Radar short-term prediction method, device, equipment and medium based on diffusion model

Abstract

The present invention provides a radar short-term prediction method, device, equipment and medium based on a diffusion model, which relates to the field of meteorological data processing technology, including: preprocessing target radar data corresponding to a target area; generating a first radar echo short-term prediction result sequence corresponding to the target area based on the preprocessed target radar data through a deterministic time series prediction model, and generating a first residual prediction value sequence corresponding to the first radar echo short-term prediction result sequence through a residual prediction model; determining the target radar echo short-term prediction result sequence corresponding to the target area based on the first radar echo short-term prediction result sequence and the first residual prediction value sequence. The present invention can alleviate the technical problems of low prediction accuracy and poor texture structure of the existing radar echo short-term prediction method.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological data processing, and in particular to a radar short-term prediction method, device, equipment and medium based on a diffusion model. Background Art

[0002] Precipitation nowcasting based on radar data plays a vital role in extreme weather forecasting and has broad significance for disaster management. Rainy weather is closely related to people's food, clothing, housing, transportation, and property safety. When encountering heavy rainfall weather, it will trigger secondary natural disasters such as mudslides and floods. Its prediction results play an important role in preventing and controlling natural disasters and protecting people's property safety.

[0003] In the actual weather forecasting process, most scenarios for short-term rainfall prediction still use traditional methods, which mainly include forecasting methods based on numerical weather prediction methods and forecasting methods based on radar echo extrapolation. Numerical Weather Prediction (NWP) is a widely used forecasting method that simulates the evolution characteristics of weather systems through calculations, but it involves solving highly complex mathematical models and has high requirements for computing power and time. Therefore, the NWP model is not suitable for short-term weather forecasts for the next few minutes to 6 hours. The forecasting method based on radar echo extrapolation uses the latest observed weather data for real-time predictions. Compared with the NWP model, it has stronger responsiveness, especially under typical meteorological unstable conditions of severe weather disasters such as thunderstorms and heavy rains, and its forecasting effect has significant advantages.

[0004] The traditional radar echo extrapolation method has certain limitations and prerequisites. When encountering mesoscale severe convective weather, its prediction effect will be greatly reduced, and it is impossible to predict the evolution of echoes. Even if (0~1) hours of extrapolation are performed, the prediction error of faster-developing echoes is still large. As time goes by, the accuracy of the extrapolation forecast will drop sharply. In response to the above problems, deep learning technology is introduced on the basis of radar echo images to learn a large amount of continuous historical radar echo data, which provides a new idea for the short-term rainfall prediction method based on radar echo extrapolation.

[0005] Deep learning methods have many application cases in radar short-term prediction. According to different prediction methods, the prediction models can be divided into deterministic models and probabilistic models. Deterministic models can capture motion trends well, but as the prediction step increases, there are problems with echo blur and the gradual disappearance of high-value echoes. Probabilistic models mainly refer to a series of models such as generative adversarial networks and diffusion models. The predicted images obtained by these methods have good texture details, but the predicted landing area is often not very accurate.

[0006] There is no effective solution to the above problems. Summary of the invention

[0007] In view of this, the purpose of the present invention is to provide a radar short-term prediction method, device, equipment and medium based on a diffusion model to alleviate the technical problems of low prediction accuracy and poor texture structure of existing radar short-term prediction methods.

[0008] In a first aspect, the present invention provides a radar short-term prediction method based on a diffusion model, comprising:

[0009] Preprocessing the target radar data corresponding to the target area;

[0010] The preprocessed target radar data is input into a pre-trained target radar short-term prediction model, which includes a deterministic time series prediction model and a residual prediction model based on a diffusion model;

[0011] Generate a first radar echo short-term prediction result sequence corresponding to the target area based on the preprocessed target radar data through a deterministic time series prediction model, and generate a first residual prediction value sequence corresponding to the first radar echo short-term prediction result sequence through a residual prediction model;

[0012] Based on the first radar echo short-term prediction result sequence and the first residual prediction value sequence, a target radar echo short-term prediction result sequence corresponding to the target area is determined.

[0013] In one embodiment, before inputting the preprocessed target radar data into the pre-trained target radar short-term prediction model, the method further includes:

[0014] Preprocess the historical radar data corresponding to the target area;

[0015] A second radar echo short-term prediction result sequence corresponding to the target area is generated based on the preprocessed historical radar data through a deterministic time series prediction model, and a residual true value sequence and a first model loss value corresponding to the deterministic time series prediction model are determined according to the second radar echo short-term prediction result sequence and the radar echo short-term true value sequence corresponding to the target area;

[0016] Generate a second residual prediction value sequence corresponding to the second radar echo short-term prediction result sequence through the residual prediction model, and determine multiple second model loss values ​​corresponding to the residual prediction model using the second residual prediction value sequence and the residual true value sequence;

[0017] A joint loss value is determined based on the first model loss value and each second model loss value, and the deterministic time series prediction model and the residual prediction model are trained using the joint loss value to obtain a target radar short-term prediction model.

