blue).
[0054] when Time: If Determined as strong (orange); if Determined as relatively strong (yellow); if Determined as normal (blue); if It is classified as weak (green).
[0055] when Time: If Determined as strong (orange); if Determined as relatively strong (yellow); if Determined as normal (blue); if It is classified as weak (green).
[0056] when Time: If Determined as relatively strong (yellow); if Determined as normal (blue); if Determined as weak (green); if It is classified as weak (green).
[0057] The final output of the short-term heavy rainfall near-term comprehensive early warning information includes three core types of information: the precipitation level at each grid point, the probability of short-term heavy rainfall, and the comprehensive early warning level. It can directly support disaster prevention and mitigation business applications such as short-term meteorological near-term early warning, urban waterlogging risk prediction, and prevention and control of mountain torrents and geological disasters.
[0058] System application effects and verification results; To quantitatively evaluate the technical effectiveness of this method, a back-calculation test was conducted based on the entire spatial scope of Jiangsu Province, and the TS score was used as a unified evaluation index for precipitation forecast effectiveness.
[0059] The determination of heavy precipitation adopts the local standard of Jiangsu Province: precipitation of ≥20mm in 1 hour is determined to be short-term heavy precipitation.
[0060] The formula for calculating the TS score for precipitation forecast is: ; In the formula, NA represents the number of grid points with correct forecasts, i.e., the number of samples where both the actual and forecasted precipitation reached the heavy precipitation threshold; NB represents the number of grid points with false alarms, i.e., the number of samples where the forecast reached the threshold but the actual precipitation did not; and NC represents the number of grid points with missed alarms, i.e., the number of samples where the actual precipitation reached the threshold but the forecast did not. Special convention: When When TS is missing, it is recorded as a missing value.
[0061] 1. Overall inspection results during the flood season; To evaluate the technical effectiveness of this method, a real-time back-calculation test was conducted across the entire Jiangsu Province, from June 1 to July 31, 2025. The test period included hourly rolling calculations for over 1400 time periods. The legally mandated 1-hour heavy precipitation threshold of 20 mm in Jiangsu Province was used as the unified criterion, and the forecast performance index (TS score) was employed. The average TS scores for different forecast products are as follows: Nationally guided forecast products: average TS score is 0.2300; Provincial-level guidance and forecast products: average TS score is 0.3724; The multi-source data spatiotemporal collaborative method of this invention has an average TS score of 0.3794.
[0062] 2. Typical process inspection results Taking a typical short-duration heavy rainfall event at 02:00 on June 7, 2025 as an example, the TS scores of various products are compared as follows: Nationally guided forecast product: TS score is 0.3095; Provincial-level guidance and forecast product: TS score is 0.2308; The spatiotemporal collaborative method for multi-source data in this invention has a TS score of 0.3548.
[0063] The test data shows that during the test period, the average TS score of the national guidance forecast products was 0.2300, and that of the provincial guidance products was 0.3724, while the TS score using the spatiotemporal collaborative and dual-chain intelligent extraction method of this invention reached 0.3794. Taking the heavy precipitation process at 02:00 on June 7, 2025 as an example, the TS score of the national guidance products was 0.3095, and that of the provincial guidance products was 0.2308. The forecast TS score of this method was improved to 0.3548. The accuracy of identifying the core convective area and the intensity extrapolation are significantly better than traditional operational products, effectively reducing the missed and false alarm rates in operational applications.
[0064] The comparison of quantitative TS scores has verified the forecast accuracy advantage of this method. This is further supported by the spatial distribution maps of the actual and predicted events during this process (corresponding to the appendix of this invention). Figure 3 To be continued Figure 6 This further demonstrates the invention's ability to precisely capture localized, short-duration heavy rainfall: Appendix Figure 3 This is a map showing the actual hourly precipitation distribution in Jiangsu Province from 2:00 AM to 3:00 AM on June 7, 2025. This event was a typical small-to-medium scale localized convective heavy precipitation event. The heavy precipitation belt was distributed in a narrow east-west direction, with the core heavy precipitation area located along the line from Yangzhou to Taizhou. The hourly rainfall at the core stations exceeded 20 mm. The rainfall intensity in the surrounding areas decreased rapidly, exhibiting significant local and sudden characteristics. This type of severe convective weather is difficult to accurately depict using conventional numerical weather prediction, and it is also a typical scenario where existing forecasting techniques generally suffer from underreporting and underestimation of intensity.
