Radar echo extrapolation method based on hypergraph neural network and related equipment

Abstract

The invention provides a radar echo extrapolation method based on a hypergraph neural network and related equipment. According to the method, after historical echo data of a weather radar is obtained, weighted echo features are obtained based on a hypergraph neural network according to the historical echo data, echo trend features are obtained based on a first radar return extrapolation network according to the weighted echo features and the historical echo data, and the echo trend features are obtained based on a second radar return extrapolation network. And obtaining texture detail features according to the historical echo data, and finally obtaining extrapolation echo prediction features of the historical echo data according to the echo trend features and the texture detail features. Compared with the prior art, the method introduces the hypergraph neural network to carry out radar echo extrapolation, and improves the accuracy.

Description

Technical Field

[0001] The present application relates to the field of weather forecasting technology, and in particular to a radar echo extrapolation method based on a hypergraph neural network and related equipment. Background Art

[0002] Currently, weather forecasting, especially short-term precipitation prediction, as one of the main tasks of meteorology, has always received extensive attention from scholars in related fields. Due to the high nonlinearity of time dynamics and the complex position characteristic patterns in the spatial domain, refined precipitation forecasting within 2 hours with high temporal and spatial resolution is an extremely challenging task.

[0003] The current general method is to use locally invariant state transitions or motion vectors to capture short-term temporal dependencies, but lacks the representation of long-term motion trends, resulting in poor prediction accuracy.

[0004] That is, the current radar echo extrapolation method has the technical problem of poor prediction accuracy and needs to be improved. Summary of the invention

[0005] The embodiments of the present application provide a radar echo extrapolation method, apparatus, computer equipment and storage medium based on a hypergraph neural network, which are used to improve the technical problem of poor prediction accuracy of the current radar echo extrapolation method.

[0006] In order to solve the above technical problems, the embodiments of the present application provide the following technical solutions:

[0007] The present application provides a radar echo extrapolation method, which includes:

[0008] Get historical echo data of weather radar;

[0009] Based on a hypergraph neural network, obtaining weighted echo features according to the historical echo data;

[0010] Based on the first radar report extrapolation network, obtaining echo trend characteristics according to the weighted echo characteristics and the historical echo data;

[0011] Based on the second radar report extrapolation network, texture detail features are obtained according to the historical echo data;

[0012] The extrapolated echo prediction features of the historical echo data are obtained according to the echo trend features and the texture detail features.

[0013] In one embodiment, the aforementioned method is based on a hypergraph neural network, and obtaining weighted echo features according to the historical echo data includes:

[0014] Obtaining hypergraph features of the historical echo data through the hypergraph neural network;

[0015] Acquiring auxiliary features of the historical echo data;

[0016] The weighted echo feature is determined according to the hypergraph feature and the auxiliary feature.

[0017] In one embodiment, the aforementioned method determines the weighted echo feature according to the hypergraph feature and the auxiliary feature, including:

[0018] Obtaining a query vector according to the hypergraph feature and the auxiliary feature;

[0019] Acquire a feature vector in a feature space including the historical echo data;

[0020] Based on the attention mechanism, the weighted echo feature is obtained according to the query vector and the feature vector.

[0021] In one embodiment, the aforementioned method obtains the hypergraph features of the historical echo data through the hypergraph neural network, including:

[0022] Dividing the historical echo data into a plurality of pixel blocks according to the latitude and longitude information;

[0023] Constructing hypergraph data according to the feature similarity between the pixel blocks;

[0024] The hypergraph data is processed to obtain the hypergraph features.

[0025] In one embodiment, the aforementioned method is based on the first radar report extrapolation network, and obtains the echo trend feature according to the weighted echo feature and the historical echo data, including:

[0026] Processing the historical echo data to obtain echo prediction data through the first radar report extrapolation network;

[0027] The echo trend feature is obtained according to the weighted echo feature and the echo prediction data.

[0028] In one embodiment, the aforementioned method obtains the extrapolated echo prediction feature of the historical echo data according to the echo trend feature and the texture detail feature, including:

[0029] The echo wave trend feature and the texture detail feature are spliced ​​to obtain an extrapolated echo feature.

[0030] In one embodiment, after obtaining the extrapolated echo prediction feature of the historical echo data according to the echo trend feature and the texture detail feature, the aforementioned method further includes:

[0031] Obtaining a true value of an echo image corresponding to the historical echo data;

[0032] The radar echo extrapolation network model is trained according to the extrapolated echo prediction features and the true value of the echo image; the radar echo extrapolation network model includes the hypergraph neural network, the first radar return extrapolation network and the second radar return extrapolation network.

[0033] At the same time, the present application also provides a radar echo extrapolation device, which includes:

[0034] Acquisition module, used to obtain historical echo data of weather radar;

[0035] A first processing module, configured to obtain weighted echo features according to the historical echo data based on a hypergraph neural network;

[0036] A second processing module, configured to obtain an echo trend feature based on the first radar report extrapolation network and the weighted echo feature and the historical echo data;

[0037] A third processing module is used to obtain texture detail features according to the historical echo data based on the second radar report extrapolation network;

[0038] The output module is used to obtain the extrapolated echo prediction features of the historical echo data according to the echo trend features and the texture detail features.

