Radar echo extrapolation method based on improved KAN convolution and visual attention network

Abstract

The invention discloses a radar echo extrapolation method based on improved KAN convolution and a visual attention network, and belongs to the technical field of weather forecast. The method comprises the steps that radar echo data are collected and preprocessed; constructing a V-KAN model, wherein the V-KAN model comprises an encoder, a VAN network and a decoder which are connected in sequence; an NSCA module and a KAN convolution layer are arranged in each of the encoder and the decoder; input data firstly pass through an encoder, each KAN convolution layer in the encoder carries out convolution operation on input features, and then space and channel information of the features is enhanced through an NSCA module; the output characteristics of the encoder are transmitted to the VAN network, and the VAN network can further capture the global dependency relationship; and finally, the output features of the VAN network are transmitted to a decoder, each KAN convolution layer and each NSCA module in the decoder process the features in sequence, and a final output result is generated. And carrying out training and optimization verification on the V-KAN model by using the preprocessed echo data, and outputting a prediction result.

Description

Technical Field

[0001] The present invention belongs to the technical field of weather forecasting, and particularly relates to a radar echo extrapolation method based on improved KAN convolution and visual attention network. Background Art

[0002] Under the background of the increasing popularity of meteorological big data, the development of artificial intelligence provides new ideas for short-term precipitation forecasting. Radar echo extrapolation methods based on deep learning have been widely applied in the meteorological field. Deep learning methods have high prediction accuracy, are suitable for high-dimensional data, can handle complex physical problems, and have high data utilization efficiency. The radar echo extrapolation problem can be analogized to the time series prediction problem in the field of computer vision. In deep learning methods, convolutional neural networks can better adapt to the diversity of radar echoes and can automatically extract features in images. Recurrent neural networks can process time series data and use historical information for prediction, which is very important for the radar echo extrapolation problem. Among them, the Long Short-Term Memory (LSTM) network performs relatively well.

[0003] Shi et al. (2015) defined radar echo extrapolation as a spatio-temporal sequence prediction problem and proposed the Convolutional LSTM Network (ConvLSTM), which innovatively combines LSTM and CNN. It can effectively capture the spatio-temporal features of radar echo data, thus predicting the future echo distribution more accurately. However, its model has a high computational complexity and is difficult to handle long-time sequences. Wang et al. (2017, 2018) successively proposed the Predictive Recurrent Neural Network (PredRNN) and the improved Predictive Recurrent Neural Network Plus Plus (PredRNN++) to enhance the capture of short-time dynamic changes and use a larger convolutional neural receptive field to capture a larger spatial change area. Wang et al. (2019) proposed MIM (Memory In Memory) in order to better learn high-order non-stationary feature information. Kevin et al. (2018) embedded an attention mechanism in the encoder-decoder of the classic convolutional network model UNet and proposed SmaAt-UNet, which uses a small architecture for nowcasting. However, due to the inherent locality characteristics of convolutional operations, it has limitations in learning the global features and time information of radar echo data. Wang et al. (2019) proposed E-3D LSTM, the core of which is the 3D LSTM unit, which can take into account both long-time and short-time information dependencies and local spatio-temporal feature extraction. However, due to the sequential characteristics of the recurrent neural network, it cannot achieve parallel operations and is very time-consuming in the backpropagation process. Xu et al. (2019) combined the Generative Adversarial Nets (GAN) with LSTM and improved the accuracy of model prediction through adversarial training. However, the problem of difficult GAN training still exists. Gao et al. (2022) proposed the SimVP model, which is completely based on CNN and is trained through end-to-end mean square error loss without introducing any additional tricks and complex strategies. However, SimVP still has insufficient representation of the long-term dynamic change information of radar echo data. Fang Wei et al. (2023) constructed a ConvLSTM prediction network based on global channel attention (Global ChannelAttention based ConvLSTM, GCA-ConvLSTM). The radar extrapolation model improved by the ensemble learning algorithm can effectively improve the fitting ability of the model. However, the complex network structure will lead to longer training and inference times.

