Laser wind finding radar low-altitude wind shear extrapolation forecasting method, device and medium

Abstract

The invention discloses a laser wind finding radar low-altitude wind shear extrapolation forecasting method and device and a medium, and the method comprises the steps: collecting low-altitude laser radar echo data of different regions, carrying out the preprocessing, and constructing a data set; a SmaAt-UNet model is constructed; the SmaAt-UNet model is trained on the basis of the data set; and deploying the trained SmaAt-UNet model, and carrying out low-altitude wind shear extrapolation forecasting. According to the low-altitude laser radar wind field extrapolation prediction method based on the SmaAt-UNet model, local changes in a wind field image can be captured more accurately, noise or detail loss possibly generated in a traditional method is avoided especially in extraction of detail features such as wind speed and rainfall, and therefore the accuracy of wind field prediction is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of low-altitude wind shear extrapolation forecasting, and particularly to a method, device and medium for low-altitude wind shear extrapolation forecasting using a lidar. Background Art

[0002] With the rapid development of artificial intelligence technology, AI technology has been widely applied in the extrapolation prediction of low-altitude lidar wind fields, especially in the meteorological and aviation fields, becoming a research hotspot. Traditional numerical model-based wind field prediction methods still face many challenges and limitations when dealing with complex and variable low-altitude wind field systems. The accuracy of numerical model prediction highly depends on the accuracy and integrity of initial data. However, due to the limitations of low-altitude observation means and technologies, it is extremely difficult to obtain high-precision and high-resolution initial data. Especially in remote or complex terrain areas, the lack and uncertainty of low-altitude observation data lead to deviations in the results of traditional numerical model predictions.

[0003] In recent years, more and more researchers have started to apply artificial intelligence and machine learning technologies to low-altitude wind field prediction. The emerging technologies can learn and extract useful information from a large amount of historical data, thus constructing a more accurate and refined wind field prediction model. Deep learning technology simulates the complex behavior of the low-altitude wind field by training neural networks, thereby achieving more accurate wind field prediction. The low-altitude lidar wind field extrapolation technology based on artificial intelligence intelligently analyzes and predicts low-altitude lidar echo data through algorithms such as deep learning and neural networks, and then inversely calculates the low-altitude wind field prediction with high spatio-temporal resolution. This method can automatically extract the features in the echo data and learn its spatio-temporal evolution law, thereby achieving accurate prediction of the future low-altitude wind field.

[0004] However, most existing technologies adopt CNN networks and have the following problems: 1) Limited learning of spatio-temporal features: Traditional CNN methods mainly rely on convolutional operations to extract local features of images. Although CNN can effectively extract spatial features, for a dynamic system like the wind field with complex spatio-temporal evolution, traditional CNN architectures may have certain limitations in capturing global spatio-temporal information. Especially when the wind field changes rapidly or there are strong local non-linearities, it is difficult for CNN to fully capture the global dependencies of the wind field. 2) Unable to effectively handle the complexity of low-altitude lidar data: Low-altitude lidar wind field data usually has high spatial resolution and complex spatio-temporal characteristics. When traditional CNN methods process high-dimensional data, it is easy to have the problem of too high computational complexity, resulting in an increase in the time cost of the training and inference processes. In addition, existing CNN models often rely on fixed convolutional kernels, which makes them less adaptable when facing data from different wind fields or different geographical regions. 3) Overfitting and insufficient generalization ability: CNN models have a strong dependence on large-scale training data, while low-altitude lidar wind field data often involves different geographical environments and meteorological conditions. In the existing technology, overfitting problems are prone to occur during the training process, especially when the data volume is insufficient, resulting in weak generalization ability of the model and being unable to effectively process unknown or changing wind field data. 4) Lack of modeling of temporal dependence: During the process of wind field extrapolation, especially for multi-time-step prediction, the existing CNN methods have relatively weak capabilities in time series modeling, lack the modeling of the temporal evolution of the wind field, and are difficult to capture the long-term dependencies between different time steps. This makes it impossible to effectively guarantee the accuracy and stability during long-term wind field prediction.