[0018] In one implementation, the calculation formula for the first model loss value is as follows:

[0019] ;

[0020] in, is the loss value of the first model; is the first in the radar echo short-term true value sequence Frame radar echo short-term true value; is the first in the sequence of short-term prediction results of the second radar echo Frame radar echo short-term prediction results; is the frame number of the radar echo short-term prediction result contained in the second radar echo short-term prediction result sequence;

[0021] The calculation formula for the loss value of the second model is as follows:

[0022] ;

[0023] in, is the loss value of the second model; is the function symbol; Represents the residual true value sequence , The residual truth value of the frame; is the second residual prediction value sequence Middle The residual prediction value of the frame The denoising state of time steps, is the second residual prediction value sequence Middle The residual prediction value of the frame, is a random normal distribution, is the noise sampled from a random normal distribution, The noise predicted by the residual prediction model;

[0024] The calculation formula of the joint loss value is as follows:

[0025] ;

[0026] in, is the joint loss value, is the weight value.

[0027] In one implementation, generating a first residual prediction value sequence corresponding to a first radar echo short-term prediction result sequence by using a residual prediction model includes:

[0028] Generate conditional feature variables based on the first radar echo short-term prediction result sequence and residual prediction values ​​generated by the residual prediction model for multiple radar echo short-term prediction results located before the current frame in the first radar echo short-term prediction result sequence;

[0029] The random normally distributed noise is input into the residual prediction model, and the residual prediction model is used to control the process of generating residuals based on noise by using conditional characteristic variables to generate residual prediction values ​​corresponding to the short-term prediction results of radar echoes in the current frame;

[0030] A first residual prediction value sequence is constructed based on the residual prediction value corresponding to each frame of the radar echo short-term prediction result in the radar echo short-term prediction result sequence.

[0031] In one embodiment, generating conditional feature variables based on a first radar echo short-term prediction result sequence and residual prediction values ​​generated by a residual prediction model for a plurality of radar echo short-term prediction results located before a current frame in the first radar echo short-term prediction result sequence includes:

[0032] Extracting the global motion trend from the first radar echo short-term prediction result sequence;

[0033] and dividing a sequence of residual prediction values ​​generated by the residual prediction model for a plurality of radar echo short-term prediction results located before the current frame in the first radar echo short-term prediction result sequence to obtain a plurality of residual segments;

[0034] The global motion trend and each residual segment are taken as conditional feature variables.

[0035] In one implementation, a sequence of residual prediction values ​​generated by the residual prediction model for multiple radar echo short-term prediction results located before the current frame in the first radar echo short-term prediction result sequence is divided to obtain multiple residual segments, including:

[0036] A sequence is constructed based on the residual prediction model for residual prediction values ​​generated for multiple radar echo short-term prediction results located before the current frame in the first radar echo short-term prediction result sequence, and the sequence is divided into multiple residual segments according to the ratio between the length of the constructed sequence and the length of the residual segment and the time order.

[0037] In one implementation, determining a target radar echo short-term prediction result sequence corresponding to a target area based on the first radar echo short-term prediction result and the first residual prediction value sequence includes:

[0038] The first radar echo short-term prediction result and the first residual prediction value sequence are subjected to frame addition processing to obtain a target radar echo short-term prediction result sequence corresponding to the target area.

[0039] In a second aspect, the present invention further provides a radar short-term prediction device based on a diffusion model, comprising:

[0040] A data preprocessing module is used to preprocess the target radar data corresponding to the target area;

[0041] A data input module is used to input the pre-processed target radar data into a pre-trained target radar short-term prediction model, where the target radar short-term prediction model includes a deterministic time series prediction model and a residual prediction model based on a diffusion model;

[0042] A preliminary prediction module is used to generate a first radar echo short-term prediction result sequence corresponding to the target area based on the preprocessed target radar data through a deterministic time series prediction model, and to generate a first residual prediction value sequence corresponding to the first radar echo short-term prediction result sequence through a residual prediction model;

[0043] The result determination module is used to determine the target radar echo short-term prediction result sequence corresponding to the target area based on the first radar echo short-term prediction result sequence and the first residual prediction value sequence.

[0044] In a third aspect, the present invention further provides an electronic device, comprising a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement any one of the methods provided in the first aspect.

[0045] In a fourth aspect, the present invention further provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement any one of the methods provided in the first aspect.