[0065] Appendix Figure 4 This product is a nationally guided precipitation forecast product for the same period. Due to limitations in model resolution and physical parameterization scheme, the overall scale of the precipitation system forecast by this product is too large, the area of heavy precipitation is significantly shifted westward, the precipitation intensity in the convective core area is significantly underestimated, and the local heavy precipitation center along the Yangzhou-Taizhou line is not identified. This fully demonstrates the inherent shortcomings of large-scale numerical models in the fine-grained forecasting of small and medium-scale convection.
[0066] Appendix Figure 5 This is a provincial-level guidance precipitation forecast product for the same period. Compared with the national guidance product, its spatial resolution has been improved, and its depiction of the overall trend of precipitation belts is closer to reality. However, it still has problems such as insufficient intensity of heavy precipitation centers and a small core area. The intensity and location of the convective core area are also deviated. The fundamental reason is that traditional model forecasts lack the dynamic constraints of real-time multi-source observations, and cannot quickly correct the errors in the formation and dissipation of convective systems.
[0067] Appendix Figure 6 The precipitation forecast results output using the multi-source data spatiotemporal collaborative method of this invention show a comprehensive improvement in forecast accuracy compared to traditional products: The predicted strong precipitation band and its core location closely match the actual situation, accurately capturing the narrow band of strong precipitation along the Yangzhou-Taizhou line without exhibiting the common problem of precipitation area offset in traditional products. Furthermore, the predicted intensity of the strong precipitation center is largely consistent with the actual situation, accurately identifying the core area of short-duration heavy precipitation exceeding 20 mm / h, effectively improving the common underestimation of heavy precipitation intensity by traditional methods. The depiction of precipitation gradients and convective cell morphology is also closer to reality, capable of distinguishing the fine structure of small- and medium-scale convection, meeting the needs for refined forecasts at the kilometer and minute levels.
[0068] The above comparison of precipitation areas corroborates the quantitative TS scoring results, and the improved accuracy essentially stems from the multi-dimensional technological innovations of this invention: the multi-source data spatiotemporal collaborative registration mechanism integrates observation data from multiple sources such as radar, satellites, and ground stations, compensating for the detection blind spots of a single data source; the three-dimensional real-time sliding deviation correction corrects the systematic errors of the numerical model in real time, improving the accuracy of the initial field; the deep learning model embedded with physical constraints retains the data-driven refined feature extraction capability while suppressing false precipitation through physical rules such as water vapor flux convergence; and the final dual-chain cross-judgment mechanism further balances the forecast intensity and the probability of occurrence, simultaneously reducing the missed rate and false alarm rate.
[0069] Example 2: This example provides a multi-source data spatiotemporal collaborative intelligent extraction system for short-term heavy precipitation near-term integrated information, including at least one processor connected in communication and a memory connected in communication with the at least one processor; the memory stores computer program instructions that can be executed by the at least one processor, and when the computer program instructions are read and executed by the at least one processor, the at least one processor performs all the steps of the intelligent extraction method for short-term heavy precipitation near-term integrated information described in Example 1.
[0070] The system is internally divided into functional modules that strictly correspond to the method steps: 1. Multi-source data access and quality control module: Configure a standard data interface array to access multi-source heterogeneous meteorological observation data from air, ground, and space, perform clutter filtering, outlier removal, and Z-score standardization processing, and output observation datasets with uniform quality.
[0071] 2. Spatiotemporal Co-registration Module: Constructs a global standard Cartesian space grid, performs interpolation mapping of missing time series and heterogeneous space grids, and generates a standardized multi-source fusion spatiotemporal feature dataset with unified spatiotemporal reference.
[0072] 3. Model Forecast Correction Module: Calculates the three-dimensional spatial sliding deviation between the mesoscale model forecast field and the actual features, updates the deviation field every 6 minutes, performs hierarchical corrections on the model forecast fields for future time periods, and outputs the three-dimensional feature correction field.
[0073] 4. Deep Learning Inference Module: Loads pre-trained model weights, inputs spatiotemporal features and 3D correction field, performs temporal feature encoding and spatiotemporal dual attention weighted decoding, and outputs grid-by-grid precipitation forecast results; among which, the physical constraint loss is only used for parameter optimization in the offline training stage and does not need to be calculated in the inference stage.
[0074] 5. Physical Probability Calculation Module: Extracts 13 environmental physical parameters from the 3D feature correction field, performs threshold Boolean judgment and weighted summation operation, and outputs the probability of short-term heavy precipitation occurrence at each grid point.