[0004] To this end, the present invention proposes a radar echo extrapolation method based on improved KAN convolution and visual attention network. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the purpose of the present invention is to provide a radar echo extrapolation method based on improved KAN convolution and visual attention network, which solves the problems in the prior art.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A radar echo extrapolation method based on improved KAN convolution and visual attention network includes the following steps:

[0008] Collect radar echo data and preprocess the echo data;

[0009] Construct a V-KAN model; the V-KAN model includes: an encoder, a VAN network, and a decoder connected in sequence; both the encoder and the decoder are provided with an NSCA module and a KAN convolution layer; the input data first passes through the encoder, and each KAN convolution layer in the encoder performs a convolution operation on the input features, and then enhances the spatial and channel information of the features through the NSCA module; the output features of the encoder are transmitted to the VAN network, and the VAN network can further capture global dependencies; finally, the output features of the VAN network are transmitted to the decoder, and each KAN convolution layer and NSCA module in the decoder processes the features in sequence to generate the final output result;

[0010] Divide the preprocessed echo data into a training set and a test set, use the training set to train the V-KAN model, and use the test set to optimize and verify the trained V-KAN model;

[0011] Input the newly collected echo data into the trained and optimized V-KAN model to output the prediction result.

[0012] Further, the echo data preprocessing includes: constant value cleaning, data screening, data clipping, generating a sample sequence with a sliding window, and normalization.

[0013] Further, the NSCA module includes: a channel attention module and a spatial attention module. The channel attention module integrates coordinate attention, captures both channel dependencies and spatial position information, and enhances the channel dimension representation ability of the features; the spatial attention module focuses on the spatial dimension of the feature map and extracts the spatial information of the key region; the two work together to enhance the features from the channel and spatial dimensions respectively.

[0014] Further, in the channel attention module, the input features are respectively subjected to average pooling in the horizontal and vertical dimensions to calculate two two-dimensional vectors, and then a concatenation operation is performed in the spatial dimension and a two-dimensional convolution with a convolution kernel size of 1×1 is used to compress the channels. Immediately afterwards, a batch normalization layer and a non-linear layer are used to encode the spatial features in the vertical and horizontal directions. Then, a segmentation operation is carried out, and two-dimensional convolutions with a convolution kernel size of 1×1 are respectively applied to obtain the same number of channels as the input features. Then, through the Sigmoid activation function, and finally, the obtained spatial information is weighted and aggregated in the channel dimension.

[0015] Further, the spatial attention M s of the spatial attention module is calculated as follows:

[0016] M s (F) = σ(f 7×7 ([AvgPool(F); MaxPool(F)]))

[0017] where F is the input tensor, AvgPool is the average pooling operation, MaxPool is the max pooling operation, the outer f 7×7 is a two-dimensional convolution with a convolution kernel size of 7×7, and the outermost σ is the Sigmoid activation function.

[0018] Further, the VAN network includes multiple VAN modules. Each VAN module includes a hybrid multi-layer perceptron network and an Attention module. The hybrid multi-layer perceptron network includes a spatial local convolution with a convolution kernel size of 3×3, a GELU activation function, and a two-dimensional convolution with a convolution kernel size of 1×1. The Attention module includes LKA, a GELU activation function, and a two-dimensional convolution with a convolution kernel size of 1×1.

[0019] The multiple groups of tensors input to the VAN network are first subjected to batch normalization operations, and then calculated through the Attention module to obtain an attention-weighted feature representation, which is added to the input tensor to obtain the first residual connection. Then, after batch normalization operations and the hybrid multi-layer perceptron network, the output result is added to the first residual connection to obtain the second residual connection, which is the output tensor of the VAN module. The sizes of the input and output tensors of each VAN module in the VAN network remain unchanged.

[0020] A radar echo extrapolation system based on an improved KAN convolution and a visual attention network includes:

[0021] A data acquisition and processing module: It acquires radar echo data and preprocesses the echo data.

