A method for predicting meteorological information and related device

Abstract

The embodiment of the application discloses a weather information prediction method and related equipment, which is used for predicting the moving trend and the birth and disappearance trend of a cloud cluster at a future moment according to the distance relationship between the cloud clusters at different moments, and improves the accuracy of weather information prediction. The method comprises the following steps: acquiring a first radar echo map of a target radar at a first moment and a second radar echo map of the target radar at a second moment; determining a first distance and a second distance for a first cloud cluster and a second cloud cluster monitored by the target radar; and predicting a third radar echo map of the target radar at a third moment.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a method for predicting meteorological information and related equipment. Background Technology

[0002] Weather forecasting is an indispensable part of smart cities. Accurate and timely forecasts enable relevant departments to know in advance about changes in various natural phenomena. In the face of extreme weather and geological disasters, early warnings can be issued to the public, thereby avoiding property damage and casualties. In practice, artificial intelligence (AI) technology can be used to forecast weather information, enabling devices to predict future conditions based on existing information.

[0003] In one meteorological forecasting method, by analyzing the spatial and temporal characteristics of radar echo maps, the movement trend of cloud clusters can be predicted, thereby predicting the probability and amount of short-term precipitation.

[0004] In this method of weather forecasting, which relies solely on the movement of cloud clusters to predict the weather, the prediction error is relatively large. Summary of the Invention

[0005] This application provides a method and related equipment for forecasting meteorological information. By analyzing the movement and formation / dissipation trends of cloud clusters, weather conditions are predicted, reducing prediction errors and improving the accuracy of meteorological information forecasting.

[0006] The first aspect of this application provides a method for forecasting meteorological information, including:

[0007] The terminal device acquires the first radar echo image and the second radar echo image detected by the target radar at a first time and a second time, respectively. The first radar echo image includes the characteristics of the cloud cluster detected by the target radar at the first time, and the second radar echo image includes the characteristics of the cloud cluster detected by the target radar at the second time. The second time is later than the first time, and the characteristics of the cloud cluster include information such as the shape and size of the cloud cluster.

[0008] For the first and second radar cloud clusters of the target radar, the terminal equipment can determine the first and second distances between the first and second cloud clusters, corresponding to the first and second time points, respectively, in the first and second radar echo maps. Finally, based on the relationship between the first and second distances, the terminal equipment can predict the third radar echo map observed by the target radar at the third time point. The third time point is later than the second time point.

[0009] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: the terminal device analyzes the movement trend and formation / dissipation trend of cloud clusters by the relationship between the first distance and the second distance, obtains the radar echo map at future times, and thus predicts the weather conditions. The technical solution of this application not only analyzes the movement trend between cloud clusters, but also analyzes the formation / dissipation trend of cloud clusters, reducing prediction errors and improving the accuracy of meteorological information prediction.

[0010] In conjunction with the first aspect, in the first implementation of the first aspect of the embodiments of this application, the terminal device can process the radar echo map to facilitate the determination of the first distance and the second distance.

[0011] First, the terminal device can divide the first radar echo image into X equal-sized first sub-regions, and the second radar echo image into Y equal-sized second sub-regions. Here, X is an integer, and X≥2. Y is an integer, and Y≥2. The number of first and second sub-regions can be the same or different; this is not limited here.

[0012] Then, the terminal device can perform pooling processing on multiple first sub-regions and multiple second sub-regions respectively, and obtain the coefficients corresponding to the first region in the pooled first radar echo image and the coefficients corresponding to the second region in the second radar echo image. The first and second regions have the same size. The first region includes m first sub-regions, and the second region includes n second sub-regions, where m is an integer and 2 ≤ m ≤ X, and n is an integer and 2 ≤ n ≤ Y.

[0013] Finally, the terminal device can determine the first target area and the second target area based on the coefficients corresponding to each area. Thus, a first distance is determined within the first target area, and a second target area is determined within the second target area. The coefficients of both the first and second target areas are greater than or equal to a preset threshold.

[0014] In this embodiment, the terminal device can segment and pool the received radar echo image, focusing on areas with denser cloud clusters by using the coefficients of each region, thus avoiding resource consumption caused by large-scale computation. Simultaneously, by reducing the data scale of the computation, the computation speed is also improved.

[0015] In conjunction with the first aspect and the first implementation of the first aspect, in the second implementation of the first aspect of this application, the terminal device can also analyze the first distance and the second distance in the feature map. The terminal device can extract the spatial features of the first radar echo map and the second radar echo map respectively through a convolutional neural network to obtain the first feature map and the second feature map. The resolution of the first feature map is smaller than that of the first radar echo map, and the resolution of the second feature map is smaller than that of the second radar echo map. Spatial features refer to information related to the spatial attributes of the cloud cluster, including the size, shape, and location of the cloud cluster.

[0016] In this embodiment, the terminal device can first obtain the feature map corresponding to the radar echo map, and then determine the first distance and the second distance from the feature map. Since the feature map has a low resolution, the consumption of the terminal device's computing resources is reduced.

[0017] In combination with the first aspect and any of the first to second implementations of the first aspect, in the third implementation of the first aspect of this application, after determining the first distance and the second distance, the terminal device can also predict the third distance between the first cloud cluster and the second cloud cluster at the third time, and generate the first reshaped feature map and the second reshaped feature map.

[0018] The terminal device can generate a reconstructed feature map in the following manner. The terminal device can look up a first target feature vector from a code table, where the first target feature vector and the feature vector in the first feature map have a first target distance less than or equal to a preset value. The terminal device can also look up a second target feature vector from the code table, where the second target feature vector and the feature vector in the second feature map have a second target distance less than or equal to a preset value. Here, the code table is a set of feature vectors. Then, the terminal device can reconstruct the first feature map based on the first target feature vector to generate a first reconstructed feature map. The terminal device can also reconstruct the second feature map based on the second target feature vector to generate a second reconstructed feature map.

[0019] Then, the terminal device can decode the first feature map and the second feature map to predict the spatial features of the first cloud cluster and the second cloud cluster at the third time. The spatial features of the cloud cluster include information such as the size, position, or shape of the cloud cluster.

[0020] Finally, the terminal device can predict the third radar echo map of the target radar at the third moment based on the spatial characteristics of the first and second cloud clusters at the third moment, as well as the third distance between the first and second cloud clusters.

[0021] In this embodiment, the terminal device can use the feature vectors in the code table to reshape the feature map, allowing the terminal device to focus on learning the main features during decoding without getting bogged down in detailed optimization, thus saving computational resources. Simultaneously, because the fitting error of the reshaped feature map is smaller, the rate of error accumulation is slowed down, improving the long-range prediction capability of the terminal device.

[0022] In conjunction with the third implementation of the first aspect, in the fourth implementation of the first aspect of this application, the terminal device can predict the third distance between the first cloud cluster and the second cloud cluster at the third moment based on the relationship between the first distance and the second distance.

[0023] If the second distance is greater than the first distance, it indicates that both the first and second cloud clusters tend to disperse over time, and the terminal device can predict that the third distance is greater than the second distance. Furthermore, the difference between the third and second distances is correlated with the time interval between the third and second moments.

[0024] If the second distance is less than the first distance, it indicates that the first and second cloud clusters tend to converge over time, and the terminal device can predict that the third distance will be less than the second distance. Furthermore, the difference between the third and second distances is correlated with the time interval between the third and second moments.

[0025] In this embodiment, the terminal device can predict the third distance between the first cloud cluster and the second cloud cluster at the third moment based on the relationship between the first distance and the second distance, thereby determining the generation and dissipation trend of the cloud cluster and making the prediction result more accurate.

[0026] In conjunction with the third or fourth implementation of the first aspect, in the fifth implementation of the first aspect of this application, the terminal device may use a convolutional neural network long-short term memory (ConvLSTM) or a convolutional neural network gate recurrent unit (ConvGRU) to decode the first reshaped feature map and the second reshaped feature map.

