Low-altitude weather forecast vertical encryption method and device based on deep learning, electronic equipment and storage medium

Abstract

The invention relates to the technical field of meteorological data processing, in particular to a low-altitude weather forecast vertical encryption method and device based on deep learning, electronic equipment and a storage medium, and the method comprises the steps: obtaining three-dimensional weather forecast field data which is outputted by a global or regional numerical weather forecast mode and has a first vertical resolution; and preprocessing the three-dimensional weather forecast field data to obtain preprocessed three-dimensional weather forecast field data. According to the invention, by means of the pre-trained deep learning model, direct and efficient nonlinear mapping from a low-vertical-resolution three-dimensional meteorological forecasting field to a high-vertical-resolution three-dimensional meteorological element forecasting field is established and realized, so that a low-resolution numerical forecasting product input in real time is quickly processed; and a refined three-dimensional meteorological field of which the number of vertical layers is remarkably increased is generated, so that key data support with higher precision is provided for low-altitude flight safety decision making.

Description

Technical Field

[0001] This invention relates to the field of meteorological data processing technology, and in particular to a method, apparatus, electronic device, and storage medium for vertical densification of low-altitude meteorological forecasts based on deep learning. Background Technology

[0002] Low-altitude airspace (generally referring to the airspace below 3000 meters above sea level) is the primary operating space for unmanned aerial vehicles (UAVs), general aviation aircraft, and urban air traffic vehicles. This atmospheric layer is directly affected by topography, surface cover, and human activities. Its meteorological elements (such as wind, temperature, and humidity) exhibit large vertical gradients and drastic changes, easily generating microscale turbulence, strong wind shear, and temperature inversion layers that endanger flight safety. Therefore, obtaining a refined three-dimensional meteorological field with high spatiotemporal resolution, especially high vertical resolution, is a key prerequisite for realizing low-altitude flight safety planning and intelligent control.

[0003] Currently, operational global or regional numerical weather prediction models (such as ECMWF and GRAPES) are the primary data source for weather forecasting. However, limited by computational resources and model theory, these models typically have low vertical resolution, with only a few layers in the near-surface to low-altitude range, making it difficult to accurately characterize the subtle evolution of meteorological structures within the boundary layer. To improve resolution, traditional methods mainly rely on increasing the number of vertical layers in the numerical model itself or performing dynamic downscaling, but this leads to an exponential increase in computational costs, failing to meet the real-time requirements of rapid meteorological information updates for low-altitude flights. Furthermore, some studies have attempted to obtain more refined wind fields by coupling mesoscale models with computational fluid dynamics models, but these methods are mostly specific to particular scenarios, computationally time-consuming, and complex, making them unsuitable for large-scale, operational forecasting services.

[0004] In recent years, data-driven deep learning methods have provided new insights into refining weather forecasts. Models such as convolutional neural networks can effectively extract spatial features from meteorological data and have shown potential in areas such as meteorological element prediction and forecast error correction. However, existing research largely focuses on improving horizontal resolution or predicting single meteorological elements. Deep learning solutions specifically designed for low-altitude flight safety, aiming to efficiently and accurately improve the vertical resolution of meteorological fields, have not yet been publicly reported. Therefore, developing an intelligent encryption method that can overcome the bottlenecks of traditional numerical models and rapidly generate high-vertical-resolution three-dimensional meteorological fields has significant application value and urgency. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method, apparatus, electronic device, and storage medium for vertical encryption of low-altitude meteorological forecasts based on deep learning.

[0006] In a first aspect, embodiments of the present invention provide a deep learning-based method for vertical densification of low-altitude meteorological forecasts, comprising: Acquire three-dimensional meteorological forecast field data with first vertical resolution output from global or regional numerical weather prediction models; The three-dimensional meteorological forecast field data is preprocessed to obtain the preprocessed three-dimensional meteorological forecast field data; The preprocessed 3D meteorological forecast field data is input into a pre-trained deep learning vertical densification model, which outputs a 3D meteorological element forecast field with a second vertical resolution. The second vertical resolution is higher than the first vertical resolution, and the deep learning vertical encryption model is used to establish a non-linear mapping relationship from the first vertical resolution to the second vertical resolution.

[0007] In conjunction with the first aspect, the deep learning vertical encryption model is based on an encoder-decoder neural network architecture; The encoder consists of multiple three-dimensional convolutional layers, which are used to extract features and downsample the input three-dimensional weather forecast field data. The decoder consists of multiple three-dimensional deconvolutional layers, which are used to upsample the features extracted by the encoder and reconstruct them into a three-dimensional meteorological element forecast field with a second vertical resolution.

[0008] In conjunction with the first aspect, a skip connection is set between the corresponding layers of the encoder and the decoder to pass the low-level spatial detail features extracted by the encoder to the decoder and fuse them with the high-level semantic features of the corresponding layer of the decoder.

[0009] In conjunction with the first aspect, an attention mechanism module is integrated into the decoder to enhance the attention weight of key vertical layers, including the boundary layer, wind shear layer, or inversion layer, when the model reconstructs the three-dimensional meteorological element forecast field.

[0010] In conjunction with the first aspect, the training methods for deep learning vertical encryption models include: Obtain the training dataset, which includes historical forecast data samples output by numerical weather prediction models and their corresponding spatiotemporal high-resolution reanalysis data labels; The historical forecast data samples and high-resolution reanalysis data labels are preprocessed to obtain preprocessed historical forecast data samples and preprocessed high-resolution reanalysis data labels. Using preprocessed historical forecast data samples as input and preprocessed high-resolution reanalysis data labels as supervision targets, a deep learning vertical encryption model is trained by optimizing the loss function.

[0011] In conjunction with the first aspect, the loss function is a joint loss function, which includes at least the mean square error loss for the wind vector field and the L1 loss for the wind speed scalar field.

