Radar echo prediction method, device, equipment and medium

Abstract

The present application provides a radar echo prediction method, apparatus, device, and medium, relating to the field of radar echo prediction technology, including: inputting four-dimensional time series data of a target area into a generative radar extrapolation echo prediction model comprising a Transformer-GAN network, a ConvLSTM network, and a fusion network to obtain a predicted image of the radar echo; the Transformer-GAN network is used to generate features from the four-dimensional time series data to obtain an initial radar echo prediction image and prediction accuracy; the ConvLSTM network is used to extract features from the initial radar echo prediction image and the corresponding prediction accuracy to obtain the spatiotemporal features of the initial radar echo prediction image; and the fusion network is used to perform feature fusion to obtain a predicted image of the radar echo. The present application can achieve accurate and clear time series prediction of radar echoes, improving the accuracy, clarity, and realism of the predicted image.

Description

Technical Field

[0001] The present application relates to the technical field of radar echo prediction, and in particular to a radar echo prediction method, apparatus, device and medium. Background Art

[0002] Existing radar echo extrapolation methods mainly include traditional radar echo extrapolation methods and deep learning-based radar echo extrapolation methods. Among them, traditional radar echo extrapolation methods are mostly based on linear extrapolation. When weather conditions are complex and changing drastically, especially when small and medium-scale convective weather occurs, the forecast error will increase significantly. While the current deep learning-based radar echo extrapolation method can to some extent make up for the shortcomings of traditional methods in their inability to perform nonlinear extrapolation, it still has the following drawbacks:

[0003] First, the RNN-based model has temporal memory functions and the characteristics of biological neural networks. The backpropagation algorithm is used to learn parameters between neural units in the recurrent neural network. When the input time series data is long, gradient explosion and vanishing phenomena will occur.

[0004] The second type, based on models based on sequence-to-sequence (Seq2Seq) architecture, is fuzzy in short-term extrapolation and has difficulty in making longer-term predictions;

[0005] The third type is that the convolutional layer in the CNN-based model is conducive to reading spatial information, but lacks the ability to learn temporal information;

[0006] Fourth, models based on generative adversarial networks (GANs) lack the ability to discriminate small-scale information.

[0007] In radar echo fields, stratiform cloud echoes have larger spatial and temporal scales and are more predictable, while strong echoes from convective clouds have smaller spatial and temporal scales and are less predictable. Furthermore, radar echo images differ significantly from conventional spatiotemporal series data, making it difficult to directly apply methods from the field of spatiotemporal series prediction to radar echo extrapolation. Due to limitations in model architecture and the unique characteristics of radar data, radar echo extrapolation still requires improvement in clarity, prediction timeliness, and distortion. The non-rigid motion characteristics of radar images cause their statistical properties to vary continuously over time, exhibiting high-order nonstationarity. This means that the echo distributions at different local locations at the same time differ significantly from those at the same local location at different times. Many current models fail to fully account for the typical characteristics of radar images and lack the ability to model high-order nonstationarity, making it impossible to accurately model long-term motion. This results in low prediction accuracy, insufficient robustness, poor adaptability, and low prediction efficiency when applied to radar echo prediction, making them unsuitable for radar echo prediction in complex scenarios.

[0008] Existing radar echo extrapolation prediction methods solely use ConvLSTMs. While this method can capture spatiotemporal features, its receptive field is limited, making it difficult to effectively capture global information. Furthermore, ConvLSTMs are prone to vanishing or exploding gradients when processing long time series data, which limits their predictive capabilities. Finally, ConvLSTMs' feature extraction capabilities are limited by the size of the convolution kernel, resulting in insufficient representation of complex spatial structures and texture information. Traditional GAN ​​generators are typically based on convolutional networks, but they are prone to mode collapse and vanishing gradients during training. Summary of the Invention

[0009] The purpose of the embodiments of the present application is to provide a radar echo prediction method, device, equipment and medium to solve the above-mentioned problems existing in the prior art and to quickly and accurately predict radar echoes.

