Space expandability forecasting method, system and device of three-dimensional cloud field and medium

Abstract

The invention relates to the technical field of three-dimensional cloud field forecasting, in particular to a three-dimensional cloud field space expandability forecasting method, system and device and a medium. The method comprises the following steps: constructing a first basic training set based on preprocessed first historical radar networking data; inputting the first basic training set into the model to obtain an initial forecasting model; inputting various special historical data into the initial forecasting model for special identification ability training to obtain a special improvement model; constructing a second basic training set based on the preprocessed second historical radar networking data, processing the second basic training set based on a special lifting model, and screening and constructing a plurality of special training sets; and inputting the second basic training set and the plurality of special training sets into a special lifting model for multi-resolution joint training to obtain a target forecasting model. The problem that a three-dimensional cloud field forecasting model is insufficient in space adaptability and special identification capability under complex weather conditions is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional cloud field prediction, in particular to a three-dimensional cloud field space scalability prediction method, system, device and readable storage medium. BACKGROUND

[0002] Accurate three-dimensional cloud field prediction has epoch-making significance for improving extreme weather warning capability, ensuring low-altitude economic safety operation and optimizing urban disaster prevention and mitigation system. As a three-dimensional presentation of atmospheric motion and phase change process, three-dimensional cloud field not only determines the distribution and intensity of precipitation, but also directly affects aviation safety, unmanned aerial vehicle logistics, urban flood control and other key fields. In recent years, the rapid development of deep learning technology has injected momentum into the field of weather prediction: convolutional neural networks extract cloud physical features through multi-level abstraction, recurrent neural networks build spatio-temporal correlation models, and generative adversarial networks can even simulate the nonlinear evolution law of atmospheric motion, making it possible to mine deep cloud field evolution rules from massive satellite remote sensing, weather radar and sounding data.

[0003] However, the current near-term prediction technology based on artificial intelligence still faces serious spatio-temporal resolution constraints. Specifically, existing models generally have regional dependence and low precision, that is, the prediction model developed for a specific region must rely on long-term accumulation of high-precision spatio-temporal observation data in that area, and this strong data dependence leads to serious lack of model migration ability; and the existing models generally have the technical problem of insufficient special recognition ability. Therefore, how to solve the problems of spatial adaptability and insufficient special recognition ability of three-dimensional cloud field prediction model under complex weather conditions has become a problem to be solved. SUMMARY

[0004] To solve the problems of spatial adaptability and insufficient special recognition ability of three-dimensional cloud field prediction model under complex weather conditions, the present application provides a three-dimensional cloud field space scalability prediction method, system, device and medium.

[0005] In a first aspect, the present application provides a three-dimensional cloud field space scalability prediction method, comprising:

[0006] constructing a first basic training set based on the preprocessed first historical radar networking data;

[0007] inputting the first basic training set into the model to obtain an initial prediction model;

[0008] inputting a plurality of special historical data into the initial prediction model for special recognition ability training to obtain a special improvement model;

[0009] construct a second basic training set based on the pretreated second historical radar networking data, and process the second basic training set based on a special promotion model to screen and construct a plurality of special training sets; wherein the data quantity of the first historical radar networking data is less than the data quantity of the second historical radar networking data;

[0010] input the second basic training set and the plurality of special training sets into the special promotion model for multi-resolution joint training to obtain the target prediction model.

[0011] In some embodiments, constructing a second basic training set based on the pretreated second historical radar networking data comprises:

[0012] acquiring second historical radar networking data covering a target area;

[0013] pretreating the second historical radar networking data; wherein the pretreatment comprises: upscaling the basic spatial resolution of the second historical radar networking data to a preset spatial resolution through a grid aggregation and averaging algorithm; and resampling the second historical radar networking data at the preset spatial resolution onto a target regular latitude-longitude grid to form the second basic training set.

[0014] In some embodiments, processing the second basic training set based on the special promotion model to screen and construct a plurality of special training sets comprises:

[0015] inputting the second basic training set into the special promotion model;

[0016] the special promotion model extracts radar echo data sequences of a typhoon system, a cold vortex system, a squall line system and a local strong convective system;

[0017] obtaining a plurality of special training sets based on the radar echo data sequences of the typhoon system, the cold vortex system, the squall line system and the local strong convective system; wherein the plurality of special training sets are respectively a typhoon training set, a cold vortex training set, a squall line training set and a local strong convective training set.

[0018] In some embodiments, the typhoon training set comprises a first spatial resolution, a first coverage range and a first continuous time sequence; the cold vortex training set comprises a second spatial resolution, a second coverage range and a second continuous time sequence; the squall line training set comprises a third spatial resolution, a third coverage range and a third continuous time sequence; and the local strong convective training set comprises a fourth spatial resolution, a fourth coverage range and a fourth continuous time sequence; wherein the first spatial resolution, the second spatial resolution, the third spatial resolution and the fourth spatial resolution are less than the preset spatial resolution.

[0019] In some embodiments, after inputting the second basic training set and the plurality of special training sets into the special promotion model for multi-resolution joint training to obtain the target prediction model, the method further comprises:

[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 configured to acquire first historical radar networking data, special historical data, and second historical radar networking data.

[0034] The data processing module is configured to construct a first basic training set, a second basic training set, and a special training set.