[0039] At the same time, the present application provides a computer device, which includes a processor and a memory, wherein the memory stores a plurality of instructions, and the instructions are suitable for the processor to load to execute the steps in the above method.

[0040] At the same time, the present application provides a computer-readable storage medium, which stores multiple instructions, and the instructions are suitable for a processor to load to execute the steps in the above method.

[0041] At the same time, the present application provides a computer program product or a computer program, which includes computer instructions, and the computer instructions are stored in a computer-readable storage medium; a processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the steps in the above method.

[0042] Beneficial effects: The present application provides a radar echo extrapolation method, device, computer equipment and storage medium; the method comprises: after obtaining the historical echo data of the weather radar, obtaining weighted echo features according to the historical echo data based on a hypergraph neural network, respectively obtaining echo trend features according to the weighted echo features and the historical echo data based on a first radar return extrapolation network, and obtaining texture detail features according to the historical echo data based on a second radar return extrapolation network, and finally obtaining the extrapolated echo prediction features of the historical echo data according to the echo trend features and the texture detail features. Compared with the current technology, the present application introduces a hypergraph neural network for meteorological forecasting, updates the feature representation of the radar image based on the hypergraph neural network, and in order to fully learn the continuous generation, dissipation, rotation and regeneration of the radar echo image, the first branch focuses on learning the overall motion trend in the radar image, the second branch focuses on describing the texture detail features of the radar echo image, and combines the feature memory module to record the long-term echo law to obtain the final echo extrapolation result as the meteorological forecast data, thereby improving the accuracy of the forecast result. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The technical solution and beneficial effects of the present application will be made apparent by describing in detail the specific implementation methods of the present application in conjunction with the accompanying drawings.

[0044] Figure 1 is a network diagram of a cloud system provided in an embodiment of the present application;

[0045] Figure 2 is a flow chart of a radar echo extrapolation method provided in an embodiment of the present application;

[0046] Figure 3 is a schematic diagram of the structure of a radar echo extrapolation device provided in an embodiment of the present application;

[0047] Figure 4 It is a data processing schematic diagram provided in an embodiment of the present application;

[0048] Figure 5 It is a schematic diagram of a data processing framework provided in an embodiment of the present application;

[0049] Figure 6 It is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of this application.

[0051] In the description of the embodiments of the present application, it is to be understood that the terms "first", "second", etc. in the specification, claims, and drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that in the drawings or descriptions. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions.

[0052] The method involved in the embodiment of the present application can be implemented by a server, specifically, by a server in a cloud system. For example, meteorological data can be predicted, stored, and called based on cloud technology to ensure the validity, timeliness, and reliability of the meteorological data and the accuracy of the prediction results.

[0053] Cloud technology refers to a hosting technology that unifies hardware, software, network and other series of resources in a wide area network or local area network to realize the calculation, storage, processing and sharing of data. Cloud technology is a general term for network technology, information technology, integration technology, management platform technology, application technology, etc. based on the application of cloud computing business model, which can form a resource pool, which is used on demand and flexible and convenient. Cloud computing technology will become an important support. The background service of the technical network system requires a large amount of computing and storage resources, such as meteorological data. With the high development and application of the Internet industry, more meteorological data may have their own identification marks in the future, and they all need to be transmitted to the background system for logical processing. Different types of data will be processed separately, and all kinds of meteorological forecasting industry data need strong system backing support, which can only be achieved through cloud computing. In this application, cloud technology can be used to realize the management of meteorological data, such as binding the real-name information and operation instructions of the query user based on the secure shell protocol (Secure Shell, SSH) and the command processor (Bourne Again Shell, Bash) to realize the real-name tracking in the weather forecast process and ensure the security and reliability of cloud data.

[0054] The method provided in the embodiment of the present application can be executed through a cloud system. For details, please refer to Figure 1 , Figure 1A network diagram of a cloud system provided in an embodiment of the present application, the system may include a data terminal 101, a storage server 102, a business server 103, a prediction server 104, a data server 105 and a user terminal 106, wherein:

[0055] The terminal and the server may be connected directly or indirectly via wired or wireless communication, which is not limited in this application.

[0056] The data terminal 101 and the user terminal 106 can be portable terminals such as smart phones, tablet computers, and laptop computers, or fixed terminals such as desktop computers and query machines, or weather radars, but are not limited thereto. The data terminal 101 and the user terminal 106 are both service ports that users can use and operate. The data terminal 101 can be used to provide users with a window for logging into the prediction server so as to upload meteorological data such as radar data (including historical echo data and corresponding meteorological prediction data); the user terminal 106 is mainly used to provide users with operation windows and display interfaces, etc., and is also used to obtain meteorological data through the business service 103, and process and display meteorological data, etc.

[0057] There may be one or more storage servers 102 , which are mainly used for storing meteorological data, etc.

[0058] There can be one or more service servers 103 , which are mainly used to process user requests and forward meteorological data.