[0012] In conjunction with the first aspect, after inputting the preprocessed 3D meteorological forecast field data into a pre-trained deep learning vertical densification model and outputting a 3D meteorological element forecast field with a second vertical resolution, the method further includes: Based on a three-dimensional meteorological element forecast field with a second vertical resolution, the vertical wind shear intensity at key points along a specified flight path is calculated. In response to the vertical wind shear intensity exceeding a preset alarm threshold, a flight risk alarm message is generated.

[0013] Secondly, embodiments of this application also provide a deep learning-based vertical encryption device for low-altitude meteorological forecasts, comprising: The data acquisition module is used to acquire three-dimensional meteorological forecast field data with first vertical resolution output by global or regional numerical weather prediction models. The preprocessing module is used to preprocess the three-dimensional weather forecast field data to obtain preprocessed three-dimensional weather forecast field data. The vertical encryption module is used to input the preprocessed three-dimensional meteorological forecast field data into the pre-trained deep learning vertical encryption model and output a three-dimensional meteorological element forecast field with a second vertical resolution. The second vertical resolution is higher than the first vertical resolution, and the deep learning vertical encryption model is used to establish a non-linear mapping relationship from the first vertical resolution to the second vertical resolution.

[0014] Thirdly, this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor runs the computer program to cause the electronic device to perform the above-described method.

[0015] Fourthly, this application provides a readable storage medium storing computer program instructions, which are read and executed by a processor to perform the above-described method.

[0016] The embodiments of the present invention bring the following beneficial effects: This application provides a method, device, electronic device, and storage medium for vertical densification of low-altitude meteorological forecasts based on deep learning. The method includes: acquiring three-dimensional meteorological forecast field data with a first vertical resolution output by a global or regional numerical weather prediction model; preprocessing the three-dimensional meteorological forecast field data to obtain preprocessed three-dimensional meteorological forecast field data; inputting the preprocessed three-dimensional meteorological forecast field data into a pre-trained deep learning vertical densification model to output a three-dimensional meteorological element forecast field with a second vertical resolution; wherein, the second vertical resolution is higher than the first vertical resolution, and the deep learning vertical densification model is used to establish a nonlinear mapping relationship from the first vertical resolution to the second vertical resolution.

[0017] This application utilizes a pre-trained deep learning model to establish and implement a direct and efficient nonlinear mapping between a low-vertical-resolution 3D meteorological forecast field and a high-vertical-resolution 3D meteorological element forecast field. This enables rapid processing of real-time low-resolution numerical forecast products, generating a refined 3D meteorological field with a significantly increased number of vertical layers, thereby providing more accurate key data support for low-altitude flight safety decisions.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This application provides a flowchart illustrating a deep learning-based method for vertical densification of low-altitude meteorological forecasts. Figure 2 This application provides a schematic diagram of the structure of a deep learning-based vertical encryption device for low-altitude meteorological forecasting. Figure 3 This is a schematic diagram of the electronic device structure provided in an embodiment of the present invention.

[0022] Figure label: 10 - Acquisition module, 20 - Preprocessing module, 30 - Vertical encryption module; 130 - Processor, 131 - Memory, 132 - Bus, 133 - Communication interface. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] To facilitate understanding of this embodiment, the technical terms used in this application will be briefly introduced below.

[0025] Vertical densification refers to the process in meteorology of improving the resolution of data or forecast products in the vertical direction (i.e., the direction of altitude or air pressure) through technical means; it is a type of downscaling in a specific direction.

[0026] A skip connection is a connection method in an encoder-decoder network that directly ("skips") the output feature map of a certain layer of the encoder to the corresponding symmetric layer of the decoder.

[0027] Vertical wind shear intensity refers to the amount or rate of change of wind speed and / or wind direction per unit distance (usually hundreds of meters or kilometers) in the vertical direction.

[0028] High-resolution reanalysis data refers to a type of gridded meteorological dataset that is spatiotemporally consistent and has high resolution, recalculated using global historical meteorological observation data (satellite, radiosonde, ground station, etc.) through advanced data assimilation techniques and numerical models. It is typically used as a ground-based reference for climate research or model training.

[0029] After introducing the technical terms used in this application, the application scenarios and design concepts of the embodiments of this application will be briefly described below.

[0030] Low-altitude flight requires high-precision three-dimensional meteorological data to ensure safety. Traditional numerical models lack sufficient vertical resolution, and improving them is costly; existing deep learning methods are not specifically designed for vertical encryption. Therefore, it is necessary to develop efficient and accurate vertical encryption technologies to fill this gap.

[0031] Based on this, this application provides a method, apparatus, electronic device, and storage medium for vertical encryption of low-altitude meteorological forecasts based on deep learning.

[0032] Example 1 This application provides a deep learning-based method for vertical densification of low-altitude meteorological forecasts, combining... Figure 1 As shown, the method includes: S110: Acquire three-dimensional meteorological forecast field data with first vertical resolution output from a global or regional numerical weather prediction model.

[0033] S120 preprocesses the three-dimensional weather forecast field data to obtain preprocessed three-dimensional weather forecast field data.

[0034] S130 inputs the preprocessed 3D meteorological forecast field data into the pre-trained deep learning vertical densification model and outputs a 3D meteorological element forecast field with a second vertical resolution.

[0035] The second vertical resolution is higher than the first vertical resolution, and the deep learning vertical encryption model is used to establish a non-linear mapping relationship from the first vertical resolution to the second vertical resolution.

[0036] This application uses a pre-trained deep learning model to quickly and intelligently encrypt the low vertical resolution products output by traditional numerical weather prediction models into high-resolution three-dimensional meteorological fields that meet the needs of low-altitude flight.

[0037] In step S110, numerical weather prediction models refer to mathematical model systems that predict future weather by solving atmospheric physics equations using supercomputers. Among them, the ECMWF (European Centre for Medium-Range Weather Forecasts) model is one of the most representative high-precision global models.