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

[0036] The model construction module is configured to construct the initial prediction model, the special improvement model, and the target prediction model.

[0037] In a third aspect, the present application provides a device for predicting the spatial scalability of a three-dimensional cloud field, comprising:

[0038] a processor;

[0039] a memory configured to store executable instructions;

[0040] The processor is configured to read the executable instructions from the memory and execute the executable instructions to implement the method for predicting the spatial scalability of a three-dimensional cloud field according to any one of the embodiments of the first aspect.

[0041] In a fourth aspect, the present application provides a computer-readable storage medium, comprising:

[0042] instructions stored therein, which, when executed by a processor, perform the method for predicting the spatial scalability of a three-dimensional cloud field according to any one of the embodiments of the first aspect.

[0043] To solve the problem of insufficient spatial adaptability and special recognition ability of a three-dimensional cloud field prediction model under complex weather conditions, the present application has the following advantages:

[0044] The present application uses first historical radar networking data with a small amount of data to quickly construct an initial prediction model, solves the efficiency problem of model cold start, then trains the special recognition ability by introducing a plurality of special historical data, so that the model obtains accurate recognition ability for specific weather phenomena, and then uses the model with special recognition ability to automatically filter a large amount of second historical radar networking data, and constructs a special training set with strong pertinence and high quality, and finally through multi-resolution joint training, the target prediction model obtained finally is significantly improved in spatial scalability, special weather prediction accuracy, and consistency of different scale predictions, and realizes the ability leap from basic prediction to fine and scalable prediction. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1A flow chart of a spatial scalability prediction method of a three-dimensional cloud field of an embodiment is shown.

[0046] Figure 2 A ConvLSTM model structure diagram of a spatial scalability prediction method of a three-dimensional cloud field of an embodiment is shown.

[0047] Figure 3 A special training set diagram of a spatial scalability prediction method of a three-dimensional cloud field of an embodiment is shown.

[0048] Figure 4 A typhoon actual observation diagram and radar diagrams of different prediction time scales of experimental data of a spatial scalability prediction method of a three-dimensional cloud field of an embodiment are shown.

[0049] Figure 5 Radar diagrams of different height layers of experimental data of a spatial scalability prediction method of a three-dimensional cloud field of an embodiment are shown.

[0050] Figure 6 A hail actual observation diagram and radar diagrams of different prediction time scales of experimental data of a spatial scalability prediction method of a three-dimensional cloud field of an embodiment are shown.

[0051] Figure 7 Radar diagrams of different height layers of experimental data of a spatial scalability prediction method of a three-dimensional cloud field of an embodiment are shown.

[0052] Figure 8 An electronic device diagram of a spatial scalability prediction method of a three-dimensional cloud field of an embodiment is shown.

[0053] Reference signs:

[0054] Electronic device 100; memory 101; processor 102; Input / Output (I / O) interface 103. DETAILED DESCRIPTION

[0055] Example embodiments will now be described more fully with reference to the accompanying drawings. Example embodiments, however, can be implemented in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of example embodiments to those skilled in the art. Like reference numerals refer to like elements throughout the drawings, and thus repeated description of the same will be omitted.

[0056] Moreover, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a thorough understanding of embodiments of the application. One skilled in the relevant art will recognize, however, that the application can be practiced without one or more of the specific details, or with other methods, components, devices, steps, etc. In other instances, well-known structures, devices, implementations, or operations are not shown or described in detail to avoid obscuring aspects of the application.

[0057] The block diagrams in the drawings show only the functionality of the features and can not imply that the functionality must be implemented in the order shown. Further, the diagrams show a sequence of operations, but this ordering is not necessary or essential. For example, it is possible to perform some operations in one order and others in a different order or concurrently. It is also possible to break some operations into sub operations not shown in the diagrams. Further, an action can be performed by hardware components or modules, using any number of separate hardware integrity testing techniques. Furthermore, a particular sequence of operations does not limit the scope of the present application. For example, an operation performed before another operation in a particular aspect can be performed after the other operation in another aspect depending upon the functional aspects of a particular implementation. Likewise, a block that is described as performing an operation can actually not perform that operation and instead can be configured to support performance of another operation. The descriptions are not intended to be limiting.

[0058] Those skilled in the art can understand that the modules or flows in the drawings are only schematic diagrams of the example embodiments, and the modules or flows in the drawings do not necessarily have to be implemented to practice the present application, and therefore cannot be used to limit the protection scope of the present application.

[0059] Accurate three-dimensional cloud field prediction has epoch-making significance for improving extreme weather warning capability, ensuring the safe operation of low-altitude economy, and optimizing the urban disaster prevention and mitigation system. As a three-dimensional presentation of atmospheric motion and phase change process, three-dimensional cloud field not only determines the distribution and intensity of precipitation, but also directly affects key fields such as aviation safety, unmanned aerial vehicle logistics, and urban flood control. In recent years, the rapid development of deep learning technology has injected momentum into the field of weather forecasting: convolutional neural networks extract cloud physical features through multi-level abstraction, recurrent neural networks build spatio-temporal correlation models, and generative adversarial networks can even simulate the nonlinear evolution law of atmospheric motion, making it possible to mine deep cloud evolution laws from massive satellite remote sensing, weather radar, and sounding data.