[0010] In a first aspect, the present invention provides a method for predicting radar echoes, the method comprising:

[0011] Acquire four-dimensional time series data of the target area; wherein the four-dimensional time series data includes meteorological data and radar data of multiple target locations in the target area at multiple consecutive moments;

[0012] Inputting the four-dimensional time series data into a pre-trained generative radar extrapolation echo prediction model to obtain a predicted image of the radar echo;

[0013] The generative radar extrapolation echo prediction model includes a Transformer-GAN network, a ConvLSTM network, and a fusion network; the Transformer-GAN network is used to generate features for the input four-dimensional time series data to obtain an initial radar echo prediction image and a corresponding prediction accuracy; the ConvLSTM network is used to extract features from the initial radar echo prediction image and the corresponding prediction accuracy output by the Transformer-GAN network to obtain the spatiotemporal features of the initial radar echo prediction image; the fusion network is used to fuse the initial radar echo prediction image with the spatiotemporal features to obtain a predicted image of the radar echo.

[0014] In an optional embodiment, the method for acquiring the four-dimensional time series data includes:

[0015] Acquire meteorological data collected at a first preset time period from a plurality of first locations within the target area, and radar data collected at a second preset time period from a plurality of second locations within the target area;

[0016] Based on the acquisition time and the second position point of the radar data, the radar data and the meteorological data are synchronized in time and space to obtain four-dimensional time series data of the target area.

[0017] In an optional embodiment, the Transformer-GAN network includes: a Transformer encoder, a Transformer generator, and a discriminator;

[0018] The Transformer encoder is used to perform feature processing on the input four-dimensional time series data to obtain first feature data;

[0019] The Transformer generator is used to perform feature extraction and mapping on the first feature data output by the Transformer encoder to obtain an initial radar echo prediction image;

[0020] The discriminator is used to determine the prediction accuracy of the initial radar echo prediction image output by the Transformer generator.

[0021] In an optional embodiment, the Transformer encoder includes: a first input layer, a self-attention layer, and a feedforward layer;

[0022] The first input layer is used to receive the four-dimensional time series data;

[0023] The self-attention layer is used to perform global feature extraction on the four-dimensional time series data to obtain second feature data;

[0024] The feedforward layer is used to perform feature processing on the second feature data to obtain first feature data.

[0025] In an optional embodiment, the Transformer generator includes: a second input layer, a multi-layer Transformer encoder, a first fully connected layer, and a first output layer;

[0026] The second input layer is used to receive the first feature data;

[0027] The multi-layer Transformer encoder is used to perform deep feature extraction on the first feature data through a multi-layer encoder to obtain third feature data;

[0028] The first fully connected layer is used to perform feature mapping on the third feature data to obtain an initial radar echo prediction image;

[0029] The first output layer is used to output the initial radar echo prediction image.

[0030] In an optional embodiment, the discriminator includes: a third input layer, a second fully connected layer, and a second output layer;

[0031] The third input layer is used to receive the initial radar echo prediction image;

[0032] The second fully connected layer is used to determine the prediction accuracy of the initial radar echo prediction image;

[0033] The second output layer is used to output the initial radar echo prediction image and the corresponding prediction accuracy.

[0034] In an optional embodiment, the ConvLSTM network includes: a fourth input layer and a ConvLSTM2D layer;

[0035] The fourth input layer is configured to receive the initial radar echo prediction image and the corresponding prediction accuracy;

[0036] The ConvLSTM2D layer is used to extract the spatiotemporal features of the initial radar echo prediction image.

[0037] In a second aspect, the present invention provides a radar echo prediction device, the device comprising:

[0038] An acquisition unit, configured to acquire four-dimensional time series data of a target area; wherein the four-dimensional time series data includes meteorological data and radar data of multiple target locations in the target area at multiple consecutive moments;

[0039] A prediction unit, configured to input the four-dimensional time series data into a pre-trained generative radar extrapolation echo prediction model to obtain a predicted image of the radar echo;

[0040] The generative radar extrapolation echo prediction model includes a Transformer-GAN network, a ConvLSTM network, and a fusion network; the Transformer-GAN network is used to generate features for the input four-dimensional time series data to obtain an initial radar echo prediction image and a corresponding prediction accuracy; the ConvLSTM network is used to extract features from the initial radar echo prediction image and the corresponding prediction accuracy output by the Transformer-GAN network to obtain the spatiotemporal features of the initial radar echo prediction image; the fusion network is used to fuse the initial radar echo prediction image with the spatiotemporal features to obtain a predicted image of the radar echo.

[0041] In a third aspect, the present invention provides an electronic device, comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0042] Memory for storing computer programs;

[0043] The processor is configured to implement any of the methods described in the foregoing embodiments when executing the program stored in the memory.