[0005] For example, the invention application with the document number CN118330780B discloses a method for predicting low-altitude wind shear at airports in complex terrain. This application scheme applies the CALMET dynamic downscaling prediction technology to calculate the horizontal wind shear and vertical wind shear within 600 meters above the airport runway reference plane for the predicted wind field respectively, effectively predicting the low-altitude wind shear in the area near the airport. However, its scheme also has the following problems: The scheme depends on traditional numerical models, has high requirements for the accuracy of initial field data, and is prone to large errors under complex terrain conditions. In addition, this method has high computational complexity and poor real-time performance, and it is difficult to meet application scenarios with high timeliness requirements. At the same time, its adaptability to wind shear phenomena with strong suddenness and small spatio-temporal scales is limited, and it is unable to capture the rapidly evolving local wind field changes in a timely manner, affecting the accuracy and practicality of wind shear warning.

[0006] Therefore, a low-altitude lidar wind field extrapolation prediction method is needed to improve the accuracy, timeliness, and precision of wind field prediction, and be able to accurately extract the spatio-temporal evolution characteristics of the wind field from radar echo data. Summary of the Invention

[0007] In view of the above problems, the purpose of the present invention is to provide a method, device and medium for extrapolating and predicting low-altitude wind shear by a lidar wind radar, which can effectively predict the change of the wind field through intelligent extrapolation of radar echo data, and overcome the limitations of traditional prediction methods under the requirements of high precision and high timeliness.

[0008] An embodiment of the present invention provides a method, device and medium for extrapolating and predicting low-altitude wind shear by a lidar wind radar.

[0009] In a first aspect: A method for extrapolating and predicting low-altitude wind shear by a lidar wind radar includes:

[0010] S1. Collect low-altitude lidar echo data from different regions and perform preprocessing to construct a data set;

[0011] S2. Construct a SmaAt-UNet model;

[0012] S3. Train the SmaAt-UNet model based on the data set;

[0013] S4. Deploy the trained SmaAt-UNet model to perform extrapolation and prediction of low-altitude wind shear.

[0014] Further, the preprocessing in S1 includes: denoising processing, normalization operation and data augmentation.

[0015] Further, the SmaAt-UNet model includes an encoder part and a decoder part, where the encoder and the decoder correspond one by one and have the same number.

[0016] Further, the encoder applies max pooling and double convolution to halve the image size and double the number of feature maps;

[0017] The decoder applies: bilinear upsampling to double the size of the feature map; skip connection to connect the decoder result feature map with the output of the previous encoder; double convolution to reduce the number of feature maps by half.

[0018] Further, the SmaAt-UNet model:

[0019] Is built-in with a CBAM module and adopts the CBAM attention mechanism; the double convolution module adopts depthwise separable convolution.

[0020] Further, the CBAM attention mechanism includes a channel attention mechanism and a spatial attention mechanism, where:

[0021] The channel attention mechanism uses global average pooling and a fully connected layer to assign weights to each channel feature; the spatial attention mechanism assigns weights to specific spatial positions through convolution and a fully connected layer.

[0022] Furthermore, the channel attention mechanism assigns weights to each channel feature using global average pooling and a fully connected layer, and the formula is expressed as:

[0023]

[0024] where is the input feature, is the activation function, and are the weight values, MaxPool represents max pooling, AvgPool represents average pooling, and MLP represents the fully connected layer operation. is the channel-weighted feature vector;

[0025] The spatial attention mechanism assigns weights to specific spatial positions through convolution and a fully connected layer, and the formula is expressed as:

[0026]

[0027] where is the input feature, is the activation function, MaxPool represents max pooling, and AvgPool represents average pooling.

[0028] Furthermore, in S3, the SmaAt-UNet model is trained based on the dataset; it includes pre-training and fine-tuning training, and through fine-tuning training, the model can better adapt to the characteristics of radar echoes in specific regions.

[0029] Second aspect: An electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, it implements the steps of the method provided in the first aspect.

[0030] Third aspect: A non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by the processor, it implements the steps of the method provided in the first aspect.

[0031] Advantages of the present invention:

[0032] 1. The present invention adopts a low-altitude lidar wind field extrapolation prediction method based on the SmaAt-UNet model. By introducing the CBAM attention module and depthwise separable convolution, it can more accurately capture local changes in the wind field image, especially in the extraction of detail features such as wind speed and precipitation, avoiding noise or detail loss that may occur in traditional methods, thereby significantly improving the accuracy of wind field prediction.

