Spatial scalability forecasting methods, systems, equipment, and media for three-dimensional cloud fields

Abstract

This invention relates to the field of three-dimensional cloud field forecasting technology, and specifically to a spatially scalable forecasting method, system, device, and medium for three-dimensional cloud fields. The method includes: constructing a first basic training set based on preprocessed first historical radar network data; inputting the first basic training set into a model to obtain an initial forecasting model; inputting various specialized historical data into the initial forecasting model to train specialized identification capabilities, obtaining a specialized enhancement model; constructing a second basic training set based on preprocessed second historical radar network data, and processing the second basic training set based on the specialized enhancement model to select and construct multiple specialized training sets; inputting the second basic training set and the multiple specialized training sets into the specialized enhancement model for multi-resolution joint training to obtain a target forecasting model. This method solves the problems of insufficient spatial adaptability and specialized identification capabilities of three-dimensional cloud field forecasting models under complex weather conditions.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional cloud field forecasting technology, and more specifically, to a spatially scalable forecasting method, system, device, and readable storage medium for three-dimensional cloud fields. Background Technology

[0002] Accurate 3D cloud field forecasting is of epoch-making significance for improving extreme weather early warning capabilities, ensuring the safe operation of the low-altitude economy, and optimizing urban disaster prevention and mitigation systems. As a three-dimensional representation of atmospheric motion and phase transition processes, 3D cloud fields not only determine precipitation distribution and intensity but also directly affect key areas such as aviation safety, drone logistics, and urban flood control. In recent years, the rapid development of deep learning technology has injected momentum into the field of meteorological forecasting: convolutional neural networks extract cloud physical features through multi-level abstraction, recurrent neural networks construct spatiotemporal correlation models, and generative adversarial networks can even simulate the nonlinear evolution of atmospheric motion, making it possible to mine deep cloud field evolution patterns from massive amounts of satellite remote sensing, weather radar, and radiosonde data.

[0003] However, current AI-based nowcasting technologies still face severe constraints in spatiotemporal resolution. Specifically, existing models generally suffer from regional dependence and low accuracy. That is, forecasting models developed for specific regions must rely on high-precision spatiotemporal observation data accumulated over a long period in those regions, leading to a severe deficiency in model transferability. Furthermore, existing models generally suffer from insufficient specific identification capabilities. Therefore, addressing the issues of spatial adaptability and specific identification capabilities of 3D cloud field forecasting models under complex weather conditions has become an urgent problem to solve. Summary of the Invention

[0004] To address the shortcomings of three-dimensional cloud field forecasting models in terms of spatial adaptability and specific identification capabilities under complex weather conditions, this invention provides a spatially scalable forecasting method, system, equipment, and medium for three-dimensional cloud fields.

[0005] In a first aspect, the present invention provides a method for predicting the spatial scalability of three-dimensional cloud fields, comprising:

[0006] The first basic training set is constructed based on the preprocessed first historical radar network data;

[0007] The first basic training set is input into the model to obtain the initial prediction model;

[0008] Various types of historical data are input into the initial forecast model to train the special identification capability, resulting in a special improvement model.

[0009] A second basic training set is constructed based on the preprocessed second historical radar network data, and the second basic training set is processed based on the special improvement model to select and construct multiple special training sets; among them, the data volume of the first historical radar network data is smaller than that of the second historical radar network data.

[0010] The second basic training set and multiple specialized training sets are input into the specialized enhancement model for multi-resolution joint training to obtain the target prediction model.

[0011] In some embodiments, constructing a second basic training set based on preprocessed second historical radar network data includes:

[0012] Acquire second historical radar network data covering the target area;

[0013] The second historical radar network data is preprocessed, including: scaling up the basic spatial resolution of the second historical radar network data to a preset spatial resolution using grid aggregation and averaging algorithms; and resampling the second historical radar network data at the preset spatial resolution onto the target regular latitude and longitude grid to form the second basic training set.

[0014] In some embodiments, processing the second basic training set based on the specialized enhancement model to select and construct multiple specialized training sets includes:

[0015] Input the second basic training set into the specialized improvement model;

[0016] The specialized enhancement model extracted radar echo data sequences from typhoon systems, cold vortex systems, squall line systems, and local severe convective systems;

[0017] Multiple specialized training sets were obtained based on radar echo data sequences from typhoon systems, cold vortex systems, squall line systems, and local severe convective systems. These specialized training sets are the typhoon training set, the cold vortex training set, the squall line training set, and the local severe convective system training set.

[0018] In some embodiments, the typhoon training set includes a first spatial resolution, a first coverage area, and a first continuous time series; the cold vortex training set includes a second spatial resolution, a second coverage area, and a second continuous time series; the squall line training set includes a third spatial resolution, a third coverage area, and a third continuous time series; and the local severe convection training set includes a fourth spatial resolution, a fourth coverage area, and a fourth continuous time series; wherein the first spatial resolution, the second spatial resolution, the third spatial resolution, and the fourth spatial resolution are smaller than a preset spatial resolution.

[0019] In some embodiments, after inputting a second basic training set and multiple specialized training sets into a specialized enhancement model for multi-resolution joint training to obtain a target prediction model, the method further includes:

[0020] Based on the preprocessed second historical radar network data, a multi-scale validation set and a test set were constructed; the multi-scale validation set includes a large-scale validation set for typhoons and cold vortices and a small-scale validation set for squall lines and strong convection.

[0021] The target prediction model was validated based on a multi-scale validation set;

[0022] The target prediction model was tested based on the test set after validation.

[0023] In some embodiments, validating the target prediction model based on a scaled validation set includes:

[0024] Historical sequence data from the large-scale validation set of typhoons and cold vortices were input into the target prediction model for validation, and the first prediction radar echo sequence was obtained.