[0027] In conjunction with the first aspect and any of the first to fifth implementations of the first aspect, in the sixth implementation of the first aspect of this application, the terminal device acquires the first radar echo map and the second radar echo map by receiving radar data from the target radar at a first moment and radar data at a second moment, and then performing median filtering on the radar data at the two moments respectively to obtain the first radar echo map and the second radar echo map. Here, radar data refers to the data obtained from radar scanning of the target radar.

[0028] In this embodiment of the application, before analyzing the temporal and spatial characteristics of the radar echo map, the terminal device can perform median filtering on the radar echo map to reduce noise interference and improve the accuracy of prediction.

[0029] In conjunction with the first aspect and any of the first to sixth implementations of the first aspect, in the seventh implementation of the first aspect of this application, the third radar echo map predicted by the terminal device includes information such as the location, shape, or size of each cloud cluster. This information can be used to predict the weather conditions of the area corresponding to the third radar echo map, such as the possibility of precipitation or the amount of precipitation.

[0030] A second aspect of this application provides a meteorological information forecasting device, comprising:

[0031] The acquisition unit is used to acquire a first radar echo image of the target radar at a first moment and a second radar echo image at a second moment, the second moment being later than the first moment. The first radar echo image includes the characteristics of the cloud clusters detected by the target radar at the first moment, and the second radar echo image includes the characteristics of the cloud clusters detected by the target radar at the second moment.

[0032] The determination unit is used to determine a first distance and a second distance for the first cloud cluster and the second cloud cluster detected by the target radar at the first time and the second time. The first distance is the distance between the first cloud cluster and the second cloud cluster in the first radar echo map, and the second distance is the distance between the first cloud cluster and the second cloud cluster in the second radar echo map.

[0033] The prediction unit is used to predict the third radar echo map of the target radar at a third moment, which is later than the second moment, based on the relationship between the first and second distances.

[0034] The weather forecasting device is used to perform the method described in the first aspect above.

[0035] The beneficial effects shown in this aspect are similar to those in the first aspect, as detailed in the first aspect, and will not be repeated here.

[0036] A third aspect of this application provides a method for training a feature map orthogonal reshaping model, including:

[0037] The terminal device receives the feature map and searches the code table for the target feature vector that is the nearest neighbor to the feature vector in the feature map. Then, using the target feature vector, the feature map is reconstructed to obtain a reconstructed feature map, which the terminal device then decodes to predict the radar echo map at the third time step.

[0038] In this embodiment, the terminal device can use a trained feature map orthogonal reshaping model to reshape the feature map, allowing the terminal device to focus on learning the main features during decoding without getting bogged down in detailed optimization, thus saving computational resources. Simultaneously, because the fitting error of the reshaped feature map is smaller, the rate of error accumulation is slowed down, improving the long-term prediction capability of the terminal device.

[0039] In conjunction with the third aspect, in the first implementation of the third aspect of this application, the terminal device can optimize the orthogonal reshaping model of the feature map, so that the distance between the feature vector in the code table and the feature vector in the feature map is continuously reduced, thereby reducing the error between the feature map and the reshaped feature map. The loss function used for optimization can be a mean squared error (MSE) function, a mean absolute error (MAE) function, a gradient loss function, a perceptual loss function, an adversarial loss function, or a weighted combination of multiple loss functions, selected according to the actual application requirements; no specific limitation is made here.

[0040] A fourth aspect of this application provides a training apparatus for a feature map orthogonal reshaping model, comprising:

[0041] The receiving unit is used to receive the feature map.

[0042] The processing unit queries the code table for the target feature vector that is the nearest neighbor to the feature vector in the feature map. Then, using the target feature vector, the feature map is reconstructed to obtain a reconstructed feature map, which the terminal device then decodes to predict the radar echo map at the third time step.

[0043] In conjunction with the fourth aspect, in the first implementation of the fourth aspect of this application, the processing unit can further optimize the orthogonal reshaping model of the feature map, so that the distance between the feature vector in the code table and the feature vector in the feature map is continuously reduced, thereby reducing the error between the feature map and the reshaped feature map. The loss function used for optimization can be a mean squared error (MSE) function, a mean absolute error (MAE) function, a gradient loss function, a perceptual loss function, an adversarial loss function, or a weighted combination of multiple loss functions, selected according to the actual application requirements; no specific limitation is made here.

[0044] A fifth aspect of this application provides a meteorological information forecasting device, including a memory and at least one processor. The memory stores program instructions, and when the at least one processor invokes the program instructions, the meteorological information forecasting device executes the methods described in the first or third aspect.

[0045] The beneficial effects shown in this aspect are similar to those in the first or third aspect, as detailed in the first or third aspect, and will not be repeated here.

[0046] A sixth aspect of this application provides a chip including at least one processor and a communication interface, the communication interface and the at least one processor being interconnected via a line, the at least one processor being used to run a computer program or instructions to perform the prediction method described in any of the first aspects to any of the possible implementations of the first aspect, or to perform the training method described in any of the third aspects to any of the possible implementations of the third aspect.

[0047] The communication interface in the chip can be an input / output interface, pins, or circuits.

[0048] In one possible implementation, the chip described above in this application further includes at least one memory storing instructions. This memory can be an internal storage unit of the chip, such as a register or cache, or it can be a storage unit of the chip itself (e.g., read-only memory, random access memory, etc.).

[0049] A seventh aspect of this application provides a computer-readable storage medium storing a program that, when executed by a computer, performs the methods described in the first or third aspect.

[0050] The beneficial effects shown in this aspect are similar to those in the first or third aspect, as detailed in the first or third aspect, and will not be repeated here. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the main framework of the artificial intelligence system in an embodiment of this application;

[0052] Figure 2 This is a schematic diagram of the structure of a meteorological information forecasting system according to an embodiment of this application;

[0053] Figure 3 This is a schematic diagram of a system architecture for a meteorological information forecasting method in the embodiments of this application;

[0054] Figure 4 This is a flowchart illustrating a meteorological information forecasting method in an embodiment of this application.

[0055] Figure 5a This is a schematic diagram of the workflow of the attention-preserving module in an embodiment of this application;

[0056] Figure 5b This is another schematic diagram of the workflow of the attention-preserving module in the embodiments of this application;

[0057] Figure 5c This is another schematic diagram of the workflow of the attention-preserving module in the embodiments of this application;

[0058] Figure 5d This is a schematic diagram of a feature map in an embodiment of this application;

[0059] Figure 6a This is another schematic diagram of the workflow of the attention-preserving module in the embodiments of this application;

[0060] Figure 6b This is a schematic diagram illustrating an application scenario of the meteorological information forecasting method in the embodiments of this application;

[0061] Figure 7 This is another flowchart illustrating the meteorological information forecasting method in this application.

[0062] Figure 8 This is a schematic diagram of the workflow of the feature map orthogonal reshaping module in an embodiment of this application;

[0063] Figure 9 This is another flowchart illustrating the meteorological information forecasting method in this application.

[0064] Figure 10 This is a schematic diagram of the meteorological information forecasting device in an embodiment of this application;

[0065] Figure 11 This is another structural schematic diagram of the meteorological information forecasting device in the embodiments of this application;

[0066] Figure 12This is a schematic diagram of a training device for a feature map orthogonal reshaping model in an embodiment of this application;

[0067] Figure 13 This is a schematic diagram of the structure of a meteorological information forecasting device in an embodiment of this application;

[0068] Figure 14 This is a schematic diagram of the chip structure in an embodiment of this application. Detailed Implementation

[0069] This application provides a method and related equipment for forecasting meteorological information. By analyzing the movement and formation / dissipation trends of cloud clusters, weather conditions are predicted, reducing prediction errors and improving the accuracy of meteorological information forecasting.