[0022] Model construction module: Construct the V-KAN model; The V-KAN model includes: an encoder, a VAN network, and a decoder connected in sequence; The encoder and decoder are provided with an NSCA module and a KAN convolutional layer; The input data first passes through the encoder, and each KAN convolutional layer in the encoder performs a convolutional operation on the input features, and then enhances the spatial and channel information of the features through the NSCA module; The output features of the encoder are transmitted to the VAN network, and the VAN network can further capture global dependencies; Finally, the output features of the VAN network are transmitted to the decoder, and each KAN convolutional layer and NSCA module in the decoder processes the features in sequence to generate the final output result;

[0023] Model training and optimization module: Divide the preprocessed echo data into a training set and a test set, use the training set to train the V-KAN model, and use the test set to optimize and verify the trained V-KAN model;

[0024] And, prediction module: Input the newly collected echo data into the trained and optimized V-KAN model to output the prediction result.

[0025] A computer storage medium stores a readable program that, when run, can execute the above-mentioned radar echo extrapolation method based on improved KAN convolution and visual attention network.

[0026] An electronic device includes: a processor, a memory, a communication interface, and a communication bus, and the processor, the memory, and the communication interface complete communication with each other through the communication bus;

[0027] The memory is used to store at least one executable instruction, and the executable instruction causes the processor to perform the operations corresponding to the above-mentioned radar echo extrapolation method based on improved KAN convolution and visual attention network.

[0028] A computer program product includes computer instructions that direct a computing device to perform the operations corresponding to the above-mentioned radar echo extrapolation method based on improved KAN convolution and visual attention network.

[0029] Advantages of the present invention:

[0030] 1. The present invention creatively introduces the KAN network into the field of weather forecasting, and discloses a new radar echo extrapolation model V-KAN that is completely based on KAN and convolutional neural networks. This network can adaptively learn the importance of different positions in the radar echo data sequence, better capture key spatio-temporal information, has a relatively fast training speed, can effectively process long radar echo data sequences, and can effectively solve the problems of difficult training and information loss of recurrent neural networks in radar echo extrapolation.

[0031] 2. The present invention proposes a new spatial channel attention module NSCA and integrates this attention module into the encoder and decoder of the V-KAN model, which can play a role in information aggregation and feature enhancement in the radar echo extrapolation algorithm, capture important features in the radar echo data, and improve the performance and accuracy of the extrapolation algorithm.

[0032] 3. The present invention embeds a visual attention network VAN into the encoder and decoder of the V-KAN model to capture dynamic features and multi-scale information in the radar echo data, which is more efficient in processing large-scale radar echo data and can perform extrapolation prediction in real time.

[0033] 4. The present invention proposes a radar echo extrapolation model integrating KAN convolution. The KAN network derived from the Kolmogorov-Arnold representation theorem has a solid theoretical foundation and can more accurately describe and explain the relationship between input data and output results. KAN convolution applies a learnable non-linear activation function to the edges of the network instead of using a fixed activation function at the nodes like traditional convolution. The V-KAN model has significant advantages in interpretability compared to previous radar echo extrapolation models. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0035] Figure 1 is the flowchart of the radar echo extrapolation method of the present invention;

[0036] Figure 2 is the schematic diagram of the structure of the Coordinate Attention module of the present invention;

[0037] Figure 3 is the detailed diagram of the V-KAN model of the present invention;

[0038] Figure 4 is the architecture diagram of the V-KAN network of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0040] Example 1

[0041] In the radar echo extrapolation task, the temporal resolution of the image is 6 minutes. Since the short-term precipitation forecast needs to extrapolate the echo distribution of the next two hours based on the radar echo of the previous hour, it is necessary to extrapolate the next 20 radar images based on the first 10 radar images. The movement of the cloud map in the radar image has a non-rigid feature, that is, in addition to the overall migration of the cloud layer, it is also accompanied by local rotation, scaling and effective evolution. This non-rigid motion feature makes the radar image sequence constantly change the data distribution characteristics over time, that is, high-order non-stationarity. To this end, in this embodiment, a radar echo extrapolation method based on an improved KAN convolution and visual attention network is proposed to model the high-order non-stationarity of the radar echo sequence and achieve more accurate extrapolation.