[0033] 2. By optimizing the network structure, the present invention reduces the number of trainable parameters. Compared with traditional deep learning models, the computing efficiency is greatly improved. In the real-time prediction of wind field extrapolation, the computing complexity is reduced, thereby reducing the prediction delay and achieving more efficient real-time wind field prediction.

[0034] 3. By adopting depthwise separable convolution and attention mechanism, the present invention greatly reduces the computing complexity of the model while maintaining the prediction accuracy, and reduces the demand for computing resources. In this way, especially in resource-constrained application scenarios, efficient wind field extrapolation prediction can be achieved, reducing the pressure on the hardware platform.

[0035] 4. By combining the pre-training and fine-tuning processes, the present invention can effectively adapt to low-altitude lidar data under different regions and meteorological conditions. Especially when dealing with "unfamiliar" wind field data, it can maintain high prediction accuracy and stability, enhance the generalization ability of the model, and improve the practicality and reliability of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic flowchart of the prediction method of the present invention;

[0037] Figure 2 It is a schematic diagram of the network structure of the SmaAt-UNet model of the present invention;

[0038] Figure 3 It is a schematic diagram of the structure of the CBAM attention module of the present invention;

[0039] Figure 4 It is a schematic diagram of the depthwise separable convolution structure of the present invention;

[0040] Figure 5 It is a schematic diagram of the structure of the electronic device of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0041] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the drawings, in which the same or similar symbols represent the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below by referring to the drawings are exemplary and are only used to explain the present invention and should not be construed as a limitation of the present invention.

[0042] Existing methods mostly use CNN models, which have problems such as limited spatio-temporal feature learning, insufficient complexity in processing data, overfitting and insufficient generalization, and lack of time-dependent modeling.

[0043] In view of the above problems, the present invention provides a method for extrapolating and predicting low-altitude wind shear by a laser wind measurement radar. Figure 1It is a schematic flowchart of the low-altitude wind shear extrapolation prediction method for the lidar wind profiler provided by the embodiment of the present invention. The method includes:

[0044] S1. Collect the low-altitude lidar echo data from different regions and perform preprocessing to construct a data set.

[0045] First, collect the low-altitude lidar echo data from different regions. The data usually contains meteorological characteristic information such as wind speed, wind shear intensity, cloud layer, etc. In order to ensure the quality of the data and improve the stability of model training, data preprocessing is carried out.

[0046] The preprocessing includes noise removal processing, normalization operation, and data augmentation, etc. The noise removal processing removes the interference information in the radar echo image. The normalization ensures that the data is within the same scale range. The data augmentation amplifies the training samples through operations such as rotation and scaling to improve the generalization ability of the model. The data set is constructed using the preprocessed data.

[0047] S2. Construct the SmaAt-UNet model.

[0048] The SmaAt-UNet model adopts an encoder-decoder structure. Among them, the encoder part gradually reduces the spatial resolution to extract features, while the decoder part gradually restores the spatial resolution through bilinear upsampling, so as to achieve high-precision segmentation.

[0049] At the same time, the SmaAt-UNet model uses skip connections between the encoder Encoder and the decoder Decoder, directly transferring the low-level features from the encoder Encoder to the decoder Decoder. This helps to capture detailed information and avoid information loss. This structure of the SmaAt-UNet model can effectively balance the image segmentation accuracy and computational efficiency, and is especially suitable for high-resolution image processing.

[0050] The present invention selects the SmaAt-UNet model as the extrapolation network. The SmaAt-UNet model introduces an attention module (CBAM) and depthwise separable convolution (DSC).

[0051] To effectively fuse and extrapolate the low-altitude lidar echo data from different regions and enable the SmaAt-UNet model to accurately perform low-altitude wind shear extrapolation prediction for the lidar wind profiler, the SmaAt-UNet model can automatically learn the associations of the low-altitude lidar echo data from different regions and fuse the spatial and temporal consistency of the products. The SmaAt-UNet model can not only be used for the acquired observation data, but also predict and deduce the radar images for future time, providing potential decision support.

[0052] The SmaAt-UNet model architecture is formed by an encoder-decoder structure in a U-shape. This structure fully combines global information and local feature information, improving image feature extraction capabilities through feature information integration. Its symmetrical U-shaped structure can better preserve all valid information in the image.