[0044] In a fourth aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described in any one of the aforementioned embodiments is implemented.

[0045] This application's generative radar extrapolation echo prediction model integrates a Transformer, a Generative Adversarial Network (GAN), and a ConvLSTM network. The Transformer's self-attention mechanism effectively captures global information and maps it to multiple spaces, enhancing the model's expressive power. The GAN network completely lacks convolution, and the generator leverages the advantages of the Transformer to improve feature resolution while also better capturing low-level textures. Combining the ConvLSTM's ability to capture spatiotemporal features with the Transformer-GAN's image generation capabilities, the resulting generative radar extrapolation echo prediction model (TG-ConvLSTM) can accurately capture spatial information features while simultaneously performing temporal radar echo prediction.

[0046] This application can achieve accurate and clear time series prediction of radar echoes, improving the precision, clarity and realism of the predicted images. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0048] Figure 1 A flow chart of a radar echo prediction method provided in an embodiment of the present application;

[0049] Figure 2 A schematic diagram of the structure of a generative radar extrapolation echo prediction model provided in an embodiment of the present application;

[0050] Figure 3 A schematic diagram of the structure of a radar echo prediction device provided in an embodiment of the present application;

[0051] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0053] This application uses Transformer instead of traditional convolutional network as the generator, which can better capture low-level texture and detail information. The self-attention mechanism of Transformer can alleviate the mode collapse problem in the GAN training process and can better improve the training stability. At the same time, the self-attention mechanism of Transformer can capture global spatial dependencies and effectively make up for the shortcomings of ConvLSTM's local receptive field.

[0054] The radar echo prediction method provided in the embodiments of the present application can be applied in a server or a terminal with strong computing capabilities. The server can be a physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal can be a user equipment (UE) such as a mobile phone, smartphone, laptop, digital broadcast receiver, personal digital assistant (PDA), tablet computer (PAD), handheld device, vehicle-mounted device, wearable device, computing device, other processing device connected to a wireless modem, mobile station (MS), mobile terminal, etc. The terminal and server can be connected directly or indirectly via wired or wireless communication methods, which is not limited in this application.

[0055] The preferred embodiments of the present application are described below in conjunction with the drawings in the specification. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application and are not used to limit the present application. In addition, the embodiments and features in the embodiments of the present application can be combined with each other if there is no conflict.

[0056] Figure 1 A schematic diagram of a radar echo prediction method provided in an embodiment of the present application. Figure 1 As shown, the method may include:

[0057] S110 , acquiring four-dimensional time series data of the target area, wherein the four-dimensional time series data includes meteorological data and radar data of multiple target location points in the target area at multiple consecutive moments.

[0058] In the embodiment of the present application, the multiple consecutive moments may be 5 consecutive moments.

[0059] In an embodiment of the present application, a method for acquiring four-dimensional time series data includes:

[0060] Acquire meteorological data collected at multiple first position points in the target area every first preset time period, and radar data collected at multiple second position points in the target area every second preset time period; based on the collection time of the radar data and the second position point, perform spatiotemporal synchronization on the radar data and the meteorological data to obtain four-dimensional time series data of the target area.

[0061] In the embodiment of the present application, the radar data and the meteorological data are spatiotemporally synchronized based on the first acquisition time and the second location point of the radar data to obtain four-dimensional time series data of the target area, including:

[0062] For any second location point, obtain a preset number of first location points closest to the second location point to obtain a third location point; use the radar data of the second location point corresponding to different first acquisition times as the first time series data; use the meteorological data of the third location point corresponding to different second acquisition times as the second time series data;

[0063] For any first collection time in the first time series data, select a preset number of second collection times closest to the first collection time from the second time series data as target collection times; perform linear interpolation on the meteorological data at the target collection time to obtain meteorological data of the third location at the first collection time;

[0064] Based on the meteorological data and radar data of different third position points at the first acquisition time, the meteorological data and radar data of the second position point at the first acquisition time are obtained; based on the meteorological data and radar data of different second position points at the first acquisition time, the four-dimensional time series data of the target area is obtained.

[0065] In an embodiment of the present application, obtaining meteorological data and radar data of a second location point at the first collection time based on meteorological data and radar data of different third location points at the first collection time includes:

[0066] The weight of each third position point is determined according to the distance between different third position points and the second position point; the meteorological data and radar data of each third position point at the first collection time are weighted and summed to obtain the meteorological data and radar data of the second position point at the first collection time.

