Radar Echo Prediction Method, Device and Electronic Equipment Based on Diffusion Model

Abstract

The present invention provides a radar echo prediction method, device and electronic device based on a diffusion model, preprocesses original radar data, and inputs the preprocessed radar data into a pre-trained radar echo prediction model (including a deterministic model, an encoding and decoding module, an adaptation model and a diffusion model); performs a first prediction on the radar data through the deterministic model, and performs a first encoding on the first prediction result through the encoding and decoding module; performs a second encoding on the radar data through the encoding and decoding module, and performs a second prediction on the second encoding result through the adaptation model; performs a fusion prediction on the first encoding result and the second prediction result through the diffusion model, and decodes the fusion prediction result through the encoding and decoding module to obtain a radar echo prediction result. Using the present invention can improve the efficiency and accuracy of radar echo prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of weather forecasting, and more particularly to a radar echo prediction method, apparatus, and electronic device based on a diffusion model. Background Art

[0002] For radar echo prediction in the short-term and nowcasting precipitation prediction task, current technical methods mainly use traditional translation methods, deterministic deep learning methods, and methods using image generation models to extrapolate radar echoes. The traditional translation method predicts the future radar echo field by translating the current radar echo field in a certain direction. This method is simple and easy to implement, but often results in an overestimated precipitation intensity because it does not consider the evolution process of the precipitation system. Deterministic deep learning methods (such as U-Net) can better capture the precipitation area, but the prediction results may be relatively fuzzy and lack sensitivity to changes in precipitation intensity. The generation model method (such as generative adversarial network) can predict local rainfall more clearly and accurately, but the model is unstable during training and difficult to adjust in real time. Summary of the Invention

[0003] In view of this, an object of the present invention is to provide a radar echo prediction method, apparatus, and electronic device based on a diffusion model to alleviate the above problems existing in the related art.

[0004] In a first aspect, an embodiment of the present invention provides a radar echo prediction method based on a diffusion model, including: preprocessing original radar data and inputting the preprocessed radar data into a pre-trained radar echo prediction model; wherein, the radar echo prediction model includes a deterministic model, an encoding and decoding module, an adaptation model, and a diffusion model; making a first prediction on the radar data through the deterministic model and performing a first encoding on the first prediction result through the encoding and decoding module; performing a second encoding on the radar data through the encoding and decoding module and making a second prediction on the second encoding result through the adaptation model; performing a fusion prediction on the first encoding result and the second prediction result through the diffusion model and decoding the fusion prediction result through the encoding and decoding module to obtain a radar echo prediction result.

[0005] In a second aspect, an embodiment of the present invention further provides a radar echo prediction device based on a diffusion model, including: a preprocessing module, configured to preprocess the original radar data and input the preprocessed radar data into a pre-trained radar echo prediction model; wherein, the radar echo prediction model includes a deterministic model, an encoding and decoding module, an adaptation model, and a diffusion model; a first prediction module, configured to perform a first prediction on the radar data through the deterministic model and perform a first encoding on the first prediction result through the encoding and decoding module; a second prediction module, configured to perform a second encoding on the radar data through the encoding and decoding module and perform a second prediction on the second encoding result through the adaptation model; a fusion prediction module, configured to perform a fusion prediction on the first encoding result and the second prediction result through the diffusion model and perform decoding on the fusion prediction result through the encoding and decoding module to obtain a radar echo prediction result.

[0006] In a third aspect, an embodiment of the present invention further provides an electronic device, including a processor and a memory, where the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the radar echo prediction method based on the diffusion model described in the first aspect above.

[0007] The radar echo prediction method, device, and electronic device provided by the embodiments of the present invention preprocess the original radar data and input the preprocessed radar data into a pre-trained radar echo prediction model (including a deterministic model, an encoding and decoding module, an adaptation model, and a diffusion model); perform a first prediction on the radar data through the deterministic model and perform a first encoding on the first prediction result through the encoding and decoding module; perform a second encoding on the radar data through the encoding and decoding module and perform a second prediction on the second encoding result through the adaptation model; perform a fusion prediction on the first encoding result and the second prediction result through the diffusion model and perform decoding on the fusion prediction result through the encoding and decoding module to obtain a radar echo prediction result. By adopting the above technology, by introducing an encoding and decoding module, an adaptation model, and a diffusion model to correct the prediction result of the deterministic model of the radar data, rapid prediction of radar echoes can be achieved, thereby improving the efficiency and accuracy of radar echo prediction.