[0038] The embodiments of this application obtain ERA5 data, a direct output product of a large-scale numerical weather prediction model (such as the ECMWF global model) widely used in operations. This type of data is the cornerstone of weather forecasting, but due to limitations in computing resources, its vertical resolution is low, i.e., the first vertical resolution is coarse. For example, ECMWF data may only have 9 standard pressure layers in the near-surface to mid-to-low-level atmosphere, with large intervals between layers (e.g., 25-100 hPa), making it difficult to depict the subtle vertical structure of the lower atmosphere.

[0039] Three-dimensional meteorological forecast field data refers to data that forms a grid structure in three dimensions: longitude, latitude, and air pressure altitude. Each grid point contains the forecast value of specific meteorological elements (such as wind and temperature).

[0040] Step S120 aims to adapt the original pattern output data to the input requirements of the deep learning model and improve the data quality to meet the input requirements of the deep learning model. This step is a key preprocessing step to ensure the effectiveness and stability of the model.

[0041] In this embodiment, preprocessing specifically includes quality control (removing or correcting obvious errors) and standardization (unifying data scale). Quality control includes boundary value checks and spatial consistency checks.

[0042] Step S120 includes: S121. Based on preset limit value constraints and spatial consistency constraints, the three-dimensional meteorological forecast field data is initially processed to obtain intermediate data.

[0043] Step S121 is the core of data cleaning and trustworthiness processing. Its purpose is to identify and process various errors in the data, producing physically reliable and spatially reasonable intermediate data. Specifically: Boundary value constraints (physical rationality filtering) are used to set rigid judgment rules based on known physical limits or climatological extremes of meteorological elements. For example, for near-surface temperatures in China, the preset boundary value can be set to -80℃ to 60℃. Any data points outside this range will be automatically marked as invalid or erroneous by the system and discarded or conservatively interpolated. This check acts as a first coarse sieve, quickly filtering out the most obvious, outrageous noise and errors.

[0044] Spatial consistency constraints (spatial rationality verification) are used to further analyze the smoothness and continuity of the spatial distribution of data, building upon data that has passed physical boundary checks. By calculating the spatial gradients of the data in the meridional and latitudinal directions (i.e., the differences between adjacent grid points), it determines whether there are local abrupt changes. For example, if the difference between the wind speed at a point and the average wind speed of all its neighboring points exceeds a preset reasonable threshold (e.g., 10 m / s), that point is identified as a spatially isolated anomaly. This check acts like a second fine sieve, used to capture hidden errors where the numerical values ​​themselves are within the physical range but are severely inconsistent with their surrounding environment.

[0045] Based on the two constraints mentioned above, the resulting data is considered intermediate data. It has largely eliminated significant errors and noise, and possesses preliminary credibility and rationality in terms of physical values ​​and spatial distribution, laying the foundation for the next stage of precise scaling. Understandably, these thresholds or judgment rules are based on general meteorological knowledge, historical statistics, or configurable parameters.

[0046] S122, standardize the intermediate data to obtain preprocessed three-dimensional weather forecast field data.

[0047] Step S122 is data scaling and model adaptation processing, the purpose of which is to convert the cleaned intermediate data into a numerical form that is suitable for efficient processing by deep learning models.

[0048] Understandably, different meteorological elements have different dimensions and numerical ranges (e.g., wind speed is measured in m / s, with a value of 0-30; temperature is measured in °C, with a value of -20 to 40). Directly inputting these vastly different raw data into the model would lead to unstable model training, slow convergence, and an imbalance in the influence of each element on the model weights. Furthermore, standardized parameters (mean and standard deviation) must be calculated based on clean data. If standardization is performed first, the extreme outliers to be removed in S121 will severely distort the calculated mean and standard deviation. For example, an abnormally high temperature value that should be removed will raise the overall mean and widen the standard deviation, causing all normal data to have distorted numerical distributions after this contaminated standardization transformation, thus polluting the entire dataset and misleading the deep learning vertical encryption model.

[0049] Therefore, after obtaining reliable intermediate data in step S121, step S122 uses mathematical transformation (standardization) to adjust the intermediate data of each meteorological element to a uniform numerical scale. The most commonly used method is Z-score standardization, which involves subtracting the mean from the data and then dividing by the standard deviation, so that the processed data distribution conforms to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0050] Subsequently, in step S130, a dedicated model that has been trained and has encryption capabilities (i.e., a deep learning vertical encryption model) is used to map the low-resolution three-dimensional atmospheric field forecast field obtained in step S110 into a high-resolution three-dimensional meteorological element forecast field.

[0051] In conjunction with the first aspect, the deep learning vertical encryption model in step S130 is a neural network architecture based on encoder-decoder; The encoder consists of multiple three-dimensional convolutional layers, which are used to extract features and downsample the input three-dimensional weather forecast field data. The decoder consists of multiple three-dimensional deconvolutional layers, which are used to upsample the features extracted by the encoder and reconstruct them into a three-dimensional meteorological element forecast field with a second vertical resolution.

[0052] In this embodiment, a neural network architecture based on 3DU-Net encoder-decoder is adopted. The encoder is used to understand the essential features of the input data, and the decoder is used to generate target data based on the essential features of the input data understood by the encoder.

[0053] The encoder consists of multiple 3D convolutional layers, meaning it's a deep network composed of stacked layers. Its core mission is to extract deep features and compress information from the preprocessed 3D meteorological field input (e.g., dimensions of [number of longitude grids, number of latitude grids, 9 pressure layers, number of meteorological elements]). Specifically: Feature extraction (3D convolution operation): Each 3D convolution layer uses a set of learnable 3D convolution kernels (e.g., 3x3x3 in size) that slide synchronously across the three dimensions of longitude, latitude, and barometric altitude of the input data. This process intelligently detects and extracts spatial-vertical joint features from local to regional scales.