[0075] 6. Dual-chain integrated analysis module: Executes 16-grid matrix mapping logic to cross-classify and match precipitation forecast values and occurrence probabilities, and outputs precipitation level, occurrence probability and color warning level.
[0076] This system achieves lightweight deployment at the architecture level, supports minute-level rolling updates on a kilometer-level grid, and can be well adapted to the business application needs of meteorological disaster prevention and mitigation.
[0077] Example 3: This example also provides a computer-readable storage medium storing computer program code. When the computer program code is executed by a processor, it implements the entire operation process of the intelligent extraction method for short-term heavy precipitation near-term comprehensive information described in Example 1. The computer-readable storage medium includes, but is not limited to, solid-state drives, read-only memory, optical discs, distributed cloud storage nodes, and other media capable of storing computer programs.
[0078] Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A method for intelligent extraction of short-term heavy precipitation near-term comprehensive information based on spatiotemporal collaboration of multi-source data, characterized in that, The method includes the following steps: S1: Acquire multi-source meteorological observation data and perform quality control processing and spatiotemporal collaborative registration to construct a standardized multi-source fusion spatiotemporal feature dataset; S2: Based on the real-world elements of the standardized multi-source fusion spatiotemporal feature dataset, the mesoscale numerical model forecast field is verified and biased to generate a three-dimensional element correction field. S3: Input the standardized multi-source fusion spatiotemporal feature dataset and the three-dimensional element correction field into the pre-trained intelligent fusion model for feature decoding, and output the grid-by-grid short-time heavy precipitation forecast value; the intelligent fusion model is a temporal deep learning model that integrates meteorological and physical consistency constraints and spatiotemporal attention mechanisms; S4: Based on the three-dimensional element correction field, extract environmental physical parameters point by point, and obtain the probability of short-term heavy precipitation by threshold judgment and weighted calculation. S5: Perform cross-level matching by combining the short-term heavy precipitation forecast value and the short-term heavy precipitation occurrence probability, and output short-term heavy precipitation near-term comprehensive early warning information. The early warning information includes precipitation level, occurrence probability and comprehensive early warning level.
2. The intelligent extraction method for short-term heavy precipitation proximity comprehensive information based on spatiotemporal collaboration of multi-source data according to claim 1, characterized in that, The process of acquiring multi-source meteorological observation data and performing quality control processing and spatiotemporal collaborative registration to construct a standardized multi-source fusion spatiotemporal feature dataset specifically includes: accessing heterogeneous meteorological observation data from multiple sources in the air, ground, and space, including radar observation data, ground rain gauge data, wind profiler detection data, meteorological satellite observation data, and atmospheric environmental field data; sequentially performing clutter filtering, outlier removal, and numerical standardization quality control processing on the multi-source meteorological observation data to obtain a uniform quality observation dataset; constructing a global standard spatial grid and setting a sliding temporal correction window; using a spatial interpolation algorithm to map the multi-source observation data to the unified spatial grid; and simultaneously completing the temporal alignment of the multi-source data within the sliding temporal correction window to generate a standardized multi-source fusion spatiotemporal feature dataset with a unified spatiotemporal benchmark.
3. The intelligent extraction method for short-term heavy precipitation proximity comprehensive information based on spatiotemporal collaboration of multi-source data according to claim 1, characterized in that, The intelligent fusion model comprises a multi-source feature encoding layer, a spatiotemporal attention weighting layer, and a feature fusion output layer connected in sequence. The multi-source feature encoding layer adopts a recurrent neural network architecture, which encodes the input time series features through a hidden state recursion mechanism to extract the temporal evolution features of the precipitation system. The spatiotemporal attention weighting layer contains serially coupled temporal attention branches and spatial attention branches, which are used to dynamically allocate spatiotemporal weights to the encoded historical time series features, generate spatiotemporal context features for the forecast time, and enable the model to focus on key spatiotemporal regions associated with heavy precipitation. The feature fusion output layer is used to output the short-term heavy precipitation forecast results for each grid point.
4. The intelligent extraction method for short-term heavy precipitation proximity comprehensive information based on spatiotemporal collaboration of multi-source data according to claim 1, characterized in that, The intelligent fusion model completes the training and optimization of network parameters by minimizing the total loss function. The total loss function is composed of a weighted combination of a data-driven loss term, a physical constraint loss term, and a temporal consistency constraint loss term. The data-driven loss term is used to constrain the deviation between the forecast and actual precipitation values, and a temporal weighting mechanism is used to assign higher loss weights to periods closer to the forecast. The physical constraint loss term is used to constrain the consistency between the physical quantities output by the model and meteorological physical laws. The temporal consistency constraint loss term is used to constrain the variation amplitude of physical quantities in adjacent time steps, ensuring the temporal smoothness of the precipitation evolution process.