[0053] The SmaAt-UNet model adopts a convolutional block attention module (CBAM), which takes an input image and applies attention to the channel and spatial dimensions sequentially, taking into account the channel as well as the spatial region of the input image.

[0054] The fewer parameters in a network model, the less likely it is to overfit. However, the model becomes simpler and less likely to approximate the distribution of the training set. The downside of reduced parameters is that the model may be too simple to learn the desired task. To reduce the number of parameters without sacrificing significant performance, the SmaAt-UNet model uses depthwise separable convolution (DSC). DSC splits the regular convolution operation into two separate operations: depthwise convolution followed by pointwise convolution. This results in fewer mathematical operations and fewer parameters than non-separable convolution.

[0055] The SmaAt-UNet model of the present invention comprises an encoder and a decoder, with the encoder and decoder having the same number of encoders. The encoder uses maximum pooling and double convolution, which can halve the image size and double the number of feature maps, respectively.

[0056] like Figure 2 As shown, each encoder module consists of two consecutive convolutional layers (double convolution) to extract local features. The convolution operation is typically followed by a nonlinear activation function (such as ReLU) to enhance the model's nonlinear expressiveness. After each encoder module, a maximum pooling operation is used to halve the spatial size of the feature map (for example, from H × W to H / 2 × W / 2). This operation reduces computational effort and enhances scale invariance to the input data.

[0057] The decoder consists of three parts: bilinear upsampling operation ( Figure 2 The green arrow in the figure doubles the feature map size and passes the skip connection ( Figure 2 The gray arrow in the figure connects the resulting feature map with the output of the previous encoder, and finally the double convolution to half the number of feature maps. The skip connection enables the model to use multiple scales of the input to generate the output. The advantage of using different scales is that it can capture objects of different sizes in the input. Finally, the last layer in the model is a 1×1 convolution ( Figure 2 ), which outputs a single feature map representing the network's predictions.

[0058] The SmaAt-UNet model adds the CBAM attention mechanism to the encoder part.

[0059] As Figure 2 shown, in CBAM, the attention mechanism is first applied to the channels of the image and then to the spatial dimension. CBAM is placed after the first double convolution and each encoder to amplify important features and suppress unimportant features at the corresponding image scale ( Figure 2 the yellow arrow in). The input to the encoder is the convolution and downsampled image from the previous encoder, so that the original image features are retained until the last encoder. The CBAM attention module is connected to the corresponding upsampling part of the decoder through a skip connection.

[0060] The SmaAt-UNet model uses depthwise separable convolutions to reduce the number of parameters. The SmaAt-UNet model replaces all convolutions in the original UNet model with depthwise separable convolutions, while in the CBAM module, regular convolutions are still used.

[0061] When in use, first input the radar echo image, perform a double convolution operation on the input image to double the number of feature maps, and then simultaneously perform the attention mechanism to extract the important features of the image after the double convolution operation and perform max pooling on the image after the double convolution operation to halve the image size. Next, perform bilinear upsampling on the image after using the attention mechanism, and then generate a new feature map through a skip connection with the result output by the previous layer of the attention mechanism. Perform a double convolution on the new feature map to halve the number of feature maps. Finally, repeat the above operations, and perform a convolution operation with a convolution kernel of 1×1 on the obtained feature map to obtain the output image of the model.

[0062] The SmaAt-UNet model adopts CBAM to enhance its decoder. This module can adaptively adjust the weights of the feature maps to ensure customized processing for specific tasks.

[0063] As Figure 3 shown, CBAM combines the channel attention mechanism and the spatial attention mechanism to solve the problem of uneven importance of features in different channels and spatial regions in convolutional neural networks. The SmaAt-UNet model dynamically weights the feature information through CBAM, and this model can emphasize key channels and spatial positions, thus improving the focus and effectiveness of the network.

[0064] Specifically, the channel attention mechanism of CBAM uses global average pooling and fully connected layers to evaluate and assign weights to the features of each channel. At the same time, the spatial attention mechanism is used to assign weights to specific spatial positions through convolution and fully connected layers, thus generating a targeted set of spatial features.