[0054] As an example, shallow convolutional kernels may learn to recognize simple patterns of local wind convergence / divergence (manifested as specific combinations of u / v wind components in the horizontal direction) or local anomalies in the vertical temperature lapse rate. Deep convolutional kernels, by combining these simple features, are able to recognize more complex structures, such as a complete boundary layer top inversion layer (manifested as a vertical profile feature where temperature increases rather than decreases with altitude) or the baroclinic structure of a mesoscale convective system (manifested as a specific configuration of the temperature and wind fields in three-dimensional space).

[0055] Downsampling and Information Abstraction: Between convolutional operations, max pooling or convolutions with a stride of 2 are used to systematically halve the size of the feature map in the three spatial dimensions (longitude, latitude, and number of layers) (e.g., from [100, 100, 9] to [50, 50, 4]). This process is called downsampling. With each downsampling operation, the receptive field of a single neuron (a point on the feature map) expands exponentially. This means that higher-level neurons can integrate and expand the recognition of atmospheric states over a larger spatial range. Simultaneously, it forces the network to discard redundant, localized details, retaining and strengthening the features most discriminative of the global structure. Ultimately, the original data is encoded into a small but highly information-dense low-dimensional latent vector, which encapsulates the essential understanding of the global weather conditions and key physical processes of the input low-resolution field.

[0056] The decoder is symmetrical to the encoder and consists of multiple 3D transposed convolutional layers. Its task is to inversely perform an upsampling process using the latent feature vectors output by the encoder to reconstruct a high-resolution 3D weather field.

[0057] Upsampling and spatial size restoration: 3D deconvolution is the inverse operation of convolution. It can intelligently fill and expand the compressed low-dimensional feature map in three dimensions, gradually restoring its spatial size (e.g., upsampling from [50, 50, 4] back to [100, 100, 16]).

[0058] Feature reconstruction and numerical generation: The key to the decoder lies not only in scaling up the size, but also in filling in the details based on the learned physical laws. It needs to translate the semantic concept of "large-scale low-pressure system accompanied by strong southwest wind" abstracted by the encoder into precise u and v wind speed and temperature values ​​at each latitude and longitude grid point on 16 specific levels of [1000, 975, 950...500]hPa.

[0059] Example: When the encoder identifies a potential clue of "low-level wind shear" in the input field, the decoder, during reconstruction, must not only generate a trend showing a significant change in wind speed with height at the corresponding vertical location (e.g., between 925 hPa and 875 hPa), but also ensure that this change is continuous and physically plausible in the horizontal direction, rather than being random noise. It learns to infer the expected wind field structure of the intermediate layer (e.g., 875 hPa) from limited low-level inputs (e.g., winds at 900 hPa and 850 hPa).

[0060] For example, assuming the input is the ECMWF9 layer forecast (1000, 950...500 hPa), the goal is to output the ERA516 layer analysis field (1000, 975...500 hPa).

[0061] Encoding process: The input data ([H, W, 9, C], where C is the number of feature channels) is convolved and downsampled layer by layer by the encoder, and finally compressed into a feature tensor ([H / 16, W / 16, 1, 512]). This 512-dimensional vector can be interpreted as a digital summary containing comprehensive information such as "the region is currently controlled by the westerly wind belt, the boundary layer is stable, and there is a weak trough line moving eastward in the middle layer".

[0062] Decoding process: The feature vector is input into the decoder, which begins deconvolution and upsampling layer by layer. In the early stages, it first reconstructs the basic framework of the large-scale circulation; as the size is restored and the low-level detailed features obtained through skip connections are incorporated, it begins to depict details: near the ground (1000-900 hPa), it finely depicts the wind speed reduction and wind direction deflection caused by topographic friction; at the top of the boundary layer (about 900-850 hPa), it clearly generates an inversion layer, causing the temperature profile to change direction here; across the entire profile, it generates a smooth and dynamically constrained vertical wind speed profile, filling in the missing intermediate layer information in the original 9 layers of data.

[0063] This encoder-decoder architecture is a powerful data-driven function approximator. Through training, it internalizes an extremely complex set of physical mapping functions from sparse vertical observations (low-resolution forecasts) to continuous vertical structures (high-resolution fields). This is completely different from traditional interpolation (which only performs mathematical smoothing) or dynamic downscaling (which requires running expensive physical models). Instead, it learns the statistical and physical laws inherent in historical data and performs fast and intelligent physical inference and detail generation on new low-resolution inputs, thereby achieving revolutionary vertical resolution encryption and providing unprecedentedly refined, real-time three-dimensional atmospheric state data for low-altitude flights.

[0064] In conjunction with the first aspect, a skip connection is set between the corresponding layers of the encoder and the decoder to pass the low-level spatial detail features extracted by the encoder to the decoder and fuse them with the high-level semantic features of the corresponding layer of the decoder.

[0065] The SkipConnection set between corresponding levels of the encoder and decoder is specifically designed to solve two core problems that may occur in the reconstruction of meteorological fields in deep encoder-decoder networks: detail blurring and non-physical abrupt changes. This improves the model encryption effect and ensures the physical rationality of the generated high-resolution field.

[0066] In a standard encoder-decoder process, after the input data is downsampled multiple times by the encoder, high-level semantic features (such as the type of weather system and large-scale circulation patterns) can be extracted. However, low-level, fine spatial details (such as subtle changes in wind speed caused by local topography and small-scale temperature gradients) are inevitably lost or attenuated during information compression. Subsequently, the decoder relies solely on these highly abstract features for upsampling and reconstruction, which often fails to recover these lost details. This can easily result in an overly smoothed meteorological field lacking fine structure, which may manifest in meteorology as blurred frontal zones and overly idealized vertical profiles, thus losing its practical application value.

[0067] This application introduces a skip connection to establish a direct high-speed information channel, thus solving the aforementioned problem. Its working principle and specific function are as follows: Skip connections directly pass (or skip) feature maps output from an intermediate layer of the encoder (e.g., after the second convolutional layer) to the corresponding layer in the decoder (e.g., before the penultimate deconvolutional layer). These feature maps from the encoder, due to their shallow network depth, have not undergone depth abstraction and downsampling, thus retaining rich low-level spatial details and finer grid-scale features. In the corresponding layer of the decoder, these raw details are concatenated along the channel dimension or added element-wise with feature maps generated by the decoder through upsampling, which contain high-level semantic information (e.g., "this is a developing low-pressure vortex").

