Method, device and equipment for short-term forecast and warning of hail based on temporal convolutional network

Abstract

The present invention provides a method, device and equipment for short-term forecasting and warning of hail based on a temporal convolutional network, which relates to the field of meteorological monitoring and early warning technology, including: determining the label and multi-dimensional feature data corresponding to each grid point in the target area based on historical radar observation data, historical automatic station observation data and historical forecast data corresponding to the target area; constructing multiple hail samples and multiple non-hail samples according to the label and multi-dimensional feature data corresponding to each grid point; performing expansion processing on the hail samples so that the ratio between the total number of hail samples after the expansion processing and the total number of non-hail samples meets a preset condition; training the temporal convolutional network using the hail samples and non-hail samples after the expansion processing, and the trained temporal convolutional network is used to perform short-term forecasting and warning of hail in the target area based on the radar observation data to be tested and the forecast data to be tested. The present invention can significantly improve the short-term forecasting and warning capability of hail.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological monitoring and early warning, and in particular to a method, device and equipment for short-term forecasting and early warning of hail based on a temporal convolutional network. Background Art

[0002] Hail is a severe weather process, which is solid water that falls from convective hail clouds. Hail occurs in a small area and for a short time, but it is fierce and often accompanied by strong winds and heavy rains. It causes great losses to agriculture, animal husbandry, industry and mining, telecommunications, transportation, and people's lives and property. It is one of the important disastrous weather. Hail is a sudden and destructive weather phenomenon, and monitoring and early warning should be strengthened to reduce the losses it brings. At present, relevant technologies propose to use weather radar to monitor and warn severe convective weather such as hail. However, when using weather radar for hail forecasting, there are usually problems such as low forecast accuracy and timeliness. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide a method, device and equipment for short-term and impending hail forecasting and warning based on a temporal convolutional network, which can significantly improve the short-term and impending hail forecasting and warning capabilities.

[0004] In a first aspect, the present invention provides a method for short-term forecasting and warning of hail based on a temporal convolutional network, comprising:

[0005] Based on the historical radar observation data, historical automatic station observation data and historical forecast data corresponding to the target area, the label and multi-dimensional feature data corresponding to each grid point in the target area are determined. The labels are divided into hail labels and non-hail labels.

[0006] According to the label and multi-dimensional feature data corresponding to each grid point, a plurality of hail samples and a plurality of non-hail samples are constructed, wherein the hail samples include the label and multi-dimensional feature data corresponding to each grid point in the first area, and the non-hail samples include the label and multi-dimensional feature data corresponding to each grid point in the second area, wherein the first area is an area composed of grid points where hail events occur in the target area, and the second area is an area composed of grid points where hail events do not occur in the target area, and the size of the first area is consistent with the size of the second area;

[0007] The hail samples are expanded so that the ratio between the total number of hail samples after the expansion process and the total number of non-hail samples meets a preset condition;

[0008] The temporal convolutional network is trained using the expanded hail samples and non-hail samples. The trained temporal convolutional network is used to provide short-term hail forecast and warning for the target area based on the radar observation data and forecast data to be tested.

[0009] In one embodiment, based on the historical radar observation data, historical automatic station observation data and historical forecast data corresponding to the target area, the label and multi-dimensional feature data corresponding to each grid point in the target area are determined, including:

[0010] The label corresponding to each grid point in the target area is determined by using the location information of the hail event described by the historical automatic station observation data corresponding to the target area;

[0011] Furthermore, using the occurrence location information of hail events described by the historical automatic station observation data corresponding to the target area, the historical radar observation data and historical forecast data corresponding to the target area are preprocessed to determine the multi-dimensional feature data corresponding to each grid point in the target area.

[0012] In one embodiment, the location information of the hail event described by the historical automatic station observation data corresponding to the target area is used to pre-process the historical radar observation data and historical forecast data corresponding to the target area to determine the multi-dimensional feature data corresponding to each grid point in the target area, including:

[0013] Using the location information of the hail event described by the historical automatic station observation data corresponding to the target area, spatial matching is performed on the historical radar observation data and historical forecast data corresponding to the target area to determine the historical radar observation data and historical forecast data corresponding to each grid point in the target area;

[0014] Based on the historical radar observation data and historical forecast data corresponding to each grid point, the multi-dimensional feature data corresponding to each grid point is extracted.