[0059] There can be one or more prediction servers 104, which are mainly used for the prediction of meteorological data, etc. Specifically, they are used to obtain the historical echo data uploaded by the data terminal 101, and after obtaining the historical echo data of the weather radar, obtain the weighted echo features based on the historical echo data based on the hypergraph neural network, obtain the echo trend features based on the weighted echo features and the historical echo data based on the first radar report extrapolation network, and obtain the texture detail features based on the historical echo data based on the second radar report extrapolation network, and finally obtain the meteorological forecast data of the historical echo data based on the echo trend features and the texture detail features. The content of the prediction server will be described in detail below.

[0060] There can be one or more data servers 105, which are mainly used for loading meteorological data, etc.

[0061] In the embodiments of the present application, the various servers may respectively be a cloud server that provides basic cloud computing services such as cloud services, cloud database, cloud computing, cloud functions, cloud storage, network services, cloud communications, domain name services, security services, CDN, as well as big data and artificial intelligence platforms, or a combination of multiple cloud servers.

[0062] In an embodiment of the present application, after obtaining the historical echo data of the weather radar, the prediction server obtains weighted echo features based on the historical echo data based on the hypergraph neural network, obtains echo trend features based on the weighted echo features and the historical echo data based on the first radar return extrapolation network, and obtains texture detail features based on the historical echo data based on the second radar return extrapolation network, and finally obtains meteorological forecast data of the historical echo data based on the echo trend features and the texture detail features. Compared with the current technology, the present application introduces a hypergraph neural network for meteorological forecasting, updates the feature representation of the radar image based on the hypergraph neural network, and in order to fully learn the continuous generation, dissipation, rotation and regeneration of the radar echo image, the first branch focuses on learning the overall motion trend in the radar image, the second branch focuses on describing the texture detail features of the radar echo image, and combines the feature memory module to record the long-term echo law to obtain the final echo extrapolation result (i.e., the extrapolated echo prediction feature mentioned above) as the meteorological forecast data, thereby improving the accuracy of the prediction result.

[0063] It should be noted that Figure 1 The networking diagram of the cloud system shown is only an example. The servers and scenarios described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Ordinary technicians in this field can know that with the evolution of the system and the emergence of new business scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0064] Combined with the above cloud system networking, the radar echo extrapolation method in this application will be introduced below. Figure 2 , Figure 2 A flow chart of a radar echo extrapolation method provided in an embodiment of the present application, such as Figure 2 As shown, the radar echo extrapolation method based on the hypergraph neural network provided in the embodiment of the present application includes the following steps:

[0065] Step 201: Acquire historical echo data of weather radar.

[0066] In the embodiments of the present application, Figure 1 As shown, the user of the user terminal 106 can set the upload path of the data terminal 101 such as the weather radar through the user terminal 106, so that the prediction server obtains the historical echo data of the weather radar.

[0067] Step 202: Based on the hypergraph neural network, obtain weighted echo features according to the historical echo data.

[0068] In the embodiment of the present application, following the previous step 201, the prediction server will perform data extrapolation after obtaining the historical echo data of the weather radar.

[0069] In some embodiments, this step includes: obtaining the hypergraph features of the historical echo data through the hypergraph neural network; obtaining the auxiliary features of the historical echo data; and determining the weighted echo features based on the hypergraph features and the auxiliary features.

[0070] In some embodiments, the weighted echo feature is determined based on the hypergraph feature and the auxiliary feature, including: obtaining a query vector based on the hypergraph feature and the auxiliary feature; obtaining a feature vector in a feature space including the historical echo data; and obtaining the weighted echo feature based on the query vector and the feature vector based on an attention mechanism.

[0071] In some embodiments, the hypergraph features of the historical echo data are obtained through the hypergraph neural network, including: dividing the historical echo data into multiple pixel blocks according to longitude and latitude information; constructing hypergraph data according to feature similarity between the pixel blocks; and processing the hypergraph data to obtain the hypergraph features.

[0072] Step 203: Based on the first radar report extrapolation network, obtain echo trend characteristics according to the weighted echo characteristics and the historical echo data.

[0073] In some embodiments, this step includes: processing the historical echo data to obtain echo prediction data through the first radar report extrapolation network; and obtaining the echo trend feature according to the weighted echo feature and the echo prediction data. The implementation of this step will be described below.

[0074] Step 204: Based on the second radar report extrapolation network, texture detail features are obtained according to the historical echo data.

[0075] The implementation of this step will be described below.

[0076] Step 205: Obtaining extrapolated echo prediction features of the historical echo data according to the echo trend features and the texture detail features.

[0077] In some embodiments, this step includes: performing splicing processing on the echo wave trend feature and the texture detail feature to obtain an extrapolated echo feature; and then obtaining the meteorological forecast data of the historical echo data according to the extrapolated echo prediction feature. The implementation of this step will be described below.

[0078] In some embodiments, after this step, the present application further includes: obtaining the true value of the echo image corresponding to the historical echo data; training the radar echo extrapolation network model according to the extrapolated echo prediction features and the true value of the echo image; the radar echo extrapolation network model includes the hypergraph neural network, the first radar return extrapolation network and the second radar return extrapolation network.

[0079] In the embodiments of the present application, a hypergraph neural network is innovatively introduced for weather forecasting, and the feature representation of the radar map is updated based on the hypergraph neural network. At the same time, in order to fully learn the continuous generation, dissipation, rotation and regeneration of non-rigid changes in the radar echo image, the first branch focuses on learning the overall motion trend in the radar image, and the second branch focuses on depicting the texture detail features of the radar echo image. The feature memory module is combined to record the long-term echo rules to obtain the final echo extrapolation result as the weather forecast data, thereby improving the accuracy of the prediction results.