[0068] When reconstructing high-resolution fields, the decoder no longer relies solely on abstract concepts but also possesses detailed sketches from the encoder. This allows the model to accurately recover minute structures that were originally present in the low-resolution input but might have been blurred by subsequent processing. For example, when reconstructing near-surface wind fields, the low-level features provided by skip connections may include initial convergence line signals caused by terrain. By combining this with the high-level semantics of the overall system velocity, the decoder can generate a finely detailed convergence line with precise location and clear structure, rather than a uniform wind field.

[0069] Changes in meteorological fields are physically continuous and conform to certain dynamic constraints. Simple decoder upsampling can sometimes produce drastic, spatially discontinuous changes that violate physical laws (i.e., non-physical abrupt changes). Low-level features transmitted via skip connections serve as anchors or constraints from the original data, guiding the decoder's generation process and ensuring that the reconstructed field maintains a consistent level of detail with the input low-resolution field, resulting in a smooth transition and effectively suppressing the generation of such non-physical noise. For example, in temperature field encryption, it can prevent the generation of an anomalously high or low temperature layer that violates atmospheric thermodynamics between two known temperature layers.

[0070] Furthermore, from the perspective of deep learning training, skip connections create short paths from shallow to deep layers of the network, which helps to transmit gradients more effectively during backpropagation, alleviates the gradient vanishing problem common in deep networks, and enables this complex model to be trained more stably and efficiently.

[0071] As an example: Suppose the model is processing a forecast field containing a terrain-induced low-level jet stream. The encoder has a deep understanding of the macroscopic semantics of "the existence of a low-level jet stream." Through skip connections, fine terrain features captured by the encoder's shallow layers, such as "initial acceleration on the leeward slope of the mountain" and "the narrowing effect of the valley passage," are directly fed to the decoder. Ultimately, the decoder fuses these two types of information, generating not only a refined field confirming the existence of a jet stream, but also precisely depicting the core height of this jet stream, the location of its maximum wind speed, and its precise spatial coupling with the terrain in both vertical and horizontal directions. These details are crucial for drones to avoid the risk of sudden strong winds.

[0072] Therefore, JumpLink organically combines data-driven detailed memory with physical law reasoning from model learning, thereby improving the generated vertical encryption products. These products are not only images with higher resolution, but also professional meteorological analysis fields that are detailed, reliable, physically consistent, and can be directly used for precise assessment of low-altitude flight safety.

[0073] In conjunction with the first aspect, an attention mechanism module is integrated into the decoder to enhance the attention weight of key vertical layers, including the boundary layer, wind shear layer, or inversion layer, when the model reconstructs the three-dimensional meteorological element forecast field.

[0074] The attention mechanism module integrated inside the decoder in the deep learning vertical encryption model provided in this embodiment aims to enable the deep learning vertical encryption model to overcome the limitations of uniform processing when reconstructing high-resolution three-dimensional meteorological fields, and dynamically identify and strengthen the modeling weights of specific vertical layers and key meteorological processes that pose a direct threat to flight safety.

[0075] Specifically, this attention mechanism operates as a learnable embedded sub-network. It first performs deep analysis on the intermediate feature maps in the decoder, generating an attention weight map corresponding to the spatial dimensions of the feature maps. This weight map quantifies the importance of each location in the feature maps (especially different vertical levels) for accurately reconstructing the target region. Subsequently, by element-wise multiplying the weight map with the original feature map, the features are recalibrated: features representing key regions (such as boundary layer turbulence, wind shear layers, or inversion layers) are significantly enhanced, while features in relatively stable regions are moderately suppressed. The process of generating the attention weight map corresponding to the spatial dimensions of the feature maps is a dynamic, data-dependent computational flow, rather than using a pre-set fixed template. This process can be decomposed into the following key operations: input and feature transformation, similarity calculation and weight generation, and normalization and weight map formation. Specifically, the input and feature transformation is performed first. The attention mechanism module receives the intermediate feature maps from the current layer of the decoder as input. This four-dimensional tensor contains rich spatial information and feature encoding. To calculate attention, the module maps the input feature map into query and key feature maps using two independent, learnable linear transformation layers (typically 1×1×1 3D convolutions). The query feature map can be understood as the representation of the target location that needs to be "added with details" in the current decoding step, while the key feature map represents all potential information sources from the input features that can be referenced and invoked. Next, similarity calculation and weight generation are performed. The core of this step is to evaluate the relevance between the query and key. Specifically, the similarity between the query and key feature maps is calculated along a specific dimension (usually the channel dimension). Common methods include calculating the dot product or using a learnable additive network. For example, in a vertical encryption scenario, the model calculates the similarity between the query vector at a vertical location (such as the feature corresponding to the boundary layer) and the key vectors at all other vertical locations. This calculation produces an initial attention score matrix, where each value represents the degree of dependence or attention to each other feature location when reconstructing the current target feature. Finally, normalization and weight map formation are performed. The generated initial attention score matrix is ​​normalized using the Softmax function. The Softmax function transforms all scores into a probability distribution, normalizing the final weight value for each location to between 0 and 1, with the sum of all weights being 1. This normalized matrix is ​​the final attention weight map. Its spatial dimensions are identical to the input feature map, and each value precisely quantifies the contribution weight that information from the corresponding location in the feature map should be assigned when reconstructing the final high-resolution weather field.Taking a specific meteorological process as an example: when the decoder processes a feature map containing potential wind shear signals, the attention mechanism, through the aforementioned dynamic calculations, may discover that features representing a specific altitude range (e.g., between 850 hPa and 700 hPa) in the feature map are highly correlated with the "strong vertical wind gradient" pattern. Therefore, in the generated weight map, the spatial regions corresponding to these altitude layers will receive extremely high weight values. In the subsequent recalibration step, the wind field feature signals from these key layers are significantly enhanced, guiding the model to reconstruct a more refined vertical wind speed profile with steeper gradient changes within this altitude range in the final encrypted output, accurately characterizing the wind shear layer. This achieves an intelligent process of automatically learning risk patterns from data and performing targeted focus encryption.