[0042] like Figure 1 As shown, the radar echo extrapolation method based on the improved KAN convolution and visual attention network includes the following steps:

[0043] S1, collect radar echo data and pre-process the echo data;

[0044] The method of collecting radar echo data is as follows: In meteorological services, radar echo data is mainly obtained through Doppler weather radar. The radar equipment operates in volume scanning mode, completes a scan every 6 minutes, and obtains key data such as reflectivity, radial velocity and spectral width. The reflectivity factor reflects the intensity and density of precipitation particles.

[0045] The echo data preprocessing process includes: first, outlier cleaning is performed to return negative reflectivity values ​​to zero to ensure the physical rationality of the data. Then, through data screening, samples with fewer echoes are eliminated and representative data are retained. On this basis, the sliding window technology is used to generate a sample sequence, and the radar echo data of continuous time steps are divided into sequences of fixed length for model training and prediction. Finally, the data is normalized to improve the training efficiency and stability of the model.

[0046] S2, build V-KAN model;

[0047] like Figure 3 and Figure 4 As shown in FIG. 1 , the V-KAN model includes an encoder, a visual attention network (VAN), and a decoder connected in sequence; both the encoder and the decoder include a spatial channel attention (NSCA) module and a KAN convolutional layer;

[0048] like Figure 3As shown in (c), after the preprocessed echo data is input into the V-KAN model, the input data first passes through the encoder. Each KAN convolutional layer in the encoder performs a convolutional operation on the input features, and then enhances the spatial and channel information of the features through the NSCA module. The output features of the encoder are transmitted to the VAN network, and the non-linear modeling ability of the VAN network can further capture global dependencies. Finally, the output features of the VAN network are transmitted to the decoder, and each KAN convolutional layer and NSCA module in the decoder process the features in turn to generate the final output result.

[0049] 1) NSCA module

[0050] For a neural network, the attention mechanism is actually a mechanism for allocating weights to input data, which enables the neural network to acquire the ability to select specific input information. As shown in (a); the NSCA module includes: a channel attention module and a spatial attention module SAM (Spatial Attention Module). The channel attention module integrates coordinate attention. By integrating coordinate attention, the channel attention module captures both channel dependencies and spatial location information simultaneously, enhancing the representational ability of the feature's channel dimension; the spatial attention module focuses on the spatial dimension of the feature map and extracts the spatial information of the key region. The two work together to enhance the features from the channel and spatial dimensions respectively, thus comprehensively improving the model's expressive ability. Figure 3

[0051] Among them, coordinate attention (Coordinate Attention) is integrated in the channel attention module, which fuses the position information with channel attention, performs one-dimensional feature encoding in two spatial dimensions respectively, cleverly aggregates global information in one dimension to capture long-range dependencies, and retains detailed information in the other dimension for precise localization, realizing the efficient integration and optimization of features; Coordinate Attention not only enhances the model's sensitivity and localization ability to key information, but also provides the model with richer and more accurate spatial information by generating direction-aware and position-sensitive attention maps. Specifically, pooling kernels with sizes of (H, 1) and (1, W) are used to perform encoding operations on the horizontal and vertical coordinates of each channel of the input X. Therefore, the output at height h of the C-th channel can be calculated by formula (1).

[0052]

[0053] Similarly, the output at width w is calculated by formula (2).

[0054] ​

[0055] Among them, H and W respectively represent the height and width of the input image, and x c is the input tensor, z is the output, and c is the channel.