[0060] However, current artificial intelligence-based nowcasting technology still faces severe spatio-temporal resolution constraints. Specifically, existing models generally have dual limitations of "regional dependence" and "data rigidity": on the one hand, a prediction model developed for a specific region must rely on long-term accumulation of high-precision spatio-temporal observation data in the region, and this strong data dependence leads to serious lack of model migration ability. On the other hand, the model architecture is deeply coupled with the resolution of the input data, and when the radar or satellite pixel is upgraded from 1km to 500m, the entire model needs to be retrained or even restructured. Such rigid constraints not only cause repeated investment of scientific research resources, but also cause a digital divide in regional prediction capabilities. Therefore, it has become an urgent need in the industry to build a new generation of intelligent prediction model with spatial adaptability and resolution flexibility.

[0061] Embodiment one

[0062] For reference Figure 1 The embodiment provides a spatial scalability prediction method for three-dimensional cloud field, comprising the following steps S10-S50:

[0063] S10, constructing a first basic training set based on the preprocessed first historical radar networking data;

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

[0065] S30, inputting a plurality of special historical data into the initial prediction model for special recognition ability training to obtain a special promotion model;

[0066] S40, constructing a second basic training set based on the preprocessed second historical radar networking data, and processing the second basic training set based on the special promotion model to screen and construct a plurality of special training sets; wherein the data amount of the first historical radar networking data is less than the data amount of the second historical radar networking data;

[0067] S50, inputting the second basic training set and the plurality of special training sets into the special promotion model for multi-resolution joint training to obtain a target prediction model.

[0068] For reference Figure 2 The initial prediction model, the special promotion model and the final target prediction model of the embodiment are all constructed and trained using a core architecture based on a convolutional long short-term memory network (CONV-LSTM). This architecture can effectively capture the spatial features and temporal dependencies of three-dimensional cloud field data. Moreover, the network uses an encoder-decoder structure, in which the encoder and the decoder each contain 2 layers of CONV-LSTM units. Specifically, the convolutional long short-term memory network model includes three gated convolutional layers, respectively, the input gate I t , the forget gate F tand output gate O t where H t-1 is the hidden state of the previous time step, H t is the hidden state of the current time step, C t-1 is the cell state of the previous time step, C t is the cell state of the current time step, X t is the input state of the current time step, and the channel number of the model input data is set to (1, 6, 512, 512), which respectively corresponds to the batch size, the time sequence length (such as 6 consecutive time sequences), and the grid height and width. During training, the batch size is 32, and the model parameter is updated through the back propagation algorithm and the gradient descent optimizer, and the total training iteration number is 1000 times to ensure that the model converges fully. This configuration is conducive to multi-resolution joint training: the unified 512x512 target regular latitude and longitude grid field allows training samples of different resolutions to be aligned in the spatial dimension, and the time sequence processing capability of the convolutional long short-term memory network can learn the cross-scale mapping relationship from the historical sequence to the future prediction. By using the first basic training set, the special historical data, and the mixed multi-scale data composed of the second basic training set and multiple special training sets in sequence for stage-by-stage training on this fixed architecture, the target prediction model finally encodes the cloud field physical evolution knowledge of different scales from the kilometer level to the hundred-meter level, thereby obtaining the spatial scalability prediction capability.

[0069] Please refer to Figure 3 In step S10, the first historical radar networking data is the networking radar data of the China Meteorological Administration, and the basic spatial resolution is 1kmx1km. In the time dimension, the data maintains the high-frequency sampling characteristics of every 6 minutes, which prepares training data for the general prediction capability of the preliminary model. The preprocessing of the first historical radar networking data includes upsizing the first historical radar networking data through the grid aggregation and averaging algorithm to form a uniform grid field with a preset spatial resolution of 8kmx8km from the basic spatial resolution of 1kmx1km, and constructing a target regular latitude and longitude grid field of 512x512. The geographical coverage range of the target regular latitude and longitude grid field is, for example, from 87 degrees east to 126 degrees east and from 17 degrees north to 55 degrees north, and the specific application is subject to the actual application, which is not limited in the present application. The geographical coverage range, for example, completely covers the entire China and part of the surrounding sea area, thereby ensuring the diversity and representativeness of the training samples in the climate zone, underlying surface type, and typical weather system.

[0070] In step S20, the first basic training set is input into the model to complete the basic training phase of the model, and an initial prediction model with basic three-dimensional cloud field prediction capability is obtained.

[0071] In step S30, the plurality of special historical data is input into the initial prediction model, so that the model not only learns the general weather evolution rule, but also solves the prediction difficulty in specific and extreme scenarios, to enhance the recognition and prediction ability of the model to specific important weather systems. Among them, the plurality of special historical data of the embodiment includes historical typhoon data, historical cold vortex data, historical squall line data and local strong convection data, and the historical typhoon data, cold vortex data, squall line data and local strong convection data are input into the initial prediction model, so that the model further improves the recognition and prediction skill of the specific weather system on the premise of the basic prediction ability, by accurately depicting the three-dimensional structure and intensity of these systems, driving the physical process (such as energy release, water vapor transport, and unstable energy triggering) inside the model, to simulate their own development and evolution, the complex interaction between systems (such as the case of typhoon and cold vortex combination to intensify the rainstorm), so as to realize the fine and quantitative prediction of the three-dimensional cloud field structure, precipitation type, intensity and falling area dominated by them, and significantly improve the prediction and early warning ability of the disastrous weather.