[0008] Other features and advantages of the present invention will be described in the following specification, and, in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention are achieved and obtained by the structures specifically pointed out in the specification, claims, and drawings.

[0009] To make the above objectives, features, and advantages of the present invention more obvious and understandable, the following specific preferred embodiments are given, and in conjunction with the accompanying drawings, the detailed description is as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0011] Figure 1 It is a schematic structural diagram of a radar echo prediction method based on a diffusion model in an embodiment of the present invention;

[0012] Figure 2 It is a schematic structural diagram of a radar echo prediction model in an embodiment of the present invention;

[0013] Figure 3 It is an example diagram of the working principle of a radar echo prediction model in an embodiment of the present invention;

[0014] Figure 4 It is an example diagram of the radar echo prediction effect in an embodiment of the present invention;

[0015] Figure 5 It is a schematic structural diagram of a radar echo prediction device based on a diffusion model in an embodiment of the present invention;

[0016] Figure 6 It is a schematic structural diagram of an electronic device in an embodiment of the present invention. Specific Embodiments

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0018] Currently, for radar echo prediction in short-term and impending precipitation prediction tasks, traditional translation methods, deterministic deep learning methods, and methods using image generation models are mainly used for radar echo extrapolation. The traditional translation method predicts the future radar echo field by translating the current radar echo field in a certain direction. This method is simple and easy to implement, but often results in an overly high intensity of the generated precipitation result because it does not consider the evolution process of the precipitation system. Deterministic deep learning methods (such as U-Net) can better capture the precipitation area, but the prediction results may be relatively fuzzy and lack sensitivity to changes in precipitation intensity. The generation model method (such as the generative adversarial network) can predict local rainfall more clearly and accurately, but the model is unstable during the training process and is difficult to adjust in real time.

[0019] Based on this, a radar echo prediction method, device, and electronic device provided by an embodiment of the present invention can alleviate the above problems existing in the related art.

[0020] For ease of understanding of this embodiment, first, a radar echo prediction method based on a diffusion model disclosed in an embodiment of the present invention will be introduced in detail. See Figure 1 As shown, the method may include the following steps:

[0021] Step S102, preprocess the original radar data and input the preprocessed radar data into a pre-trained radar echo prediction model.

[0022] Among them, the original radar data may be a continuous time series of radar echo signal intensity distribution maps.

[0023] See Figure 2 As shown, the radar echo prediction model may include a deterministic model 21, an encoding and decoding module 22, an adaptation model 23, and a diffusion model 24. The deterministic model 21 may adopt SmaAt-Unet, a sequential model, or a Transformer, etc., and is not limited thereto.

[0024] Step S104, perform a first prediction on the radar data through the deterministic model and perform a first encoding on the first prediction result through the encoding and decoding module.

[0025] Step S106, perform a second encoding on the radar data through the encoding and decoding module and perform a second prediction on the second encoding result through the adaptation model.

[0026] Step S108, perform a fusion prediction on the first encoding result and the second prediction result through the diffusion model and perform decoding on the fusion prediction result through the encoding and decoding module to obtain the radar echo prediction result.

[0027] A radar echo prediction method based on a diffusion model provided by an embodiment of the present invention preprocesses original radar data and inputs the preprocessed radar data into a pre-trained radar echo prediction model (including a deterministic model, an encoding and decoding module, an adaptation model, and a diffusion model); performs a first prediction on the radar data through the deterministic model and performs a first encoding on the first prediction result through the encoding and decoding module; performs a second encoding on the radar data through the encoding and decoding module and performs a second prediction on the second encoding result through the adaptation model; performs a fusion prediction on the first encoding result and the second prediction result through the diffusion model, and decodes the fusion prediction result through the encoding and decoding module to obtain a radar echo prediction result. By adopting the above technology, by introducing an encoding and decoding module, an adaptation model, and a diffusion model to correct the prediction result of the deterministic model of radar data, rapid prediction of radar echoes can be realized, thereby improving the efficiency and accuracy of radar echo prediction.

[0028] As a possible implementation, as shown in Figure 2 and Figure 3 the diffusion model 24 may include a plurality of denoising modules (i.e., the denoising blocks in Figure 3 connected in sequence); based on this, the fusion prediction of the first encoding result and the second prediction result through the diffusion model in the above step S108 may include: fusing the first encoding result and the second prediction result, and gradually denoising the fusion result by using the input conditional information and reference information through a plurality of denoising modules, and then taking the gradually denoised result as the fusion prediction result.