[0025] Extract the orbital path, intensity characterization, and precipitation distribution at the first and second spatial resolutions from the first forecast radar echo sequence;

[0026] The path, intensity characterization, and precipitation distribution are compared with the first actual observation data to calculate the path error, intensity error, and precipitation score.

[0027] Historical sequence data from squall lines and small-scale strong convection validation sets were input into the target prediction model for validation, resulting in the second prediction radar echo sequence.

[0028] The trigger time, movement vector, strong center position, and fine structure features at the third and fourth spatial resolutions were extracted from the second forecast radar echo sequence.

[0029] The trigger time, movement vector, strong center position, and fine structural features are compared with the second actual observation data to calculate the trigger time error, movement error, position error, and structural overlap index.

[0030] In some embodiments, after inputting a second basic training set and multiple specialized training sets into a specialized enhancement model for multi-resolution joint training to obtain a target prediction model, the method further includes:

[0031] Input the target spatial resolution into the target prediction model and output the radar echo sequence of the target spatial resolution.

[0032] Secondly, the present invention provides a spatial scalability forecasting system for three-dimensional cloud fields. This system is applied to the spatial scalability forecasting method for three-dimensional cloud fields provided in any embodiment of the first aspect. The spatial scalability forecasting method system for three-dimensional cloud fields includes:

[0033] The data acquisition module is used to acquire the first historical radar network data, special historical data, and the second historical radar network data.

[0034] The data processing module is used to construct the first basic training set, the second basic training set, and the specialized training set;

[0035] The model training module is used to train an initial prediction model based on the first basic training set, train a special improvement model based on historical data of multiple special projects, and train a target prediction model based on the second basic training set and multiple special project training sets.

[0036] The model building module is used to build the initial forecast model, the special enhancement model, and the target forecast model.

[0037] Thirdly, the present invention provides a spatial scalability forecasting device for three-dimensional cloud fields, comprising:

[0038] processor;

[0039] Memory, used to store executable instructions;

[0040] The processor is configured to read executable instructions from memory and execute the executable instructions to implement the spatial scalability prediction method for three-dimensional cloud fields according to any embodiment of the first aspect.

[0041] Fourthly, the present invention provides a computer-readable storage medium, comprising:

[0042] The system stores instructions that, when executed by a processor, perform a spatial scalability prediction method for a three-dimensional cloud field according to any embodiment of the first aspect.

[0043] To address the shortcomings of three-dimensional cloud field forecasting models in terms of spatial adaptability and specific identification capabilities under complex weather conditions, this invention offers the following advantages:

[0044] This invention utilizes relatively small amounts of first-generation historical radar network data to rapidly construct an initial forecast model, solving the efficiency problem of model cold start. Then, by introducing various specialized historical data for specialized identification training, the model gains the ability to accurately identify specific weather phenomena. Next, the model, already possessing specialized identification capabilities, automatically filters massive amounts of second-generation historical radar network data to construct a highly targeted and high-quality specialized training set. Finally, through multi-resolution joint training, the final target forecast model achieves significant improvements in spatial scalability, specialized weather forecast accuracy, and consistency across different scales, realizing a leap from basic forecasting to refined and scalable forecasting capabilities. Attached Figure Description

[0045] Figure 1A flowchart illustrating an embodiment of a spatial scalability prediction method for three-dimensional cloud fields is shown.

[0046] Figure 2 A ConvLSTM model structure diagram of a spatial scalability prediction method for three-dimensional cloud fields according to an embodiment is shown;

[0047] Figure 3 A schematic diagram of a specific training set for a spatial scalability prediction method for three-dimensional cloud fields according to an embodiment is shown;

[0048] Figure 4 The experimental data of a spatially scalable cloud field forecasting method according to an embodiment are shown, including an actual observation map of a typhoon and radar images with different forecast lead times.

[0049] Figure 5 The experimental data of a spatially scalable cloud field forecasting method according to an embodiment are shown in radar images of different altitudes of a typhoon.

[0050] Figure 6 Experimental data of a spatially scalable cloud field forecasting method according to one embodiment are shown, including actual hail observation maps and radar maps with different forecast lead times.

[0051] Figure 7 The experimental data of a spatially scalable cloud field forecasting method according to an embodiment are shown: radar images of hail at different altitudes.

[0052] Figure 8 A schematic diagram of an electronic device is shown for an embodiment of a spatial scalability prediction method for three-dimensional cloud fields.

[0053] Figure label:

[0054] Electronic device 100; memory 101; processor 102; input / output (I / O) interface 103. Detailed Implementation

[0055] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, they are provided so that the invention will be thorough and complete, and the concept of the exemplary embodiments will be fully conveyed to those skilled in the art. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0056] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full understanding of embodiments of the invention. However, those skilled in the art will recognize that the technical solutions of the invention can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of the invention.

[0057] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all contents and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be decomposed, while others can be combined or partially combined; therefore, the actual order of execution may change depending on the actual situation. It should be understood that although the terms first, second, third, etc., may be used herein to describe various components, these components should not be limited by these terms. These terms are used to distinguish one component from another. Therefore, the first component discussed below may be referred to as the second component without departing from the teachings of the present invention. As used herein, the term "and / or" includes any and all combinations of any one or more of the associated listed items.

[0058] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of exemplary embodiments, and the modules or processes in the drawings are not necessarily essential for implementing the present invention, and therefore cannot be used to limit the scope of protection of the present invention.