[0070] First, the concepts that may be involved in the embodiments of this application will be explained.

[0071] (1) Long short-term memory network (LSTM).

[0072] Long Short-Term Memory (LSTM) networks are a deep learning technique used to process cyclic data, capable of handling and predicting important events with very long intervals and delays in time series. They have wide applications in natural language processing, video understanding, object detection, and deep reinforcement learning.

[0073] (2) Convolutional neural network (CNN).

[0074] Convolutional neural networks (CNNs) are a type of deep learning technique used to learn visual features of images. They are constructed by mimicking the visual and perceptual mechanisms of biological organisms. Their artificial neurons can respond to a portion of the surrounding units within their coverage area, and they perform exceptionally well in large-scale image processing.

[0075] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. Simply put, AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making. Figure 1 A schematic diagram of an artificial intelligence framework is shown, which describes the overall workflow of an artificial intelligence system and is applicable to general artificial intelligence domain needs.

[0076] The above-mentioned artificial intelligence framework will be elaborated from two dimensions: "intelligent information chain" (horizontal axis) and "IT value chain" (vertical axis).

[0077] The "intelligent information chain" reflects a series of processes from data acquisition to processing. For example, it could be a general process of intelligent information perception, intelligent information representation and formation, intelligent reasoning, intelligent decision-making, and intelligent execution and output. In this process, data undergoes a condensation process of "data—information—knowledge—wisdom."

[0078] The "IT value chain" reflects the value that artificial intelligence brings to the information technology industry, from the underlying infrastructure of human intelligence, information (provided and processed by technology) to the industrial ecosystem of systems.

[0079] (1) Infrastructure.

[0080] The infrastructure provides computing power to support artificial intelligence systems, enabling communication with the external world and providing support through a basic platform. Communication with the outside world is achieved through sensors; computing power is provided by intelligent chips, including central processing units (CPUs), neural network processing units (NPUs), graphics processing units (GPUs), application-specific integrated circuits (ASICs), or field-programmable gate arrays (FPGAs) and other hardware acceleration chips. The basic platform includes distributed computing frameworks and related platform guarantees and support, which may include cloud storage and computing, interconnected networks, etc. For example, sensors communicate with the outside world to acquire data, and this data is provided to intelligent chips in the distributed computing system provided by the basic platform for computation.

[0081] (2) Data.

[0082] The data at the next layer of infrastructure is used to represent the data sources in the field of artificial intelligence. The data involves graphics, images, voice, text, and IoT data from traditional devices, including business data from existing systems and sensor data such as force, displacement, liquid level, temperature, and humidity.

[0083] (3) Data processing.

[0084] Data processing typically includes methods such as data training, machine learning, deep learning, search, reasoning, and decision-making.

[0085] Among them, machine learning and deep learning can perform intelligent information modeling, extraction, preprocessing, and training on data, including symbolization and formalization.

[0086] Reasoning refers to the process in which, in a computer or intelligent system, the machine thinks and solves problems by simulating human intelligent reasoning, based on reasoning control strategies and using formalized information. Typical functions include search and matching.

[0087] Decision-making refers to the process of making decisions based on intelligent information after reasoning, and it typically provides functions such as classification, sorting, and prediction.

[0088] (4) General ability.

[0089] After the data processing mentioned above, the results of the data processing can be used to form some general capabilities, such as algorithms or a general system, for example, translation, text analysis, computer vision processing, speech recognition, image recognition, etc.

[0090] (5) Smart products and industry applications.

[0091] Intelligent products and industry applications refer to products and applications of artificial intelligence systems in various fields. They encapsulate overall artificial intelligence solutions, productize intelligent information decision-making, and realize practical applications. Their application areas mainly include: intelligent manufacturing, intelligent transportation, smart home, intelligent healthcare, intelligent security, autonomous driving, safe city, and intelligent terminals.

[0092] The meteorological information forecasting system provided in the embodiments of this application is described below. Please refer to [link / reference]. Figure 2 , Figure 2 This is a system architecture diagram of a meteorological information forecasting system provided in an embodiment of this application. Figure 2 In the embodiment shown, the meteorological information forecasting system includes an execution device 210, a training device 220, a database 230, a terminal device 240, and a data storage system 250. The execution device 210 includes a computing module 211 and an input / output (I / O) interface.

[0093] The data acquisition device 260 can send the acquired data to the database 230. Corresponding to the meteorological information prediction method in this embodiment, the data acquisition device 260 can be a radar, and the acquired data can be a radar echo map. The execution device 210 can acquire the radar echo map from the database 230 or the data storage system 250, and preprocess the radar echo map, analyzing its spatial and temporal characteristics to obtain a feature map. The spatial characteristics of the radar echo map include information such as the size, shape, and location of cloud clusters, while the temporal characteristics include information such as the movement trend of cloud clusters. Then, the execution device 210 can send the feature map to the database 230 via the I / O interface 212, or it can send the feature map to the terminal device 240 via the I / O interface 212, which will then send the feature map to the database 230. The training device 220 can acquire at least two feature maps from the database 230 as training data for the target model / rule 201. The training of the feature map reshaping module is completed when the difference between the radar echo prediction map and the sample radar echo map obtained by the execution device 210 based on the data output by the target model / rule 201 meets the preset conditions. In this embodiment, the target model / rule 201 can be a feature map orthogonal reshaping model.

[0094] The radar echo map predicted by the execution device 210 can be sent to the terminal device 240 for display via the I / O interface 212. Users or staff can learn about the weather conditions, such as the possibility of precipitation or the amount of precipitation, based on the predicted radar echo map, and issue forecasts.

[0095] It should be noted that the training device 220 does not necessarily obtain training data entirely from the database 230; it can also obtain training data from the cloud or other devices for model training. This is not specifically limited here. The execution device 210 can be a terminal, such as a laptop or tablet, or it can be a cloud service or server. This is not specifically limited here.

[0096] It is important to note that Figure 2 This is merely a schematic diagram of a system architecture provided in an embodiment of this application, and does not constitute a limitation on the positional relationships between the devices, components, modules, etc. shown in the diagram. For example, in Figure 2 In the illustrated embodiment, the data storage system 250 is independent of the execution device 210. In some possible embodiments, the data storage system 250 may also be placed within the execution device 210.

[0097] The system architecture of this application embodiment is briefly described below. Please refer to [link / reference]. Figure 3 , Figure 3 This is a system architecture for a meteorological information prediction method in the embodiments of this application.

[0098] Radar monitors weather information in real time and sends radar echoes to the weather forecasting system. In practical applications, because a single radar needs to scan a large area, it cannot scan the entire area in a short time. Therefore, there is a certain time interval between the radar echoes sent to the weather forecasting system. A common interval between two radar echo images is 5 or 6 minutes. After receiving the radar echo images, the weather forecasting system analyzes the distances between cloud clusters in multiple radar echo images and analyzes the spatial and temporal characteristics of the radar echo images to predict the radar echo images for future times. Based on the predicted radar echo images, staff can determine future weather information, issue early forecasts for extreme weather, and reduce casualties and property damage.

[0099] In practical applications, the accuracy of short-term weather forecasts is better than that of long-term forecasts, especially short-term forecasts of 0 to 2 hours, which have a wider range of applications. The weather information forecasting method provided in this application can also be applied to short-term forecasts.

[0100] Next, the meteorological information prediction method provided in the embodiments of this application will be introduced. In the embodiments of this application, the terminal device can analyze the distance relationship between cloud clusters in the radar echo image, and it can also analyze the distance relationship between cloud clusters in the feature image, which will be described separately below.

[0101] I. Analyze the distance relationships between cloud clusters in the feature map of the terminal device.