[0046] The present invention provides a radar short-term prediction method, device, equipment and medium based on a diffusion model. First, target radar data corresponding to a target area is preprocessed; then the preprocessed target radar data is input into a pre-trained target radar short-term prediction model, wherein the target radar short-term prediction model comprises a deterministic time series prediction model and a residual prediction model based on a diffusion model, thereby generating a first radar echo short-term prediction result sequence corresponding to the target area based on the preprocessed target radar data through the deterministic time series prediction model, and generating a first residual prediction value sequence corresponding to the first radar echo short-term prediction result sequence through the residual prediction model; finally, based on the first radar echo short-term prediction result sequence and the first residual prediction value sequence, the target radar echo short-term prediction result sequence corresponding to the target area is determined. The above method starts from the perspective that the deterministic prediction model is accompanied by echo ambiguity and disappearance of high-value echoes, considers strengthening the texture details of the radar echo short-term prediction result sequence through the diffusion model, and establishes a radar short-term prediction model algorithm based on the diffusion model. The residual distribution of the radar echo short-term prediction result sequence output by the deterministic prediction model is learned through the diffusion model, which improves the texture structure of the radar echo short-term prediction result, solves the problem of echo ambiguity and disappearance of high-value echoes in the deterministic time series prediction model, and thus achieves the technical effect of improving the accuracy and reliability of the radar short-term prediction algorithm.

[0047] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0048] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 A schematic flow chart of a radar short-term prediction method based on a diffusion model provided in an embodiment of the present invention;

[0051] Figure 2 A schematic diagram of a flow chart of a training target radar short-term prediction model provided by an embodiment of the present invention;

[0052] Figure 3 A schematic structural diagram of a radar short-term prediction device based on a diffusion model provided in an embodiment of the present invention;

[0053] Figure 4 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described in combination with the embodiments below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] At present, the existing radar short-term prediction models are divided into two types: deterministic models and probabilistic models. As the prediction step increases, the deterministic model is accompanied by the problem of echo ambiguity and the gradual disappearance of high-value echoes; the prediction landing area of ​​the probabilistic model is often not very accurate. Based on this, the present invention implements a method, device, equipment and medium for radar short-term prediction based on a diffusion model to alleviate the technical problems of low prediction accuracy and poor texture structure of the existing radar short-term prediction method.

[0056] To facilitate understanding of this embodiment, firstly, a radar short-term prediction method based on a diffusion model disclosed in an embodiment of the present invention is described in detail, see Figure 1 The flowchart of a radar short-term prediction method based on a diffusion model is shown, and the method mainly includes the following steps S102 to S108:

[0057] Step S102, preprocessing the target radar data corresponding to the target area, wherein the preprocessing includes analysis processing, quality control processing, projection processing and networking processing.

[0058] Step S104, inputting the preprocessed target radar data into a pre-trained target radar short-term prediction model, where the target radar short-term prediction model includes a deterministic time series prediction model and a residual prediction model based on a diffusion model.

[0059] The deterministic time series prediction model can adopt a prediction model of any structure, whose input is the preprocessed target radar data, and whose output is a radar short-term prediction result sequence, including multiple frames of radar short-term prediction results. In order to distinguish the results output by the model in the training stage and the prediction stage, the embodiment of the present invention records the results output by the model in the prediction stage as the first radar echo short-term prediction result sequence, and records the results output by the model in the training stage as the second radar echo short-term prediction result sequence.

[0060] The residual prediction model based on the diffusion model can adopt a UNet network, whose input is random normally distributed noise, and whose output is a residual prediction value sequence, including the residual prediction value corresponding to the short-term prediction result of each frame of radar echo. Similarly, in order to distinguish the results output by the model in the training stage and the prediction stage, the embodiment of the present invention records the results output by the model in the prediction stage as the first residual prediction value sequence, and records the results output by the model in the training stage as the second residual prediction value sequence.

[0061] Step S106, generating a first radar echo short-term prediction result sequence corresponding to the target area based on the preprocessed target radar data through a deterministic time series prediction model, and generating a first residual prediction value sequence corresponding to the first radar echo short-term prediction result sequence through a residual prediction model.

[0062] Firstly, the preprocessed target radar data is input into the deterministic time series prediction model, which is used to predict the global motion trend of the target radar data and output the corresponding first radar echo short-term prediction result sequence.

[0063] Then, conditional characteristic variables are produced based on the first radar echo short-term prediction result sequence and used as feature inputs of the residual prediction model; random normally distributed noise is input into the residual prediction model, and the conditional generation of residuals is controlled by using the aforementioned conditional characteristic variables to obtain the first residual prediction value sequence.

[0064] Step S108: determining a target radar echo short-term prediction result sequence corresponding to the target area based on the first radar echo short-term prediction result sequence and the first residual prediction value sequence.

[0065] In one example, the first residual prediction value sequence is used to correct the first radar echo short-term prediction result sequence to obtain a target radar echo short-term prediction result sequence with higher accuracy. Optionally, the target radar echo short-term prediction result sequence can be obtained by adding the first radar echo short-term prediction result sequence and the first residual prediction value sequence frame by frame.