[0059] Accurate 3D cloud field forecasting is of epoch-making significance for improving extreme weather early warning capabilities, ensuring the safe operation of the low-altitude economy, and optimizing urban disaster prevention and mitigation systems. As a three-dimensional representation of atmospheric motion and phase transition processes, 3D cloud fields not only determine precipitation distribution and intensity but also directly affect key areas such as aviation safety, drone logistics, and urban flood control. In recent years, the rapid development of deep learning technology has injected momentum into the field of meteorological forecasting: convolutional neural networks extract cloud physical features through multi-level abstraction, recurrent neural networks construct spatiotemporal correlation models, and generative adversarial networks can even simulate the nonlinear evolution of atmospheric motion, making it possible to mine deep cloud field evolution patterns from massive amounts of satellite remote sensing, weather radar, and radiosonde data.

[0060] However, current AI-based nowcasting technologies still face severe constraints in terms of spatiotemporal resolution. Specifically, existing models generally suffer from the dual limitations of "regional dependence" and "data rigidity": on the one hand, forecasting models developed for specific regions must rely on high-precision spatiotemporal observation data accumulated over a long period in that region, and this strong data dependence leads to a serious lack of model transferability. On the other hand, the model architecture is deeply coupled with the resolution of the input data; when the radar or satellite pixel resolution increases from 1km to 500m, the entire model needs to be retrained or even reconstructed. This rigid constraint not only results in the duplication of research resources but also creates a digital divide in forecasting capabilities between regions. Therefore, building a new generation of intelligent forecasting models with spatial adaptability and resolution flexibility has become an urgent need for the industry.

[0061] Example 1

[0062] Please refer to Figure 1 This embodiment provides a method for predicting the spatial scalability of a three-dimensional cloud field, including the following steps S10-S50;

[0063] S10, Construct the first basic training set based on the preprocessed first historical radar network data;

[0064] S20, Input the first basic training set into the model to obtain the initial prediction model;

[0065] S30: Input various types of historical data for specific purposes into the initial forecast model to train the specific identification capability and obtain the specific improvement model;

[0066] S40, a second basic training set is constructed based on the preprocessed second historical radar network data, and the second basic training set is processed based on the special improvement model to select and construct multiple special training sets; among them, the amount of data of the first historical radar network data is less than the amount of data of the second historical radar network data.

[0067] S50, the second basic training set and multiple specialized training sets are input into the specialized enhancement model for multi-resolution joint training to obtain the target prediction model.

[0068] Please refer to Figure 2 In this embodiment, the initial forecast model, the specialized enhancement model, and the final target forecast model are all constructed and trained using a core architecture based on a Convolutional Long Short-Term Memory (CONV-LSTM) network. This architecture can effectively capture both the spatial features and temporal dependencies of 3D cloud field data. Furthermore, the network employs an encoder-decoder structure, where both the encoder and decoder contain two layers of CONV-LSTM units. Specifically, the convolutional long short-term memory network model includes three gated convolutional layers, namely the input gate I... t Forgotten Gate F tand output gate O t Among them, H t-1 H is the hidden state from the previous time step. t C represents the hidden state at the current time step. t-1 For the cell state at the previous time step, C t X represents the cell state at the current time step. t This represents the input state at the current time step; the specific number depends on the actual application. The number of channels for the model input data is set to (1, 6, 512, 512), corresponding to the batch size, time series length (e.g., 6 consecutive time series), and grid height and width, respectively. During training, the batch size is 32, and model parameters are updated using the backpropagation algorithm and gradient descent optimizer. The total number of training iterations is 1000 to ensure sufficient model convergence. This configuration is beneficial for multi-resolution joint training: the unified 512×512 target rule latitude and longitude grid field allows training samples of different resolutions to be aligned with the input in spatial dimensions, while the temporal processing capability of the convolutional long short-term memory network can learn the cross-scale mapping relationship from historical sequences to future forecasts. By using this fixed architecture, the model is trained in stages using the first basic training set, specialized historical data, and mixed multi-scale data consisting of the second basic training set and multiple specialized training sets. Ultimately, the target forecast model internally encodes cloud field physical evolution knowledge at different scales, from kilometers to hundreds of meters, thereby achieving spatially scalable forecasting capabilities.

[0069] Please refer to Figure 3 Step S10 uses networked radar data from the China Meteorological Administration, with a basic spatial resolution of 1km×1km. In terms of time, the data maintains a high-frequency sampling characteristic of once every 6 minutes, providing training data for the initial establishment of the model's general forecasting capabilities. The preprocessing of the first historical radar network data includes upscaling the data using grid aggregation and averaging algorithms. This transforms the basic spatial resolution of 1km×1km into a uniform grid field with a preset spatial resolution of 8km×8km, and constructs a 512×512 target regular latitude and longitude grid field. The geographical coverage of this target regular latitude and longitude grid field is, for example, from 87°E to 126°E and 17°N to 55°N, depending on the actual application. This geographical coverage, for example, completely encompasses the entire territory of China and surrounding sea areas, thus ensuring the diversity and representativeness of the training samples in terms of climate zones, underlying surface types, and typical weather systems.

[0070] In step S20, the first basic training set is input into the model to complete the basic training phase of the model, resulting in an initial forecast model with basic three-dimensional cloud field forecasting capabilities.

[0071] In step S30, various specialized historical data are input into the initial forecast model. This allows the model to not only learn general weather evolution patterns but also overcome forecasting difficulties in specific and extreme scenarios, thereby enhancing the model's ability to identify and forecast specific important weather systems. In this embodiment, the various specialized historical data include historical typhoon data, historical cold vortex data, historical squall line data, and local severe convection data. Inputting these historical typhoon, cold vortex, squall line, and local severe convection data into the initial forecast model enhances its ability to identify and forecast specific weather systems, building upon its basic forecasting capabilities. By accurately characterizing the three-dimensional structure and intensity of these systems, the model drives internal physical processes (such as energy release, water vapor transport, and unstable energy triggering) to simulate their own development and evolution, and complex interactions between systems (such as the combination of typhoons and cold vortices exacerbating heavy rainfall). This enables refined and quantitative forecasts of the three-dimensional cloud field structure, precipitation type, intensity, and location dominated by these systems, significantly improving the forecasting and early warning capabilities for severe weather.