[0102] Please see Figure 4 , Figure 4 This is a flowchart illustrating a meteorological information forecasting method in an embodiment of this application, including:

[0103] 401. Terminal equipment receives radar data.

[0104] Radar can scan objects within its coverage area using radio waves to obtain radar data. Terminal devices receive a large amount of radar data. This embodiment uses radar data detected by the same radar at two different times as an example for illustration. Both the radar data at the first time and the radar data at the second time show a first cloud cluster and a second cloud cluster, with the second time being later than the first time.

[0105] 402. Terminal equipment acquires radar echo map.

[0106] After receiving the radar data at the first and second moments, the terminal device processes this data to obtain a first radar echo map and a second radar echo map, making them more closely reflect the needs of actual applications. The processing methods may include the following:

[0107] Optionally, in practical applications, radar may also scan for buildings, trees, etc. on the ground, and the resulting radar echo images often contain noise interference, affecting the accuracy of predictions. Under normal circumstances, cloud-related data changes relatively smoothly, while other data changes abruptly compared to cloud data. Terminal devices can use median filtering to filter out abrupt pixel changes in the radar echo image, thereby reducing interference from noise data.

[0108] Optionally, the terminal device can apply a mask of varying size to the filtered data to simulate radar data loss. This is because in practical applications, each radar covers a wide area, and completing a full scan cycle takes a long time. To ensure timely prediction, the terminal device uses a portion of the received radar data for prediction. For example, predicting weather information for a large area might require 12 radars working together. However, in practical applications, only radar data from 8 radars might be received within a preset time period. In this case, the terminal device will use the radar echo maps from these 8 radars for prediction. In this situation, to make the experimental environment of this embodiment as close to the real environment as possible and improve the reliability of the technical solution in practical applications, even if radar echo maps from 12 radars are received within a preset time period, the filtered data will be masked with a mask of varying size to simulate radar data loss, making the experimental data closer to reality.

[0109] It should be noted that, in the embodiments of this application, the simulated data can also be filtered after the simulated data is lost.

[0110] 403. The terminal device acquires the feature map.

[0111] After the terminal device preprocesses the radar echo image, it can input the preprocessed radar echo image into a convolutional neural network to extract the spatial features of the radar echo image, thus obtaining a feature map. The first radar echo image corresponds to the first feature map, and the second radar echo image corresponds to the second feature map. The spatial features include information such as the shape, size, or location of the cloud cluster.

[0112] 404. The terminal device determines the distance relationship between cloud clusters in the feature map.

[0113] When determining the distance relationship between cloud clusters, the terminal device performs the same processing on the feature map. This application embodiment takes the processing of the first feature map as an example to illustrate the process by which the terminal device determines the distance relationship between cloud clusters in the feature map.

[0114] Because radar echo maps cover a wide area, the actual areas with echoes are relatively sparse. This means that not all areas in the feature map contain cloud information. Therefore, in this embodiment, a volume-preserving attention (VPA) module is used to focus on processing the areas with radar echoes. For example, the VPA module's workflow can be as follows: Figure 5a As shown.

[0115] First, the terminal device can divide the feature map into several sub-regions of the same size and rearrange these sub-regions according to the depth direction. It should be noted that this embodiment does not limit the number of sub-regions after dividing different feature maps, as long as the size of each sub-region is the same within the same feature map. For example, the terminal device can divide the first radar echo map into 16 first sub-regions, each with a size of 5px × 5px. Then, the second radar echo map can be divided into 4 second sub-regions, each with a size of 10px × 10px. Here, px is an abbreviation for pixel.

[0116] The purpose of the terminal device in organizing the feature map along the depth direction is to structurally reconstruct the arrangement of image data, enabling smooth data processing. The size of the sub-region is represented by the pixel block size of the pixel matrix. The size of the pixel block is determined by the management granularity of the weather forecasting system. A finer management granularity, such as 1px × 1px per pixel block, yields more accurate results. A coarser management granularity, such as 20px × 20px per pixel block, saves computational resources. The management granularity of the weather forecasting system is selected based on the needs of the actual application; no specific limitation is made here.

[0117] Then, the VPA module can simulate the visual characteristics of the human brain to quickly capture the "where" and "what" information in the radar echo map and describe the distance relationship between cloud clusters.

[0118] Here, "where" refers to the area with radar echoes. The VPA module can adaptively focus on areas with radar echoes, thus avoiding large-scale computations. In the spatial adaptive selection 501 stage, the VPA module can use a dual-pooling fusion method to determine whether a certain area contains valid radar echo information. Pooling treats several sub-regions as a single region. Different pooling methods have different beneficial effects. Average pooling results in the average value of radar echoes from several sub-regions, helping to determine the overall echo intensity level of a certain area. Max pooling results in the maximum value among radar echoes from several sub-regions, helping to determine whether a certain area contains small but high-intensity echoes.

[0119] It should be noted that although in practical applications, the size of the sub-regions in different feature maps obtained by segmenting different feature maps by the terminal device may be different, the size of the region processed by the terminal device during pooling is the same, which facilitates the terminal device to perform calculations uniformly.

[0120] For example, the process of splitting and pooling can be simply understood as Figure 5b or Figure 5c As shown in the process, the terminal device can divide the first feature map into 16 first sub-regions, each with a size of 5px × 5px. Then, it divides the second feature map into 4 second sub-regions, each with a size of 10px × 10px. During pooling, the terminal device can perform pooling on the first regions corresponding to the 8 first sub-regions and pooling on the second regions corresponding to the 2 second sub-regions, ensuring that the sizes of the first and second regions are the same, and that the positions of the first and second regions in the first and second feature maps are the same.

[0121] The benefit of pooling is that it combines multiple sub-regions for processing, reducing computational resources. After pooling, the coefficients for each region can be output through a fully connected network. Typically, the coefficients range from 0 to 1. The larger the coefficient, the more attention should be paid to that region during computation, providing a basis for subsequent allocation of computational resources.

[0122] For example, such as Figure 5d As shown, the pooled feature map can include nine regions, with y1 to y9 representing the coefficients corresponding to each region. In practical applications, the preset threshold can be manually set according to the management granularity of the meteorological information prediction system. The finer the management granularity, the smaller the preset threshold, and the more accurate the result. Conversely, the coarser the management granularity, the larger the preset threshold, which saves computational resources.

[0123] “what” corresponds to Figure 5aThe regional relation description 502 in the model is performed on the target region focused during the adaptive selection phase to characterize the local fine structure of the cloud clusters, including the distance relationship between the first and second cloud clusters. The coefficient of the target region is greater than or equal to a preset threshold. The VPA module uses a spatial self-attention mechanism to model regional correlation for the target region, obtaining a self-attention feature map. Regional correlation refers to the distance relationship between cloud clusters.

[0124] For example, the self-attention mechanism in the VPA module can be as follows: Figure 6a As shown. According to Figure 6a As can be seen, the VPA module stores convolution feature maps. Different feature maps are convolved using different functions and different convolution kernels, and then transposed to obtain attention maps. After that, a softmax activation model is used to obtain self-attention feature maps.

[0125] Finally, the terminal device needs to arrange the sub-regions with coefficients less than a preset threshold and the self-attention feature maps in the order of the first feature map, and then reassemble them so that the stitched image maintains the correspondence with the first feature map. It is important to note that segmentation and reassembly do not change the size of the first feature map.

[0126] For example, such as Figure 6b As shown, after the terminal device performs the above processing on the received radar echo map 601, it can obtain the self-attention feature map 602.

[0127] 405. Temporal characteristics of feature maps acquired by terminal devices.