[0072] In the embodiment, the construction of the second basic training set based on the preprocessed second historical radar networking data in step S40 includes:

[0073] Obtaining second historical radar networking data covering the target area;

[0074] Preprocessing the second historical radar networking data; wherein the preprocessing includes: performing upscaled processing of the basic spatial resolution of the second historical radar networking data to a preset spatial resolution by grid aggregation and averaging algorithm; and resampling the second historical radar networking data of the preset spatial resolution to a target regular latitude-longitude grid to form a second basic training set.

[0075] It can be understood that the second historical radar network data also uses the network radar data of the China Meteorological Administration. The basic spatial resolution of the second historical radar network data is 1 km x 1 km, the preset spatial resolution is 8 km x 8 km, and the target regular latitude and longitude grid is 512 x 512. That is, the second historical radar network data with a basic spatial resolution of 1 km x 1 km is up-scaled to a preset spatial resolution of 8 km x 8 km through a grid aggregation and averaging algorithm. And the second historical radar network data with a preset spatial resolution of 8 km x 8 km is resampled to a target regular latitude and longitude grid of 512 x 512, and the actual application is specific. In the time dimension, the data maintains a high-frequency sampling characteristic of every 6 minutes, and the data amount of the second historical radar network data is greater than that of the first historical radar network data. The second historical radar network data may be historical multi-year and continuous time series data, and the first historical radar network data may be historical one-year time series data, which provides a basis for further implementing a refined data expansion strategy, so that the model can select a more comprehensive special training set from the second basic training set.

[0076] It can be understood that reducing the data to 8 km preset spatial resolution in the embodiment can greatly reduce the data amount and model calculation burden, and at the same time, since the movement and evolution of large-scale weather systems (typhoons and cold vortices) are mainly controlled by macro dynamic and thermal processes, the 8 km preset spatial resolution can effectively capture the main structure and movement trend, thereby ensuring the effectiveness of the initial model learning.

[0077] Please refer to Figure 3 In the embodiment, the processing of the second basic training set based on the special promotion model in step S40 further includes the following steps:

[0078] The second basic training set is input into the special promotion model;

[0079] The special promotion model extracts radar echo data sequences of typhoon systems, cold vortex systems, squall line systems and local strong convective systems;

[0080] A plurality of special training sets are obtained based on the radar echo data sequences of the typhoon systems, the cold vortex systems, the squall line systems and the local strong convective systems; wherein the plurality of special training sets are respectively a typhoon training set, a cold vortex training set, a squall line training set and a local strong convective training set.

[0081] It can be understood that the special improvement model obtained by step S30, for example, is an intermediate prediction model, can filter and construct a special training set from a large amount of technical data, further strengthen the special training of the model based on the special training set, and improve the recognition and prediction ability of the model to specific important weather systems. Based on the special improvement model, a plurality of special training sets can be filtered and constructed to improve the efficiency and objectivity of data filtering, avoid subjective bias and heavy labor of manual screening, and also ensure that the constructed special training set accurately contains complete and typical life history data of typhoon systems, cold vortex systems, squall line systems and local strong convective systems, and ensures the accuracy of subsequent model prediction.

[0082] Further, the typhoon system and the cold vortex system are large-scale systems, and the squall line system and the local strong convective system are mesoscale or small-scale systems, wherein the horizontal coverage range of the large-scale system can reach hundreds to thousands of kilometers, and the life history is measured in days, while the horizontal coverage range of the mesoscale and small-scale system is several kilometers to several hundred kilometers, and the life history is measured in hours, so that the special improvement model and the subsequent target prediction model can learn the potential or mapping relationship inside the system affecting precipitation and clouds based on the large-scale system, the mesoscale or the small-scale system, further improving the accuracy of the target prediction model.

[0083] In the embodiment, the typhoon training set includes a first spatial resolution, a first coverage range and a first continuous time sequence, wherein the first spatial resolution is 3 km, the first coverage range is 1536x1536 km, and the first continuous time sequence is more than 36 h; the cold vortex training set includes a second spatial resolution, a second coverage range and a second continuous time sequence, wherein the second spatial resolution is 2 km, the second coverage range is 1024x1024 km, and the second continuous time sequence is more than 24 h; the squall line training set includes a third spatial resolution, a third coverage range and a third continuous time sequence, wherein the third spatial resolution is 1 km, the third coverage range is 512x512 km, and the third continuous time sequence is more than 12 h; the local strong convective training set includes a fourth spatial resolution, a fourth coverage range and a fourth continuous time sequence, wherein the fourth spatial resolution is 500 m, the fourth coverage range is 256x256 km, and the fourth continuous time sequence is more than 6 h; wherein the first spatial resolution, the second spatial resolution, the third spatial resolution and the fourth spatial resolution are less than 8 km, and such a multi-resolution architecture enables the model to grasp the overall structure of the typhoon eye wall or the moving trend of the cold vortex cloud system from a macroscopic scale, and to understand the outflow boundary of the squall line front or the core structure evolution inside the convective cell from a microscopic scale.