[0072] In this embodiment, step S40, which involves constructing the second basic training set based on the preprocessed second historical radar network data, includes:

[0073] Acquire second historical radar network data covering the target area;

[0074] The second historical radar network data is preprocessed, including: scaling up the basic spatial resolution of the second historical radar network data to a preset spatial resolution using grid aggregation and averaging algorithms; and resampling the second historical radar network data at the preset spatial resolution onto the target regular latitude and longitude grid to form the second basic training set.

[0075] Understandably, the second historical radar network data also uses network radar data from the China Meteorological Administration. The basic spatial resolution of the second historical radar network data is 1km×1km, the preset spatial resolution is 8km×8km, and the target regular latitude and longitude grid is 512×512. That is, the second historical radar network data with a basic spatial resolution of 1km×1km is upscaled to the preset spatial resolution of 8km×8km through grid aggregation and averaging algorithms. Then, the second historical radar network data with the preset spatial resolution of 8km×8km is resampled onto the target regular latitude and longitude grid of 512×512, depending on the actual application. In the time dimension, the data maintains a high-frequency sampling characteristic of once every 6 minutes, and the data volume of the second historical radar network data is greater than that of the first historical radar network data. The second historical radar network data can be, for example, continuous time series data over many years, while the first historical radar network data can be, for example, a one-year time series data. This provides a foundation for further implementing refined data expansion strategies, enabling the model to select a more comprehensive specialized training set from the second basic training set.

[0076] Understandably, in this embodiment, reducing the data to a preset spatial resolution of 8km can greatly reduce the amount of data and the computational burden on the model. At the same time, since the movement and evolution of large-scale weather systems (typhoons and cold vortices) are mainly controlled by macroscopic dynamic and thermodynamic processes, the preset spatial resolution of 8km can effectively capture their main structure and movement trend, thereby ensuring the effectiveness of the initial model learning.

[0077] Please refer to Figure 3 In this embodiment, the process of processing the second basic training set based on the specialized enhancement model in step S40 to select and construct multiple specialized training sets further includes the following steps:

[0078] Input the second basic training set into the specialized improvement model;

[0079] The specialized enhancement model extracted radar echo data sequences from typhoon systems, cold vortex systems, squall line systems, and local severe convective systems;

[0080] Multiple specialized training sets were obtained based on radar echo data sequences from typhoon systems, cold vortex systems, squall line systems, and local severe convective systems. These specialized training sets are the typhoon training set, the cold vortex training set, the squall line training set, and the local severe convective system training set.

[0081] Understandably, the specialized enhancement model obtained in step S30, such as an intermediate forecast model, can be used to select and construct specialized training sets from a large amount of technical data. Based on these specialized training sets, the model can be further trained to improve its ability to identify and forecast specific important weather systems. Using specialized enhancement models to select and construct multiple specialized training sets can improve the efficiency and objectivity of data selection, avoiding the subjective bias and tedious work of manual selection. Furthermore, it ensures that the constructed specialized training sets accurately include complete and typical lifecycle data for typhoon systems, cold vortex systems, squall line systems, and local severe convective systems, guaranteeing the accuracy of subsequent model forecasts.

[0082] Furthermore, typhoon systems and cold vortex systems are large-scale systems, while squall line systems and local severe convective systems are mesoscale or small-scale systems. Among them, large-scale systems have a horizontal coverage range of hundreds to thousands of kilometers and a lifespan measured in days, while mesoscale and small-scale systems have a horizontal coverage range of several kilometers to hundreds of kilometers and a lifespan measured in hours. This allows the special enhancement model and subsequent target forecast models to learn the potential or mapping relationships that affect precipitation and clouds based on large-scale, mesoscale, or small-scale systems, thereby further improving the accuracy of the target forecast models.

[0083] In this embodiment, the typhoon training set includes a first spatial resolution, a first coverage area, and a first continuous time series, wherein the first spatial resolution is 3km, the first coverage area is 1536×1536km, and the first continuous time series is 36 hours or more; the cold vortex training set includes a second spatial resolution, a second coverage area, and a second continuous time series, wherein the second spatial resolution is 2km, the second coverage area is 1024×1024km, and the second continuous time series is 24 hours or more; the squall line training set includes a third spatial resolution, a third coverage area, and a third continuous time series, wherein the third spatial resolution is 1km, the third coverage area is 51 km, and the third continuous time series is 51 km. The training set for local severe convection includes a fourth spatial resolution, a fourth coverage area, and a fourth continuous time series. The fourth spatial resolution is 500m, the fourth coverage area is 256×256km, and the fourth continuous time series is over 6 hours. The first, second, third, and fourth spatial resolutions are all less than 8km. This multi-resolution architecture allows the model to grasp the overall structure of the typhoon eyewall or the movement trend of the cold vortex cloud system at a macro scale, and to observe the evolution of the outflow boundary of the squall line front or the core structure inside the convective cell at a micro scale.

[0084] Understandably, spatial resolution refers to the actual horizontal spatial distance represented by each grid point in a regular latitude and longitude grid field. The smaller the resolution value, the denser the grid points, and the more details of the weather system can be described. Coverage refers to the size of the actual geographical area corresponding to the training data. Continuous time series refers to a sequence of radar observation data that is continuous and uninterrupted in time for a single training sample.