[0128] After the terminal device determines the first distance between the first cloud cluster and the second cloud cluster in the first radar echo image, as well as the second distance between the first cloud cluster and the second cloud cluster in the first radar echo image, it can encode the re-stitched first feature map and the second feature map. Based on ConvLSTM, ConvGRU or other more advanced spatiotemporal feature extraction modules, the temporal features of the feature map are determined. The temporal features include the trend of the first distance and the second distance changing over time, as well as the trend of each cloud cluster moving over time.

[0129] 406. Predictive radar echo map of terminal equipment.

[0130] Terminal devices can use ConvLSTM, ConvGRU, or other more advanced timing generation modules to decode the encoded feature map and predict the third radar echo map at the third time step. In the third echo map, the distance between the first and second cloud clusters is the third distance. The third time step is later than the second time step, and the interval between the third and second time steps is N times the interval between the first and second time steps, where N is an integer greater than or equal to 1.

[0131] If the first distance is greater than the second distance, it indicates that the first and second cloud clusters are closing in on each other over time and are likely to coalesce into a larger cloud cluster in the future. The terminal device can determine that the third distance is less than the second distance, and the difference between the third and second distances is correlated with the time interval between the third and second moments. For example, if the first distance is 500 meters and the distance at the second moment is 400 meters, assuming the cloud cluster is moving at a constant speed, and the interval between the third and second moments is the same as the interval between the second and first moments, then the terminal device can predict that the third distance is 300 meters.

[0132] If the first distance is less than the second distance, it indicates that the distance between the first and second cloud clusters is increasing over time and is likely to disperse in the future. The terminal device can determine that the third distance is greater than the second distance, and the difference between the third and second distances is correlated with the time interval between the third and second moments. For example, if the first distance is 300 meters and the distance at the second moment is 500 meters, assuming the cloud clusters are moving at a constant speed, and the interval between the third and second moments is the same as the interval between the second and first moments, then the terminal device can predict that the third distance is 700 meters.

[0133] In this embodiment, the terminal device analyzes the movement and formation / dissipation trends of cloud clusters based on the relationship between a first distance and a second distance, obtaining a radar echo map for future times, thereby predicting weather conditions. This technical solution not only analyzes the movement trends of cloud clusters but also their formation / dissipation trends, reducing prediction errors and improving the accuracy of meteorological information forecasting.

[0134] Furthermore, the terminal device performs calculations on the radar echo information in the feature map, which saves computing resources compared to calculating the massive amount of raw radar echo data. Simultaneously, the terminal device allocates computing resources primarily to areas with denser cloud clusters, further reducing the data size required for processing and improving processing speed.

[0135] In this embodiment, after extracting the temporal sequence of the feature maps, the terminal device can further reshape the feature maps based on an orthogonal code table, allowing the spatiotemporal encoder to focus on learning the main features without getting bogged down in detailed optimization. The process of reshaping the feature maps is described below; please refer to [link to relevant documentation]. Figure 7 , Figure 7 This is a flowchart illustrating a meteorological information forecasting method in an embodiment of this application, including:

[0136] 701. Terminal equipment receives radar data.

[0137] 702. Terminal equipment acquires radar echo map.

[0138] 703. Terminal device acquires feature map.

[0139] 704. The terminal device determines the distance relationship between cloud clusters in the feature map.

[0140] 705. Temporal characteristics of feature maps acquired by terminal devices.

[0141] In this embodiment, steps 701 to 705 are... Figure 4 Steps 401 to 405 in the illustrated embodiment are similar and will not be repeated here.

[0142] 706. The terminal device acquires the reshaped feature map.

[0143] Terminal devices can input the encoded feature map into a feature map orthogonal reshaping model to obtain a reshaped feature map. The reshaping process involves querying the code table for the closest target feature vector to each feature vector of the feature map, and then using the target feature vector to re-represent the feature map, resulting in the reshaped feature map. This process can be used... Figure 8 The illustrated embodiment is used for illustration. It should be noted that reshaping the feature map does not change its size.

[0144] The initial values ​​of the orthogonal reshaping model for the feature map are feature vectors of a certain data volume, which can be randomly assigned manually. Through training, the distance between the feature vectors of the input feature map and the nearest neighbor feature vectors in the code table is continuously reduced, so that the final reshaped feature map is as similar as possible to the input feature map. When optimizing the orthogonal reshaping model, orthogonal constraints can be imposed on the feature vectors in the code table as a regularization term introduced into the objective function. Applying orthogonal constraints can prevent feature vector redundancy in the code table, ensuring that each feature vector in the code table represents information of different dimensions. The code table of the trained orthogonal reshaping model includes a series of feature vectors, each with a specific physical meaning, such as line orientation, echo intensity, or region connectivity. The value and meaning of each feature vector are automatically assigned by the data-driven learning process.

[0145] 707. Predictive radar echo map of terminal equipment.

[0146] After acquiring the reconstructed feature map, the terminal device can use ConvLSTM, ConvGRU, or other more advanced time-series generation modules to decode the encoded feature map and predict the third radar echo map at the third time step. In the third echo map, the distance between the first and second cloud clusters is the third distance. The third time step is later than the second time step, and the interval between the third and second time steps is N times the interval between the first and second time steps, where N is an integer greater than or equal to 1.

[0147] If the first distance is greater than the second distance, it means that the first and second cloud clusters are getting closer over time and are likely to merge into a large cloud cluster in the future. The terminal device can determine that the third distance is less than the second distance.

[0148] If the first distance is less than the second distance, it means that the distance between the first cloud cluster and the second cloud cluster has increased over time and is likely to disperse in the future. The terminal device can determine that the third distance is greater than the second distance.

[0149] After predicting the radar echo map, the terminal device can calculate the error between the third radar echo map and the actual radar echo map at the third time point, and use different loss functions to optimize the meteorological information prediction model. The loss function used for optimization can be the MSE function, MAE function, gradient loss function, perceptual loss function, adversarial loss function, or a weighted combination of multiple loss functions, selected according to the actual application needs, and is not limited here.

[0150] In this embodiment, the terminal device analyzes the movement and formation / dissipation trends of cloud clusters based on the relationship between a first distance and a second distance, obtaining a radar echo map for future times, thereby predicting weather conditions. This technical solution not only analyzes the movement trends of cloud clusters but also their formation / dissipation trends, reducing prediction errors and improving the accuracy of meteorological information forecasting.

[0151] Furthermore, reconstructing the feature map involves extracting the most basic unit of the image space. Compared to an unreconstructed feature map, it can better approximate the regression value through the decoder, allowing the spatiotemporal encoder to focus on learning key features without getting bogged down in detailed optimization. Simultaneously, because the fitting error to the target value is small, it also slows down the rate of error accumulation, improving the long-term forecasting capability of the weather forecasting system. Here, the most basic unit of the image space refers to the physical meaning contained in the radar image, including the actual size, edge orientation, or shape of cloud clusters.

[0152] II. Terminal equipment analyzes the distance relationships between cloud clusters in radar echo images.

[0153] When the radar echo image has a low resolution, contains a small amount of data, or the terminal device has sufficient computing resources, the terminal device can also directly analyze the distance relationships between cloud clusters in the radar echo image. This situation is explained below; please refer to [link / reference]. Figure 9 , Figure 9 This is a flowchart illustrating a meteorological information forecasting method in an embodiment of this application, including:

[0154] 901. Terminal equipment receives radar data.

[0155] 902. Terminal equipment acquires radar echo map.

[0156] In this embodiment, steps 901 and 902 are... Figure 4 Steps 401 and 402 in the illustrated embodiment are similar and will not be described again here.

[0157] 903. The terminal equipment determines the distance relationship between cloud clusters in the radar echo image.

[0158] In this embodiment, step 903 and Figure 4 Step 404 in the illustrated embodiment is similar and will not be repeated here. The difference lies in... Figure 4 In the illustrated embodiment, step 404 processes the feature map, while in this embodiment, the preprocessed radar echo map is analyzed.