[0029] Among them, the input conditional information and reference information may be the latent space representations of the echo sequence to be corrected (i.e., the preliminary radar echo prediction result, that is, the echo sequence before correction) and the latent space representation of the actual radar echo sequence (i.e., the real-time radar echo sequence), respectively.

[0030] The gradual denoising may include: denoising the current data to be denoised by using the conditional information and reference information through the current denoising module, and taking the current denoising result as the data to be denoised and inputting it into the next denoising module, so as to take the current denoising result as the data to be denoised of the next denoising module and denoise the data to be denoised by using the conditional information and reference information through the next denoising module, and the current data to be denoised is the fusion result for the first denoising module.

[0031] Exemplarily, see Figure 2 and Figure 3As shown, the preprocessed real-time radar echo sequence is input into the prediction block (i.e., the deterministic model 21 at this time) for prediction to obtain a preliminary radar echo prediction result. The preliminary radar echo prediction result is encoded by an encoder (i.e., converted from the pixel space to the latent space) to obtain a first encoded result. The preprocessed real-time radar echo sequence is input into another encoder for encoding (i.e., converted from the pixel space to the latent space) to obtain a second encoded result. The second encoded result is input into the adaptation block (i.e., the adaptation model 23 at this time) for prediction to obtain a second prediction result. Then, the first encoded result and the second prediction result are added together, and the added data, along with the known conditional information and reference information, are input into the diffusion model 24 to perform step-by-step noise reduction in the latent space through a plurality of sequentially connected denoising blocks. The noise is sequentially processed by different denoising blocks using the conditional information and reference information, that is, the input data of each denoising block includes conditional information, reference information, and the output data X of the previous denoising block n-i (i = 0, 1, …, n, and the input data of the first denoising block includes conditional information, reference information, and the added data X n ). Then, the output data X0 of the last denoising block is used as the fusion prediction result, so that the subsequent fusion prediction result can be decoded by a decoder (i.e., converted from the latent space to the pixel space) to obtain a high-precision radar echo prediction result.

[0032] As a possible implementation, referring to Figure 2 and Figure 3 shown, each denoising module (i.e., the denoising block in Figure 3 ) may include a corresponding patch embedding module, a first DiT module, an MLP module, a second DiT module, and a linear decoding module; based on this, the above step of using the conditional information and reference information by the corresponding denoising module to denoise the corresponding data to be denoised may include: performing patch embedding on the corresponding data to be denoised by the corresponding patch embedding module using the conditional information, and performing a first process on the patch embedding result by the corresponding first DiT module; performing a second process on the first process result and the reference information by the corresponding MLP module, and performing a third process on the second process result by the corresponding second DiT module; performing a fifth encoding on the third process result by the corresponding linear decoding module.

[0033] Continuing the previous example, referring to Figure 2 and Figure 3As shown, for a certain denoising block, the input data of the denoising block includes autoregressive variables, conditional information, and reference information, and the output data of the denoising block has new autoregressive variables. The autoregressive variables can be input into the denoising block, and the conditional information can be input into the denoising block. Then, a patch embedding operation is performed and calculations are carried out through a DiT (Diffusion Transformer) module. After that, the reference information and the calculation result of the DiT module are jointly input into an MLP for processing, and calculations are carried out through another DiT module. The calculation result of this DiT module is decoded by a linear decoder to obtain new autoregressive variables. The new autoregressive variables, the conditional information, and the reference information can be jointly used as the input data of the next denoising block to perform a corresponding denoising process through the next denoising block.

[0034] For each denoising block, the autoregressive variable input into the denoising block is a vector with the same dimension as the output vector (i.e., the output result) of the denoising block. Here, the autoregressive variable is the representation of the output result in the latent space. The denoising process of each denoising block can be regarded as an autoregressive process; at the beginning of each autoregressive process, the output vector of the current denoising block is the result after adding noise to the corrected echo sequence (the closer the autoregressive process is to the front, the closer the input autoregressive variable is to pure Gaussian noise). Each autoregressive process uses the output of the previous autoregressive process (i.e., the output vector of the previous denoising block) as the input of the current denoising block to denoise the output vector of the previous denoising block through the current denoising block, and so on. After multiple autoregressive processes (i.e., after denoising through multiple denoising blocks), the denoised vector is finally restored as the step-by-step denoising result.