[0076] In this embodiment, the addition of the attention mechanism module facilitates targeted encryption. The deep learning vertical encryption model, by learning the correlation patterns of severe weather in historical data, can automatically focus more computational attention and representational capabilities on signals in the input features that predict strong vertical wind shear, significant temperature inversions, or complex boundary layer disturbances. Therefore, in the final high-resolution product, not only is the overall number of vertical layers increased, but also the level of detail recovery and physical realism at critical security levels such as the height of wind shear and the interface where the inversion layer is located far surpasses that of uniform encryption. For example, for a terrain-induced low-level wind shear process, the attention mechanism can guide the model to more accurately depict the abrupt changes in wind speed within the shear layer in the encrypted wind field, rather than smoothing it out.

[0077] Therefore, the integration of the attention mechanism module essentially involves deeply integrating prior knowledge of low-altitude flight safety into the model's learning and reasoning process in a data-driven manner. This ensures that the generated vertical encryption product is not only spatially more refined data, but also a specialized decision support field that directly serves safety risk assessment and provides high-confidence early warnings for potentially hazardous areas, greatly enhancing the business application value and reliability of the entire technical solution.

[0078] In conjunction with the first aspect, the training methods for deep learning vertical encryption models include: S210, Obtain the training dataset, which includes historical forecast data samples output by numerical weather prediction models and their corresponding high-resolution reanalysis data labels in time and space.

[0079] The training dataset is constructed as follows: for the same historical moment, forecast field data with a forecast lead time of T hours from the output of numerical weather prediction models are collected as historical forecast data samples; simultaneously, high-resolution reanalysis data at the corresponding analysis time (i.e., historical moment + T hours) are collected as labels. This method ensures that the forecast data and label data in each sample pair are aligned in physical time, simulating real-time forecasting application scenarios.

[0080] High-resolution reanalysis data labels are high-quality data that precisely correspond to each historical forecast sample in time (the same forecast time) and space (the same geographic region), typically using reanalysis products such as ERA5. This type of data fuses global observations and has a higher second vertical resolution (e.g., 16 layers), and is widely regarded in meteorology as the analytical field or ground reality that most closely approximates the true atmospheric conditions. Here, it serves as a supervised target, i.e., the standard answer that the model strives to approximate.

[0081] In this way, by providing a massive number of question-answer pairs, the model can discover and internalize the stable, physically consistent correspondence between the two through statistical learning.

[0082] S220, preprocess the historical forecast data samples and high-resolution reanalysis data labels respectively to obtain preprocessed historical forecast data samples and preprocessed high-resolution reanalysis data labels.

[0083] Understandably, the collected training data may contain errors and noise, and have different dimensions and numerical ranges. If used directly for training without processing, the model may learn noisy patterns or have difficulty converging due to differences in data scale. The preprocessing process is exactly the same as the preprocessing procedure in step S120, and will not be described in detail here.

[0084] S230 uses preprocessed historical forecast data samples as input and preprocessed high-resolution reanalysis data labels as supervision targets to train a deep learning vertical encryption model by optimizing the loss function.

[0085] The deep learning vertical encryption model repeatedly executes an iterative loop of forward propagation, loss calculation, backpropagation, and parameter update on a large amount of data, constituting a complete training epoch. Each loop specifically includes: Forward propagation: A batch of preprocessed historical forecast data samples is input into the model. The data flows through the model's encoder, decoder, and all integrated sub-modules (such as skip connections and attention mechanisms), ultimately outputting the model's predictions for high-resolution fields.

[0086] In one specific implementation of this embodiment, the deep learning vertical encryption model adopts the 3D U-Net architecture, and its specific construction parameters are as follows: The encoder consists of four downsampling stages. Each stage comprises two 3×3×3 convolutional layers (each followed by a batch normalization layer and a ReLU activation function) and a 3×3×3 convolutional layer with a stride of 2 (used for downsampling). The number of output channels for the four stages are 64, 128, 256, and 512, respectively.

[0087] The decoder consists of four upsampling stages, symmetrical to the encoder. Each stage comprises a 2×2×2 three-dimensional deconvolutional layer (for upsampling, followed by batch normalization and ReLU), a channel concatenation operation with the corresponding encoder feature map from the skip connection, and two 3×3×3 three-dimensional convolutional layers.

[0088] Skip connection: The output feature map of the last convolutional layer of each stage of the encoder is concatenated with the deconvolutional output feature map of the corresponding stage of the decoder in the channel dimension.

[0089] Attention Mechanism Module: Integrated into each upsampling stage of the decoder. This module employs a channel-space hybrid attention mechanism. Specifically, the input feature map is first subjected to global average pooling and global max pooling respectively. The resulting two feature vectors are then input into a shared two-layer fully connected network to generate channel attention weights. Simultaneously, a 1×1×1 convolutional layer is used to generate a spatial attention weight map in the spatial dimension. The two are then combined and multiplied element-wise with the original feature map to achieve feature recalibration.

[0090] Output layer: The last convolutional layer of the decoder is followed by a 1×1×1 convolutional layer, which maps the number of channels to the number of target meteorological elements (e.g., U-wind, V-wind, and temperature, a total of 3 channels). Loss calculation: The model's predicted values ​​and the corresponding preprocessed high-resolution reanalysis data labels (i.e., ground truth values) are simultaneously input into a predefined loss function. The loss function precisely quantifies the differences between the two across all grid points and all meteorological elements.