[0159] 904. The terminal equipment acquires the spatiotemporal characteristics of the radar echo map.

[0160] After the terminal device determines the first distance between the first cloud cluster and the second cloud cluster in the first radar echo image, and the second distance between the first cloud cluster and the second cloud cluster in the second radar echo image, it can encode the re-stitched first and second radar echo images. Based on ConvLSTM, ConvGRU, or other more advanced spatiotemporal feature extraction modules, it can analyze the spatial features of the first and second cloud clusters, determine the trends of the first and second distances changing over time, and the trends of the movement of each cloud cluster over time, and obtain the first feature map and the second feature map.

[0161] 905. The terminal device acquires the reshaped feature map.

[0162] In this embodiment, step 905 and Figure 7 Step 706 in the illustrated embodiment is similar and will not be repeated here.

[0163] It's important to note that step 905 can be executed first, followed by step 904. In this case, step 905 reshapes not the feature map, but rather the preprocessed radar echo map. After reshaping the radar echo map, its temporal and spatial characteristics are analyzed to obtain the reshaped feature map. However, in practice, due to the large amount of data contained in radar echo maps, reshaping them consumes excessive resources. To conserve computational resources, this method is rarely used in practical applications to obtain the reshaped feature map.

[0164] 906. Predictive radar echo map of terminal equipment.

[0165] In this embodiment, step 905 and Figure 7 Step 707 in the illustrated embodiment is similar and will not be repeated here.

[0166] In this embodiment, the terminal device analyzes the movement and formation / dissipation trends of cloud clusters based on the relationship between a first distance and a second distance, obtaining a radar echo map for future times, thereby predicting weather conditions. This technical solution not only analyzes the movement trends of cloud clusters but also their formation / dissipation trends, reducing prediction errors and improving the accuracy of meteorological information forecasting.

[0167] Furthermore, the terminal device can first determine the distance relationship between cloud clusters in the radar echo image, and then analyze the spatial characteristics of the radar echo image. At the same time, the terminal device can analyze the temporal and spatial characteristics of the radar echo image simultaneously, or it can analyze the spatial characteristics first and then the temporal characteristics, which improves the flexibility of the technical solution of this application.

[0168] The meteorological information forecasting device provided in the embodiments of this application will be described below. Please refer to... Figure 10 , Figure 10 One embodiment of the meteorological information forecasting device 1000 in this application includes:

[0169] The acquisition unit 1001 is used to acquire a first radar echo image of the target radar at a first moment and a second radar echo image at a second moment, the second moment being later than the first moment. The first radar echo image includes the characteristics of the cloud clusters detected by the target radar at the first moment, and the second radar echo image includes the characteristics of the cloud clusters detected by the target radar at the second moment.

[0170] The determining unit 1003 is used to determine a first distance and a second distance for the first cloud cluster and the second cloud cluster detected by the target radar at a first time and a second time. The first distance is the distance between the first cloud cluster and the second cloud cluster in the first radar echo map, and the second distance is the distance between the first cloud cluster and the second cloud cluster in the second radar echo map.

[0171] The prediction unit 1005 is used to predict the third radar echo map of the target radar at a third time point based on the relationship between the first distance and the second distance point, where the third time point is later than the second time point.

[0172] In some optional embodiments, the weather information forecasting device 1000 further includes a processing unit 1002, for:

[0173] Segment the first radar echo map to obtain X first sub-regions. Each of the X first sub-regions has the same size, where X is an integer and X≥2.

[0174] Segment the second radar echo map to obtain Y second sub-regions. Each of the Y second sub-regions has the same size. X is an integer and Y≥2.

[0175] Pool X first sub-regions and obtain the coefficients corresponding to the pooled first region. The pooled first region includes m first sub-regions, where m is an integer and 2≤m≤X.

[0176] Pool Y second sub-regions and obtain the coefficients corresponding to the pooled second regions. The pooled second regions include n second sub-regions, where n is an integer and 2≤n≤Y. The second regions have the same size as the first regions.

[0177] Unit 1003 is specifically used for:

[0178] In the first target region, a first distance is determined, the first target region is contained within the first region, and the coefficient corresponding to the first target region is greater than or equal to a preset threshold.

[0179] In the second target region, a second distance is determined, the second target region is contained within the second region, and the coefficient corresponding to the second target region is greater than or equal to a preset threshold.

[0180] In some optional embodiments, the first radar echo map includes a first feature map, the second radar echo map includes a second feature map, the resolution of the first feature map is less than the resolution of the first radar echo map, the distance between the first cloud cluster and the second cloud cluster in the first feature map is a first distance, the resolution of the second feature map is less than the resolution of the second radar echo map, and the distance between the first cloud cluster and the second cloud cluster in the second feature map is a second distance.

[0181] In some optional embodiments, the weather information forecasting device 1000 further includes a generation unit 1004.

[0182] Generation unit 1004 is specifically used for:

[0183] The code table searches for the first target feature vector whose first target distance is less than or equal to a preset value. The first target distance is the distance between the feature vector in the first feature map and the first target feature vector. The code table is a set of feature vectors.

[0184] The code table is used to find the second target feature vector whose distance to the second target is less than or equal to a preset value. The second target distance is the distance between the feature vector in the second feature map and the second target feature vector.

[0185] A first reshaped feature map is generated based on the first target feature vector.

[0186] A second reshaped feature map is generated based on the second target feature vector.

[0187] The prediction unit 1005 is also used to predict the third distance between the first cloud cluster and the second cloud cluster at the third time based on the first distance and the second distance.

[0188] Decode the first and second reshaped feature maps to predict the spatial features of the first and second cloud clusters at the third time point. The spatial features include the size, location, or shape of the cloud clusters.

[0189] Based on the spatial characteristics of the first cloud cluster, the spatial characteristics of the second cloud cluster, and the third distance, predict the third radar echo map of the target radar at the third moment.

[0190] In some optional embodiments, the prediction unit 1005 is specifically used for:

[0191] If the second distance is greater than the first distance, then the predicted third distance is greater than the second distance, and the difference between the third distance and the second distance is related to the time interval between the third time and the second time.

[0192] If the second distance is less than the first distance, then the predicted third distance is less than the second distance, and the difference between the third distance and the second distance is related to the time interval between the third time and the second time.

[0193] In some optional embodiments, the prediction unit 1005 is specifically used to decode the first reshaped feature map and the second reshaped feature map using ConvLSTM or ConvRGU.

[0194] In some optional embodiments, the acquisition unit 1001 is specifically used for:

[0195] Receive radar data from the target radar at the first moment and radar data from the target radar at the second moment.

[0196] The radar data at the first moment is subjected to median filtering to obtain the first radar echo map.

[0197] The radar data at the second time point is subjected to median filtering to obtain the second radar echo map.

[0198] In some alternative embodiments, the third radar echo map includes the location, shape, or size of individual cloud clusters, used to predict meteorological information for the region corresponding to the third radar echo map, such as the likelihood of precipitation and the amount of precipitation.

[0199] In this embodiment of the application, the meteorological information forecasting device 1000 can perform the aforementioned... Figure 2 The meteorological information forecasting system shown in the embodiment is... Figure 3 The weather forecasting system shown in the embodiment is... Figure 4 The terminal device in the illustrated embodiment, Figure 7 The terminal device in the illustrated embodiment, or Figure 9 The specific operations performed by the terminal device in the illustrated embodiment will not be described in detail here.

[0200] Please see Figure 11 , Figure 11 This is an embodiment of the meteorological information forecasting device 1100 in this application, comprising:

[0201] Receiver module 1101 is used to receive radar echo maps sent by radar equipment and can perform... Figure 10 The operations performed by the acquisition unit 1001 in the illustrated embodiment will not be described in detail here.