[0066] The radar short-term prediction method based on the diffusion model provided by the embodiment of the present invention starts from the perspective that the deterministic prediction model is accompanied by echo blur and disappearance of high-value echoes, considers strengthening the texture details of the radar short-term prediction result sequence through the diffusion model, and establishes a radar short-term prediction model algorithm based on the diffusion model. The residual distribution of the radar echo short-term prediction result sequence output by the deterministic prediction model is learned through the diffusion model, the texture structure of the radar short-term prediction result is improved, and the problem of echo blur and disappearance of high-value echoes in the deterministic time series prediction model is solved, thereby achieving the technical effect of improving the accuracy and reliability of the radar short-term prediction algorithm.

[0067] In one implementation, the target radar short-term prediction model needs to be trained in advance. The embodiment of the present invention provides a specific implementation of training the target radar short-term prediction model, see Figure 2 The flowchart of a training target radar short-term prediction model shown in FIG. 1 includes the following steps S202 to S208:

[0068] Step S202, preprocessing the historical radar data corresponding to the target area, wherein the preprocessing includes analysis processing, quality control processing, projection processing and networking processing.

[0069] In one example, the purpose of the analysis process is to extract the radar echo intensity signal from the historical radar data. The analysis process is as follows: by decoding and converting the original signal received by the radar (that is, the historical radar data), the radar echo intensity signal is extracted, and by analyzing the intensity distribution of the radar echo, the density and size of the precipitation particles can be inferred, and then the precipitation intensity can be estimated.

[0070] Specifically, there is a certain empirical relationship between precipitation intensity (R, unit: mm / h) and radar reflectivity factor (Z, unit: dBZ), which is usually expressed as the ZR relationship:

[0071] ; Among them, a≈200, b is between 1.5 and 2.

[0072] In one example, the purpose of quality control processing is to filter out outliers and errors in the data. The quality control process is as follows: attenuation correction is performed based on the empirical formula and dual-polarization radar method, and data cleaning is performed on the parsed historical radar data based on the denoising algorithm, outlier detection algorithm, and missing value supplementation algorithm.

[0073] In one example, the purpose of projection processing is to convert the quality control processed data into a unified coordinate system. The projection processing process is as follows: Based on the polar coordinate conversion formula, the radar projection polar coordinates are converted into Cartesian coordinates. The coordinate conversion formula is as follows:

[0074] ;

[0075] ;

[0076] ;

[0077] Among them, (r, , ) is the polar coordinate of the target point, r is the distance, is the azimuth, is the elevation angle, and (X, Y, Z) is the converted Cartesian coordinate.

[0078] In one example, the purpose of network processing is to integrate the projected data into a three-dimensional grid. The process of network processing is as follows: by using the observation data of multiple radar stations, combined with differential phase processing, Doppler velocity measurement and other technologies, the radar signal is reconstructed in three dimensions to provide more accurate precipitation distribution information.

[0079] Step S204, using a deterministic time series prediction model, based on the preprocessed historical radar data, generates a second radar echo short-term prediction result sequence corresponding to the target area, and determines a residual true value sequence and a first model loss value corresponding to the deterministic time series prediction model according to the second radar echo short-term prediction result sequence and the radar echo short-term true value sequence corresponding to the target area. The residual true value sequence includes a residual true value corresponding to each frame of the radar echo short-term prediction result, and the residual true value is the residual between the radar echo short-term prediction result and the radar echo short-term true value.

[0080] In one embodiment, the deterministic prediction model is a general module and can be a prediction model of any structure. The purpose of the deterministic prediction model is to make a rough prediction of the historical radar data and obtain the preliminary prediction results of the historical radar data. The training process is as follows: based on the preprocessed historical radar data, the training set and validation set are divided, and the appropriate input size, batch size, etc. are selected. The second radar echo short-term prediction result sequence and the radar echo short-term true value sequence of the deterministic prediction model are used to establish a loss function. The loss function is used to calculate the loss value of the first model, and then train the deterministic time series prediction model.

[0081] Specifically, the calculation formula for the loss value of the first model is as follows:

[0082] ;

[0083] in, is the loss value of the first model; is the first in the radar echo short-term true value sequence Frame radar echo short-term true value; is the first in the sequence of short-term prediction results of the second radar echo Frame radar echo short-term prediction results; is the frame number of the radar echo short-term prediction result contained in the second radar echo short-term prediction result sequence;

[0084] Specifically, the calculation formula of the residual true value sequence is as follows:

[0085] ;

[0086] in, is the residual true value sequence, is the first in the radar echo short-term true value sequence Frame radar echo short-term true value, is the first in the sequence of short-term prediction results of the second radar echo Frame radar echo short-term prediction results.

[0087] Step S206, generating a second residual prediction value sequence corresponding to the second radar echo short-term prediction result sequence through the residual prediction model, and determining a plurality of second model loss values ​​corresponding to the residual prediction model using the second residual prediction value sequence and the residual true value sequence.

[0088] Among them, the specific process of generating the second residual prediction value sequence corresponding to the second radar echo short-term prediction result sequence through the residual prediction model can be referred to the subsequent explanation (three), and the process of generating the residual prediction value sequence through the residual prediction model in the embodiment of the present invention will not be further described.