5. The intelligent extraction method for short-term heavy precipitation proximity comprehensive information based on spatiotemporal collaboration of multi-source data according to claim 1, characterized in that, The physical constraint loss term takes the physical relationship of water vapor flux convergence as the core constraint basis. It calculates the predicted value of water vapor flux convergence based on the specific humidity field and wind vector field output by the model, and uses the deviation between the predicted value and the actual value of water vapor flux convergence as the loss constraint term, so that the model output conforms to the physical law of water vapor advection.
6. The intelligent extraction method for short-term heavy precipitation proximity comprehensive information based on spatiotemporal collaboration of multi-source data according to claim 1, characterized in that, The actual elements of the standardized multi-source fusion spatiotemporal feature dataset are used to perform actual verification and bias correction on the mesoscale numerical model forecast field. Specifically, this includes: obtaining historical hourly forecast products of the mesoscale numerical model; using a point-to-point moving average bias algorithm to calculate the moving average bias value between the forecast products and actual elements at each grid point during the historical period, and constructing a three-dimensional moving bias field; and correcting the bias of the model forecast field for subsequent time periods based on the three-dimensional moving bias field to generate the three-dimensional element correction field.
7. The intelligent extraction method for short-term heavy precipitation proximity comprehensive information based on spatiotemporal collaboration of multi-source data according to claim 1, characterized in that, The process of obtaining the probability of short-duration heavy precipitation through threshold determination and weighted calculation specifically includes: selecting multiple types of water vapor, thermal, and dynamic environmental physical parameters that are strongly correlated with the occurrence of short-duration heavy precipitation, and setting corresponding determination thresholds and weight coefficients for each type of physical parameter; determining whether each physical parameter meets the corresponding threshold conditions point by point, taking the valid determination value when it meets the conditions, and taking the invalid determination value when it does not meet the conditions; and weighting and summing the determination values of each physical parameter with the corresponding weight coefficients to obtain the probability of short-duration heavy precipitation occurring point by point.
8. The intelligent extraction method for short-term heavy precipitation proximity comprehensive information based on spatiotemporal collaboration of multi-source data according to claim 1, characterized in that, The short-term heavy precipitation forecast value and the short-term heavy precipitation occurrence probability are cross-graded and matched, using a dual-chain comprehensive judgment method of deep learning forecast and physical threshold discrimination. Specifically, this includes: dividing the short-term heavy precipitation forecast value output by deep learning into multiple intensity level intervals, and dividing the short-term heavy precipitation occurrence probability calculated by physical threshold into multiple probability level intervals; through the cross-mapping of intensity level intervals and probability level intervals, multi-level warning levels are generated, and the corresponding comprehensive warning level is output grid-by-grid.
9. A multi-source data spatiotemporal collaborative intelligent extraction system for short-term heavy precipitation near-term integrated information, executing the intelligent extraction method for short-term heavy precipitation near-term integrated information as described in any one of claims 1 to 8, characterized in that, The multi-source data access and quality control module is used to acquire multi-source meteorological observation data and perform quality control processing and spatiotemporal collaborative registration to construct a standardized multi-source fusion spatiotemporal feature dataset. The model forecast correction module is used to perform real-world verification and bias correction on the mesoscale numerical model forecast field based on the real-world elements of the standardized multi-source fusion spatiotemporal feature dataset, and generate a three-dimensional element correction field. The deep learning inference module is used to input the standardized multi-source fusion spatiotemporal feature dataset and the three-dimensional feature correction field into the pre-trained intelligent fusion model for feature decoding, and output grid-by-grid short-time heavy precipitation forecast values; the intelligent fusion model is a temporal deep learning model that integrates meteorological and physical consistency constraints and spatiotemporal attention mechanisms; The physical probability calculation module is used to extract environmental physical parameters point by point based on the three-dimensional element correction field, and obtain the probability of short-term heavy precipitation through threshold judgment and weighted calculation. The dual-chain integrated analysis module is used to perform cross-level matching of the short-term heavy precipitation forecast value and the short-term heavy precipitation occurrence probability, and output short-term heavy precipitation imminent comprehensive early warning information. The early warning information includes precipitation level, occurrence probability and comprehensive early warning level.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the steps of the intelligent extraction method for short-term heavy precipitation proximity integrated information as described in any one of claims 1 to 8.