[0202] The spatial feature encoding module 1102 is used to perform convolution processing on the preprocessed radar echo image, analyze the spatial features in the radar echo image, and obtain a feature map.

[0203] The cloud cluster relationship modeling module 1103 is used to process radar echo maps or feature maps to determine the distance relationships between various cloud clusters. It can perform... Figure 10 The specific operations performed by the processing unit 1002 or the determining unit 1003 in the illustrated embodiment will not be described in detail here.

[0204] The spatiotemporal feature coding module 1104 is used to analyze the temporal features in radar echo maps or feature maps. It should be noted that the spatiotemporal feature coding module 1104 can be coupled with the spatial feature coding module 1102 into a single module, which is usually referred to as the spatiotemporal feature coding module.

[0205] The feature map orthogonal reshaping module 1105 is used to reshape the feature map and can perform... Figure 10 The specific operations performed by the generation unit 1004 in the illustrated embodiment will not be described in detail here.

[0206] The spatiotemporal feature decoding module 1106 is used to decode the first and second feature maps, combine them with the third distance, and predict the radar echo map at the third time point. It can then execute... Figure 10 The specific operations performed by the prediction unit 1005 in the illustrated embodiment will not be described here.

[0207] In this embodiment of the application, the meteorological information forecasting device 1100 can perform the aforementioned... Figure 2 The meteorological information forecasting system shown in the embodiment is... Figure 3 The weather forecasting system shown in the embodiment is... Figure 4 The terminal device in the illustrated embodiment, Figure 7 The terminal device in the illustrated embodiment, or Figure 9 The specific operations performed by the terminal device in the illustrated embodiment will not be described in detail here.

[0208] Please see Figure 12 , Figure 12 This is an embodiment of the training device 1200 for the feature map orthogonal reshaping model in this application, comprising:

[0209] The receiving unit 1201 is used to receive the feature map.

[0210] The processing unit 1202 is used to query the target feature vector that is the nearest neighbor of the feature vector in the feature map from the code table. Then, the feature map is reconstructed using the target feature vector to obtain a reconstructed feature map, which is then decoded by the terminal device to predict the radar echo map at the third time moment.

[0211] In some optional embodiments, the processing unit 1202 can also be used to optimize the orthogonal reshaping model of the feature map, so that the distance between the feature vectors in the code table and the feature vectors in the feature map is continuously reduced, thereby reducing the error between the feature map and the reshaped feature map. The loss function used for optimization can be the MSE function, MAE function, gradient loss function, perceptual loss function, adversarial loss function, or a weighted combination of multiple loss functions, selected according to the actual application needs; no specific limitation is made here.

[0212] Please see Figure 13 , Figure 13 This is an embodiment of the meteorological information forecasting device 1300 in this application. The meteorological information forecasting device 1300 may include one or more central processing units (CPUs) 1301 and a memory 1302, in which one or more applications or data are stored.

[0213] The memory 1302 can be volatile or persistent storage. The program stored in the memory 1302 can include one or more modules, each of which can be used to execute a series of operations performed by the meteorological information forecasting device 1300. Furthermore, the processor 1301 can be configured to communicate with the memory 1302 and execute a series of instruction operations stored in the memory 1302 on the meteorological information forecasting device 1300.

[0214] The weather forecasting device 1300 may also include one or more power supplies 1305, one or more wired or wireless interfaces 1303, one or more input / output interfaces 1304, and / or one or more operating systems, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0215] Central processing unit 1301 can execute the aforementioned Figure 2 The meteorological information forecasting system shown in the embodiment is... Figure 3 The weather forecasting system shown in the embodiment is... Figure 4 The terminal device in the illustrated embodiment, Figure 7 The terminal device in the illustrated embodiment, or Figure 9 The specific operations performed by the terminal device in the illustrated embodiment will not be described in detail here.

[0216] Please see Figure 14 , Figure 14 This is one embodiment of the chip in this application. The chip can be represented as a neural network processor 1400, which is mounted as a coprocessor on the main CPU and its tasks are assigned by the host CPU. The core of the neural network processor 1400 is the arithmetic circuit 1403. The controller 1404 can control the arithmetic circuit 1403 to extract data from the weight memory 1402 or the input memory 1401 and perform calculations.

[0217] In some implementations, the arithmetic circuit 1403 internally includes multiple process engines (PEs). In some implementations, the arithmetic circuit 1403 can be a two-dimensional pulsating array. The arithmetic circuit 1403 can also be a one-dimensional pulsating array or other electronic circuitry capable of performing mathematical operations such as multiplication and addition. In some implementations, the arithmetic circuit 1403 is a general-purpose matrix processor.

[0218] For example, suppose we have an input matrix A, a weight matrix B, and an output matrix C. The arithmetic circuit retrieves the corresponding data of matrix B from the weight memory 1402 and caches it in each PE of the arithmetic circuit. The arithmetic circuit retrieves the data of matrix A from the input memory 1401 and performs matrix operations with matrix B. The partial result or the final result of the obtained matrix is ​​stored in the accumulator 1408.

[0219] Unified memory 1406 is used to store input and output data. Weight data is directly transferred to weight memory 1402 via direct memory access controller (DMAC) 1405. Input data is also transferred to unified memory 1406 via DMAC.

[0220] The bus interface unit 1410 (BIU) is used to enable interaction between the main CPU, DMAC, and instruction fetch buffer 1409 (IFB) via a bus. The instruction fetch buffer 1409 is used to store instructions used by the controller 1404.

[0221] The vector computation unit 1407 includes multiple arithmetic processing units that further process the output of the computation circuit as needed, such as vector multiplication, vector addition, exponential operations, logarithmic operations, size comparisons, etc. It is mainly used for computations in non-convolutional / fully connected layers of neural networks, such as pixel-level summation and upsampling of feature planes.

[0222] in, Figures 4 to 9 The operations of each layer in the neural network shown in the corresponding embodiment can be performed by the operation circuit 1403 or the vector calculation unit 1407.

[0223] The processor mentioned above can be a central processing unit, a microprocessor, or one or more integrated circuits used to control the execution of a program in the first aspect of the method.

[0224] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0225] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0226] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0227] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0228] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0229] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for forecasting meteorological information, characterized in that, include: Acquire a first radar echo image of the target radar at a first moment and a second radar echo image at a second moment, the second moment being later than the first moment. The first radar echo image includes the characteristics of the cloud clusters detected by the target radar at the first moment, and the second radar echo image includes the characteristics of the cloud clusters detected by the target radar at the second moment. For the first cloud cluster and the second cloud cluster detected by the target radar at the first time and the second time, a first distance and a second distance are determined. The first distance is the distance between the first cloud cluster and the second cloud cluster in the first radar echo map, and the second distance is the distance between the first cloud cluster and the second cloud cluster in the second radar echo map. Based on the relationship between the first distance and the second distance, a third radar echo map of the target radar is predicted at a third time point, which is later than the second time point.

2. The method according to claim 1, characterized in that, Before determining the first distance and the second distance, the method further includes: The first radar echo image is segmented to obtain X first sub-regions, each of the X first sub-regions having the same size, where X is an integer and X≥2; The second radar echo map is segmented to obtain Y second sub-regions, each of the Y second sub-regions having the same size, where Y is an integer and Y≥2; Pool the X first sub-regions to obtain the coefficients corresponding to the pooled first region. The pooled first region includes m first sub-regions, where m is an integer and 2≤m≤X. Pool the Y second sub-regions to obtain the coefficients corresponding to the pooled second region. The pooled second region includes n second sub-regions, where n is an integer and 2≤n≤Y. The second region has the same size as the first region. Determining the first distance and the second distance includes: In the first target region, the first distance is determined, the first target region is contained within the first region, and the coefficient corresponding to the first target region is greater than or equal to a preset threshold. In the second target region, the second distance is determined, the second target region is contained within the second region, and the coefficient corresponding to the second target region is greater than or equal to a preset threshold.