[0085] Understandably, this embodiment sets training resolutions and coverage areas that best match the physical scale of different types of weather systems. The coverage areas are all within China, encompassing typical regions from south to north and from west to east, ensuring optimal visibility and clarity of the input data when the model learns from different systems. For example, typhoon systems are large and move slowly; therefore, a first spatial resolution of 3km and a first coverage area of ​​1536×1536km for typhoons can fully capture their macroscopic circulation structure while ensuring computational efficiency. Local severe convective systems, on the other hand, are small in scale and develop rapidly; a fourth spatial resolution of 500m and a fourth coverage area of ​​256×256km can accurately characterize their internal fine structure and development core area. Furthermore, based on the stepped first, second, third, and fourth spatial resolutions, the spatial scalability range required for model learning is clearly defined. This means the model can learn how to recover or deduce fine structures up to 500 meters from a relatively coarse 8km base field, thereby improving the forecast accuracy of the target forecast model.

[0086] Meanwhile, a continuous time series of no less than its typical lifespan is set for each system, namely typhoon > 36h, cold vortex > 24h, squall line > 12h, and strong convection > 6h. The specific time series is subject to actual application and no specific limitation is made in this invention. This ensures that the model can learn the complete dynamic evolution process of this type of system from occurrence, development to decay, rather than fragments, thereby improving the consistency and physical rationality of its forecast.

[0087] In this embodiment, after step S50, which involves inputting the second basic training set and multiple specialized training sets into the specialized enhancement model for multi-resolution joint training to obtain the target prediction model, the method further includes:

[0088] Based on the preprocessed second historical radar network data, a multi-scale validation set and a test set were constructed; the multi-scale validation set includes a large-scale validation set for typhoons and cold vortices and a small-scale validation set for squall lines and strong convection.

[0089] The target prediction model was validated based on a multi-scale validation set;

[0090] The target prediction model was tested based on the test set after validation.

[0091] Understandably, the multi-scale validation mechanism implemented in this paper ensures the comprehensiveness, relevance, and impartiality of the target forecast model. By constructing two validation sets—a large-scale validation set for typhoons and cold vortices and a small-scale validation set for squall lines and severe convection—the forecasting capability of the model on different types and scales of weather systems can be independently quantified, avoiding evaluation biases that may occur when using a single mixed validation set.

[0092] In this embodiment, validating the target prediction model based on a scaled validation set includes:

[0093] Historical sequence data from the large-scale validation set of typhoons and cold vortices were input into the target prediction model for validation, and the first prediction radar echo sequence was obtained.

[0094] Extract the orbital path, intensity characterization, and precipitation distribution at the first and second spatial resolutions from the first forecast radar echo sequence;

[0095] The path, intensity characterization, and precipitation distribution are compared with the first actual observation data to calculate the path error, intensity error, and precipitation score.

[0096] Historical sequence data from squall lines and small-scale strong convection validation sets were input into the target prediction model for validation, resulting in the second prediction radar echo sequence.

[0097] The trigger time, movement vector, strong center position, and fine structure features at the third and fourth spatial resolutions were extracted from the second forecast radar echo sequence.

[0098] The trigger time, movement vector, strong center position, and fine structural features are compared with the second actual observation data to calculate the trigger time error, movement error, position error, and structural overlap index.

[0099] Understandably, the evaluation focuses on typhoons and cold vortices, assessing their forecasting capabilities for track, intensity, and precipitation distribution at spatial resolutions of 3km and 2km. For squall lines and severe convection, the evaluation examines their ability to capture system triggering timing, movement direction, and strong center location at spatial resolutions of 1km and 500m. The selected evaluation metrics include track error, triggering time error, and first-order structure overlap, rather than generic machine learning loss functions. These are core performance indicators recognized in meteorological forecasting operations and directly reflecting forecast accuracy. This allows the performance of the target forecasting model to be intuitively understood and directly benchmarked against the operational performance of existing numerical weather prediction models or extrapolation algorithms.

[0100] Furthermore, test results show that while maintaining forecast stability across all scales, the model exhibits particular advantages at resolvable convection scales. Its 500-meter resolution output can clearly reproduce the fine structure of thunderstorm outflow boundaries and gust fronts, a feat difficult to achieve with traditional extrapolation algorithms and single-scale models. This capability essentially stems from the comprehensive coverage of the physical processes of various weather systems from macroscopic to microscopic levels in the training data, as well as the cross-scale mapping relationships established by the model itself during the encoding-decoding process. Therefore, the integrated approach from data strategy to model architecture provides a complete technical path and feasibility verification for developing next-generation adaptive, multi-scale fusion meteorological forecasting models.

[0101] In this embodiment, after inputting the second basic training set and multiple specialized training sets into the specialized enhancement model for multi-resolution joint training to obtain the target prediction model, the method further includes:

[0102] Input the target spatial resolution into the target prediction model and output the radar echo sequence of the target spatial resolution.

[0103] Understandably, based on the target prediction time, multiple continuous time series and historical data of target spatial resolution are obtained and input into the target prediction model, and the radar echo sequence of target prediction time and target spatial resolution is output. The radar echo sequence contains information on typhoons, cold vortices, squall lines and local severe convection, and based on the above information, information such as precipitation can be obtained. The specifics depend on the actual application.

[0104] Understandably, based on this meticulously constructed training system that combines breadth and depth, the final target forecasting model breaks through the limitations of traditional single-resolution models, transforming into a powerful tool with excellent universality and refined application capabilities. Firstly, at the operational forecasting level, the model can seamlessly adapt to different application needs from regional to local levels. Whether it's a 72-hour typhoon track forecast requiring a macroscopic understanding at the provincial level, or a short-term severe convective weather warning requiring precision down to the city street level, the model can provide matching forecast products by calling data inputs of the appropriate resolution, greatly improving the flexibility of operational applications. Secondly, its value is even more prominent at the level of cutting-edge scientific exploration and specific industry services. For researchers, the model can generate high-resolution, physically consistent three-dimensional cloud and wind field data, providing a new digital experimental field for in-depth research on the mechanisms of weather system occurrence and development. For the booming low-altitude economy, the model's refined forecasts at the 500-meter to 1-kilometer level can provide crucial meteorological risk decision support for UAV logistics route planning and urban air traffic (UAM) management, effectively mitigating flight risks caused by sudden weather phenomena such as low-level wind shear and microbursts. Ultimately, the core contribution of this work lies in its innovative data organization and training paradigm, which successfully integrates broad-spectrum forecasting capabilities with precise analytical capabilities, laying a crucial technological foundation for building next-generation intelligent, adaptive, and multi-scale unified meteorological forecasting models.