[0035] As a possible implementation, refer to Figure 2 and Figure 3 shown, the adaptation model 23 (i.e., Figure 3 the adaptation block in

[0036] Continuing with the previous example, refer to Figure 2 and Figure 3As shown, after obtaining the latent space radar echo sequence as the second coding result, the latent space radar echo sequence can be input into an adaptation block with an MHSA (Multi-Head Self-Attention) mechanism and a U-Net. The U-Net is used for feature extraction and feature attention calculation is performed to obtain an adapted radar echo sequence (i.e., the second prediction result at this time), so as to subsequently fuse the first coding result and the second prediction result and sequentially perform step-by-step noise reduction and decoding to obtain a high-precision radar echo prediction result.

[0037] As a possible implementation, refer to Figure 2 and Figure 3 As shown, the deterministic model 21 may include a feature encoding module, a feature decoding module, a prediction encoding module, and a prediction decoding module; based on this, the first prediction of the radar data through the deterministic model in the above step S104 may include: performing third coding on the radar data through the feature encoding module, and performing first decoding on the third coding result through the feature decoding module to obtain the radar signal strength feature and the radar dynamic feature; performing fourth coding on the radar data and the radar signal strength feature through the prediction encoding module, and performing second decoding on the fourth coding result and the radar dynamic feature through the prediction decoding module to obtain the first prediction result.

[0038] Continuing with the previous example, refer to Figure 2 and Figure 3 As shown, after the real-time radar echo sequence is preprocessed, it is first encoded by the feature encoder, and then decoded by the intensity feature decoder and the dynamic feature decoder respectively to obtain the radar signal strength feature and the radar dynamic feature. Then, the result of the preprocessed real-time radar echo sequence is added to the radar signal strength feature and input into the prediction encoder for encoding, and the radar dynamic feature and the output data of the prediction encoder are input into the prediction decoder for decoding to obtain a preliminary radar echo prediction result.

[0039] As a possible implementation, refer to Figure 2 and Figure 3 As shown, the encoding and decoding module 22 may include a first encoder, a second encoder, and a decoder; based on this, the first encoding of the first prediction result through the encoding and decoding module in the above step S104 may include: performing first encoding on the first prediction result through the first encoder. Correspondingly, the second encoding of the radar data through the encoding and decoding module in the above step S106 may include: performing second encoding on the radar data through the first encoder. The decoding of the fused prediction result through the encoding and decoding module in the above step S108 may include: performing decoding on the fused prediction result through the decoder.

[0040] As a possible implementation, refer to Figure 2and Figure 3 As shown, the preprocessing of the original radar data in the above step S102 may include: converting the original radar data into multiple data slices of equal size; detecting outliers in the obtained data slices and deleting the detected outliers; and normalizing the data after deleting the outliers.

[0041] For ease of understanding, the implementation manner of the above radar echo prediction method based on the diffusion model is described exemplarily below by taking a specific application as an example.

[0042] See Figure 2 and Figure 3 As shown, the radar echo prediction model may include a deterministic model 21, an encoder-decoder (i.e., an encoding and decoding module 22), an adaptation network (i.e., an adaptation model 23), and a diffusion model (i.e., a diffusion model 24); the above radar echo prediction method based on the diffusion model mainly includes: a data preprocessing step, a deterministic model construction step, an encoder-decoder training step, an adaptation network construction step, a diffusion model construction step, an overall fine-tuning training step, and a radar echo prediction step. Through the data preprocessing step, quality control, screening, and normalization are performed on the radar echo data; through the deterministic model construction step, the generation, disappearance, and motion characteristics of the radar echo are learned, and a basic radar echo extrapolation result is formed; through the encoder-decoder training step, the mutual conversion between the pixel space and the latent space is realized; through the adaptation network construction step and the diffusion model construction step, the prediction of large values and high-frequency details of the radar echo is realized, and the basic radar echo extrapolation result is corrected; through the overall fine-tuning step, the overall performance of the radar echo prediction model is optimized; through the radar echo prediction step, high-precision prediction of the radar echo is realized.

[0043] See Figure 3 As shown, the above radar echo prediction method based on the diffusion model may include the following implementation steps:

[0044] Step 0, data preprocessing.

[0045] The original radar echo data can be sliced into a size of 512×512 to meet the input requirements of the model, and this step ensures the consistency and processability of the radar echo data.

[0046] Quality control can also be performed on the clutter in the radar echo to reduce the influence of noise on subsequent analysis. This may include using filters or other signal processing techniques to identify and remove outliers; for example, filtering out small echoes to reduce interference and noise in the data and improve the data quality.