[0065] For the channel attention module, the input feature F comes from the max pooling (MaxPool) and average pooling (AvgPool) operations. Then the feature is processed through a shared fully connected layer to obtain a combined and weighted feature vector, which is expressed by the formula:

[0066]

[0067] where input feature, activation function, and weight value.

[0068] The spatial attention module uses the feature vector of the channel attention module as its input. After post-processing through max pooling and average pooling, the results are combined and further refined through a standard method. And it is further refined through a standard convolution to finally generate the spatial attention weight. The program formula of the spatial attention is expressed as:

[0069]

[0070] The CBAM model can adaptively identify and emphasize the basic features across channels and spatial domains, which greatly enhances the performance of the SmaAt-UNet model. In addition, its low computational burden allows for seamless integration into existing network frameworks.

[0071] As Figure 4 shown, the SmaAt-UNet model adopts depthwise separable convolutions (including depth convolution and pointwise convolution). Depthwise separable convolutions can perform personalized feature processing on each input channel and then combine these features, greatly reducing the model's parameter count and computational requirements. The SmaAt-UNet model adopts a streamlined and efficient design of depthwise separable convolutions, making it a lightweight but powerful model for complex vision tasks.

[0072] S3. Train the SmaAt-UNet model based on the dataset.

[0073] After completing data preprocessing, pre-train the model based on at least two years of historical radar echo data. The goal of pre-training is to provide a learning basis for the model through historical data so that it can identify and understand the basic features of radar echo images. Through the pre-trained model, the network can initially learn the spatio-temporal features and variation laws in the radar echo, laying a foundation for subsequent fine-tuning training.

[0074] After the pre-training is completed, fine-tuning training is carried out. The fine-tuning training collects the latest and representative radar echo data, and constructs a fine-tuning training set based on these data. The pre-trained model is further trained using this fine-tuning set, enabling the model to better adapt to the characteristics of radar echoes in a specific region (such as the Chinese region), thereby enhancing the prediction ability for typical wind shear processes.

[0075] S4. Deploy the trained SmaAt-UNet model for low-altitude wind shear extrapolation forecasting.

[0076] Use the fine-tuned SmaAt-UNet model for extrapolation prediction of wind shear. At this stage, input the radar echo image at the current moment, and the SmaAt-UNet model will automatically calculate the dynamic convolution kernel, extract features according to the characteristics of the input image, and predict the wind shear changes at future moments. The model can accurately predict the changes in wind shear and output wind shear features such as future wind speed gradients and wind direction changes, which is applicable to tasks such as wind shear monitoring and early warning.

[0077] The application of the SmaAt-UNet model in the extrapolation forecasting of low-altitude wind shear by lidar can significantly improve the prediction accuracy and timeliness of low-altitude wind shear. This technology effectively improves the prediction ability for low-altitude wind shear changes by integrating the high-precision wind speed and wind direction data of lidar with other meteorological information and using the SmaAt-UNet model for deep learning and extrapolation. Especially in areas with drastic wind speed changes, complex weather systems, or variable terrains, the joint analysis of lidar data and other meteorological data can make up for the deficiencies of single data sources and provide more detailed and comprehensive wind shear early warnings. The application of the SmaAt-UNet model in the extrapolation forecasting of low-altitude wind shear not only improves the accuracy of meteorological early warnings but also provides more scientific and efficient decision-making support for industries such as aviation and transportation in the prevention and emergency response of low-altitude wind shear disasters, greatly ensuring safety and operation efficiency.

[0078] By combining the SmaAt-UNet model, through the introduction of the attention mechanism and depthwise separable convolution, it can better capture the spatio-temporal relationships and global dependencies in wind field data. Compared with traditional CNN methods, the SmaAt-UNet model can enhance the prediction ability of the model under different wind field conditions through a more refined feature extraction mechanism.

[0079] The SmaAt-UNet model can process high-resolution low-altitude lidar data. Through adaptive feature extraction and network optimization, it reduces the computational complexity, effectively utilizes limited training data, improves the adaptability to low-altitude lidar data, and avoids the high computational overhead problem of the CNN method in high-dimensional data processing. By introducing the attention module of the SmaAt-UNet model, the present invention can enhance the adaptability of the model in the face of different geographical regions and meteorological conditions and improve the generalization ability of the model. The attention mechanism enables the model to automatically focus on key regions, effectively reduces the overfitting problem, and improves the generalization ability of the model.