[0015] In one embodiment, the multi-dimensional characteristic data includes one or more of maximum reflectivity factor, vertically integrated liquid water content, vertically integrated liquid water content density, echo top height, strong echo centroid height, zero-degree layer height, and convective effective potential energy.

[0016] In one embodiment, the hail sample is expanded and processed, including:

[0017] For each hail sample, multiple adjacent hail samples are determined according to the Euclidean distance between the hail sample and other hail samples, multiple target adjacent hail samples are randomly selected from the adjacent hail samples, and multiple new hail samples are generated based on the hail sample and the target adjacent hail samples to achieve expansion processing of the hail samples.

[0018] In one embodiment, generating a plurality of new hail samples based on the hail sample and target adjacent hail samples includes:

[0019] Determine a difference hail sample between the hail sample and a target adjacent hail sample;

[0020] The random number function is used to adjust the difference hail samples;

[0021] The adjusted difference hail samples are fused with the target adjacent hail samples to obtain new hail samples.

[0022] In one embodiment, the method further comprises:

[0023] Based on the radar observation data to be measured and the forecast data to be measured corresponding to the target area, the multi-dimensional feature data to be measured corresponding to each grid point in the target area is extracted;

[0024] Through the trained temporal convolutional network, short-term hail forecast and warning are carried out for the target area based on the multi-dimensional feature data to be tested, and the short-term hail forecast and warning results corresponding to the target area are obtained.

[0025] In a second aspect, the present invention further provides a short-term forecast and warning device for hail based on a temporal convolutional network, comprising:

[0026] A data processing module is used to determine the label and multi-dimensional feature data corresponding to each grid point in the target area based on the historical radar observation data, historical automatic station observation data and historical forecast data corresponding to the target area, and the labels are divided into hail labels and non-hail labels;

[0027] A sample construction module is used to construct a plurality of hail samples and a plurality of non-hail samples according to the label and multi-dimensional feature data corresponding to each grid point, wherein the hail samples include the label and multi-dimensional feature data corresponding to each grid point in the first area, and the non-hail samples include the label and multi-dimensional feature data corresponding to each grid point in the second area, wherein the first area is an area composed of grid points where hail events occur in the target area, and the second area is an area composed of grid points where hail events do not occur in the target area, and the size of the first area is consistent with the size of the second area;

[0028] The sample expansion module is used to expand the hail samples so that the ratio between the total number of hail samples after expansion and the total number of non-hail samples meets the preset conditions;

[0029] The network training module is used to train the temporal convolutional network using the expanded hail samples and non-hail samples. The trained temporal convolutional network is used to provide short-term forecast and warning of hail in the target area based on the radar observation data to be tested and the forecast data to be tested.

[0030] In a third aspect, the present invention further provides an electronic device, comprising a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement any one of the methods provided in the first aspect.

[0031] In a fourth aspect, the present invention further provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement any one of the methods provided in the first aspect.

[0032] The present invention provides a method, device and equipment for short-term forecasting and warning of hail based on a temporal convolutional network. First, based on historical radar observation data, historical automatic station observation data and historical forecast data corresponding to the target area, the label and multi-dimensional feature data corresponding to each grid point in the target area are determined, and the label is divided into a hail label and a non-hail label; then, according to the label and the multi-dimensional feature data corresponding to each grid point, a plurality of hail samples and a plurality of non-hail samples are constructed. The hail samples include the label and the multi-dimensional feature data corresponding to each grid point in the first area, and the non-hail samples include the label and the multi-dimensional feature data corresponding to each grid point in the second area. The first area is the area composed of grid points where hail events occur in the target area, and the second area is the area composed of grid points where hail events do not occur in the target area. The size of the first area is consistent with the size of the second area. The hail samples are then expanded so that the ratio of the total number of expanded hail samples to the total number of non-hail samples meets the preset conditions. Finally, the expanded hail samples and non-hail samples are used to train the temporal convolutional network. The trained temporal convolutional network is used to perform short-term forecast and warning of hail in the target area based on the radar observation data to be tested and the forecast data to be tested. The above method integrates multi-source data such as historical radar observation data, historical automatic station observation data and historical forecast data to generate labels and multi-dimensional feature data corresponding to each grid point in the target area, which are used to construct hail samples and non-hail samples, and use the expanded hail samples and non-hail samples to train the temporal convolutional network. This can not only enable the temporal convolutional network to better learn the mapping relationship between labels and multi-dimensional feature data, but also reduce the problem of poor training effect of the temporal convolutional network due to uneven samples, thereby significantly improving the short-term forecast and warning capabilities of hail.