[0091] Backpropagation and parameter update: Calculate the gradient of the loss function value with respect to each trainable parameter of the model. The gradient indicates the direction and magnitude by which each parameter should be adjusted to reduce the loss. Subsequently, an optimizer (such as the Adam optimizer) is used to efficiently and stably update all parameters of the model based on this gradient information.

[0092] In conjunction with the first aspect, the loss function is a joint loss function, which includes at least the mean square error loss for the wind vector field and the L1 loss for the wind speed scalar field.

[0093] The joint loss function directly determines the optimization direction of model learning and the practicality of the final product during model training. This loss function is based on the core requirements of low-altitude flight safety assurance and the physical characteristics of meteorological elements, and performs multi-dimensional and refined measurement and constraint on model prediction error.

[0094] Specifically, in this embodiment, the loss function aims to accurately constrain the model's predictive ability regarding the coordinated changes in wind direction and speed. The wind vector field (typically composed of two components: zonal wind U and meridional wind V) is the core vector describing atmospheric motion. The mean squared error loss is calculated by measuring the model's predicted (U... pred V pred ) and Real Labels (U true V true The average of the squared differences between the components is used to penalize the error by squared error. This makes the model extremely sensitive to large vector deviations (whether directional errors or severe inconsistencies in intensity), thus forcing it to prioritize corrections of major wind field prediction errors that could lead to significant deviations in aircraft attitude or drastic changes in energy.

[0095] For low-altitude flight, the direction of the wind directly affects flight path planning and crosswind tolerance, while the composite wind vector determines the relationship between the aircraft's airspeed and ground speed. This loss ensures that the encrypted wind field output by the model is highly consistent with the real situation in the joint distribution of "wind direction-wind speed," providing a reliable vector basis for dynamic route planning and wind shear identification.

[0096] The L1 loss for the wind speed scalar field focuses on directly optimizing the prediction accuracy of wind speed magnitude. The wind speed scalar field is calculated from the modulus of the wind vector field (wind speed = sqrt(U...). 2 +V 2 L1 loss calculates the average of the absolute differences between the predicted and actual wind speeds. Compared to the squared penalty of mean squared error, L1 loss is relatively insensitive to outliers and more robust, enabling the model's predicted wind speed values ​​to converge more stably to near the actual values, avoiding over-adjustment due to individual extreme errors.

[0097] Wind speed is a direct basis for determining flight safety levels, calculating wind loads, and assessing wind energy resources. Overly high wind speed forecasts may lead to unnecessary route avoidance and capacity losses, while underly high forecasts may mask risks. This loss directly impacts the accuracy of wind speed values ​​and is crucial for safety operations such as setting flight speed envelopes and triggering wind speed threshold alarms (e.g., gust warnings).

[0098] The total loss of the model during training. It is the weighted sum of these two losses plus losses from other factors such as temperature. That is: ;in, To account for the mean square error loss of the wind vector field, the mean square errors of the zonal wind (U) and meridional wind (V) components are calculated separately and then summed. To calculate the L1 loss in the wind speed scalar field, the mean absolute error between the wind speed scalars (i.e., modulus) obtained from the predicted wind vector and the actual wind vector is calculated. This represents the mean square error loss of the temperature field; , , These are the weighting coefficients, and in a preferred embodiment, their values ​​are set to 1.0, 0.5, and 0.8, respectively. By optimizing this joint loss function, the model can accurately predict wind speed and temperature while maintaining the accuracy of the wind field vector structure.

[0099] The optimizer drives the model to improve its wind field infiltration capability from two dimensions simultaneously by minimizing this joint loss: Vector accuracy guides macroscopic structure: Wind vector MSE loss ensures that the model learns to correctly organize wind flow patterns (such as cyclones, anticyclones, and shear lines) in three-dimensional space, which is the basis for the physical rationality of the wind field; Scalar accuracy anchors key strengths: Wind speed L1 loss ensures that the model provides accurate wind speed values ​​under the correct flow pattern, which is crucial for quantifying risk assessment.

[0100] Understandably, this joint loss function design essentially translates key indicators in low-altitude flight safety application scenarios directly into mathematical language that drives model optimization. It forces the model not only to accurately depict the wind field's morphology but also to calibrate its intensity, thereby ensuring that the final vertical encryption product meets the stringent requirements of business applications in terms of both physical consistency and quantitative accuracy. This is one of the important technical guarantees that this method can provide reliable security support.

[0101] In conjunction with the first aspect, after step S130, the following also includes: S140 calculates the vertical wind shear intensity at key points along a specified flight path based on a three-dimensional meteorological forecast field with a second vertical resolution.

[0102] S150 generates flight risk warning information in response to vertical wind shear intensity exceeding a preset alarm threshold.

[0103] The high-resolution three-dimensional meteorological element forecast field generated in step S130 provides a core data foundation for subsequent risk quantification and decision-making applications. The key application process of calculating vertical wind shear intensity and generating flight risk warnings based on this encrypted field realizes a complete value loop from fine weather forecasting to proactive safety warnings.

[0104] Specifically, firstly, based on a high-resolution three-dimensional wind field (U and V components), densely distributed wind vector sequences in the vertical direction are extracted for key waypoints along a pre-defined flight path. Using this high-density vertical profile, numerical difference or curve fitting methods can be used to accurately calculate the rate of change of wind speed and / or wind direction with altitude within a specific altitude layer (e.g., the near-surface 50-500 meter layer, which is crucial for low-altitude flight safety), thereby obtaining a quantitative estimate of vertical wind shear intensity. The accuracy of this calculation directly depends on the high vertical resolution data provided by the aforementioned encryption steps.

[0105] Subsequently, the system compares the calculated vertical wind shear intensity with pre-set alarm thresholds based on aviation safety standards, aircraft performance, and operational specifications in real time. Once the intensity exceeds the safety threshold, the system immediately and automatically triggers the alarm generation process. The generated structured flight risk alarm information includes the risk location (latitude, longitude, and altitude), risk type (strong vertical wind shear), intensity level, expected duration, and specific avoidance suggestions (such as adjusting flight altitude, modifying flight path, or postponing the mission). This information is then pushed to the flight control system or management platform in real time via a data interface, providing operators or autonomous flight algorithms with immediate and actionable decision support.