[0047] Data normalization processing can also be performed to make the data have a unified scale, which helps the model to learn and generalize better.

[0048] In the actual application process, the operation mode of the data preprocessing step can be as follows: access the radar echo combined reflectivity distribution map, and slice the radar echo combined reflectivity distribution map to obtain slice data. The slice data is a continuous radar observation sequence of a 512×512 image. Then, filter out the data with a radar echo combined reflectivity below 5 dBz, and normalize the filtered data.

[0049] Step 1: Construct a deterministic model.

[0050] Construct a deterministic model with UNet or Transformer as the backbone network to realize the basic extrapolation of radar echoes (i.e., predict the generation, disappearance, and movement laws of radar echoes) through the deterministic model, and use the preprocessed data in Step 0 to train the deterministic model. Figure 3 The prediction block shown in is a form of an available deterministic model. The above radar echo prediction method based on the diffusion model can be adapted to various deterministic models, such as radar echo extrapolation models like SmaAt-Unet, sequential models, and Transformer. Through sufficient training, the deterministic model can basically correctly predict the changes in radar echoes and improve the accuracy of radar echo prediction.

[0051] In the actual application process, the radar echo extrapolation model selected in the deterministic model construction step can be a radar echo prediction model with Transformer or U-Net as the backbone network, can be any model with sequence extrapolation function, or can also be the generator of a trained adversarial generative network (GAN) with radar echo prediction function.

[0052] Step 2: Train the encoder-decoder.

[0053] Figure 3 shows the structure of the encoder-decoder, including two encoders and one decoder. The output end of one encoder is connected to the input end of the adaptation block, and the input end of the other encoder is connected to the output end of the prediction block. Models such as VAE (Variational Auto Encoder) or VQGAN (Vector Quantized Generative Adversarial Network) can be used as the encoder-decoder and train the encoder-decoder to realize the mutual conversion between the pixel space and the latent space through the encoder-decoder. The encoder-decoder can be trained with single-channel single-frame radar data. The encoder extracts features from the input high-dimensional vector to convert the high-dimensional vector into a feature map of a low-dimensional dense vector, and the decoder restores the low-dimensional feature map to a feature map in the high-dimensional pixel space.

[0054] In the actual application process, the encoder and decoder selected in the codec training step can be restricted to an input image size of 512×512, an input image channel number of 1, an output image size of 64×64, and an output image channel number of 4.

[0055] Step 3: Construction of the adaptation network.

[0056] See Figure 3 As shown, the adaptation network has an MHSA mechanism and a U-Net, and can process the latent space radar echo sequence into an adapted radar echo sequence through the adaptation network. The adaptation network can actually be a radar live encoder based on UNet + multi-head attention mechanism. When constructing the adaptation network, the radar live features are encoded to the target size for effective feature extraction and subsequent processing. The U-Net structure is used as the backbone network, and the details are gradually restored during the decoding process to ensure that the details are retained. The adaptation network takes UNet as the backbone and incorporates the MHSA mechanism, and is divided into three parts: an encoder (for downsampling), a decoder (for upsampling), and a skip connection. Through multi-level feature fusion and attention enhancement, the compression and detail restoration of high-resolution radar features are achieved, and the multi-head attention mechanism is added to further enhance the model's ability to capture complex relationships between features. In this way, the features of the radar live are effectively encoded to the target size, providing high-quality input for the subsequent diffusion model.

[0057] Step 4: Construction of the diffusion model.

[0058] See Figure 3 As shown, the diffusion model includes multiple layers of denoising blocks, and each layer of denoising block has two DiT modules. Gradual denoising can be performed through the multiple layers of denoising blocks, and in the process of gradual denoising, the inference of each denoising process is carried out through two layers of DiT modules. The DiT module uses the Transformer module to perform iterative denoising on the radar echo data. Each iteration combines the result of encoding the radar live data (i.e., the denoising result obtained from the previous iteration) and the conditional information and reference information as the basis, so as to gradually restore the clarity and details of the data. After denoising for a specified number of iteration steps, the radar echo with increased details and large values can be restored.

[0059] In the actual application process, the number of layers of the DiT module selected in the diffusion model construction step can be set to 6 to 12 layers to balance the generation effect and inference speed. The input of the diffusion model is set to the radar echo inference results and the actual radar echo sequence for the same time span (such as 3 hours), and the output of the diffusion model is set to the corrected radar echo sequence. The optimal number of autoregressive iteration steps of the diffusion model is between 50 and 100, and the loss function of the diffusion model uses L1 loss or L2 loss.