3. The method according to claim 1 or 2, characterized in that, The first radar echo image includes a first feature image, and the second radar echo image includes a second feature image. The resolution of the first feature image is less than the resolution of the first radar echo image. The distance between the first cloud cluster and the second cloud cluster in the first feature image is the first distance. The resolution of the second feature image is less than the resolution of the second radar echo image. The distance between the first cloud cluster and the second cloud cluster in the second feature image is the second distance.

4. The method according to claim 3, characterized in that, After determining the first distance and the second distance, the method further includes: The code table is used to find the first target feature vector whose first target distance is less than or equal to a preset value. The first target distance is the distance between the feature vector in the first feature map and the first target feature vector. The code table is a set of feature vectors. The code table is used to find the second target feature vector whose second target distance is less than or equal to a preset value. The second target distance is the distance between the feature vector in the second feature map and the second target feature vector. Based on the first target feature vector, a first reconstructed feature map is generated; Based on the second target feature vector, a second reconstructed feature map is generated; Based on the first distance and the second distance, predict the third distance between the first cloud cluster and the second cloud cluster at the third time. The step of predicting the third radar echo map of the target radar at the third moment based on the relationship between the first distance and the second distance includes: Decode the first reconstructed feature map and the second reconstructed feature map, and predict the spatial features of the first cloud cluster and the second cloud cluster at the third time, wherein the spatial features include the size, position or shape of the cloud cluster; Based on the spatial characteristics of the first cloud cluster, the spatial characteristics of the second cloud cluster, and the third distance, the third radar echo map of the target radar at the third time is predicted.

5. The method according to claim 4, characterized in that, The step of predicting the third distance between the first cloud cluster and the second cloud cluster at the third time point based on the first distance and the second distance includes: If the second distance is greater than the first distance, then the third distance is predicted to be greater than the second distance, and the difference between the third distance and the second distance is related to the time interval between the third time moment and the second time moment; If the second distance is less than the first distance, then the third distance is predicted to be less than the second distance, and the difference between the third distance and the second distance is related to the time interval between the third time moment and the second time moment.

6. The method according to claim 4, characterized in that, Decoding the first reconstructed feature map and the second reconstructed feature map includes: The first and second reshaped feature maps are decoded using a Convolutional Long Short-Term Memory (ConvLSTM) network or a Convolutional Gated Recurrent Unit (ConvRGU).

7. The method according to claim 1 or 2, characterized in that, The acquisition of the target radar's first radar echo image at a first moment and the second radar echo image at a second moment includes: Receive radar data from the target radar at the first moment; The radar data at the first moment is subjected to median filtering to obtain the first radar echo map; Receive radar data from the target radar at the second moment; The radar data at the second time point is subjected to median filtering to obtain the second radar echo map.

8. The method according to claim 1 or 2, characterized in that, The third radar echo map includes the location, shape, or size of each cloud cluster, and is used to predict meteorological information for the region corresponding to the third radar echo map.

9. A weather information forecasting device, characterized in that, include: The acquisition unit is used to acquire a first radar echo image of the target radar at a first moment and a second radar echo image at a second moment, the second moment being later than the first moment. The first radar echo image includes the characteristics of the cloud clusters detected by the target radar at the first moment, and the second radar echo image includes the characteristics of the cloud clusters detected by the target radar at the second moment. The determining unit is used to determine a first distance and a second distance for the first cloud cluster and the second cloud cluster detected by the target radar at the first time and the second time, wherein the first distance is the distance between the first cloud cluster and the second cloud cluster in the first radar echo map, and the second distance is the distance between the first cloud cluster and the second cloud cluster in the second radar echo map; The prediction unit is configured to predict a third radar echo pattern of the target radar at a third time point, which is later than the second time point, based on the relationship between the first distance and the second distance.

10. The apparatus according to claim 9, characterized in that, The device further includes: a processing unit; The processing unit is specifically used for: The first radar echo image is segmented to obtain X first sub-regions, each of the X first sub-regions having the same size, where X is an integer and X≥2; The second radar echo map is segmented to obtain Y second sub-regions, each of the Y second sub-regions having the same size, where X is an integer and Y≥2; Pool the X first sub-regions to obtain the coefficients corresponding to the pooled first region. The pooled first region includes m first sub-regions, where m is an integer and 2≤m≤X. Pool the Y second sub-regions to obtain the coefficients corresponding to the pooled second region. The pooled second region includes n second sub-regions, where n is an integer and 2≤n≤Y. The second region has the same size as the first region. The determining unit is specifically used for: In the first target region, the first distance is determined, the first target region is contained within the first region, and the coefficient corresponding to the first target region is greater than or equal to a preset threshold. In the second target region, the second distance is determined, the second target region is contained within the second region, and the coefficient corresponding to the second target region is greater than or equal to a preset threshold.

11. The apparatus according to claim 9 or 10, characterized in that, The first radar echo image includes a first feature image, and the second radar echo image includes a second feature image. The resolution of the first feature image is less than the resolution of the first radar echo image. The distance between the first cloud cluster and the second cloud cluster in the first feature image is the first distance. The resolution of the second feature image is less than the resolution of the second radar echo image. The distance between the first cloud cluster and the second cloud cluster in the second feature image is the second distance.

12. The apparatus according to claim 11, characterized in that, The apparatus further includes: a generation unit; The generation unit is specifically used for: The code table is used to find the first target feature vector whose first target distance is less than or equal to a preset value. The first target distance is the distance between the feature vector in the first feature map and the first target feature vector. The code table is a set of feature vectors. The code table is used to find the second target feature vector whose second target distance is less than or equal to a preset value. The second target distance is the distance between the feature vector in the second feature map and the second target feature vector. Based on the first target feature vector, a first reconstructed feature map is generated; Based on the second target feature vector, a second reconstructed feature map is generated; The prediction unit is also used for: Based on the first distance and the second distance, predict the third distance between the first cloud cluster and the second cloud cluster at the third time. Decode the first reconstructed feature map and the second reconstructed feature map, and predict the spatial features of the first cloud cluster and the second cloud cluster at the third time, wherein the spatial features include the size, position or shape of the cloud cluster; Based on the spatial characteristics of the first cloud cluster, the spatial characteristics of the second cloud cluster, and the third distance, the third radar echo map of the target radar at the third time is predicted.

13. The apparatus according to claim 12, characterized in that, The prediction unit is specifically used for: If the second distance is greater than the first distance, then the third distance is predicted to be greater than the second distance, and the difference between the third distance and the second distance is related to the time interval between the third time moment and the second time moment; If the second distance is less than the first distance, then the third distance is predicted to be less than the second distance, and the difference between the third distance and the second distance is related to the time interval between the third time moment and the second time moment.

14. The apparatus according to claim 12, characterized in that, The prediction unit is specifically used to decode the first reshaped feature map and the second reshaped feature map using ConvLSTM or ConvRGU.

15. The apparatus according to claim 9 or 10, characterized in that, The acquisition unit is specifically used for: Receive radar data from the target radar at the first moment; The radar data at the first moment is subjected to median filtering to obtain the first radar echo map; Receive radar data from the target radar at the second moment; The radar data at the second time point is subjected to median filtering to obtain the second radar echo map.

16. The apparatus according to claim 9 or 10, characterized in that, The third radar echo map includes the location, shape, or size of each cloud cluster, and is used to predict meteorological information for the region corresponding to the third radar echo map.

17. A weather information forecasting device, characterized in that, include: A memory and at least one processor, wherein the memory is used to store program instructions, and the at least one processor, when invoking the program instructions, implements the method according to any one of claims 1-8.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains a program that, when executed by the computer, performs the method according to any one of claims 1 to 8.

Images