[0080] In order to implement the content of the above embodiment, the embodiment of the present application provides a radar echo extrapolation device, which can be set in any cloud server or in a server cluster of any combination of these cloud servers. The radar echo extrapolation device is used to execute the radar echo extrapolation method provided in the above method embodiment, please refer to Figure 3 , the device comprises:

[0081] An acquisition module 301 is used to acquire historical echo data of weather radar;

[0082] A first processing module 302 is used to obtain weighted echo features according to the historical echo data based on a hypergraph neural network;

[0083] A second processing module 303 is used to obtain an echo trend feature based on the first radar report extrapolation network and the weighted echo feature and the historical echo data;

[0084] A third processing module 304 is configured to obtain texture detail features according to the historical echo data based on a second radar report extrapolation network;

[0085] The output module 305 is used to obtain the extrapolated echo prediction feature of the historical echo data according to the echo trend feature and the texture detail feature.

[0086] In one embodiment, the first processing module 302 is specifically used to: obtain the hypergraph features of the historical echo data through the hypergraph neural network; obtain the auxiliary features of the historical echo data; and determine the weighted echo features based on the hypergraph features and the auxiliary features.

[0087] In one embodiment, the first processing module 302 is specifically used to: obtain a query vector based on the hypergraph feature and the auxiliary feature; obtain a feature vector in a feature space including the historical echo data; and obtain the weighted echo feature based on the query vector and the feature vector based on an attention mechanism.

[0088] In one embodiment, the first processing module 302 is specifically used to: divide the historical echo data into multiple pixel blocks according to the longitude and latitude information; construct hypergraph data according to the feature similarity between the pixel blocks; and process the hypergraph data to obtain the hypergraph features.

[0089] In one embodiment, the second processing module 303 is specifically used to: process the historical echo data to obtain echo prediction data through the first radar report extrapolation network; and obtain the echo trend feature according to the weighted echo feature and the echo prediction data.

[0090] In one embodiment, the output module 305 is specifically used to: perform splicing processing on the echo wave trend feature and the texture detail feature to obtain an extrapolated echo feature.

[0091] In one embodiment, the aforementioned prediction device also includes a training module, which is used to: obtain the true value of the echo image corresponding to the historical echo data; train the radar echo extrapolation network model according to the extrapolated echo prediction features corresponding to the meteorological forecast data and the true value of the echo image; the radar echo extrapolation network model includes the hypergraph neural network, the first radar report extrapolation network and the second radar report extrapolation network.

[0092] The present application will now be further described in conjunction with specific embodiments.

[0093] This application scenario is explained using short-term precipitation prediction as an example.

[0094] In practical applications, radar echo refers to the various echo images formed when the radar emits an electromagnetic wave signal and the signal encounters a target or obstacle, and its reflected and scattered signals are received by the radar receiver. Among them, the meteorological radar echo is the signal emitted and received by the meteorological radar. The electromagnetic wave backscattering caused by the clouds in the atmosphere, the water vapor in the rainfall, and the condensation of various substances, as well as the drastic changes in some meteorological elements such as temperature, humidity, and air pressure in the atmosphere are the direct factors in the formation of meteorological echoes. Meteorological radar is one of the most basic active atmospheric remote sensing instruments. The radar echo it detects reflects the meteorological conditions and can help forecasters identify and classify weather systems. By using weather radar echo extrapolation technology, the intensity and distribution of echoes in the future period can be predicted based on previous weather radar observations, and the approaching movement and evolution of weather systems can be predicted. Therefore, radar echo extrapolation has become one of the important means of weather forecasting. As an atmospheric remote sensing platform, meteorological radar receives transmission signals backscattered by cloud, rain, snow and other particles, and records the local weather conditions as echo intensity. The predicted echo intensity can intuitively present future weather and serve as a quantitative reference for forecasting meteorological factors, such as precipitation estimation.

[0095] Doppler weather radar echo data has the ability to record dynamic information about the evolution of clouds and rain in the atmosphere. A large amount of weather radar echo data has a wide coverage, which can achieve refined measurements of the airspace and obtain the three-dimensional structure of wind, cloud, and rain targets, thereby further improving the monitoring and early warning capabilities of disastrous weather. In recent years, precipitation prediction through extrapolation of radar echo images has gradually become the mainstay of short-term precipitation prediction systems. Especially with the rapid development of artificial intelligence technology, the combination of neural networks and meteorology has become closer, and a series of extrapolation methods based on deep learning have been proposed. The use of satellite multi-instrument and multi-band data to observe precipitation has become a research hotspot in the current field of atmospheric remote sensing and climate change. By utilizing the spatiotemporal correlation in the echo sequence, it is possible to model complex motion patterns.

[0096] Considering that radar echo is a comprehensive spatiotemporal evolution process, its formation and evolution are affected by many conditions such as temperature, humidity, circulation, thermodynamics and dynamics. The current mainstream motion-based extrapolation method is limited to the research field of dynamics and it is difficult to effectively model the entire echo cycle.