[0033] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0034] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0036] Figure 1 A schematic diagram of a flow chart of a method for short-term forecasting and warning of hail based on a temporal convolutional network provided by an embodiment of the present invention;

[0037] Figure 2 A schematic diagram of a specific process of a method for short-term forecasting and warning of hail based on a temporal convolutional network provided by an embodiment of the present invention;

[0038] Figure 3 A schematic diagram of the structure of a short-term forecast and warning device for hail based on a temporal convolutional network provided by an embodiment of the present invention;

[0039] Figure 4 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described in combination with the embodiments below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] At present, the existing hail forecasting technology usually has problems such as low forecast accuracy and timeliness. Based on this, the present invention implements a method, device and equipment for short-term forecasting and warning of hail based on a temporal convolutional network, which can significantly improve the short-term forecasting and warning capabilities of hail.

[0042] To facilitate understanding of this embodiment, firstly, a method for short-term forecasting and warning of hail based on a temporal convolutional network disclosed in an embodiment of the present invention is described in detail. Figure 1 The flowchart of a method for short-term forecasting and warning of hail based on a temporal convolutional network is shown, and the method mainly includes the following steps S102 to S108:

[0043] Step S102, based on the historical radar observation data, historical automatic station observation data and historical forecast data corresponding to the target area, determine the label and multi-dimensional feature data corresponding to each grid point in the target area.

[0044] Among them, the historical forecast data is specifically the EC forecast data. The EC forecast data is the meteorological forecast data provided by the Atmospheric Initial Condition Field Forecast System (AIFS) of the European Center for Medium-Range Weather Forecasts (ECMWF), which includes forecasts of various meteorological elements around the world. Labels are divided into hail labels and non-hail labels. For example, when the label value is set to 1, it indicates that a hail event has occurred at the grid point, that is, it is marked with a hail label. When the label value is set to 0, it indicates that no hail event has occurred at the grid point, that is, it is marked with a non-hail label. Multi-dimensional feature data may include one or more of the maximum reflectivity factor, vertically integrated liquid water content, vertically integrated liquid water content density, echo top height, strong echo centroid height, zero-degree layer height, and convective effective potential energy.

[0045] In one example, the location information of hail events represented by historical automatic station observation data can be used to automatically label each grid point; at the same time, the location information of hail events represented by historical automatic station observation data can be used to spatially match historical radar observation data and historical EC forecast data, and then the multi-dimensional feature data corresponding to each grid point can be extracted using the spatially matched historical radar observation data and historical EC forecast data.

[0046] Step S104, constructing a plurality of hail samples and a plurality of non-hail samples according to the label and multi-dimensional feature data corresponding to each grid point.

[0047] Among them, the hail samples include the labels and multi-dimensional feature data corresponding to each grid point in the first area, and the non-hail samples include the labels and multi-dimensional feature data corresponding to each grid point in the second area. The first area is the area composed of grid points where hail events occur in the target area, and the second area is the area composed of grid points where hail events do not occur in the target area. The size of the first area is consistent with the size of the second area.

[0048] In one example, a slice size can be pre-configured based on the occurrence range of a hail event, and the target area can be sliced ​​according to the slice size to obtain multiple first areas and multiple second areas. Combined with the label corresponding to each grid point and the multi-dimensional feature data, multiple hail samples and multiple non-hail samples can be constructed.

[0049] Step S106, performing expansion processing on the hail samples so that the ratio between the total number of the expanded hail samples and the total number of the non-hail samples meets a preset condition.

[0050] The preset condition may be a preset ratio threshold. In one example, multiple algorithms such as SMOTE (Synthetic Minority Over-sampling Technique) may be used to implement expansion processing of hail samples, and the original hail samples and the new hail samples are used together as expanded hail samples.

[0051] Step S108, using the expanded hail samples and non-hail samples to train the temporal convolutional network, the trained temporal convolutional network is used to perform short-term forecast and warning of hail in the target area based on the radar observation data to be tested and the forecast data to be tested.

[0052] In one example, the multi-dimensional feature data corresponding to each grid point contained in the expanded hail samples and non-hail samples are used as the input of the temporal convolutional network, and the labels corresponding to each grid point contained in the expanded hail samples and non-hail samples are used as training labels. The temporal convolutional network is continuously trained to learn the mapping relationship between the multi-dimensional feature data and the labels, so that the trained temporal convolutional network can better realize the short-term forecast and warning of hail.