[0056] As Figure 2 shown, the input features respectively apply average pooling (Avg Pool) in the horizontal dimension and vertical dimension to calculate two two-dimensional vectors, perform a concatenation (Concat) operation in the spatial dimension and a two-dimensional convolution with a convolution kernel size of 1×1 to compress the channels. Immediately afterwards, through a batch normalization (Batch Normalization, BN) layer and a non-linear layer (Non-Linear) to encode the spatial features in the vertical and horizontal directions. Then, a split operation is performed, and a two-dimensional convolution with a convolution kernel size of 1×1 is respectively applied to obtain the same number of channels as the input features. Then, through the Sigmoid activation function, and finally, the obtained spatial information is weighted and aggregated in the channel dimension.

[0057] In the NSCA module, in order to pay more attention to the key regions in the radar echo map, immediately after the channel attention module is a spatial attention module SAM (Spatial Attention Module). The spatial attention calculation of SAM is shown in formula (3):

[0058] M s (F) = σ(f 7×7 ([AvgPool(F); MaxPool(F)])) (3)

[0059] Among them, F is the input tensor, AvgPool is the average pooling operation, MaxPool is the maximum pooling operation, and the outer f 7×7 is a two-dimensional convolution with a convolution kernel size of 7×7, and the outermost σ is the Sigmoid activation function. M s is the calculated spatial attention.

[0060] 1) Visual Attention Network VAN

[0061] Embed the visual attention network VAN between the encoder and decoder of the V-KAN model. Due to the non-linear modeling ability of the internal hybrid multi-layer perceptron and the feature extraction ability of the large kernel attention LKA (Large Kernel Attention, LKA) mechanism, it has advantages such as parallel computing and multi-scale feature fusion compared with RNN in the radar echo extrapolation algorithm, and can better process large-scale radar data and improve the prediction accuracy.

[0062] The LKA attention mechanism combines the advantages of self-attention mechanism and large-kernel convolution, and captures the long-term dependence of radar echo data by delicately decomposing the large-kernel convolution. It is mainly decomposed into three modules: the spatial local convolution module, the spatial long-range convolution module, and the channel convolution module, as shown in formulas (4)-(5).

[0063] Attention = Conv 1×1 (DW-D-Conv(DW-Conv(F))) (4)

[0064]

[0065] Among them, F represents the input feature, DW-Conv is the spatial local convolution, which is a depthwise separable convolution that realizes independent convolution processing for each channel of the input tensor and is a special grouped convolution. DW-D-Conv is the spatial long-range convolution, which is a depthwise separable dilated convolution. By introducing the dilation rate in the convolution kernel, it further expands the receptive field of each channel convolution and increases the perception range of the model. Attention is the decomposed large-kernel attention. The outermost Conv 1×1 is a two-dimensional convolution with a convolution kernel size of 1×1. means applying the decomposed large-kernel convolutional attention to the input feature F.

[0066] The VAN network in the V-KAN model consists of multiple VAN modules, as Figure 3 shown in (b) of. The VAN module includes: MixMlp (mixed multi-layer perceptron network) and Attention module; MixMlp consists of a spatial local convolution with a convolution kernel size of 3×3, a GELU activation function, and a two-dimensional convolution with a convolution kernel size of 1×1; The Attention module consists of LKA, a GELU activation function, and a two-dimensional convolution with a convolution kernel size of 1×1.

[0067] The L groups of tensors input to the VAN network first undergo batch normalization operations, and then through Attention calculation, the attention-weighted feature representation is obtained and added to the input tensor to get the first residual connection. Then, after batch normalization operations and the mixed multi-layer perceptron network, the output result is added to the first residual connection to get the second residual connection, which is the output tensor of VAN; The input and output tensor sizes of each VAN module in the VAN network remain unchanged.

[0068] In radar echo extrapolation, although traditional ordinary convolution operations can extract certain spatial features, their inherent limitations often restrict the depth and accuracy of feature extraction. When the ordinary convolution kernel traverses the input tensor, it performs a fixed weighted sum on each local area. This operation method may not be able to fully capture the subtle changes and key features in the data when dealing with complex and variable radar echo data. Therefore, the present invention introduces KAN convolution in the encoder and decoder of the model. KAN convolution does not directly apply the dot product between the convolution kernel and the corresponding pixels in the image. Instead, it applies a learnable non-linear activation function to each element and is parameterized using a spline function, which can dynamically learn and optimize the complex patterns in the data. Then, these activated values are aggregated together to generate a new feature map.