[0097] In response to the above problems, this application proposes a short-term precipitation prediction method based on radar echo extrapolation. The model uses a hypergraph to construct high-order relationships between different radar echo areas, integrates multi-dimensional features, and uses hypergraph convolution to update node features; MotionRNN and CNN are used to jointly construct a dual-branch network, and each branch learns the overall trend of the echo and the specific texture details and completes the extrapolation. Compared with previous methods, the method proposed in this application is superior in terms of information aggregation and information transmission capabilities.

[0098] Specifically, the main contents of this application include:

[0099] Set up historical weather radar echo sequence (i.e. the historical echo data mentioned above) is composed of N radar The echo data at time t can be represented by a tensor To express it, where WI, HE and CH are the spatial width, height and number of channels respectively. In this application, the input channel number CH is 1. The characteristic value of the echo data is the logarithmic physical index (dBZ) of the radar reflection factor, which represents the intensity of the meteorological object.

[0100] like Figure 4 As shown, the goal of radar echo extrapolation is to obtain a given historical echo sequence. Learn to predict future echo sequences of length T And compared with the real echo situation Satisfy the same distribution, that is:

[0101]

[0102] This application uses a double-branch radar echo extrapolation network model based on a hypergraph to implement f model .like Figure 5 As shown in Figure 1, the model (i.e., the radar echo extrapolation network model mentioned above) is composed of a regular evolution module based on a hypergraph neural network and a dual-branch extrapolation module based on MotionRNN. Input echo sequence The radar echo image at the future moment is predicted through the entire network model.

[0103] Among them, the description of the regular evolution module based on the hypergraph neural network is as follows:

[0104] The changes in echoes are real-time and complex, and simple point-to-point relationships are not enough to fully describe this high-order relationship. The hypergraph structure makes full use of the high-order relationships between nodes. Its basic concept is to use the point-hyperedge-point structure to transmit information. To construct a hypergraph for radar echo images, you first need to divide the image into sub-regions (i.e., pixel blocks in the above text) according to the longitude and latitude information; then, combine the KNN algorithm to calculate the feature similarity between pixel blocks, and use the sub-regions divided in the first step as nodes to build connections between pixel blocks with higher similarity; finally, the hyperedge is I radar is the hyperedge index set. Based on the point set Hyperedge ε radar and the hyperedge weight set W radar , the hypergraph can be represented as The features obtained by the hypergraph neural network will be combined with auxiliary features to generate the query vector q, and combined with the attention mechanism to generate weighted memory features.

[0105] Among them, the description of the dual-branch extrapolation module based on MotionRNN is as follows:

[0106] The core idea of ​​radar echo extrapolation is to use the statistical characteristics of historical echo sequences to predict future radar echo images, among which analyzing the detailed spatiotemporal information in radar echo images is the key to accurately predict future echo images. Considering that the movement of radar echoes has both overall evolution trends and texture detail features such as formation, dissipation and regeneration, it is particularly important to study the dual-branch network and model the overall evolution trends and specific texture details in a targeted manner.

[0107] Due to its excellent spatiotemporal modeling capabilities, the Recurrent Neural Network (RNN) is a commonly used module in time series data prediction and analysis tasks, and has been widely recognized in tasks such as Natural Language Processing (NLP). In the radar echo extrapolation task, without the guidance of high-level semantic information, the prediction error will continue to accumulate as the extrapolation network deepens. Moreover, in order to reduce the computational load of the model, it is generally necessary to downsample the input features, which will inevitably cause the loss of high-frequency details in the input information. However, fast and accurate prediction is the core of the task in short-term precipitation tasks, especially in severe convection and other disaster weather. A slight error in prediction is enough to cause immeasurable losses.

[0108] Taking into account the computational efficiency, information redundancy and rational use, this application introduces the MotionRNN model to support the dual-branch radar echo extrapolation working module. MotionRNN is highly flexible and focuses on modeling the internal changes of motion. It learns its internal transient changes based on the spatiotemporal information carried by the echo information itself, and remembers the motion trend of the data itself in a unified way in the previous accumulation. The codec constructed in this way fully takes into account the long-term and short-term information carried by the spatiotemporal sequence. In the experiment, the model will pay more attention to important information and discard redundant information. In the dual-branch extrapolation network, information such as echo motion texture details changes quickly, and feature information updates quickly. It is necessary to use higher temporal resolution to process features and capture precise rules; for information that changes slowly, such as overall trends, use higher spatial resolution to retain details and avoid fuzzy predictions.

[0109] According to the above description, combined with Figure 5 A specific implementation process of the present application is described, but the scope of protection of the present application is not limited to the implementation process.

[0110] Step 1: Hypergraph Neural Network Features u st to obtain.

[0111] u st It stands for spatio-temporal echo features, which means using a hypergraph neural network to obtain the correlation between regions across distances. The starting point for this feature solution is that networks like CNN are generally translation invariant, that is, they naturally assume that the correlation between adjacent regions is stronger. However, in weather forecasting tasks, the correlation between two adjacent regions may not be strong due to changes in the terrain.