[0060] Step 5, overall fine-tuning training.

[0061] The goal of the overall fine-tuning training step is to optimize the performance of the entire network while maintaining the learned feature representations. By freezing the parameters of the encoder-decoder, the remaining parameters are then continued to be learned. This strategy further adjusts the network to adapt to specific tasks or datasets while maintaining the original feature extraction ability.

[0062] In the actual application process, the overall fine-tuning step requires freezing the parameters of the encoder-decoder, and training the deterministic model, the adaptation model, and the diffusion model by alternately freezing / thawing the parameters of the deterministic model, the adaptation model, and the diffusion model respectively (that is, freezing the parameters of other models and thawing the parameters of the model being trained), to ensure that the output result of the deterministic model (the intermediate result of the entire network) is still the radar echo, so as to adapt to operations such as stitching and fusion in the business.

[0063] Step 6, radar echo prediction.

[0064] The real-time radar echo sequence of the current period can be input into the entire network after overall fine-tuning training, so that the entire network can calculate and output the high-precision radar echo prediction result for the future period according to the Figure 3 shown structure and working principle.

[0065] Using the above radar echo prediction method based on the diffusion model, by adding a conditional latent diffusion model on the basis of the radar echo extrapolation of the deterministic model, and correcting the large values and details of the radar echo based on the live observation, the accuracy of the large value prediction of the radar echo is improved while retaining the characteristics of the radar echo movement and generation / elimination learned by the deterministic model.

[0066] The advantages of the above radar echo prediction method based on the diffusion model mainly include:

[0067] (1) A correction scheme for the radar echo prediction results of a deterministic model based on the diffusion model is proposed, which solves the problem that the radar echo prediction results in the radar echo prediction task become more and more average and the large values disappear.

[0068] (2) Through the diffusion model, it has better scalability than the traditional GAN and can meet the needs of continuously optimizing and training the model in the business process.

[0069] Based on the above two advantages, the above radar echo prediction method based on the diffusion model can output a radar echo sequence with both large values and details and a small deviation in the position of the large values through the calculation of the entire network. As Figure 4 shown, Figure 4 in which the output of the entire network (i.e., the predicted echo sequence) is the radar echo prediction sequence at 18:00 pm on July 20, 2024 Beijing time. The radar echo from 17:36 to 18:00 on July 20, 2024 Beijing time (i.e., the actual echo sequence) is used as the input of the entire network, and the real radar echo three hours after 18:00 on July 20, 2024 Beijing time is used as the ground truth. The entire network calculates the input actual echo sequence and separately outputs the intermediate prediction result of the deterministic model (i.e., the result of the deterministic model) and the final prediction result of the entire network (i.e., the corrected result). Both the result of the deterministic model and the corrected result are 512km×512km slices at the same position three hours after 18:00 on July 20, 2024 Beijing time; According to Figure 4 it can be seen that the corrected result can accurately restore the position where the large value appears compared with the result of the deterministic model, and the entire network has greater practical value than the deterministic model.

[0070] Based on the above radar echo prediction method based on the diffusion model, the embodiment of the present invention also provides a radar echo prediction device based on the diffusion model. Refer to Figure 5 shown, this device may include the following modules:

[0071] A preprocessing module 502, configured to preprocess the original radar data and input the preprocessed radar data into a pre-trained radar echo prediction model; wherein, the radar echo prediction model includes a deterministic model, an encoding and decoding module, an adaptation model, and a diffusion model.

[0072] A first prediction module 504, configured to perform a first prediction on the radar data through the deterministic model and perform a first encoding on the first prediction result through the encoding and decoding module.

[0073] A second prediction module 506, configured to perform a second encoding on the radar data through the encoding and decoding module and perform a second prediction on the second encoding result through the adaptation model.

[0074] The fusion prediction module 508 is configured to perform fusion prediction on the first encoded result and the second prediction result through the diffusion model, and decode the fusion prediction result through the encoding and decoding module to obtain a radar echo prediction result.

[0075] By adopting the above-mentioned radar echo prediction device based on the diffusion model, the deterministic model prediction result of radar data can be corrected by introducing an encoding and decoding module, an adaptation model and a diffusion model, so as to realize the rapid prediction of radar echoes, thereby improving the efficiency and accuracy of radar echo prediction.