[0069] In the V-KAN model, the shape of the input tensor is B×T×C×H×W, where B is the batch size, T is the sequence length, C is the number of channels, and H and W are the height and width of the input image, respectively. In the encoder, the input tensor is reshaped into a four-dimensional tensor with the shape of (B×T)×C×H×W. Each frame is regarded as a separate sample, focusing on the single-frame features without considering the temporal changes. The number of KAN convolution layers and the size of the convolution kernel in the encoder can be adjusted according to the size of the input image. In the present invention, 4 KAN convolution layers are set, and the size of the convolution kernel is 3×3, and the stride sequence is [2, 1, 2, 1]. That is, the encoder needs to go through two downsamplings. The output result of each KAN convolution layer in the encoder is shown in formula (6).

[0070]

[0071] Among them, Norm2d is the normalization layer, KANConv2d is the KAN convolution, and Z i-1 The initial value is the input tensor, σ is the LeakyRelu activation function, and Attention is the NSCA attention module. Indicates applying the NSCA module to the input tensor, and N s Is the number of KAN convolution layers in the encoder.

[0072] The output tensor obtained in the V-SimVP encoder is reshaped into a tensor with the shape of B×(T×C)×H×W and passed through the VAN network as shown in formula (7).

[0073] Z j = VANblock(Z j-1 ), 1≤j≤N t (7)

[0074] Among them, VANblock is the VAN module, and Z j-1 The initial value is the output tensor of the encoder, and N tIt is the number of VAN modules, which is set to 4 in the present invention. The output tensor obtained from the VAN module is reshaped into a tensor with the shape of (B×T)×C×H×W and input into the decoder of V-KAN. The present invention also sets 4 KAN convolutional layers, and the size of the convolutional kernel is also 3×3, and the stride sequence is [2, 1, 2, 1]. The decoder also needs to go through upsampling twice. The output result of each KAN convolutional layer of the decoder is shown in formula (8).

[0075]

[0076] Among them, KANConv2d is the KAN convolution, and Z k-1 The initial value is the output tensor of the VAN network, σ is the LeakyRelu activation function, and Attention is the NSCA attention module. Indicates that the NSCA module is applied to the input tensor, and N s is the number of KAN convolutional layers of the decoder. Finally, the output tensor is reshaped into a tensor with the shape of B×T×C×H×W.

[0077] S3. Divide the preprocessed echo data into a training set and a test set, use the training set to train the V-KAN model, and use the test set to optimize and verify the trained V-KAN model;

[0078] Training steps: First, divide the preprocessed echo data into a training set and a test set; then, initialize the V-KAN model parameters; then, use the training set data to optimize the model parameters through multiple iterations of training, and adjust the hyperparameters according to the training effect; finally, save the trained model weights for subsequent verification and testing.

[0079] The process of validating and optimizing the V-KAN model can be summarized as: First, use the test set data to evaluate the trained model, and measure the model performance by calculating indicators such as accuracy and false alarm rate; then, adjust the hyperparameters of the model according to the evaluation results, such as the learning rate, model architecture, etc., to find the optimal configuration; finally, regularization processing is also required to prevent overfitting, to prevent the model from overfitting, and to ensure that the model has good generalization ability while maintaining high performance.

[0080] S4. Input the newly collected echo data into the trained and optimized V-KAN model to output the prediction result.

[0081] Based on a similar inventive concept, an embodiment of the present invention also provides a computer storage medium storing a readable program, which can execute the above-mentioned radar echo extrapolation method based on the improved KAN convolution and visual attention network when the program runs.

[0082] Based on a similar inventive concept, an embodiment of the present invention provides an electronic device, including: a processor, a memory, a communication interface, and a communication bus. The processor, the memory, and the communication interface complete communication with each other through the communication bus;

[0083] The memory is used to store at least one executable instruction, and the executable instruction causes the processor to perform the operations corresponding to the above-mentioned radar echo extrapolation method based on the improved KAN convolution and visual attention network.