[0089] In an embodiment of the present invention, noise is added by physical means, and then the corresponding noise reduction process is learned through a residual prediction model based on a diffusion model (such as a UNet network), and the diffusion model is trained based on the residual true value sequence r as supervisory information. Specifically: First, a UNet network is constructed based on a diffusion model, and the residual true value sequence between the second radar echo short-term prediction result sequence and the radar echo short-term true value sequence is used as input to enable the UNet network to learn the residual distribution. Then, the corresponding physical noise removal process is fitted based on the UNet network, and noise is added based on the calculated residual. The residual sequence of the previous segment and the global motion trend are used as conditional variables of the residual prediction model (such as a UNet network), and the loss function is established using the denoised image and the real image. The loss function is used to calculate the loss value of the second model. Finally, a conditional feature variable can be made based on the first radar echo short-term prediction result sequence, which is used as a feature input of the trained UNet network to control the generation conditions of the residual. Among them, the calculation formula of the loss value of the second model is as follows:

[0090] ;

[0091] in, is the loss value of the second model; is the function symbol; Represents the residual true value sequence , The residual truth value of the frame; at time step When it is 0, is the first The residual truth value of the frame; at time step When it is not 0, is the second residual prediction value sequence Middle The residual prediction value of the frame The denoising state of the time step; is the second residual prediction value sequence Middle The residual prediction value of the frame, is a random normal distribution, is the noise sampled from a random normal distribution, is the noise predicted by the residual prediction model.

[0092] Step S208, determining a joint loss value based on the first model loss value and each second model loss value, and using the joint loss value to train the deterministic time series prediction model and the residual prediction model to obtain a target radar short-term prediction model.

[0093] In one implementation, the deterministic time series prediction model and the residual prediction model based on the diffusion model are cascaded and jointly trained in an end-to-end manner. The calculation formula of the joint loss value is as follows:

[0094] ;

[0095] in, is the joint loss value, is the weight value.

[0096] Based on the above embodiments, the trained target radar short-term prediction model can be used to implement radar short-term prediction. Based on this, the embodiment of the present invention provides a specific implementation of a radar short-term prediction method based on a diffusion model.

[0097] (I) Preprocessing the target radar data corresponding to the target area. The specific process of preprocessing the target radar data can be referred to the aforementioned step S202, and the embodiment of the present invention will not elaborate on the preprocessing.

[0098] (ii) Input the preprocessed target radar data into the pre-trained target radar short-term prediction model.

[0099] (III) Generate a first radar echo short-term prediction result sequence corresponding to the target area based on the preprocessed target radar data through a deterministic time series prediction model, and generate a first residual prediction value sequence corresponding to the first radar echo short-term prediction result sequence through a residual prediction model. For details, see (1) to (3) below:

[0100] (1) Based on the first radar echo short-term prediction result sequence and the residual prediction value generated by the residual prediction model for multiple radar echo short-term prediction results before the current frame in the first radar echo short-term prediction result sequence, conditional feature variables are generated. Specifically, it includes the following (1.1) to (1.3):

[0101] (1.1) Extracting the global motion trend from the first radar echo short-term prediction result sequence. In one example, a layer of Conv-RNN network is used to extract the global motion trend of the first radar echo short-term prediction result sequence. The Conv-RNN network consists of a downsampling operator, a ConvGRU operator, and a Resnet operator.

[0102] (1.2) Dividing a sequence of residual prediction values ​​generated by the residual prediction model for multiple radar echo short-term prediction results located before the current frame in the first radar echo short-term prediction result sequence to obtain multiple residual segments.

[0103] In an embodiment of the present invention, the residual sequence r is divided into a plurality of residual segments s, and the residual of the previous sequence segment is used as a conditional characteristic variable for residual prediction of the next sequence segment and the global motion trend. In one example, a sequence is constructed based on the residual prediction model for the residual prediction values ​​generated for the plurality of radar echo short-term prediction results before the current frame in the first radar echo short-term prediction result sequence, and the sequence is divided into a plurality of residual segments according to the ratio between the length of the constructed sequence and the length of the residual segment and the time sequence.

[0104] In the specific implementation, based on the residual sequence, it is divided into multiple segments according to the input and output sequence ratio and time order. The division formula is: , where s represents the partition frame, r represents the residual sequence, K represents the length of each segment, j=0,1,…, , Represents the length of the residual sequence.

[0105] (1.3) The global motion trend and each residual segment are taken as conditional feature variables.

[0106] (2) The random normally distributed noise is input into the residual prediction model, and the residual prediction model is used to control the process of generating residuals based on noise by using conditional feature variables to generate residual prediction values ​​corresponding to the short-term prediction results of the radar echo of the current frame. Preferably, a temporal attention mechanism is added between the residual segments.