[0076] The radar echo prediction device based on the diffusion model provided by the embodiments of the present invention has the same implementation principle and technical effects as those of the foregoing embodiments of the radar echo prediction method based on the diffusion model. For brief description, for the parts not mentioned in the embodiments of the radar echo prediction device based on the diffusion model, reference may be made to the corresponding content in the foregoing embodiments of the radar echo prediction method based on the diffusion model.

[0077] Embodiments of the present invention also provide an electronic device, as Figure 6 shown, is a schematic structural diagram of the electronic device. Among them, the electronic device includes a processor 61 and a memory 60. The memory 60 stores computer-executable instructions that can be executed by the processor 61, and the processor 61 executes the computer-executable instructions to implement the above-mentioned radar echo prediction method based on the diffusion model.

[0078] In Figure 6 the shown embodiment, the electronic device further includes a bus 62 and a communication interface 63. Among them, the processor 61, the communication interface 63 and the memory 60 are connected through the bus 62.

[0079] Among them, the memory 60 may include a high-speed random access memory (RAM, Random Access Memory), and may also include a non-volatile memory, such as at least one disk memory. Through at least one communication interface 63 (which can be wired or wireless), a communication connection between the system network element and at least one other network element can be realized, and the Internet, wide area network, local area network, metropolitan area network, etc. can be used. The bus 62 may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 62 can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience of representation,Figure 6 It is represented by only one bidirectional arrow in the figure, but it does not mean that there is only one bus or one type of bus.

[0080] The processor 61 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor 61 or the instructions in the form of software. The above-mentioned processor 61 may be a general-purpose processor, including a central processing unit (CPU for short), a network processor (NP for short), etc.; it may also be a digital signal processor (DSP for short), an application specific integrated circuit (ASIC for short), a field programmable gate array (FPGA for short), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the radar echo prediction method based on the diffusion model disclosed in the embodiments of the present invention can be directly embodied as being executed and completed by the hardware decoding processor, or executed and completed by a combination of the hardware and software modules in the decoding processor. The software module may be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, and other well-known storage media in the art. This storage medium is located in the memory, and the processor 61 reads the information in the memory and combines its hardware to complete the steps of the radar echo prediction method based on the diffusion model in the foregoing embodiments.

[0081] Unless otherwise specifically stated, the relative steps, numerical expressions, and values of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0082] When the above-mentioned functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium executable by a processor. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. The 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 the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0083] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation to the present invention. In addition, the terms "first", "second", "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0084] Finally, it should be noted that the above-mentioned embodiments are only specific embodiments of the present invention, used to illustrate the technical solutions of the present invention, rather than limiting them. The protection scope of the present invention is not limited thereto. 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 any person skilled in the art within the technical scope disclosed by the present invention can still modify the technical solutions described in the foregoing embodiments, or can easily think of changes, or make equivalent replacements for some of the technical features; and these modifications, changes 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, and should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the protection scope of the claims.

Claims

1. A radar echo prediction method based on a diffusion model, characterized in that, Including: Preprocessing the original radar data and inputting the preprocessed radar data into a pre-trained radar echo prediction model; wherein, the radar echo prediction model includes a deterministic model, an encoding and decoding module, an adaptation model, and a diffusion model, and the adaptation model includes a backbone network and a multi-head self-attention module; Performing a first prediction on the radar data through the deterministic model and performing a first encoding on the first prediction result through the encoding and decoding module; Performing a second encoding on the radar data through the encoding and decoding module and performing a second prediction on the second encoding result through the adaptation model; Performing a fusion prediction on the first encoding result and the second prediction result through the diffusion model and decoding the fusion prediction result through the encoding and decoding module to obtain a radar echo prediction result; The diffusion model includes a plurality of denoising modules connected in sequence; performing a fusion prediction on the first encoding result and the second prediction result through the diffusion model includes: fusing the first encoding result and the second prediction result, and gradually denoising the fusion result by using the input conditional information and reference information through a plurality of denoising modules, and then using the gradually denoised result as the fusion prediction result; wherein, the gradual denoising includes: denoising the current data to be denoised by using the conditional information and the reference information through the current denoising module, and using the current denoising result as the data to be denoised input to the next denoising module, so as to use the current denoising result as the data to be denoised of the next denoising module and denoise the data to be denoised by using the conditional information and the reference information through the next denoising module, and the current data to be denoised is the fusion result for the first denoising module; Each denoising module includes a corresponding patch embedding module, a first DiT module, an MLP module, a second DiT module, and a linear decoding module; denoising the corresponding data to be denoised by using the conditional information and the reference information through the corresponding denoising module includes: performing patch embedding on the corresponding data to be denoised by using the conditional information through the corresponding patch embedding module, and performing a first processing on the patch embedding result through the corresponding first DiT module; performing a second processing on the first processing result and the reference information through the corresponding MLP module, and performing a third processing on the second processing result through the corresponding second DiT module; performing a fifth encoding on the third processing result through the corresponding linear decoding module.