[0067] In an embodiment of the present application, the four-dimensional time series data of the target area includes: meteorological data and radar data of each second position point in the target area at different first acquisition times.

[0068] In an embodiment of the present application, the meteorological data includes: radar combined reflectivity, first brightness temperature data, and second brightness temperature data; wherein the first water vapor content is the brightness temperature value of FY4B AGRI channel 10, i.e., the middle-layer water vapor channel; the second water vapor content is the brightness temperature value of FY4B AGRI channel 11, i.e., the low-layer water vapor channel; the radar data is radar combined emissivity data provided by the meteorological station, and the prediction factor is a four-dimensional data. For example, the meteorological data is for an area with a spatial range of 20 meters * 20 meters for three consecutive time periods, and the dimension of the prediction factor is expressed as 5*20*20*3, where 5 represents five consecutive time periods, 20*20 represents the spatial area range of the data, and 3 represents the radar combined reflectivity, the first brightness temperature data, and the second brightness temperature data.

[0069] In practical applications, the horizontal resolution of radar data is 0.1°*0.1° and the temporal resolution is 6 minutes; the spatial resolution of meteorological data is 4 km and the temporal resolution is 15 minutes. Due to the different temporal and spatial resolutions of the two, it is necessary to unify the temporal and spatial resolutions of meteorological data and radar data.

[0070] For example, meteorological data is collected every 15 minutes for every 20m*20m area, and radar data is collected every 6 minutes for every 1m*1m area. Each 1m*1m area of ​​radar data is considered as a location point. For any location point (i.e., the second location point), four adjacent location points (i.e., multiple third location points) are selected from the meteorological data. The four third location points form a square or rectangle, and the target point is located inside it.

[0071] For each third location, the radar data corresponding to the second location are collected at the 0th minute, the 6th minute, and the 12th minute, respectively; the meteorological data corresponding to the third location are collected at the 0th minute and the 15th minute, respectively. For the 6th minute, linear interpolation is performed on the meteorological data at the 0th minute and the 15th minute to obtain the meteorological data for the third location at the 6th minute.

[0072] The weights of the four third position points are determined based on the distances between them and any second position point. Interpolation calculation is performed based on the meteorological data of the four third position points at the 6th minute and the weights of the four third position points to obtain the four-dimensional time series data of the second position point at the 6th minute. The four-dimensional time series data of all second position points at all first acquisition times are the four-dimensional time series data of the target area.

[0073] Unlike other methods that solely use radar echoes as input, this application incorporates data from two infrared channels of the satellite, which are typically used to observe atmospheric temperature and cloud top characteristics at different altitudes. Channel 10 primarily observes surface and cloud top temperatures, and is sensitive to mid- and low-level clouds and surface temperature. Channel 11 primarily observes water vapor distribution and cloud top characteristics in the mid- and high-level atmosphere, and is sensitive to high-level clouds and water vapor content. The brightness temperature data from Channels 10 and 11 can help distinguish between precipitating and non-precipitating clouds. By combining this with radar data, precipitating clouds can be more accurately separated from non-precipitating clouds, thereby improving inversion accuracy.

[0074] S120: Input the four-dimensional time series data into a pre-trained generative radar extrapolation echo prediction model to obtain a predicted image of the radar echo.

[0075] In an embodiment of the present application, the generative radar extrapolation echo prediction model includes a Transformer-GAN network, a ConvLSTM network and a fusion network.

[0076] In the embodiment of the present application, the Transformer-GAN network includes: a Transformer encoder, a Transformer generator, and a discriminator;

[0077] Transformer encoder, used to perform feature processing on the input four-dimensional time series data to obtain first feature data;

[0078] A Transformer generator is used to extract and map the first feature data output by the Transformer encoder to obtain an initial radar echo prediction image;

[0079] The discriminator is used to determine the prediction accuracy of the initial radar echo prediction image output by the Transformer generator.

[0080] In an embodiment of the present application, a Transformer encoder includes: a first input layer, a self-attention layer, and a feedforward layer;

[0081] The first input layer is used to receive four-dimensional time series data;

[0082] The self-attention layer is used to extract global features from the four-dimensional time series data to obtain the second feature data;

[0083] The feedforward layer is used to perform feature processing on the second feature data to obtain the first feature data.