[0105] Furthermore, to verify the model's forecasting performance in different regions, two regions were randomly selected as the validation set and test set, respectively, within the coverage of the training dataset, for forecasting verification. These regions were a typhoon event in South China in a certain year and a hail event in Southwest China in a certain year.

[0106] Experimental Data 1:

[0107] Please refer to Figure 4 , Figure 4The image shows the typhoon forecast for a specific region. The leftmost image displays the actual observation results, while the right sides show the forecasts from the target forecast model for 6-minute, 1-hour, 2-hour, and 3-hour lead times. The 6-minute lead time results are highly consistent with the actual observations; the typhoon eye location, spiral rainband structure, and intensity distribution are very close to reality. The typhoon center location is accurate, the eyewall structure is clear, and the distribution of the outer spiral rainbands is also largely consistent with observations. This indicates that the target forecast model performs excellently in very short-term forecasts. The overall structure of the 1-hour lead time results remains good, with the typhoon center location and main rainband distribution being basically accurate. However, a slightly stronger reflectance intensity trend can be observed, especially in the eyewall and core regions, where the 50-60 dBZ range is slightly larger than observed. The stronger intensity trend is even more pronounced in the 2-hour lead time, with the high reflectance value distribution range in the eyewall region further expanding. However, the typhoon's overall spatial position and spiral structure remained largely intact, and the direction and distribution of the main precipitation bands generally matched the actual situation. The 3-hour forecast showed the largest lead time deviation, with the intensity significantly overestimated. It is worth noting, however, that the model was still able to roughly capture the typhoon's location and the distribution characteristics of the main rainbands.

[0108] The model's typhoon forecast accuracy continues to decrease with increasing lead time. As the forecast lead time lengthens, the model tends to overestimate precipitation intensity, which may be related to the model parameters. However, the typhoon's structural characteristics remain relatively good with longer forecast leads. Even in 3-hour forecasts, the model can still reasonably depict the typhoon's spiral structure and the location of the main precipitation bands, indicating that the model's grasp of the typhoon's dynamic structure is quite accurate. Considering the model's excellent forecast performance within 0-1 hour, it is also suitable for short-term, nowcast typhoon warnings.

[0109] Please refer to Figure 5 , Figure 5 The image shows a comparison of actual observations (top row, left to right) and forecasts (bottom row, left to right) of the target forecast model at different altitudes: 3.5km, 5km, 6km, and 7km. At 3.5km, actual observations show a small and compact typhoon eye with a clear yellow-orange strong echo band in the eyewall. The overall forecast shows high agreement, accurate eye location, and similar spiral rainband morphology, with only a slightly stronger eyewall. At 5km, actual observations show a slightly enlarged typhoon eye with moderate echo intensity. This altitude level shows the best forecast performance, with the typhoon structure, intensity distribution, and spiral rainband details all consistent with the observed altitude. At 6km, observations show a significantly enlarged typhoon eye and a substantial decrease in echo intensity, dominated by weak green echoes. While the forecast captured the eye expansion trend, the intensity was significantly stronger than predicted, with more areas of medium to strong yellow echoes. Observations at a height of 7km revealed a large and clear eye to the typhoon, dominated by weak blue-green echoes and exhibiting a loose structure. Forecast results showed a similar structure but a higher intensity.

[0110] The target forecast model exhibits a significant high dependence, with accurate forecasts at the low levels (3.5-5km) and a systematic overestimation of intensity at the mid-to-high levels (6-7km). However, the model successfully captures the evolution of the typhoon's horizontal and vertical spatial structure, including the expansion of the typhoon eye with height, the weakening of echo intensity with height, the location of the typhoon center, and the spatial distribution of the main rainbands.

[0111] Overall, the target forecasting model demonstrated a good ability to predict the three-dimensional structure of typhoons, especially excelling in the 3.5-5 km altitude range. The model accurately captured the overall structure, evolution trend, and key characteristics of typhoons, providing reliable technical support for refined typhoon forecasting.

[0112] Experimental Data 2:

[0113] Figure 6 This image shows the forecast of a hail event in a certain region in a certain year. The leftmost image is the actual observation result, and the rightmost images show the forecast results of the target forecast model at 6-minute, 30-minute, 60-minute, and 90-minute lead times. The observation results show that at 15:00, multiple strong convective cells appeared in the central and western parts of the region, mainly concentrated in one area and its surrounding areas. The strongest echo reached 50-60 dBZ, exhibiting typical strong convection characteristics. The convective system was distributed in a northeast-southwest direction, with multiple convective cells relatively independent. The 6-minute forecast result is highly consistent with the observations; the location, intensity, and morphology of the main convective cells are very close to the actual situation. The center of the strong echo in this region is accurate, and the distribution of the surrounding weaker echoes is also basically consistent. This indicates that the target forecast model performs excellently in ultra-short-term forecasts. The 30-minute forecast result shows that the overall convective structure remains good, and the location of the main strong echo area is basically accurate. However, an overestimation trend began to emerge, particularly in the high reflectivity area (50-55 dBZ) near the region, which expanded compared to actual conditions. Some areas with weak green echoes also showed slight enhancement. The overestimation was even more pronounced in the 60-minute forecast, with not only the main strong echo area expanding but also the surrounding previously weaker convective cells showing significant intensity increases. While the spatial distribution pattern remained largely unchanged, the convective coverage expanded significantly, indicating that the model tended to overestimate convective development. The 90-minute forecast showed the largest lead time deviation, with a significant increase in strong echo areas and an excessively large convective system coverage. Although the main convective center was still identifiable, the overall intensity and extent were significantly overestimated, with considerable loss of detailed features.