[0084] It can be understood that the spatial resolution refers to the actual horizontal spatial distance represented by each grid point in a regular latitude-longitude grid field, the smaller the resolution value, the denser the grid points, and the more detailed the weather system can be described; the coverage refers to the actual geographical area size corresponding to the training data; the continuous time sequence refers to the continuous and uninterrupted radar observation data sequence used for one training sample in time.

[0085] It can be understood that the embodiment sets the training resolution and coverage for different types of weather systems that best match their physical scales, wherein the coverage is within the Chinese region, covering typical regions from south to north and from west to east, etc., ensuring that the model learns different systems with optimal input data field of view and clarity. That is, the typhoon system has a large range and moves slowly, so the first spatial resolution of the typhoon is 3 km, and the first coverage is 1536x1536 km, which can fully capture the macro circulation structure of the typhoon while ensuring computational efficiency. The local strong convective system has a small scale and develops rapidly, so the fourth spatial resolution is 500 m, and the fourth coverage is 256x256 km, which can accurately depict the internal fine structure and development core of the system. Based on the first, second, third and fourth spatial resolutions in a step-like manner, the spatial expandable range required by the model is clearly defined, i.e., the model can learn how to restore or deduce fine structures up to 500 meters from a relatively rough 8 km base field, thereby improving the prediction accuracy of the target prediction model.

[0086] At the same time, a continuous time sequence not less than the typical life length of each system is set, i.e., typhoon > 36h, cold vortex > 24h, squall line > 12h, strong convection > 6h, and the actual application is specific, which is not limited in the present application, ensuring that the model can learn the complete dynamic evolution process of the system from occurrence, development to decline, rather than a fragment, thereby improving the continuity and physical reasonableness of the prediction.

[0087] In the embodiment, after the second base training set and the plurality of special training sets are input into the special promotion model in step S50 for multi-resolution joint training to obtain the target prediction model, the method further comprises:

[0088] Based on the preprocessed second historical radar networking data, a scale verification set and a test set are constructed; wherein the scale verification set includes a typhoon and cold vortex large-scale verification set and a squall line and strong convection small-scale verification set;

[0089] The target prediction model is verified based on the scale verification set;

[0090] The target prediction model after verification is tested based on the test set.

[0091] It can be understood that the sub-scale verification mechanism of the embodiment ensures the comprehensiveness, pertinence and fairness of the target prediction model. By constructing two sets of verification sets of typhoon and cold vortex large-scale verification set and squall line and strong convection small-scale verification set, the prediction ability of the model on different types and different scales of weather systems can be quantified independently, and the evaluation deviation that may occur when a single mixed verification set is used can be avoided.

[0092] In the embodiment, the verification of the target prediction model based on the sub-scale verification set comprises:

[0093] The historical sequence data in the typhoon and cold vortex large-scale verification set is input into the target prediction model for verification to obtain a first predicted radar echo sequence;

[0094] The running path, intensity characteristic quantity and precipitation distribution under the first spatial resolution and the second spatial resolution are extracted from the first predicted radar echo sequence;

[0095] The running path, intensity characteristic quantity and precipitation distribution are compared with the first actual observation data to calculate the path error, intensity error and precipitation score;

[0096] The historical sequence data in the squall line and strong convection small-scale verification set is input into the target prediction model for verification to obtain a second predicted radar echo sequence;

[0097] The trigger time, moving vector, strong center position and fine structure feature under the third spatial resolution and the fourth spatial resolution are extracted from the second predicted radar echo sequence;

[0098] The trigger time, moving vector, strong center position and fine structure feature are compared with the second actual observation data to calculate the trigger time error, moving error, position error and structure coincidence degree index.

[0099] It can be understood that for the key evaluation of typhoon and cold vortex, the prediction skills of path, intensity and precipitation distribution under 3km and 2km spatial resolution are evaluated; for squall line and strong convection, the capture ability of system trigger time, moving direction and strong center position under 1km and 500m spatial resolution is verified. The evaluation indexes selected are path error, trigger time error and first-level structure coincidence degree, which are not general machine learning loss functions, but core evaluation indexes recognized in meteorological prediction business, which directly reflect the prediction accuracy, so that the performance of the target prediction model can be intuitively understood and directly compared with the business performance of existing numerical prediction models or extrapolation algorithms.

[0100] Further, the test results show that the model maintains the stability of each scale prediction, and especially shows advantages on the convective resolvable scale. The 500-meter resolution output can clearly reproduce the fine structure of the thunderstorm outflow boundary and the gust front, which is difficult to achieve by traditional extrapolation algorithms and single-scale models. The realization of this ability is essentially due to the full coverage of the physical processes of various weather systems from macro to micro in the training data, as well as the cross-scale mapping relationship established by the model itself in the encoding-decoding process. Therefore, the integrated solution from data strategy to model architecture will provide a complete technical path and feasibility verification for the development of the next generation of adaptive, multi-scale fusion weather prediction models.

[0101] In the embodiment, after obtaining the target prediction model by inputting the second basic training set and the plurality of special training sets into the special promotion model for multi-resolution joint training, the method further includes:

[0102] Inputting the target spatial resolution into the target prediction model outputs the radar echo sequence of the target spatial resolution.