[0084] In an embodiment of the present application, the Transformer generator includes: a second input layer, a multi-layer Transformer encoder, a first fully connected layer, and a first output layer;

[0085] A second input layer, configured to receive first feature data;

[0086] a multi-layer Transformer encoder, configured to perform deep feature extraction on the first feature data through the multi-layer encoder to obtain third feature data;

[0087] The first fully connected layer is used to perform feature mapping on the third feature data to obtain an initial radar echo prediction image;

[0088] The first output layer is used to output the initial radar echo prediction image.

[0089] In one embodiment of the present application, the multi-layer Transformer encoder is a multi-layer Transformer encoder with a local self-attention mechanism, thereby enhancing the ability to capture local details; a local self-attention mechanism is introduced in each layer of the encoder, focusing only on information within a local time window, reducing computational complexity and improving the ability to capture local dependencies; a sliding window or a sparse attention matrix is ​​used to limit the attention range; a feedforward neural network in each layer of the encoder performs a nonlinear transformation on the features; residual connections and layer normalization are used in each layer of the encoder to stabilize the training process; and Dropout or L2 regularization is introduced to prevent overfitting.

[0090] In one embodiment of the present application, the above-mentioned self-attention mechanism can be replaced by a sparse attention mechanism (such as Longformer, BigBird) or a local attention mechanism (such as Reformer) to reduce the computational complexity of the self-attention mechanism.

[0091] In the embodiment of the present application, the discriminator includes: a third input layer, a second fully connected layer, and a second output layer;

[0092] The third input layer is used to receive the initial radar echo prediction image;

[0093] The second fully connected layer is used to determine the prediction accuracy of the initial radar echo prediction image;

[0094] The second output layer is used to output the initial radar echo prediction image and the corresponding prediction accuracy.

[0095] In the embodiment of the present application, the ConvLSTM network includes: a fourth input layer and a ConvLSTM2D layer;

[0096] The fourth input layer is used to receive the initial radar echo prediction image and the corresponding prediction accuracy;

[0097] The ConvLSTM2D layer is used to extract the spatiotemporal features of the initial radar echo prediction image.

[0098] In an embodiment of the present application, a fusion network is used to perform feature fusion on the spatiotemporal features of the initial radar echo prediction image output by the first output layer of the Transformer generator and the initial radar echo prediction image output by the ConvLSTM2D layer; wherein the initial radar echo prediction image captures the global information and low-level texture details of the radar echo; and the spatiotemporal features of the initial radar echo prediction image capture the dynamic changes of the radar echo in time and space.

[0099] In the embodiment of the present application, the fusion network uses attention mechanism fusion to perform feature fusion, including:

[0100] The correlation between the initial radar echo prediction image and the spatiotemporal features of the initial radar echo prediction image is calculated to generate attention weights; the spatiotemporal features of the initial radar echo prediction image and the initial radar echo prediction image are weightedly fused according to the attention weights to obtain the radar echo prediction image.

[0101] In the embodiment of the present application, the predicted image of the radar echo is also four-dimensional data, including: the predicted value of the radar echo of each second position point in the target area within a future preset time period.

[0102] In an embodiment of the present application, the number of predicted time points in a future preset time period is a multiple of the number of time points covered by multiple consecutive moments of the four-dimensional time series data, ensuring coverage of a wider time range.

[0103] For example, you can input meteorological data and radar data for 5 consecutive moments to obtain the radar echo prediction value within a future preset time period consisting of the next 10 consecutive moments.

[0104] This application uses Transformer instead of traditional convolutional network as the generator, which can better capture low-level texture and detail information. The self-attention mechanism of Transformer can alleviate the mode collapse problem in the GAN training process and can better improve the training stability. At the same time, the self-attention mechanism of Transformer can capture global spatial dependencies and effectively make up for the shortcomings of ConvLSTM's local receptive field.

[0105] In one embodiment of the present application, the training of the generative radar extrapolation echo prediction model is distributed to multiple GPUs or nodes to accelerate the training process.

[0106] In an embodiment of the present application, a radar echo prediction method includes:

[0107] (1) Data selection:

[0108] (1) China radar mosaic data uses radar basic emissivity mosaic data provided by the Central Meteorological Observatory, with a horizontal resolution of 0.1°*0.1° and a temporal resolution of 6 minutes.