[0112] Specifically, it is known that based on hypergraph From its adjacency matrix H G , the hypergraph Laplacian matrix Δ can be expressed as:

[0113]

[0114] Where I is the identity matrix, D v and D e is the degree matrix of vertices and hyperedges, W is the weight matrix, and H is the adjacency matrix. Based on graph convolutional neural networks, in order to reduce computational complexity, this application converts spatial signals into frequency domain signals through Fourier transform, and converts them back to spatial domain through Fourier transform after the operation. That is, for the input signal x=(x1,x2,…,x n ), hypergraph convolution kernel g, then:

[0115] g*x=Φ T ((Φ T g)⊙(Φ T x))=Φ T g(∧)Φ T x,

[0116] In order to further reduce the computational complexity, the above formula is approximated by truncated Chebyshev expansion:

[0117]

[0118] where θ k is a learnable parameter, for ChebyShev expansion, λ max is set to 2, the recursive Chebyshev first-order expansion is further expressed as:

[0119]

[0120] Among them, θ0 and θ1 are node features during model training. For the convenience of calculation, this application uses a single parameter to represent:

[0121]

[0122] Based on the above settings, the hypergraph convolution can be further expressed as:

[0123]

[0124] Where (W+I) is the hypergraph weight matrix. Based on the above information, construct the hypergraph convolution:

[0125]

[0126] Where σ is the activation function, Θ (l-1) is the parameter matrix learned in the (l-1)th layer, X (l-1) is the output of the (l-1)th layer of the graph neural network. Finally, based on the features extracted by the hypergraph neural network (i.e., X (l) The simple processing result after the fully connected layer) is recorded as u st .

[0127] Step 2: Obtain the query vector q.

[0128] In order to fully and effectively utilize historical information, the historical echo sequence will first be reconstructed into a feature space FS. As a complete tensor space, each feature in it is an element in the space. FS will be combined with the attention mechanism to jointly guide the extrapolation of subsequent radar echoes.

[0129] Specifically, in the hypergraph feature u st and auxiliary feature u aux Based on the (mainly including time and weather category features), the query vector q (a component of the attention mechanism used to generate weighted echo features) can be expressed as:

[0130] q = tanh[W st ·fc st (u st )+W aux ·(u aux )+b],

[0131] Here, fc st For fully connected operation, W st and W aux are two modulation matrices respectively, and b is the bias matrix.

[0132] Step 3: Weighted echo feature F echo Generation of.

[0133] After that, q will generate the weighted echo feature F from the feature vector in FS with the help of the attention mechanism. echo(The weighted feature represents the feature-level radar echo information, and the subsequent extrapolation work will continue to be based on this feature):

[0134]

[0135] F echot Indicates F echo The expression at time t, ai is its weighted weight, and its calculation formula is:

[0136] a i =softmax(CS[g(F echoi ),q]),

[0137] Where g(F echoi ) represents the global average pooling operation on channel i, CS[m,q] represents the calculation of the cosine similarity between m and q. If the feature space is recorded as FS, the above two equations can be simplified to:

[0138] F echo =Attention(FS,q),

[0139] Afterwards, F echo The output of the regular evolution module will be used to guide the extrapolation work of subsequent modules.

[0140] Step 4: Feature-level historical echo sequence F dual Generation of.

[0141] Once the historical echo sequence is input into the dual-branch network, it will first be encoded and downsampled to convert the pixel-level historical echo sequence into a feature-level historical echo sequence F dual .

[0142] Step 5: Encoding feature processing.

[0143] Subsequently, the Encoder-Decoder constructed by MotionRNN will begin to extrapolate the radar echo features, starting with the encoding of the features by the Encoder structure:

[0144]

[0145] Here F t l Represents the modeling of motion, Represents overall motion trend modeling. is the input of the lth layer. Represents the potential module of the prediction unit. The last formula is the modeling of Motion Highway. tis the output gate of the prediction unit. After encoding, the dual-branch network decodes the echo features through different networks in turn to obtain two types of extrapolated information: overall echo trend and specific texture details.

[0146] Step 6: First branch decoding.

[0147] For the first branch network (i.e., the first radar return extrapolation network mentioned above), the decoder decoding result can be expressed as:

[0148]

[0149] Here cat[] represents a full connection operation, W mh and W mc Represent two different convolution kernels respectively. It represents the element-by-element multiplication operation between matrices.

[0150] Step 7: Second branch decoding.

[0151] Similarly, for the second branch network (i.e., the second radar return extrapolation network mentioned above), the decoder decoding result can also be expressed as:

[0152]

[0153] Here λ represents the weight coefficient, W xλ and W hλ It also represents two convolution kernels of different sizes.

[0154] Step 8: Conversion of feature-level radar echo sequence to pixel-level radar echo image

[0155] Based on this, and The extrapolated echo features that are jointly constructed (e.g., by splicing, etc.) will be upsampled and decoded to obtain the pixel-level radar echo image at the future i-th time step. (i.e., the extrapolated echo prediction features corresponding to the meteorological forecast data of the historical echo data mentioned above).

[0156] Step 9: Calculation of network loss.

[0157] In order to improve the prediction accuracy, the neural network is dynamically updated. Specifically, based on the prediction results, the experiment sets the L2 loss constraint network training:

[0158]

[0159] in, It represents the pixel-level radar echo image at i time steps, R T+iIt represents the true value of the echo image at that moment (that is, the true value of the echo image corresponding to the historical echo data in the above text).