[0053] The method for short-term hail forecasting and warning based on a temporal convolutional network provided by an embodiment of the present invention integrates multi-source data such as historical radar observation data, historical automatic station observation data and historical forecast data to generate labels and multi-dimensional feature data corresponding to each grid point in the target area, so as to construct hail samples and non-hail samples, and use the hail samples and non-hail samples after expansion processing to train the temporal convolutional network, which can not only enable the temporal convolutional network to better learn the mapping relationship between labels and multi-dimensional feature data, but also reduce the problem of poor training effect of the temporal convolutional network due to uneven samples, thereby significantly improving the short-term hail forecasting and warning capabilities.

[0054] For ease of understanding, the present invention provides a specific implementation method of a short-term forecast and warning method for hail based on a temporal convolutional network. It uses radar observation data, automatic station observation data, and numerical forecast data combined with deep learning to perform short-term forecast and warning analysis on hail, and compares it with other short-term forecast and warning methods for hail. It continuously optimizes and improves the hail forecasting capability, which can be applied to future actual business. For details, see Figure 2 The specific process diagram of a method for short-term forecasting and warning of hail based on a temporal convolutional network is shown, including:

[0055] S1: Obtain the historical radar observation data, historical automatic station observation data and historical EC forecast data three hours before the occurrence of hail in a certain area in recent years.

[0056] S2: Based on the historical radar observation data, historical automatic station observation data and historical forecast data corresponding to the target area, determine the label and multi-dimensional feature data corresponding to each grid point in the target area, and build a feature factor library. The feature factor library includes the multi-dimensional feature data corresponding to each grid point.

[0057] In one example, the process of determining the label corresponding to each grid point is as follows: the label corresponding to each grid point in the target area is determined using the location information of the hail event described by the historical automatic station observation data corresponding to the target area. Exemplarily, the historical automatic station observation data describes the location information of the hail event, and the location information can be a longitude and latitude range. Using the location information of the hail event as the retrieval condition, extract the grid point that matches the location information, which is the grid point where the hail event occurred, and set its label value to 1 (hail label), and set the label value of other grid points in the target area to 0 (non-hail label).

[0058] In one example, the process of determining the multi-dimensional feature data corresponding to each grid point is as follows: using the location information of the hail event described by the historical automatic station observation data corresponding to the target area, the historical radar observation data and historical forecast data corresponding to the target area are preprocessed to determine the multi-dimensional feature data corresponding to each grid point in the target area.

[0059] Specifically, the occurrence location information of hail events described by the historical automatic station observation data corresponding to the target area is used to spatially match the historical radar observation data and the historical forecast data corresponding to the target area to determine the historical radar observation data and the historical forecast data corresponding to each grid point in the target area; based on the historical radar observation data and the historical forecast data corresponding to each grid point, the multi-dimensional feature data corresponding to each grid point are extracted, and the multi-dimensional feature data include the maximum reflectivity factor of the storm cell, the vertically integrated liquid water content, the vertically integrated liquid water content density, the echo top height, the strong echo centroid height, the zero-degree layer height and the convective effective potential energy.

[0060] The embodiment of the present invention provides a calculation formula for the maximum reflectivity factor:

[0061] ;

[0062] in To calculate the maximum value of all grid points in the vertical height layer, that is, the maximum reflectivity matrix. The greater the radar reflectivity value, the greater the probability of hail. The radar reflectivity value can be provided by historical radar observation data.

[0063] The embodiment of the present invention provides a calculation formula for vertically integrated liquid water content:

[0064] ;

[0065] In the formula is the number of sampling layers, and are the radar reflectivity values ​​at the lower and upper locations of the sampling layer, is the thickness of the sampling layer. Although the vertical cumulative liquid water content was originally designed to show precipitation potential, higher vertical cumulative liquid water content values ​​often indicate the occurrence of hail with a high probability. When used for hail detection, there is no clear threshold for vertical cumulative liquid water content. The threshold for hail when the storm is in cold air is smaller than when it is in warm air. Therefore, the vertical cumulative liquid water content cannot often be used independently for hail warning.

[0066] The present invention provides an implementation method for determining the vertically integrated liquid water content density: if the convective storm has a long vertical extension, it will tend to accumulate more liquid water, but this is not directly equivalent to having a higher reflectivity factor; on the contrary, some storms may contain solid particles, but due to their limited vertical scale, it is impossible to measure a significant amount of liquid water. Combined with standardization to solve the above problems, vertical integration of liquid water content density The specific definition is:

[0067] .