[0109] (2) FY4B AGRI channels 10 and 11 represent the mid-layer water vapor and low-layer water vapor channels, respectively, with a spatial resolution of 4 km and a temporal resolution of 15 minutes.

[0110] (2) Data fusion:

[0111] Multi-source data fusion refers to the process of integrating and comprehensively utilizing data from different sources across time, space, or feature dimensions. This fusion enables the acquisition of richer, more comprehensive information from multiple perspectives and levels, thereby improving the accuracy and reliability of data analysis and processing. In the temporal dimension, fusing data from different time points can capture dynamic trends; in the spatial dimension, combining data from different locations helps construct more detailed and comprehensive spatial distribution maps; and in the feature dimension, integrating multiple types of feature information can enhance understanding and prediction of complex phenomena.

[0112] Linear interpolation is used to unify the spatial and temporal resolutions of radar and FY4B data. The data used in this paper is a time-series 4D data set (5*400*400*3), which includes the temporal dimension of the observation data. Here, 5 represents five consecutive 6-minute intervals of data, and 3 represents satellite channel 10, channel 11, and radar combined reflectivity data. Each sample contains observation data from five consecutive moments. Experience has shown that using five consecutive moments of observation data in deep learning models achieves a good balance between computation time and forecasting performance.

[0113] (3) Model architecture design:

[0114] Considering the complexity of radar echo data, a radar echo extrapolation model combining Transformer-GAN--ConvLSTM (TG-ConvLSTM) is proposed, such as Figure 2 As shown in Figure 2, the Transformer's self-attention mechanism effectively captures global information and maps it into multiple spaces, thereby enhancing the model's expressive power. A GAN architecture is constructed that completely excludes convolutions. The generator in this architecture leverages the advantages of the Transformer, improving feature resolution while better capturing low-level textures. Combining the spatiotemporal feature capture capabilities of ConvLSTM with the image generation capabilities of Transformer-GAN, the final generative radar echo extrapolation model (TG-ConvLSTM) is obtained.

[0115] Transformer-GAN consists of three parts: the Transformer encoder, the Transformer generator, and the discriminator. The Transformer encoder includes an input layer, a self-attention layer, and a feedforward layer. The input layer receives preprocessed 4D temporal data, the self-attention layer extracts global features, and the feedforward layer further processes the features.

[0116] The Transformer generator includes an input layer, a multi-layer Transformer encoder, a fully connected layer, and an output layer. The input layer is used to receive processed features. The multi-layer Transformer encoder performs deep feature extraction through multi-layer encoders and uses a fully connected layer to map the features. The output layer generates a high-resolution radar echo image.

[0117] The discriminator consists of an input layer, a fully connected layer, and an output layer. The input layer receives real or generated radar echo images, the fully connected layer is used to evaluate the authenticity of the image, and the output layer outputs the authenticity score of the image.

[0118] The ConvLSTM network includes an input layer and a ConvLSTM2D layer, which are used to receive time series four-dimensional data and capture the spatiotemporal features of images respectively.

[0119] The high-resolution image features output by the Transformer generator and the spatiotemporal features output by the ConvLSTM are fused, and further convolution is performed on the representative features to generate the final radar echo forecast image. This radar echo forecast image is a 10*400*400*1 data set, where 10 represents the radar combined reflectivity results at a continuous interval of 6 minutes. This means that this application can extrapolate and forecast the radar echo value for the next hour using 5 frames of data (satellite channels 10 and 11, radar). (That is, 10 6-minute data frames).

[0120] In the embodiment of the present application, the authenticity score is not a fixed indicator. It can be understood as an evaluation indicator of whether the discriminator evaluates whether the input image is a real image. The range is usually between 0 and 1, which is a probability value. A value close to 0 indicates that the input image is very different from the real image, and a value close to 1 indicates that the input image is very different from the real image.