[0160] Compared with the prior art, this application has the following obvious advantages and beneficial effects:

[0161] 1. A short-term precipitation prediction method based on radar echo image extrapolation is proposed, which has the ability to predict radar images in the future period based on known radar echo images.

[0162] 2. This application introduces the hypergraph neural network for weather forecasting for the first time, and fully explores the advantages of hypergraph in modeling the correlation between traditional meteorological observation data and radar echo images. The hypergraph neural network is introduced into the radar echo extrapolation method to update the feature representation of the radar image. Considering the similar meteorological environment in the same area, the echo image is divided into multiple sub-areas, and a hypergraph is established based on feature similarity.

[0163] 3. In order to fully learn the continuous non-rigid changes in radar echo images such as generation, dissipation, rotation and regeneration, a dual-branch radar echo extrapolation network based on MotionRNN is proposed: the first branch focuses on learning the overall motion trend in the radar image, and the second branch focuses on characterizing the texture detail characteristics of the radar echo image, and combines the feature memory module to record the long-term echo rules to obtain the final echo extrapolation result.

[0164] In summary, this application provides a short-term precipitation prediction method based on radar echo extrapolation based on deep learning technology. First, according to the latitude and longitude information of the echo image, the input image is subdivided into sub-areas containing multiple nodes, and the similarity between pixel blocks across distances is mined to construct a hypergraph structure. The MotionRNN is introduced to construct the Encoder-Decoder structure to extract time dependency and complete the image extrapolation work. The first branch focuses on capturing the overall evolution laws such as the echo morphology distribution, and the second branch focuses on retaining the detailed features such as the echo texture and intensity.

[0165] Correspondingly, an embodiment of the present application also provides a computer device, which includes a server or a terminal, etc.

[0166] like Figure 6 As shown, the computer device may include a processor 601 having one or more processing cores, a memory 602 including one or more computer-readable storage media, an input unit 603, a display unit 604, a wireless fidelity (WiFi) module 605, a power supply 606, and a radio frequency (RF) circuit 607. Those skilled in the art will appreciate that Figure 6The computer device structure shown in the figure does not constitute a limitation on the computer device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently. Among them:

[0167] The processor 601 is the control center of the computer device, which uses various interfaces and lines to connect various parts of the entire mobile phone, and executes various functions of the computer device and processes data by running or executing software programs and / or modules stored in the memory 602, and calling data stored in the memory 602.

[0168] The memory 602 can be used to store software programs and modules, and the processor 601 executes various functional applications and data processing by running the software programs and modules stored in the memory 602. The input unit 603 can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control.

[0169] The display unit 604 may be used to display information input by a user or information provided to a user and various graphical user interfaces of the computer device. These graphical user interfaces may be composed of graphics, texts, icons, videos and any combination thereof.

[0170] WiFi is a short-range wireless transmission technology. Computer devices can help users send and receive emails, browse web pages, and access streaming media through WiFi module 605. It provides users with wireless broadband Internet access. Figure 6 A WiFi module 605 is shown, but it is understandable that it is not a necessary component of the computer device and can be omitted as needed without changing the essence of the application.

[0171] The computer device also includes a power supply 606 (such as a battery) for supplying power to various components. Preferably, the power supply can be logically connected to the processor 601 through a power management system, so that the power management system can manage charging, discharging, power consumption and other functions.

[0172] The radio frequency circuit 607 may be used for receiving and sending signals during information transmission or communication. In particular, after receiving downlink information from the base station, it is processed by one or more processors 601 ; in addition, uplink data is sent to the base station.

[0173] Although not shown, the computer device may also include a Bluetooth module, etc., which will not be described in detail herein. Specifically in this embodiment, the processor 601 in the computer device will load the executable files corresponding to the processes of one or more application programs into the memory 602 according to the following instructions, and the processor 601 will run the application programs stored in the memory 602.

[0174] For example, the computer device is set on the server side to achieve the following functions:

[0175] Get historical echo data of weather radar;

[0176] Based on a hypergraph neural network, obtaining weighted echo features according to the historical echo data;

[0177] Based on the first radar report extrapolation network, obtaining echo trend characteristics according to the weighted echo characteristics and the historical echo data;

[0178] Based on the second radar report extrapolation network, texture detail features are obtained according to the historical echo data;

[0179] The extrapolated echo prediction features of the historical echo data are obtained according to the echo trend features and the texture detail features.

[0180] In the above embodiments, the description of each embodiment has its own emphasis. For the part that is not described in detail in a certain embodiment, please refer to the detailed description above and will not be repeated here.

[0181] A person of ordinary skill in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be completed by instructions, or by controlling related hardware through instructions. The instructions may be stored in a computer-readable storage medium and loaded and executed by a processor.

[0182] To this end, an embodiment of the present application provides a computer-readable storage medium, in which a plurality of instructions are stored, and the instructions can be loaded by a processor and used on the server side to implement the following functions:

[0183] Get historical echo data of weather radar;

[0184] Based on a hypergraph neural network, obtaining weighted echo features according to the historical echo data;

[0185] Based on the first radar report extrapolation network, obtaining echo trend characteristics according to the weighted echo characteristics and the historical echo data;

[0186] Based on the second radar report extrapolation network, texture detail features are obtained according to the historical echo data;

[0187] The extrapolated echo prediction features of the historical echo data are obtained according to the echo trend features and the texture detail features.