[0068] In addition, the aforementioned echo top height and strong echo centroid height can be calculated based on historical radar observation data. The echo top height reflects the maximum height reached by the cloud cluster. Here, three top heights of 30, 45, and 55dBZ are selected, representing the maximum heights of the cloud cluster as a whole, the core, and the strong core, respectively. The specific calculation process of the echo top height and the strong echo centroid height will not be repeated in the embodiment of the present invention; the aforementioned zero-degree layer height and convective effective potential energy can be directly provided by historical EC forecast data.

[0069] S3: Construct multiple hail samples and multiple non-hail samples according to the label and multi-dimensional feature data corresponding to each grid point.

[0070] S4: Use the SMOTE oversampling algorithm to expand the hail samples and balance the hail samples and non-hail samples to obtain a trainable data set. Specifically, it includes the following S4.1 to S4.3:

[0071] S4.1, determining a plurality of adjacent hail samples based on the Euclidean distance between the hail sample and other hail samples. For small sample events, all hail samples Constitute the minority class sample set. For each hail sample , calculate the hail sample The Euclidean distance between all hail samples in the minority class sample set is calculated, and multiple k neighboring hail samples are determined in the order of Euclidean distance from small to large.

[0072] S4.2, randomly select multiple target adjacent hail samples from the adjacent hail samples. In one example, the sampling ratio can be set with reference to the sample imbalance ratio. , sampling rate This is the preset ratio threshold mentioned above, which can be used to calculate the total number of non-hail samples and the sampling rate. The product of is taken as the total number of denoised samples after the expansion processing, and the difference between the total number and the number of original hail samples is the number of selected target neighboring hail samples.

[0073] S4.3, generating multiple new hail samples based on the hail sample and the target adjacent hail sample. In one embodiment, first, determine the difference hail sample between the hail sample and the target adjacent hail sample; then, use a random number function to adjust the difference hail sample; finally, fuse the adjusted difference hail sample with the target adjacent hail sample to obtain a new hail sample.

[0074] In a specific embodiment, for each randomly selected target adjacent hail sample , calculate the new hail samples. The formula is as follows:

[0075] ;

[0076] in, For the new hail sample, is the target adjacent hail sample, For this hail sample, the range of rand(0, 1) is between [0, 1], which is used to control the position of the synthetic sample. For each minority class sample, the calculation is repeated to generate a corresponding number of synthetic samples to balance the data set.

[0077] S5: Construct a temporal convolutional network, input the training set into the temporal convolutional network, perform multi-scale extraction and recognition on multi-dimensional feature data, and output short-term hail forecast and warning results.

[0078] The temporal convolutional network consists of the following modules:

[0079] (1) Causal Convolution:

[0080] Causal convolution only considers the one-dimensional convolution operation of the past time state. The output of each time step depends only on the time step before it, not on the future. This is achieved through proper padding so that the convolution of each layer does not span future time steps. Hail forecast results at the moment , only depends on , without depending on future input At the same time, the further back in time the historical information is traced, the more hidden layers are needed.

[0081] (2) Dilated convolution:

[0082] When tasks require longer histories, causal convolution can only review a history that is linear in size in the network depth, which cannot meet the needs. Dilated convolution is used to achieve an exponentially large receptive field. Dilated convolution expands the sampling interval of the convolution kernel, so that the output feature map has a larger receptive field and higher expressive power. The sampling rate is controlled by the hyperparameter expansion rate.

[0083] For a one-dimensional sequence input and a filter , the expansion and dilation convolution operation on the sequence elements is defined as:

[0084] ;

[0085] in is the expansion factor, is the filter size, explains the direction of past development. Therefore, dilation is equivalent to introducing a fixed step between every two adjacent filter sections. When , the dilated convolution is a regular convolution. Using a larger dilation allows the output of the top layer to represent a wider range of inputs, effectively expanding the receptive field of the convolutional neural network. Choosing a larger filter size and increase the expansion factor The receptive field of the TCN can be increased. This ensures that the filter can contain valid data in the historical data.

[0086] (3) Residual Connection (ResNet):

[0087] Residual connections contain transformations Add input Modules:

[0088] ;

[0089] This effectively allows the layers to learn modifications to the identity mapping rather than the entire transformation, which has been shown many times to benefit very deep networks.