[0121] Corresponding to the above method, the embodiment of the present application also provides a radar echo prediction device, such as Figure 3 As shown, the radar echo prediction device includes:

[0122] An acquisition unit 310 is configured to acquire four-dimensional time series data of a target area, wherein the four-dimensional time series data includes meteorological data and radar data of multiple target locations in the target area at multiple consecutive moments;

[0123] The prediction unit 320 is used to input the four-dimensional time series data into a pre-trained generative radar extrapolation echo prediction model to obtain a predicted image of the radar echo;

[0124] Among them, the generative radar extrapolation echo prediction model includes a Transformer-GAN network, a ConvLSTM network and a fusion network; the Transformer-GAN network is used to generate features for the input four-dimensional time series data to obtain the initial radar echo prediction image and the corresponding prediction accuracy; the ConvLSTM network is used to extract features from the initial radar echo prediction image and the corresponding prediction accuracy output by the Transformer-GAN network to obtain the spatiotemporal features of the initial radar echo prediction image; the fusion network is used to fuse the initial radar echo prediction image and the spatiotemporal features to obtain the radar echo prediction image.

[0125] The functions of each functional unit of the radar echo prediction device provided in the above embodiment of the present application can be achieved through the above method steps. Therefore, the specific working process and beneficial effects of each unit in the radar echo prediction device provided in the embodiment of the present application will not be repeated here.

[0126] The present application also provides an electronic device, such as Figure 4 As shown, it includes a processor 410 , a communication interface 420 , a memory 430 and a communication bus 440 , wherein the processor 410 , the communication interface 420 , and the memory 430 communicate with each other via the communication bus 440 .

[0127] Memory 430, for storing computer programs;

[0128] The processor 410 is configured to execute the program stored in the memory 430 by performing the following steps:

[0129] Acquire four-dimensional time series data of the target area; wherein the four-dimensional time series data includes meteorological data and radar data of multiple target locations in the target area at multiple consecutive moments;

[0130] Input the four-dimensional time series data into the pre-trained generative radar extrapolation echo prediction model to obtain the predicted image of the radar echo;

[0131] Among them, the generative radar extrapolation echo prediction model includes a Transformer-GAN network, a ConvLSTM network and a fusion network; the Transformer-GAN network is used to generate features for the input four-dimensional time series data to obtain the initial radar echo prediction image and the corresponding prediction accuracy; the ConvLSTM network is used to extract features from the initial radar echo prediction image and the corresponding prediction accuracy output by the Transformer-GAN network to obtain the spatiotemporal features of the initial radar echo prediction image; the fusion network is used to fuse the initial radar echo prediction image and the spatiotemporal features to obtain the radar echo prediction image.

[0132] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. This communication bus can be divided into address buses, data buses, and control buses. For ease of illustration, the figure uses only one thick line, but this does not mean that there is only one bus or only one type of bus.

[0133] The communication interface is used for communication between the above electronic device and other devices.

[0134] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage. Alternatively, the memory may be at least one storage device located away from the processor.

[0135] The above-mentioned processor 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 gate or transistor logic devices, and discrete hardware components.

[0136] The implementation methods and beneficial effects of the various components of the electronic device in the above embodiments to solve the problems can be found in Figure 1The various steps in the embodiment shown are implemented, therefore, the specific working process and beneficial effects of the electronic device provided by the embodiment of the present application are not repeated here.

[0137] In another embodiment provided in the present application, a computer-readable storage medium is further provided, wherein instructions are stored in the computer-readable storage medium. When the computer-readable storage medium is executed on a computer, the computer executes the radar echo prediction method described in any one of the above embodiments.

[0138] In another embodiment provided by the present application, a computer program product including instructions is further provided, which, when executed on a computer, enables the computer to execute the radar echo prediction method described in any one of the above embodiments.

[0139] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the embodiments of the present application may take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware. Furthermore, the embodiments of the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0140] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0141] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0142] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0143] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they become aware of the basic creative concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present invention.

[0144] Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims and their equivalents, the embodiments of the present application are also intended to include these modifications and variations.

Claims

1. A radar echo prediction method, characterized in that: The method comprises: Acquire four-dimensional time series data of the target area; wherein the four-dimensional time series data includes meteorological data and radar data of multiple target locations in the target area at multiple consecutive moments; the meteorological data includes: radar combined reflectivity, first brightness temperature data, and second brightness temperature data; Inputting the four-dimensional time series data into a pre-trained generative radar extrapolation echo prediction model to obtain a predicted image of the radar echo; The generative radar extrapolation echo prediction model includes a Transformer-GAN network, a ConvLSTM network, and a fusion network; the Transformer-GAN network is used to generate features for the input four-dimensional time series data to obtain an initial radar echo prediction image and a corresponding prediction accuracy; the ConvLSTM network is used to extract features from the initial radar echo prediction image and the corresponding prediction accuracy output by the Transformer-GAN network to obtain spatiotemporal features of the initial radar echo prediction image; the fusion network is used to fuse the initial radar echo prediction image with the spatiotemporal features to obtain a radar echo prediction image; The Transformer-GAN network includes: a Transformer encoder, a Transformer generator, and a discriminator; the Transformer encoder is a multi-layer Transformer encoder with a local self-attention mechanism; The Transformer encoder is used to perform feature processing on the input four-dimensional time series data to obtain first feature data; The Transformer generator is used to perform feature extraction and mapping on the first feature data output by the Transformer encoder to obtain an initial radar echo prediction image; The discriminator is used to determine the prediction accuracy of the initial radar echo prediction image output by the Transformer generator.