[0106] In summary, this extended process from data encryption to risk alerts not only verifies the practical business value of high-resolution vertical encryption products, but also upgrades a weather forecasting technology into a complete intelligent risk proactive perception and early warning system for low-altitude flight safety, achieving a leap from passively viewing the weather to proactively managing risks.

[0107] Secondly, embodiments of this application also provide a deep learning-based vertical encryption device for low-altitude meteorological forecasting, combined with... Figure 2 As shown, the device includes: a data acquisition module 10, a preprocessing module 20, and a vertical encryption module 30.

[0108] The data acquisition module 10 is used to acquire three-dimensional meteorological forecast field data with a first vertical resolution output by a global or regional numerical weather prediction model.

[0109] The preprocessing module 20 is used to preprocess the three-dimensional meteorological forecast field data to obtain the preprocessed three-dimensional meteorological forecast field data.

[0110] The vertical encryption module 30 is used to input the preprocessed three-dimensional meteorological forecast field data into the pre-trained deep learning vertical encryption model and output a three-dimensional meteorological element forecast field with a second vertical resolution.

[0111] The second vertical resolution is higher than the first vertical resolution, and the deep learning vertical encryption model is used to establish a non-linear mapping relationship from the first vertical resolution to the second vertical resolution.

[0112] Thirdly, embodiments of this application provide an electronic device, combined with Figure 3 As shown, the electronic device includes a memory 131 and a processor 130. The memory 131 stores a computer program, and the processor 130 runs the computer program to make the electronic device perform the above-described method.

[0113] Furthermore, combined Figure 3 The electronic device shown also includes a bus 132 and a communication interface 133, with the processor 130, the communication interface 133 and the memory 131 connected via the bus 132.

[0114] The memory 131 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 133 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 132 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0115] Processor 130 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 130 or by instructions in software form. Processor 130 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 131, and processor 130 reads the information in memory 131 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0116] Fourthly, embodiments of this application provide a readable storage medium storing computer program instructions, which are read and executed by a processor to perform the above-described method.

[0117] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0118] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0119] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0120] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0121] Finally, it should be noted that the above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A deep learning-based method for vertical densification of low-altitude meteorological forecasts, characterized in that, include: Acquire three-dimensional meteorological forecast field data with first vertical resolution output from global or regional numerical weather prediction models; The three-dimensional meteorological forecast field data is preprocessed to obtain preprocessed three-dimensional meteorological forecast field data; The preprocessed three-dimensional meteorological forecast field data is input into a pre-trained deep learning vertical encryption model, and the output is a three-dimensional meteorological element forecast field with a second vertical resolution. Wherein, the second vertical resolution is higher than the first vertical resolution, and the deep learning vertical encryption model is used to establish a non-linear mapping relationship from the first vertical resolution to the second vertical resolution.

2. The method according to claim 1, characterized in that, The deep learning vertical encryption model is based on an encoder-decoder neural network architecture; The encoder consists of multiple three-dimensional convolutional layers, used to extract features and downsample the input three-dimensional weather forecast field data; The decoder consists of multiple three-dimensional deconvolutional layers, used to upsample and reconstruct the features extracted by the encoder into the three-dimensional meteorological element forecast field with a second vertical resolution.

3. The method according to claim 2, characterized in that, A skip connection is provided between the corresponding layers of the encoder and the decoder to pass the low-level spatial detail features extracted by the encoder to the decoder and fuse them with the high-level semantic features of the corresponding layer of the decoder.

4. The method according to claim 2, characterized in that, The decoder integrates an attention mechanism module, which enables the model to increase the attention weight of key vertical layers, including the boundary layer, wind shear layer, or inversion layer, when reconstructing the three-dimensional meteorological element forecast field.

5. The method according to any one of claims 1 to 4, characterized in that, The training method for the deep learning vertical encryption model includes: Obtain a training dataset, which includes historical forecast data samples output by numerical weather prediction models and their corresponding spatiotemporal high-resolution reanalysis data labels; The historical forecast data samples and high-resolution reanalysis data labels are preprocessed to obtain preprocessed historical forecast data samples and preprocessed high-resolution reanalysis data labels. Using the preprocessed historical forecast data samples as input and the preprocessed high-resolution reanalysis data labels as the supervision target, the deep learning vertical encryption model is trained by optimizing the loss function.

6. The method according to claim 5, characterized in that, The loss function is a joint loss function, which includes at least the mean square error loss for the wind vector field and the L1 loss for the wind speed scalar field.

7. The method according to claim 1, characterized in that, After inputting the preprocessed 3D meteorological forecast field data into a pre-trained deep learning vertical densification model and outputting a 3D meteorological element forecast field with a second vertical resolution, the method further includes: Based on the three-dimensional meteorological element forecast field with the second vertical resolution, the vertical wind shear intensity at key points of the specified flight route is calculated. In response to the vertical wind shear intensity exceeding a preset alarm threshold, a flight risk alarm message is generated.

8. A deep learning-based vertical encryption device for low-altitude meteorological forecasting, characterized in that, include: The data acquisition module is used to acquire three-dimensional meteorological forecast field data with first vertical resolution output by global or regional numerical weather prediction models. The preprocessing module is used to preprocess the three-dimensional weather forecast field data to obtain preprocessed three-dimensional weather forecast field data. The vertical encryption module is used to input the preprocessed three-dimensional meteorological forecast field data into a pre-trained deep learning vertical encryption model and output a three-dimensional meteorological element forecast field with a second vertical resolution. Wherein, the second vertical resolution is higher than the first vertical resolution, and the deep learning vertical encryption model is used to establish a non-linear mapping relationship from the first vertical resolution to the second vertical resolution.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images