[0084] Based on a similar inventive concept, an embodiment of the present invention further provides a computer program product, including computer instructions, and the computer instructions instruct a computing device to perform the operations corresponding to the above-mentioned radar echo extrapolation method based on the improved KAN convolution and visual attention network.

[0085] Embodiment 2

[0086] According to the radar echo extrapolation method based on the improved KAN convolution and visual attention network proposed in Embodiment 1, in this embodiment, a radar echo extrapolation system based on the improved KAN convolution and visual attention network is proposed, specifically including:

[0087] Data acquisition and processing module: Collect radar echo data and preprocess the echo data;

[0088] Model construction module: Construct a V-KAN model; the V-KAN model includes, in sequence: an encoder, a VAN network, and a decoder; an NSCA module and a KAN convolution layer are provided in the encoder and the decoder; the input data first passes through the encoder, and each KAN convolution layer in the encoder performs a convolution operation on the input features, and then enhances the spatial and channel information of the features through the NSCA module; the output features of the encoder are transmitted to the VAN network, and the nonlinear modeling ability of the VAN network can further capture global dependencies; finally, the output features of the VAN network are transmitted to the decoder, and each KAN convolution layer and NSCA module in the decoder process the features in sequence to generate the final output result;

[0089] Model training and optimization module: Divide the preprocessed echo data into a training set and a test set, use the training set to train the V-KAN model, and use the test set to optimize and verify the trained V-KAN model;

[0090] And a prediction module: Input the newly collected echo data into the trained and optimized V-KAN model and output a prediction result.

[0091] The method of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CDROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored in a remote recording medium or a non-transitory machine-readable medium and downloaded through a network and to be stored in a local recording medium, so that the method described herein can be stored on such a recording medium for software processing using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It will be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code, and when the software or computer code is accessed and executed by the computer, processor, or hardware, the method described herein is implemented. In addition, when a general-purpose computer accesses the code for implementing the method shown herein, the execution of the code converts the general-purpose computer into a dedicated computer for executing the method shown herein.

[0092] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and the description in the specification are only illustrative of the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and all such changes and improvements fall within the scope of the present invention claimed.

Claims

1. A radar echo extrapolation method based on an improved KAN convolution and a visual attention network, characterized in that It includes the following steps: Collect radar echo data and preprocess the echo data; Construct a V-KAN model; the V-KAN model includes: an encoder, a VAN network, and a decoder connected in sequence; both the encoder and the decoder are provided with an NSCA module and a KAN convolutional layer; the input data first passes through the encoder, and each KAN convolutional layer in the encoder performs a convolutional operation on the input features, and then enhances the spatial and channel information of the features through the NSCA module; the output features of the encoder are transmitted to the VAN network, and the VAN network can further capture global dependencies; finally, the output features of the VAN network are transmitted to the decoder, and each KAN convolutional layer and NSCA module in the decoder processes the features in sequence to generate the final output result; Divide the preprocessed echo data into a training set and a test set, use the training set to train the V-KAN model, and use the test set to optimize and verify the trained V-KAN model; Input the newly collected echo data into the trained and optimized V-KAN model to output the prediction result.

2. The radar echo extrapolation method based on the improved KAN convolution and visual attention network according to claim 1, wherein The echo data preprocessing includes: constant value cleaning, data screening, data cropping, generating a sample sequence with a sliding window, and normalization.

3. The radar echo extrapolation method based on the improved KAN convolution and visual attention network according to claim 1, characterized in that The NSCA module includes: a channel attention module and a spatial attention module. The channel attention module integrates coordinate attention, simultaneously captures channel dependencies and spatial position information, and enhances the channel dimension representation ability of the features; the spatial attention module focuses on the spatial dimension of the feature map and extracts the spatial information of the key region; the two work together to enhance the features from the channel and spatial dimensions respectively.