[0114] From 6 minutes to 90 minutes, forecast accuracy gradually decreases. As the forecast lead time increases, the model tends to overestimate the intensity and extent of convection, which is common in severe convective weather forecasts. However, the model's spatial positioning capability is good; even in a 90-minute forecast, the location of the main convective system remains largely accurate. Forecasts within 0-1 hours perform excellently, making them suitable for operational nowcasting of severe convective weather such as hail.

[0115] Please refer to Figure 7 , Figure 7 The image shows a comparison of actual observations (top row, left to right) and forecasts (bottom row, left to right) from the target forecast model at different altitudes for this hail event, specifically 3.5 km, 5 km, 6 km, and 7 km. At all altitudes, the model was able to accurately predict the spatial distribution characteristics of the main hail systems in the region. Particularly at the mid-to-low altitudes (3.5 km and 5 km), the model demonstrated excellent ability to identify strong convective cells in the region, exhibiting a clear reflectance structure and location that closely matched the observation altitude, showing strong spatial consistency.

[0116] At the main reflectivity centers in the region, both forecasts and observations exhibit strong echo characteristics, indicating that the model has a good predictive ability for the organization and structure of strong convective systems in the region. Notably, the model shows high spatial consistency across all altitude levels in the region. Although slightly overestimated in the intensity of upper-level echoes, the model performs well overall in predicting the reflectivity structure of the region at altitudes of 3.5 km–7 km, demonstrating its ability to predict the three-dimensional structure of complex convective processes in the region.

[0117] Example 2

[0118] This embodiment provides a spatial scalability forecasting model for three-dimensional cloud fields. This model is applied to the spatial scalability forecasting method for three-dimensional cloud fields in Embodiment 1. The temporal scalability forecasting model for three-dimensional cloud fields includes:

[0119] The data acquisition module is used to acquire the first historical radar network data, special historical data, and the second historical radar network data.

[0120] The data processing module is used to construct the first basic training set, the second basic training set, and the specialized training set;

[0121] The model training module is used to train an initial prediction model based on the first basic training set, train a special improvement model based on historical data of multiple special projects, and train a target prediction model based on the second basic training set and multiple special project training sets.

[0122] The model building module is used to build the initial forecast model, the special enhancement model, and the target forecast model.

[0123] Example 3

[0124] This embodiment provides a spatial scalability forecasting device for three-dimensional cloud fields, including:

[0125] processor;

[0126] Memory, used to store executable instructions;

[0127] The processor is configured to read the executable instructions from the memory and execute the executable instructions to implement the spatial scalability prediction method for three-dimensional cloud fields in Embodiment 1 above.

[0128] Example 4

[0129] like Figure 8 As shown, one embodiment of the present invention provides an electronic device 100. The electronic device 100 includes a memory 101, a processor 102, and an input / output (I / O) interface 103. The memory 101 is used to store instructions. The processor 102 is used to execute the spatial scalability prediction method for three-dimensional cloud fields according to embodiments of the present invention by calling the instructions stored in the memory 101. The processor 102 is connected to both the memory 101 and the I / O interface 103, for example, via a bus system and / or other forms of connection mechanisms (not shown). The memory 101 can be used to store programs and data, including the program for the spatial scalability prediction method for three-dimensional cloud fields according to embodiments of the present invention. The processor 102 executes various functional applications and data processing of the electronic device 100 by running the program stored in the memory 101.

[0130] In this embodiment of the invention, the processor 102 can be implemented using at least one of the following hardware forms: digital signal processor (DSP), field-programmable gate array (FPGA), and programmable logic array (PLA). The processor 102 can be one or a combination of several of the following: central processing unit (CPU) or other processing units with data processing capability and / or instruction execution capability.

[0131] The memory 101 in this embodiment of the invention may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD).

[0132] In this embodiment of the invention, the I / O interface 103 can be used to receive input instructions (such as numeric or character information, and key signal inputs related to user settings and function control of the electronic device 100), and can also output various information (such as images or sounds) to the outside. In this embodiment of the invention, the I / O interface 103 may include one or more of the following: a physical keyboard, function keys (such as volume control keys, power buttons, etc.), a mouse, a joystick, a trackball, a microphone, a speaker, and a touch panel.

[0133] It is understood that although operations are described in a specific order in the accompanying drawings in the embodiments of the present invention, this should not be construed as requiring these operations to be performed in the specific order or serial order shown, or requiring all of the shown operations to be performed to obtain the desired result. In certain environments, multitasking and parallel processing may be advantageous.

[0134] The methods and apparatuses involved in the embodiments of the present invention can be implemented using standard programming techniques, and various method steps can be implemented using rule-based logic or other logic. It should also be noted that the terms "apparatus" and "module" as used herein and in the claims are intended to include implementations using one or more lines of software code and / or hardware implementations and / or devices for receiving input.

[0135] Any step, operation, or procedure described herein may be performed or implemented using one or more hardware or software modules, either alone or in combination with other devices. In one embodiment, the software module is implemented using a computer program product comprising a computer-readable medium containing computer program code, which is executable by a computer processor to perform any or all of the described steps, operations, or procedures.

[0136] Those skilled in the art will understand that the above embodiments are specific examples of implementing the present invention, and in practical applications, various changes can be made in form and detail without departing from the scope of the present invention.