[0188] The specific implementation of the above operations can be found in the previous embodiments, which will not be described in detail here.

[0189] The storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0190] Since the instructions stored in the storage medium can execute the steps in any method provided in the embodiments of the present application, the beneficial effects that can be achieved by any method provided in the embodiments of the present application can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0191] At the same time, the embodiment of the present application provides a computer program product or a computer program, which includes a computer instruction, and the computer instruction is stored in a computer-readable storage medium. The processor of the computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device performs the method provided in the above various optional implementations. For example, the following functions can be implemented:

[0192] Get historical echo data of weather radar;

[0193] Based on a hypergraph neural network, obtaining weighted echo features according to the historical echo data;

[0194] Based on the first radar report extrapolation network, obtaining echo trend characteristics according to the weighted echo characteristics and the historical echo data;

[0195] Based on the second radar report extrapolation network, texture detail features are obtained according to the historical echo data;

[0196] The extrapolated echo prediction features of the historical echo data are obtained according to the echo trend features and the texture detail features.

[0197] The above is a detailed introduction to a radar echo extrapolation method, device, computer equipment and storage medium based on a hypergraph neural network provided in an embodiment of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the technical solution and its core idea of ​​the present application, but not to limit it. Although the present invention is described in detail with reference to the aforementioned embodiments, ordinary technicians in this field should understand that it is still possible to modify the technical solutions recorded in the aforementioned embodiments, or to replace some of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solution deviate from the scope of the technical solution of each embodiment of the present application.

Claims

1. A radar echo extrapolation method based on a hypergraph neural network, characterized in that: include: Get historical echo data of weather radar; Based on a hypergraph neural network, obtaining weighted echo features according to the historical echo data; Based on the first radar report extrapolation network, obtaining echo trend characteristics according to the weighted echo characteristics and the historical echo data; Based on the second radar report extrapolation network, texture detail features are obtained according to the historical echo data; The extrapolated echo prediction features of the historical echo data are obtained according to the echo trend features and the texture detail features.

2. The radar echo extrapolation method according to claim 1, characterized in that: Based on the hypergraph neural network, a weighted echo feature is obtained according to the historical echo data, including: Obtaining hypergraph features of the historical echo data through the hypergraph neural network; Acquiring auxiliary features of the historical echo data; The weighted echo feature is determined according to the hypergraph feature and the auxiliary feature.

3. The radar echo extrapolation method according to claim 2, characterized in that: Determining the weighted echo feature according to the hypergraph feature and the auxiliary feature includes: Obtaining a query vector according to the hypergraph feature and the auxiliary feature; Acquire a feature vector in a feature space including the historical echo data; Based on the attention mechanism, the weighted echo feature is obtained according to the query vector and the feature vector.

4. The radar echo extrapolation method according to claim 2, characterized in that: Obtaining the hypergraph features of the historical echo data through the hypergraph neural network includes: Dividing the historical echo data into a plurality of pixel blocks according to the latitude and longitude information; Constructing hypergraph data according to the feature similarity between the pixel blocks; The hypergraph data is processed to obtain the hypergraph features.

5. The radar echo extrapolation method according to claim 1, characterized in that: Based on the first radar report extrapolation network, the echo trend feature is obtained according to the weighted echo feature and the historical echo data, including: Processing the historical echo data to obtain echo prediction data through the first radar report extrapolation network; The echo trend feature is obtained according to the weighted echo feature and the echo prediction data.

6. The radar echo extrapolation method according to claim 1, characterized in that: The extrapolated echo prediction features of the historical echo data are obtained according to the echo trend features and the texture detail features, including: The echo wave trend feature and the texture detail feature are spliced ​​to obtain an extrapolated echo feature.

7. The radar echo extrapolation method according to any one of claims 1 to 6, characterized in that: After obtaining the extrapolated echo prediction feature of the historical echo data according to the echo trend feature and the texture detail feature, the method further includes: Obtaining a true value of an echo image corresponding to the historical echo data; The radar echo extrapolation network model is trained according to the extrapolated echo prediction features and the true value of the echo image; the radar echo extrapolation network model includes the hypergraph neural network, the first radar return extrapolation network and the second radar return extrapolation network.

8. A radar echo extrapolation device based on a hypergraph neural network, characterized in that: include: Acquisition module, used to obtain historical echo data of weather radar; A first processing module, configured to obtain weighted echo features according to the historical echo data based on a hypergraph neural network; A second processing module, configured to obtain an echo trend feature based on the first radar report extrapolation network and the weighted echo feature and the historical echo data; A third processing module is used to obtain texture detail features according to the historical echo data based on the second radar report extrapolation network; The output module is used to obtain the extrapolated echo prediction features of the historical echo data according to the echo trend features and the texture detail features.

9. A computer device, characterized in that: The computer device comprises a processor and a memory, wherein the memory stores a plurality of instructions, and the instructions are suitable for being loaded by the processor to execute the steps in 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 a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the steps in the method according to any one of claims 1 to 7.

Images