2. The method according to claim 1, wherein The method for acquiring four-dimensional time series data includes: Acquire meteorological data collected at a first preset time period from a plurality of first locations within the target area, and radar data collected at a second preset time period from a plurality of second locations within the target area; Based on the acquisition time and the second position point of the radar data, the radar data and the meteorological data are synchronized in time and space to obtain four-dimensional time series data of the target area.

3. The method according to claim 1, wherein The Transformer encoder includes: a first input layer, a self-attention layer, and a feedforward layer; The first input layer is used to receive the four-dimensional time series data; The self-attention layer is used to perform global feature extraction on the four-dimensional time series data to obtain second feature data; The feedforward layer is used to perform feature processing on the second feature data to obtain first feature data.

4. The method according to claim 3, wherein The Transformer generator includes: a second input layer, a multi-layer Transformer encoder, a first fully connected layer, and a first output layer; The second input layer is used to receive the first feature data; The multi-layer Transformer encoder is used to perform deep feature extraction on the first feature data through a multi-layer encoder to obtain third feature data; The first fully connected layer is used to perform feature mapping on the third feature data to obtain an initial radar echo prediction image; The first output layer is used to output the initial radar echo prediction image.

5. The method according to claim 4, wherein The discriminator includes: a third input layer, a second fully connected layer, and a second output layer; The third input layer is used to receive the initial radar echo prediction image; The second fully connected layer is used to determine the prediction accuracy of the initial radar echo prediction image; The second output layer is used to output the initial radar echo prediction image and the corresponding prediction accuracy.

6. The method according to claim 1, wherein The ConvLSTM network includes: a fourth input layer and a ConvLSTM2D layer; The fourth input layer is configured to receive the initial radar echo prediction image and the corresponding prediction accuracy; The ConvLSTM2D layer is used to extract the spatiotemporal features of the initial radar echo prediction image.

7. A radar echo prediction device, characterized in that: The device comprises: an acquisition unit, configured to acquire four-dimensional time series data of a target area; wherein the four-dimensional time series data includes meteorological data and radar data of multiple target locations in the target area at multiple consecutive moments; the meteorological data includes radar combined reflectivity, first brightness temperature data, and second brightness temperature data; A prediction unit, configured to input the four-dimensional time series data into a pre-trained generative radar extrapolation echo prediction model to obtain a predicted image of the radar echo; The generative radar extrapolation echo prediction model includes a Transformer-GAN network, a ConvLSTM network, and a fusion network; the Transformer-GAN network is used to generate features for the input four-dimensional time series data to obtain an initial radar echo prediction image and a corresponding prediction accuracy; the ConvLSTM network is used to extract features from the initial radar echo prediction image and the corresponding prediction accuracy output by the Transformer-GAN network to obtain spatiotemporal features of the initial radar echo prediction image; the fusion network is used to fuse the initial radar echo prediction image with the spatiotemporal features to obtain a radar echo prediction image; The Transformer-GAN network includes: a Transformer encoder, a Transformer generator, and a discriminator; the Transformer encoder is a multi-layer Transformer encoder with a local self-attention mechanism; The Transformer encoder is used to perform feature processing on the input four-dimensional time series data to obtain first feature data; The Transformer generator is used to perform feature extraction and mapping on the first feature data output by the Transformer encoder to obtain an initial radar echo prediction image; The discriminator is used to determine the prediction accuracy of the initial radar echo prediction image output by the Transformer generator.

8. An electronic device, characterized in that: The electronic device includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; Memory for storing computer programs; A processor, configured to implement the method according to any one of claims 1 to 6 when executing a program stored in a memory.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images