[0107] In the specific implementation, the residual prediction model based on the diffusion model is constructed, the physical forward diffusion process is constructed, the noise of random normal distribution is gradually increased, the inverse denoising process is constructed based on the Unet network, the temporal attention mechanism is constructed, and the Conv-RNN network is constructed to extract the global motion trend of the first radar echo short-term prediction result sequence. The residual sequence and global motion trend of the previous segment are used as the conditional feature variables of the Unet network, and the conditional feature variables are used to control the generation of the residual prediction value corresponding to the short-term prediction result of the radar echo of the current frame.

[0108] (3) Constructing a first residual prediction value sequence based on the residual prediction value corresponding to each frame of the radar echo short-term prediction result in the first radar echo short-term prediction result sequence.

[0109] (iv) Determine the target radar echo short-term prediction result sequence corresponding to the target area based on the first radar echo short-term prediction result sequence and the first residual prediction value sequence. In one embodiment, the first radar echo short-term prediction result and the first residual prediction value sequence may be processed by frame addition to obtain the target radar echo short-term prediction result sequence corresponding to the target area. The target radar echo short-term prediction result sequence may be determined according to the following formula:

[0110] ;

[0111] in, is the target radar echo short-term prediction result sequence, is the short-term prediction result of the first radar echo, is the first residual prediction value sequence.

[0112] In summary, the radar short-term prediction method based on the diffusion model provided in the embodiment of the present invention achieves the purpose of using the diffusion model to improve the texture structure of the prediction result of the deterministic prediction model, thereby solving the problem of echo ambiguity and disappearance of high-value echoes in the deterministic time series prediction model, thereby achieving the technical effect of improving the accuracy and reliability of the radar short-term prediction algorithm.

[0113] Based on the above embodiments, the present invention provides a radar short-term prediction device based on a diffusion model. Figure 3 The structure diagram of a radar short-term prediction device based on a diffusion model is shown in FIG. The device mainly includes the following parts:

[0114] The data preprocessing module 302 is used to preprocess the target radar data corresponding to the target area;

[0115] A data input module 304 is used to input the pre-processed target radar data into a pre-trained target radar short-term prediction model, where the target radar short-term prediction model includes a deterministic time series prediction model and a residual prediction model based on a diffusion model;

[0116] The preliminary prediction module 306 is used to generate a first radar echo short-term prediction result sequence corresponding to the target area based on the preprocessed target radar data through a deterministic time series prediction model, and generate a first residual prediction value sequence corresponding to the first radar echo short-term prediction result sequence through a residual prediction model;

[0117] The result determination module 308 is used to determine the target radar echo short-term prediction result sequence corresponding to the target area based on the first radar echo short-term prediction result sequence and the first residual prediction value sequence.

[0118] The radar short-term prediction method based on the diffusion model provided by the embodiment of the present invention starts from the perspective that the deterministic prediction model is accompanied by echo blur and disappearance of high-value echoes, considers strengthening the texture details of the radar short-term prediction result sequence through the diffusion model, and establishes a radar short-term prediction model algorithm based on the diffusion model. The residual distribution of the radar echo short-term prediction result sequence output by the deterministic prediction model is learned through the diffusion model, the texture structure of the radar short-term prediction result is improved, and the problem of echo blur and disappearance of high-value echoes in the deterministic time series prediction model is solved, thereby achieving the technical effect of improving the accuracy and reliability of the radar short-term prediction algorithm.

[0119] In one embodiment, a training module is further included for:

[0120] Preprocess the historical radar data corresponding to the target area;

[0121] A second radar echo short-term prediction result sequence corresponding to the target area is generated based on the preprocessed historical radar data through a deterministic time series prediction model, and a residual true value sequence and a first model loss value corresponding to the deterministic time series prediction model are determined according to the second radar echo short-term prediction result sequence and the radar echo short-term true value sequence corresponding to the target area;

[0122] Generate a second residual prediction value sequence corresponding to the second radar echo short-term prediction result sequence through the residual prediction model, and determine multiple second model loss values ​​corresponding to the residual prediction model using the second residual prediction value sequence and the residual true value sequence;

[0123] A joint loss value is determined based on the first model loss value and each second model loss value, and the deterministic time series prediction model and the residual prediction model are trained using the joint loss value to obtain a target radar short-term prediction model.

[0124] In one implementation, the calculation formula for the first model loss value is as follows:

[0125] ;

[0126] in, is the loss value of the first model; is the first in the radar echo short-term true value sequence Frame radar echo short-term true value; is the first in the sequence of short-term prediction results of the second radar echo Frame radar echo short-term prediction results; is the frame number of the radar echo short-term prediction result contained in the second radar echo short-term prediction result sequence;

[0127] The calculation formula for the loss value of the second model is as follows:

[0128] ;

[0129] in, is the loss value of the second model; is the function symbol; Represents the residual true value sequence , The residual truth value of the frame; is the second residual prediction value sequence Middle The residual prediction value of the frame The denoising state of time steps, is the second residual prediction value sequence Middle The residual prediction value of the frame, is a random normal distribution, is the noise sampled from a random normal distribution, The noise predicted by the residual prediction model;

[0130] The calculation formula of the joint loss value is as follows:

[0131] ;

[0132] in, is the joint loss value, is the weight value.