2. The radar echo prediction method based on a diffusion model according to claim 1, wherein Performing a second prediction on the second encoding result through the adaptation model includes: Extracting features from the second encoding result through the backbone network and performing attention calculation on the extracted features through the multi-head self-attention module to obtain the second prediction result.

3. The radar echo prediction method based on a diffusion model according to claim 1, wherein The deterministic model includes a feature encoding module, a feature decoding module, a prediction encoding module, and a prediction decoding module; Performing a first prediction on the radar data through the deterministic model includes: Performing a third encoding on the radar data through the feature encoding module and performing a first decoding on the third encoding result through the feature decoding module to obtain a radar signal strength feature and a radar dynamic feature; The fourth encoding is performed on the radar data and the radar signal strength feature by the prediction encoding module, and the second decoding is performed on the fourth encoding result and the radar dynamic feature by the prediction decoding module to obtain the first prediction result.

4. The radar echo prediction method based on a diffusion model according to claim 3, wherein The deterministic model adopts SmaAt-Unet, a sequential model or Transformer.

5. The radar echo prediction method based on a diffusion model according to claim 1, characterized in that: The encoding and decoding module includes a first encoder, a second encoder and a decoder; performing the first encoding on the first prediction result by the encoding and decoding module includes: performing the first encoding on the first prediction result by the first encoder; Performing the second encoding on the radar data by the encoding and decoding module includes: performing the second encoding on the radar data by the first encoder; Performing decoding on the fusion prediction result by the encoding and decoding module includes: performing decoding on the fusion prediction result by the decoder.

6. The radar echo prediction method based on the diffusion model according to claim 1, wherein, Preprocessing the original radar data includes: Converting the original radar data into a plurality of data slices of equal size; Performing outlier detection on the obtained data slices and deleting the detected outliers; Normalizing the data after deleting the outliers.

7. A radar echo prediction device based on a diffusion model, characterized in that, Including: A preprocessing module for preprocessing the original radar data and inputting the preprocessed radar data into a pre-trained radar echo prediction model; wherein, the radar echo prediction model includes a deterministic model, an encoding and decoding module, an adaptation model and a diffusion model, and the adaptation model includes a backbone network and a multi-head self-attention module; A first prediction module for performing a first prediction on the radar data by the deterministic model and performing a first encoding on the first prediction result by the encoding and decoding module; A second prediction module for performing a second encoding on the radar data by the encoding and decoding module and performing a second prediction on the second encoding result by the adaptation model; A fusion prediction module for performing a fusion prediction on the first encoding result and the second prediction result by the diffusion model and performing decoding on the fusion prediction result by the encoding and decoding module to obtain a radar echo prediction result; The diffusion model includes a plurality of denoising modules connected in sequence; performing a fusion prediction on the first encoding result and the second prediction result by the diffusion model includes: fusing the first encoding result and the second prediction result, and gradually denoising the fusion result by a plurality of denoising modules using the input conditional information and reference information, and then using the gradually denoised result as the fusion prediction result; wherein, the gradual denoising includes: denoising the current data to be denoised by the current denoising module using the conditional information and the reference information, and using the current denoising result as the data to be denoised input to the next denoising module, so as to use the current denoising result as the data to be denoised of the next denoising module and denoise the data to be denoised by the next denoising module using the conditional information and the reference information, and the current data to be denoised is the fusion result for the first denoising module; Each denoising module includes a corresponding patch embedding module, a first DiT module, an MLP module, a second DiT module, and a linear decoding module; denoising the corresponding data to be denoised by the corresponding denoising module using the conditional information and the reference information includes: performing patch embedding on the corresponding data to be denoised by the corresponding patch embedding module using the conditional information, and performing a first process on the patch embedding result by the corresponding first DiT module; performing a second process on the first process result and the reference information by the corresponding MLP module, and performing a third process on the second process result by the corresponding second DiT module; performing a fifth encoding on the third process result by the corresponding linear decoding module.

8. An electronic device, characterized in that, It includes a processor and a memory, the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the radar echo prediction method based on the diffusion model according to any one of claims 1 to 6.

Images