[0103] It can be understood that based on the target prediction time, a plurality of continuous time sequences and historical data of the target spatial resolution are obtained and input into the target prediction model to output the radar echo sequence of the target prediction time and the target spatial resolution. The radar echo sequence contains typhoon, cold vortex, squall line and local severe convective information, and based on the above information, precipitation information can be obtained, and the specific actual application is subject to the actual application.

[0104] It can be understood that based on this set of carefully constructed training system with breadth and depth, the final trained target prediction model breaks through the limitations of traditional single resolution models and evolves into a powerful tool with excellent universality and fine application ability. First, in the business prediction layer, the model can seamlessly adapt to different application needs from regional to local. Whether it is a provincial range of 72-hour typhoon path prediction that needs to be grasped macroscopically, or a short-time severe convective weather warning that requires precision to the city street, the model can provide matching prediction products by calling the corresponding resolution data input, greatly improving the flexibility of business application. Secondly, in the field of frontier scientific exploration and specific industry services, its value is more prominent. For researchers, the model can generate high-resolution, physically consistent three-dimensional cloud and wind field data, providing a new digital experiment field for in-depth study of the occurrence and development mechanism of weather systems. For the booming low-altitude economy, the fine prediction of 500 meters to 1 kilometer provided by the model can provide key meteorological risk decision support for unmanned aerial vehicle logistics route planning and urban air traffic (UAM) management, effectively avoiding flight risks caused by low-altitude wind shear, microburst and other sudden weather phenomena. Finally, the core contribution of this work is that it successfully integrates broad-spectrum prediction ability and precise analytical ability through an innovative data organization and training paradigm, laying a key technical cornerstone for building the next generation of intelligent, adaptive, and full-scale unified weather prediction model.

[0105] Further, to test the prediction effect of the model in different regions, two regions were randomly selected as the validation set and test set respectively for prediction verification based on the range covered by the training data set, which were a typhoon process in South China in a certain year and a hail process in Southwest China in a certain year.

[0106] Test data one:

[0107] Please refer to Figure 4 , Figure 4For the typhoon landing in a certain area, the leftmost is the actual observation result, and the right side is the prediction result of the target prediction model with 6-minute, 1-hour, 2-hour, and 3-hour prediction time. From the results, the 6-minute prediction time result is highly consistent with the actual observation, and the typhoon eye position, spiral rain belt structure, and intensity distribution are very close to the actual situation. The typhoon center position is accurate, the eye wall structure is clear, and the distribution of the peripheral spiral rain belt is basically consistent with the observation. This shows that the target prediction model performs excellently in very short-term prediction. The overall structure of the 1-hour prediction time result remains good, and the typhoon center position and main rain belt distribution are basically accurate. However, it can be observed that the reflectivity intensity starts to show a strong trend, especially in the typhoon eye wall and core area, with a 50-60 dBZ range slightly larger than the actual situation. The 2-hour prediction time result shows a more obvious intensity bias, and the high reflectivity value distribution range in the eye wall area further expands. However, the overall spatial position and spiral structure of the typhoon are still well maintained, and the main precipitation belt trend and distribution are basically consistent with the actual situation. The 3-hour prediction time result has the largest prediction error, and the intensity is significantly biased. However, it is worth noting that the model can still roughly capture the position and main rain belt distribution characteristics of the typhoon.

[0108] The prediction accuracy of the model for typhoon still decreases with time, and as the prediction time increases, the model tends to overestimate the precipitation intensity, which may be related to the model parameters. However, as the prediction time increases, the typhoon structure characteristics remain good, and even in the 3-hour prediction, the model can still reasonably depict the spiral structure and main precipitation belt position of the typhoon, indicating that the model has a good grasp of the dynamic structure of the typhoon. Considering the excellent prediction performance of the model within 0-1 hour, it is also suitable for short-term warning business applications of typhoon.

[0109] Please refer to Figure 5 , Figure 5 For the actual observation result of the target prediction model for typhoon at different height layers (upper row from left to right), and the prediction result (lower row from left to right), the different height layers are 3.5 km, 5 km, 6 km, and 7 km. The actual observation result at 3.5 km height layer shows a small and compact typhoon eye, with a clear yellow-orange strong echo band in the eye wall. The prediction result has high overall consistency, with accurate typhoon eye position, similar spiral rain belt shape, and only slightly strong eye wall intensity. The actual observation result at 5 km height layer shows a slightly expanded typhoon eye with moderate echo intensity. This height layer has the best prediction effect, with consistent typhoon structure, intensity distribution, and spiral rain belt details with the observation. The observation at 6 km height layer shows a significantly expanded typhoon eye with a large decrease in echo intensity, mainly with green weak echoes. The prediction captures the expansion trend of the eye area, but the intensity is significantly biased, with more yellow medium-strong echo regions. The observation at 7 km height layer shows a large and clear typhoon eye with overall blue-green weak echoes and loose structure. The prediction result has similar structure but larger intensity.

[0110] The target prediction model exhibits significant height dependence, with high accuracy in the lower layer (3.5-5 km) and systematic overestimation in the mid-high layer (6-7 km). 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 typhoon center position, and the spatial distribution of main rain bands.