[0133] In one implementation, the preliminary prediction module 306 is specifically used to:

[0134] Generate conditional feature variables based on the first radar echo short-term prediction result sequence and residual prediction values ​​generated by the residual prediction model for multiple radar echo short-term prediction results located before the current frame in the first radar echo short-term prediction result sequence;

[0135] The random normally distributed noise is input into the residual prediction model, and the residual prediction model is used to control the process of generating residuals based on noise by using conditional characteristic variables to generate residual prediction values ​​corresponding to the short-term prediction results of radar echoes in the current frame;

[0136] A first residual prediction value sequence is constructed based on the residual prediction value corresponding to each frame of the radar echo short-term prediction result in the radar echo short-term prediction result sequence.

[0137] In one implementation, the preliminary prediction module 306 is specifically used to:

[0138] Extracting the global motion trend from the first radar echo short-term prediction result sequence;

[0139] and dividing a sequence of residual prediction values ​​generated by the residual prediction model for a plurality of radar echo short-term prediction results located before the current frame in the first radar echo short-term prediction result sequence to obtain a plurality of residual segments;

[0140] The global motion trend and each residual segment are taken as conditional feature variables.

[0141] In one implementation, the preliminary prediction module 306 is specifically used to:

[0142] A sequence is constructed based on the residual prediction model for residual prediction values ​​generated for multiple radar echo short-term prediction results located before the current frame in the first radar echo short-term prediction result sequence, and the sequence is divided into multiple residual segments according to the ratio between the length of the constructed sequence and the length of the residual segment and the time order.

[0143] In one implementation, the result determination module 308:

[0144] The first radar echo short-term prediction result and the first residual prediction value sequence are subjected to frame addition processing to obtain a target radar echo short-term prediction result sequence corresponding to the target area.

[0145] The device provided in the embodiment of the present invention has the same implementation principle and technical effects as those of the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the device embodiment, reference may be made to the corresponding contents in the aforementioned method embodiment.

[0146] An embodiment of the present invention provides an electronic device. Specifically, the electronic device includes a processor and a storage device. The storage device stores a computer program, and when the computer program is executed by the processor, it executes the method described in any one of the above-mentioned implementation methods.

[0147] Figure 4 A structural diagram of an electronic device provided in an embodiment of the present invention, the electronic device 100 includes: a processor 40, a memory 41, a bus 42 and a communication interface 43, wherein the processor 40, the communication interface 43 and the memory 41 are connected via the bus 42; the processor 40 is used to execute an executable module stored in the memory 41, such as a computer program.

[0148] The memory 41 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 43 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

[0149] The bus 42 may be an ISA bus, a PCI bus, or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 4 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0150] Among them, the memory 41 is used to store programs, and the processor 40 executes the program after receiving the execution instruction. The method executed by the device for flow process definition disclosed in any embodiment of the above-mentioned embodiment of the present invention can be applied to the processor 40 or implemented by the processor 40.

[0151] The processor 40 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 hardware integrated logic circuit or software instructions in the processor 40. The above processor 40 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present invention can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiment of the present invention can be directly embodied as a hardware decoding processor to be executed, or the hardware and software modules in the decoding processor can be executed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 41, and the processor 40 reads the information in the memory 41 and completes the steps of the above method in combination with its hardware.

[0152] The computer program product of the readable storage medium provided in the embodiment of the present invention includes a computer-readable storage medium storing program code, and the instructions included in the program code can be used to execute the methods described in the previous method embodiments. The specific implementation can be referred to the previous method embodiments, which will not be repeated here.

[0153] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the 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, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0154] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; 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 be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A radar short-term prediction method based on a diffusion model, characterized in that: include: Preprocessing the target radar data corresponding to the target area; Inputting the preprocessed target radar data into a pre-trained target radar short-term prediction model, wherein the target radar short-term prediction model includes a deterministic time series prediction model and a residual prediction model based on a diffusion model; Generate a first radar echo short-term prediction result sequence corresponding to the target area based on the preprocessed target radar data by using the deterministic time series prediction model, and generate a first residual prediction value sequence corresponding to the first radar echo short-term prediction result sequence by using the residual prediction model; Based on the first radar echo short-term prediction result sequence and the first residual prediction value sequence, a target radar echo short-term prediction result sequence corresponding to the target area is determined.