[0080] Through the design of the SmaAt-UNet model, the present invention can better perform time series modeling, capture the time dependence in the wind field evolution process, and strengthen the time-dependent modeling ability. The model can extract the temporal features of the wind field from data at multiple time steps, effectively perform multi-time step prediction, and improve the accuracy and stability of wind field prediction.

[0081] The present invention also provides an electronic device. Figure 5 As shown in the structural schematic diagram of the electronic device provided by the embodiment of the present invention, Figure 5 the electronic device may include: a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete mutual communication through the communication bus. The processor can call the logical instructions in the memory, for example, to execute the following method:

[0082] S1. Collect low-altitude lidar echo data from different regions and perform preprocessing to construct a data set;

[0083] S2. Construct a SmaAt-UNet model;

[0084] S3. Train the SmaAt-UNet model based on the data set;

[0085] S4. Deploy the trained SmaAt-UNet model to perform low-altitude wind shear extrapolation forecasting.

[0086] In addition, when the logical instructions in the above-mentioned memory are implemented in the form of 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 the present invention, in essence, or the part that contributes to the prior art, or a part of this 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 for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, external hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0087] An embodiment of the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is used to execute the methods provided in the above-mentioned various embodiments, for example, including:

[0088] S1. Collect low-altitude lidar echo data from different regions and perform preprocessing to construct a data set;

[0089] S2. Construct a SmaAt-UNet model;

[0090] S3. Train the SmaAt-UNet model based on the data set;

[0091] S4. Deploy the trained SmaAt-UNet model to perform low-altitude wind shear extrapolation forecasting.

[0092] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative efforts.

[0093] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for extrapolating and predicting low-altitude wind shear by a lidar wind profiler, characterized in that, Including: S1. Collect low-altitude lidar echo data from different regions, preprocess it, and construct a dataset; S2. Construct the SmaAt-UNet model; S3. Train the SmaAt-UNet model based on the dataset; S4. Deploy the trained SmaAt-UNet model for low-altitude wind shear extrapolation forecasting.

2. The method for extrapolating and predicting low-altitude wind shear of a lidar wind profiler according to claim 1, wherein, The preprocessing in S1 includes: noise removal processing, normalization operation, and data augmentation.

3. The prediction method according to claim 1, characterized in that, The SmaAt-UNet model includes an encoder part and a decoder part, where the encoders and decoders are in one-to-one correspondence and have the same number.

4. The prediction method according to claim 3, wherein The encoder applies max pooling and double convolution to halve the image size and double the number of feature maps; The decoder application includes: bilinear upsampling to double the size of the feature map; skip connection to connect the decoder result feature map with the output of the previous encoder; double convolution to reduce the number of feature maps by half.

5. The prediction method according to claim 3, characterized in that, The SmaAt-UNet model: Built-in with a CBAM module, adopting the CBAM attention mechanism; the double convolution module adopts depthwise separable convolution.

6. The prediction method according to claim 5, characterized in that, The CBAM attention mechanism includes a channel attention mechanism and a spatial attention mechanism, where: The channel attention mechanism uses global average pooling and a fully connected layer to assign weights to each channel feature; the spatial attention mechanism assigns weights to specific spatial positions through convolution and a fully connected layer.

7. The prediction method according to claim 6, wherein The channel attention mechanism uses global average pooling and a fully connected layer to assign weights to each channel feature, and the formula is expressed as: ； Among them, input features, activation function, , weight value, MaxPool represents max pooling, AvgPool represents average pooling, and MLP represents fully connected layer operation; is the channel weighted feature vector; The spatial attention mechanism assigns weights to specific spatial positions through convolution and a fully connected layer, and the formula is expressed as: ； Among them, input features activation function, MaxPool represents max pooling, and AvgPool represents average pooling.

8. The prediction method according to claim 1, wherein In S3, training the SmaAt-UNet model based on the dataset includes pre-training and fine-tuning training, where through fine-tuning training, the model can better adapt to the characteristics of radar echoes in a specific region.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the forecasting method according to any one of claims 1 to 8.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the forecasting method according to any one of claims 1 to 8.

Images