[0111] Overall, the target prediction model exhibits good prediction capability for the three-dimensional structure of typhoons, especially in the 3.5-5 km height layer. The model accurately grasps the overall structure, evolution trend, and key features of typhoons, providing reliable technical support for the fine prediction of typhoons.

[0112] Test data two:

[0113] Figure 6 The leftmost side is the actual observation result, and the right side is the prediction result of the target prediction model with 6-minute, 30-minute, 60-minute, and 90-minute prediction times. The observation result shows that at 15 o'clock, multiple strong convective cells appear in the central and western parts of the region, mainly concentrated in a region and its surrounding areas. The strongest echo reaches 50-60 dBZ, showing typical strong convective characteristics. The convective system is distributed in a northeast-southwest direction, with multiple convective cells relatively independent. The 6-minute prediction time result is highly consistent with the observation, with the position, intensity, and shape of the main convective cells very close to the actual situation. The position of the strong echo center in the region is accurate, and the distribution of the surrounding weak echoes is also basically consistent. This indicates that the target prediction model performs excellently in ultra-short-term prediction. The 30-minute prediction time result maintains good overall convective structure, with the position of the main strong echo zone basically accurate. However, there is a tendency to overestimate the intensity, especially the range of the reflectivity high value zone (50-55 dBZ) near the region has expanded compared to the actual situation, and some green weak echo zones have also slightly strengthened. The 60-minute prediction time result shows more obvious overestimation of intensity, not only does the main strong echo zone in the region expand, but the intensity of the surrounding weak convective cells also significantly increases. The spatial distribution pattern is generally maintained, but the convective coverage range is significantly expanded, indicating that the model tends to overestimate convective development. The 90-minute prediction time result has the largest deviation, with a significant increase in the strong echo area and an overlarge convective system coverage range. Although the main convective center position can still be identified, the overall intensity and range are significantly overestimated, and many detailed features are lost.

[0114] The prediction accuracy gradually decreases from 6 minutes to 90 minutes. With the extension of the prediction time, the model tends to overestimate the convective intensity and range, which is common in severe convective weather prediction. However, the model has good spatial positioning ability, and the location of the main convective system is still basically accurate even in the 90-minute prediction. The prediction within 0-1 hour performs well and is suitable for near warning business applications of severe convective weather such as hail.

[0115] Please refer to Figure 7 , Figure 7 The actual observation results (upper row from left to right) and prediction results (lower row from left to right) of different height layers of the target prediction model for this hail process are compared. Different height layers are 3.5 km, 5 km, 6 km and 7 km. In different height layers, the model can accurately predict the spatial distribution characteristics of the main hail system in the region, especially in the middle and low layers (3.5 km and 5 km). The model has excellent recognition ability for the strong convective cell in the region, the reflectivity structure is clear, and the position is consistent with the observation, showing strong spatial consistency.

[0116] In the main reflectivity center of the region, both the prediction and the observation show obvious strong echo characteristics, indicating that the model has good prediction ability for the organization structure of the strong convective system in the region. It is worth noting that the model shows high spatial consistency at each height layer in the region. Although the predicted intensity of the high layer echo is slightly strong, the model performs well in predicting the reflectivity structure at each height layer from 3.5 km to 7 km, showing its ability to predict the three-dimensional structure of the complex convective process in the region.

[0117] Embodiment Two

[0118] The embodiment provides a spatial scalability prediction model of a three-dimensional cloud field, which is applied to the three-dimensional cloud field spatial scalability prediction method in embodiment one. The three-dimensional cloud field time scalability prediction model comprises:

[0119] A data acquisition module is configured to acquire first historical radar networking data, special historical data, and second historical radar networking data.

[0120] A data processing module is configured to construct a first basic training set, a second basic training set, and a special training set.

[0121] A model training module is configured to train an initial prediction model based on the first basic training set, train a special improvement model based on a plurality of special historical data, and train a target prediction model based on the second basic training set and a plurality of special training sets.

[0122] A model construction module is configured to construct the initial prediction model, the special improvement model, and the target prediction 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 the embodiments of the present application can include one or more computer program products, which can include various forms of computer readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM), cache memory, and / or the like. The non-volatile memory may, for example, include read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), and / or the like.

[0132] In the embodiments of the present application, the I / O interface 103 can be used to receive input instructions (such as digital or character information, and generate key signal inputs related to user settings and function control of the electronic device 100, etc.), and can also output various information to the outside (such as images or sounds, etc.). The I / O interface 103 in the embodiments of the present application can include one or more of a physical keyboard, function keys (such as volume control keys, on-off keys, etc.), a mouse, a joystick, a trackball, a microphone, a speaker, and a touch panel, etc.

[0133] It can be understood that, although the operations are described in a particular order in the accompanying drawings in the embodiments of the present application, it should not be understood as requiring the operations to be performed in the particular order shown or in a serial order, or requiring all of the operations to be performed to obtain a desired result. In a particular environment, multitasking and parallel processing can be advantageous.

[0134] The methods and apparatuses involved in the embodiments of the present application can be completed 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 words "apparatus" and "module" 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 of the steps, operations, or procedures described herein can be performed or implemented using one or more hardware or software modules, alone or in combination with other devices. In one embodiment, a software module is implemented using a computer program product including a computer readable medium containing computer program code, which can be executed by a computer processor for performing any or all of the steps, operations, or procedures described.