4. The radar echo extrapolation method based on the improved KAN convolution and visual attention network according to claim 3, wherein In the channel attention module, the input features are respectively subjected to average pooling in the horizontal dimension and the vertical dimension to calculate two two-dimensional vectors, a concatenation operation is performed in the spatial dimension and a two-dimensional convolution with a convolution kernel size of 1×1 is used to compress the channels. Immediately afterwards, a batch normalization layer and a non-linear layer are used to encode the spatial features in the vertical and horizontal directions. Then, a splitting operation is performed, and two-dimensional convolutions with a convolution kernel size of 1×1 are respectively applied to obtain the same number of channels as the input features. Then, through the Sigmoid activation function, and finally, the obtained spatial information is weighted and aggregated in the channel dimension.

5. The radar echo extrapolation method based on the improved KAN convolution and visual attention network according to claim 3, characterized in that, The spatial attention M of the spatial attention module s is calculated as follows: M s (F) = σ(f 7×7 ([AvgPool(F); MaxPool(F)])) Among them, F is the input tensor, AvgPool is the average pooling operation, MaxPool is the max pooling operation, and the outer f 7×7 is a two-dimensional convolution with a convolutional kernel size of 7×7, and the outermost σ is the Sigmoid activation function.

6. The radar echo extrapolation method based on the improved KAN convolution and visual attention network according to claim 1, wherein The VAN network includes multiple VAN modules, and the VAN module includes: a hybrid multi-layer perceptron network and an Attention module; the hybrid multi-layer perceptron network includes: a spatial local convolution with a convolution kernel size of 3×3, a GELU activation function, and a two-dimensional convolution with a convolution kernel size of 1×1; the Attention module includes: LKA, a GELU activation function, and a two-dimensional convolution with a convolution kernel size of 1×1; Multiple groups of tensors input into the VAN network first undergo batch normalization operations, and then are calculated through the Attention module to obtain an attention-weighted feature representation, which is added to the input tensor to obtain the first residual connection. Then, after batch normalization operations and a hybrid multi-layer perceptron network, the output result is added to the first residual connection to obtain the second residual connection, which is the output tensor of the VAN module; the sizes of the input and output tensors of each VAN module in the VAN network remain unchanged.

7. A radar echo extrapolation system based on an improved KAN convolution and a visual attention network, characterized in that, Including: Data acquisition and processing module: Collect radar echo data and preprocess the echo data. Model construction module: Construct a V-KAN model; the V-KAN model includes, connected in sequence: an encoder, a VAN network, and a decoder; the encoder and decoder are provided with an NSCA module and a KAN convolutional layer; the input data first passes through the encoder, and each KAN convolutional layer in the encoder performs a convolutional operation on the input features, and then enhances the spatial and channel information of the features through the NSCA module; the output features of the encoder are transmitted to the VAN network, which can further capture global dependencies; finally, the output features of the VAN network are transmitted to the decoder, and each KAN convolutional layer and NSCA module in the decoder processes the features in sequence to generate the final output result. Model training and optimization module: Divide the preprocessed echo data into a training set and a test set, use the training set to train the V-KAN model, and use the test set to optimize and verify the trained V-KAN model. And a prediction module: Input the newly collected echo data into the V-KAN model after training and optimization, and output the prediction result.

8. A computer storage medium storing a readable program, characterized in that, When the program runs, it can execute the radar echo extrapolation method based on the improved KAN convolution and visual attention network according to any one of claims 1-6.

9. An electronic device, characterized in that, Including: A processor, a memory, a communication interface, and a communication bus, and the processor, the memory, and the communication interface complete mutual communication through the communication bus; The memory is used to store at least one executable instruction, and the executable instruction causes the processor to perform the operations corresponding to the radar echo extrapolation method based on the improved KAN convolution and visual attention network according to any one of claims 1-6.

10. A computer program product, comprising computer instructions, characterized in that, The computer instruction instructs the computing device to perform the operations corresponding to the radar echo extrapolation method based on the improved KAN convolution and visual attention network according to any one of claims 1-6.

Images