Claims

1. A method for predicting the spatial scalability of three-dimensional cloud fields, characterized in that, include: The first basic training set is constructed based on the preprocessed first historical radar network data; The first basic training set is input into the model to obtain the initial prediction model; The initial prediction model is a convolutional long short-term memory network model; Multiple types of specific historical data are input into the initial forecast model to train its specific identification capabilities, resulting in a specific improvement model; wherein, the specific historical data includes historical typhoon data, historical cold vortex data, historical squall line data, and local severe convection data; A second basic training set is constructed based on the preprocessed second historical radar network data, and the second basic training set is processed based on the special improvement model to filter and construct multiple special training sets; wherein, the data volume of the first historical radar network data is smaller than the data volume of the second historical radar network data. The second basic training set and multiple specialized training sets are input into the specialized enhancement model for multi-resolution joint training to obtain the target prediction model.

2. The spatial scalability prediction method for three-dimensional cloud fields according to claim 1, characterized in that, The second basic training set, constructed based on the preprocessed second historical radar network data, includes: Acquire the second historical radar network data covering the target area; The second historical radar network data is preprocessed; wherein, the preprocessing includes: upscaling the basic spatial resolution of the second historical radar network data to a preset spatial resolution using a grid aggregation and averaging algorithm; and resampling the second historical radar network data at the preset spatial resolution onto a target regular latitude and longitude grid to form the second basic training set.

3. The spatial scalability prediction method for three-dimensional cloud fields according to claim 2, characterized in that, Based on the aforementioned specialized enhancement model, the second basic training set is processed to select and construct multiple specialized training sets, including: The second basic training set is input into the specialized enhancement model; The specific enhancement model extracts radar echo data sequences from typhoon systems, cold vortex systems, squall line systems, and local severe convective systems. Multiple specialized training sets are obtained based on the radar echo data sequences of the typhoon system, cold vortex system, squall line system, and local severe convection system; wherein, the multiple specialized training sets are the typhoon training set, the cold vortex training set, the squall line training set, and the local severe convection training set.

4. The spatial scalability prediction method for three-dimensional cloud fields according to claim 3, characterized in that, The typhoon training set includes a first spatial resolution, a first coverage area, and a first continuous time series; the cold vortex training set includes a second spatial resolution, a second coverage area, and a second continuous time series; the squall line training set includes a third spatial resolution, a third coverage area, and a third continuous time series; the local severe convection training set includes a fourth spatial resolution, a fourth coverage area, and a fourth continuous time series; wherein the first spatial resolution, the second spatial resolution, the third spatial resolution, and the fourth spatial resolution are smaller than the preset spatial resolution.

5. The spatial scalability prediction method for three-dimensional cloud fields according to claim 4, characterized in that, After inputting the second basic training set and multiple specialized training sets into the specialized enhancement model for multi-resolution joint training to obtain the target prediction model, the process further includes: Based on the preprocessed second historical radar network data, a multi-scale validation set and a test set are constructed; wherein, the multi-scale validation set includes a large-scale validation set for typhoons and cold vortices and a small-scale validation set for squall lines and strong convection. The target prediction model is validated based on the scaled validation set. The target prediction model, after verification, is tested based on the test set.

6. The spatial scalability prediction method for three-dimensional cloud fields according to claim 5, characterized in that, The validation of the target prediction model based on the scaled validation set includes: The historical sequence data of the large-scale validation set of typhoons and cold vortices are input into the target prediction model for validation, and the first prediction radar echo sequence is obtained. Extract the trajectory, intensity characterization, and precipitation distribution at the first spatial resolution and the second spatial resolution from the first forecast radar echo sequence; The running path, intensity characterization, and precipitation distribution are compared with the first actual observation data to calculate the path error, intensity error, and precipitation score. The historical sequence data of the squall line and the small-scale validation set of strong convection are input into the target prediction model for verification to obtain the second prediction radar echo sequence; The trigger time, movement vector, strong center position, and fine structure features at the third and fourth spatial resolutions are extracted from the second forecast radar echo sequence. The trigger time, movement vector, strong center position, and fine structural features are compared with the second actual observation data to calculate the trigger time error, movement error, position error, and structural overlap index.

7. The spatial scalability prediction method for three-dimensional cloud fields according to claim 1, characterized in that, After inputting the second basic training set and multiple specialized training sets into the specialized enhancement model for multi-resolution joint training to obtain the target prediction model, the process further includes: The target spatial resolution is input into the target prediction model, and the radar echo sequence of the target spatial resolution is output.

8. A spatially scalable prediction method system for three-dimensional cloud fields, characterized in that, The spatial scalability forecasting method system for three-dimensional cloud fields is applied to the spatial scalability forecasting method for three-dimensional cloud fields according to any one of claims 1-7, wherein the spatial scalability forecasting method system for three-dimensional cloud fields comprises: The data acquisition module is used to acquire the first historical radar network data, special historical data, and the second historical radar network data. The data processing module is used to construct the first basic training set, the second basic training set, and the specialized training set; The model training module is used to train an initial prediction model based on a first basic training set, train a special improvement model based on various special historical data, and train a target prediction model based on a second basic training set and multiple special training sets. The model building module is used to build the initial forecast model, the special enhancement model, and the target forecast model.

9. A method and device for predicting the spatial scalability of three-dimensional cloud fields, characterized in that, include: processor; Memory, used to store executable instructions; The processor is configured to read the executable instructions from the memory and execute the executable instructions to implement the spatial scalability prediction method for three-dimensional cloud fields as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that... The system stores instructions that, when executed by a processor, perform the spatial scalability prediction method for three-dimensional cloud fields as described in any one of claims 1-7.

Images