[0136] Those of ordinary skill in the art can understand that the above embodiments are specific cases for implementing the present application, and in actual applications, various changes can be made in form and details without departing from the scope of the present application.

Claims

1. A method for spatial scalability prediction of a three-dimensional cloud field, characterized in that, The method comprises: constructing a first basic training set based on preprocessed first historical radar networking data; inputting the first basic training set into a model to obtain an initial prediction model; inputting multiple special historical data into the initial prediction model for special recognition ability training to obtain a special improvement model; constructing a second basic training set based on preprocessed second historical radar networking data, and processing the second basic training set based on the special improvement model to screen and construct multiple special training sets; wherein the data quantity of the first historical radar networking data is less than the data quantity of the second historical radar networking data; inputting the second basic training set and the multiple special training sets into the special improvement model for multi-resolution joint training to obtain a target prediction model.

2. The method of claim 1, wherein the second basic training set is constructed based on the preprocessed second historical radar networking data, comprising: obtaining the second historical radar networking data covering the target area; preprocessing the second historical radar networking data; wherein the preprocessing comprises: upscaling the basic spatial resolution of the second historical radar networking data to a preset spatial resolution through a grid aggregation and averaging algorithm; and resampling the second historical radar networking data at the preset spatial resolution onto a target regular latitude-longitude grid to form the second basic training set.

3. The method of claim 2, wherein the second basic training set is processed based on the special improvement model to screen and construct multiple special training sets, comprising: inputting the second basic training set into the special improvement model; the special improvement model extracts radar echo data sequences of typhoon systems, cold vortex systems, squall line systems and local strong convective systems; based on the radar echo data sequences of the typhoon systems, cold vortex systems, squall line systems and local strong convective systems, multiple special training sets are obtained; wherein the multiple special training sets are respectively a typhoon training set, a cold vortex training set, a squall line training set and a local strong convective training set.

4. The method of claim 3, wherein the typhoon training set comprises a first spatial resolution, a first coverage range and a first continuous time sequence; the cold vortex training set comprises a second spatial resolution, a second coverage range and a second continuous time sequence; the squall line training set comprises a third spatial resolution, a third coverage range and a third continuous time sequence; and the local strong convective training set comprises a fourth spatial resolution, a fourth coverage range and a fourth continuous time sequence; wherein the first spatial resolution, the second spatial resolution, the third spatial resolution and the fourth spatial resolution are less than the preset spatial resolution.

5. The method of claim 4, wherein ​ ​ ​ The second basic training set and the plurality of special training sets are input into the special promotion model for multi-resolution joint training to obtain a target prediction model, and the method further comprises: constructing a scale verification set and a test set based on the preprocessed second historical radar networking data; wherein the scale verification set comprises a typhoon and cold vortex large-scale verification set and a squall line and strong convection small-scale verification set; verifying the target prediction model based on the scale verification set; testing the target prediction model based on the test set.

6. The method according to claim 5, wherein the verifying the target prediction model based on the scale verification set comprises: inputting historical sequence data in the typhoon and cold vortex large-scale verification set into the target prediction model for verification to obtain a first predicted radar echo sequence; extracting running path, intensity characteristic quantity and precipitation distribution under the first spatial resolution and the second spatial resolution from the first predicted radar echo sequence; comparing the running path, intensity characteristic quantity and precipitation distribution with first actual observation data to calculate path error, intensity error and precipitation score; inputting historical sequence data in the squall line and strong convection small-scale verification set into the target prediction model for verification to obtain a second predicted radar echo sequence; extracting trigger time, moving vector, strong center position and fine structure feature under the third spatial resolution and the fourth spatial resolution from the second predicted radar echo sequence; comparing the trigger time, moving vector, strong center position and fine structure feature with second actual observation data to calculate trigger time error, moving error, position error and structure coincidence degree index.

7. The method according to claim 1, wherein the method further comprises: inputting a target spatial resolution into the target prediction model to output a radar echo sequence of the target spatial resolution. The method system of claim 1-7, wherein the method system comprises: a data acquisition module configured to acquire first historical radar networking data, special historical data and second historical radar networking data; 8. A method system for spatial scalability forecasting of three-dimensional cloud fields, characterized in that, a data processing module configured to construct a first basic training set, a second basic training set and a special training set; a model training module configured to train an initial prediction model based on the first basic training set, train a special promotion model based on a plurality of special historical data, and train a target prediction model based on the second basic training set and the plurality of special training sets; a model construction module configured to construct the initial prediction model, the special promotion model and the target prediction model. a processor; a memory configured to store executable instructions; 9. A method of spatial scalability prediction of a three-dimensional cloud field, characterized by, ​ ​ ​ The processor is configured to read the executable instructions from the memory and execute the executable instructions to implement the method for predicting spatial extensibility of a three-dimensional cloud field according to any one of claims 1-7. 10.A computer readable storage medium, comprising instructions stored thereon, wherein the instructions, when executed by a processor, perform the method for predicting spatial extensibility of a three-dimensional cloud field according to any one of claims 1-7. 10.A computer readable storage medium, comprising instructions stored thereon, wherein the instructions, when executed by a processor, perform the method for predicting spatial extensibility of a three-dimensional cloud field according to any one of claims 1-7.

Images