2. The radar short-term prediction method based on the diffusion model according to claim 1 is characterized in that: Before inputting the preprocessed target radar data into the pre-trained target radar short-term prediction model, the method further includes: Preprocess the historical radar data corresponding to the target area; Generate a second radar echo short-term prediction result sequence corresponding to the target area based on the preprocessed historical radar data through the deterministic time series prediction model, and determine a residual true value sequence and a first model loss value corresponding to the deterministic time series prediction model according to the second radar echo short-term prediction result sequence and the radar echo short-term true value sequence corresponding to the target area; Generate a second residual prediction value sequence corresponding to the second radar echo short-term prediction result sequence through the residual prediction model, and determine a plurality of second model loss values ​​corresponding to the residual prediction model according to the second residual prediction value sequence and the residual true value sequence; A joint loss value is determined based on the first model loss value and each of the second model loss values, and the deterministic time series prediction model and the residual prediction model are trained using the joint loss value to obtain a target radar short-term prediction model.

3. The radar short-term prediction method based on the diffusion model according to claim 2 is characterized in that: The calculation formula of the loss value of the first model is as follows: ； in, is the loss value of the first model; is the first Frame radar echo short-term true value; is the first in the sequence of short-term prediction results of the second radar echo Frame radar echo short-term prediction results; is the frame number of the radar echo short-term prediction result contained in the second radar echo short-term prediction result sequence; The calculation formula of the loss value of the second model is as follows: ； in, is the loss value of the second model; is the function symbol; Represents the first , The residual truth value of the frame; is the second residual prediction value sequence Middle The residual prediction value of the frame The denoising state of time steps, is the second residual prediction value sequence Middle The residual prediction value of the frame, is a random normal distribution, is the noise sampled from a random normal distribution, is the noise predicted by the residual prediction model; The calculation formula of the joint loss value is as follows: ； in, is the joint loss value, is the weight value.

4. The radar short-term prediction method based on the diffusion model according to claim 1 is characterized in that: Generating a first residual prediction value sequence corresponding to the first radar echo short-term prediction result sequence by using the residual prediction model includes: Generate a conditional feature variable based on the first radar echo short-term prediction result sequence and the residual prediction value generated by the residual prediction model for a plurality of radar echo short-term prediction results located before a current frame in the first radar echo short-term prediction result sequence; Inputting random normally distributed noise into the residual prediction model, and controlling the process of generating residuals based on noise by using the conditional characteristic variables through the residual prediction model to generate a residual prediction value corresponding to the short-term prediction result of the radar echo of the current frame; A first residual prediction value sequence is constructed based on the residual prediction value corresponding to the short-term prediction result of each frame of the radar echo in the first radar echo short-term prediction result sequence.

5. The radar short-term prediction method based on the diffusion model according to claim 4 is characterized in that: Generating conditional feature variables based on the first radar echo short-term prediction result sequence and the residual prediction value generated by the residual prediction model for a plurality of radar echo short-term prediction results located before a current frame in the first radar echo short-term prediction result sequence, including: Extracting the global motion trend from the first radar echo short-term prediction result sequence; and dividing a sequence of residual prediction values ​​generated by the residual prediction model for a plurality of radar echo short-term prediction results located before a current frame in the first radar echo short-term prediction result sequence to obtain a plurality of residual segments; The global motion trend and each of the residual segments are used as conditional feature variables.

6. The radar short-term prediction method based on the diffusion model according to claim 5 is characterized in that: A sequence consisting of residual prediction values ​​generated by the residual prediction model for multiple radar echo short-term prediction results located before the current frame in the first radar echo short-term prediction result sequence is divided to obtain multiple residual segments, including: A sequence is constructed based on the residual prediction model for residual prediction values ​​generated for multiple radar echo short-term prediction results located before the current frame in the first radar echo short-term prediction result sequence, and the sequence is divided into multiple residual segments according to the ratio between the length of the constructed sequence and the length of the residual segment and the time order.

7. The radar short-term prediction method based on the diffusion model according to claim 1 is characterized in that: Determining a target radar echo short-term prediction result sequence corresponding to the target area based on the first radar echo short-term prediction result and the first residual prediction value sequence, including: The first radar echo short-term prediction result and the first residual prediction value sequence are subjected to frame addition processing to obtain a target radar echo short-term prediction result sequence corresponding to the target area.

8. A radar short-term prediction device based on a diffusion model, characterized in that: include: A data preprocessing module is used to preprocess the target radar data corresponding to the target area; A data input module, used for inputting the pre-processed target radar data into a pre-trained target radar short-term prediction model, wherein the target radar short-term prediction model includes a deterministic time series prediction model and a residual prediction model based on a diffusion model; A preliminary prediction module, configured to generate a first radar echo short-term prediction result sequence corresponding to the target area based on the preprocessed target radar data by using the deterministic time series prediction model, and to generate a first residual prediction value sequence corresponding to the first radar echo short-term prediction result sequence by using the residual prediction model; The result determination module is used to determine the target radar echo short-term prediction result sequence corresponding to the target area based on the first radar echo short-term prediction result sequence and the first residual prediction value sequence.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the method according to any one of claims 1 to 7.

Images