[0090] Each layer of a temporal convolutional network consists of multiple filters for feature extraction. Therefore, in the designed general temporal convolutional network model, a general residual module is used instead of the convolutional layer. In a residual block, the temporal convolutional network has two layers of dilated causal convolutions and nonlinearities, for which rectified linear units (ReLU) are used. For normalization, weight normalization is applied to the convolutional filters. In addition, spatial dropout is added after each dilated convolution for regularization: the entire channel is zeroed at each training step. However, in a standard ResNet, the input is directly added to the output of the residual function, while in TCN (and ConvNets in general), the input and output may have different widths. To account for the difference in input-output width, an additional 1x1 convolution is used to ensure that element-wise additions receive tensors of the same shape.

[0091] S6: Use the test set to optimize the temporal convolutional network and test the performance of the temporal convolutional network. Specifically, the specific steps for testing the performance of the temporal convolutional network include:

[0092] S6.1: The data set is divided into a training set and a test set. Based on the observation data of the storm cell at the hail station, a label is clearly assigned to each storm cell area, where 1 represents that hail weather has occurred in the area, and 0 represents that hail weather has not occurred.

[0093] S6.2: In the process of training the temporal convolutional network, the cross entropy loss function is selected to guide model learning, and the model parameters are tuned to optimize performance. The formula of the cross entropy loss function is shown below, which effectively measures the difference between the model prediction and the actual hail occurrence, thereby obtaining an accurate temporal convolutional network for short-term hail forecast and warning (referred to as the short-term hail forecast and warning model).

[0094] ;

[0095] in, For discrete variables The probability distribution of .

[0096] S6.3: Use Adam to minimize the loss function and add L2 regularization. The specific steps include:

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] ;

[0102] ;

[0103] Among them, the gradient is the loss function right Find the partial derivative, for exist The next order distance estimation in the form of moment momentum, for exist Second-order moment estimation in the form of moment momentum, and For the first-order and second-order distance estimates after bias correction, the above parameter updates are calculated. get , , , , , These are the network parameters of the temporal convolutional network.

[0104] S6.4: Based on the actual observed hail results and the hail results predicted by the model, multiple scoring indicators of the short-term hail forecast and warning model are calculated to comprehensively evaluate its performance. These target indicators include: critical success index (TS), hit probability (POD) and false alarm rate (FAR). The calculation formulas of each indicator have been clearly given, aiming to accurately measure the accuracy and reliability of the model in the short-term hail forecast and warning tasks through quantitative analysis.

[0105] ;

[0106] ;

[0107] ;

[0108] in, represents the number of true observations and true predictions, represents the number of false observations and true predictions, Represents the number of true observations and false predictions.

[0109] S6.5: Based on the target indicators, determine a model suitable for the short-term and impending hail forecast and warning, namely the short-term and impending hail forecast and warning model.

[0110] S7: Through the trained temporal convolutional network, based on the radar observation data and forecast data to be tested, the short-term forecast and warning of the hail time point to be tested is realized.

[0111] Specifically, firstly, based on the radar observation data to be measured and the forecast data to be measured corresponding to the target area, the multi-dimensional feature data to be measured corresponding to each grid point in the target area is extracted; then, through the trained temporal convolutional network, the target area is subjected to a short-term forecast and warning of hail based on the multi-dimensional feature data to be measured, and a short-term forecast and warning result of hail corresponding to the target area is obtained. The process of extracting the multi-dimensional feature data to be measured can be referred to the aforementioned S2, and the embodiment of the present invention will not be described in detail.

[0112] In summary, the short-term forecast and warning method for hail based on a time convolutional network provided by an embodiment of the present invention includes: obtaining radar, automatic station and EC forecast data three hours before the occurrence of hail in a certain area in recent years, and preprocessing the data. Extract the characteristic data of the storm cell based on radar, automatic station and EC forecast data, and build a characteristic factor library. Hail events are small sample events relative to non-hail events. The SMOTE oversampling algorithm is used to expand the data volume of hail weather, and balance the positive and negative samples to obtain a trainable data set. The data set processed by the SMOTE algorithm is divided into a training set and a test set. Construct a time convolutional network (TCN) model, input the training set into the model, perform multi-scale extraction and recognition on the characteristic factors, and output the classification results. Use the test set to optimize the model and test the model performance. Input the data to be tested into the trained model to realize the short-term forecast and warning of the hail time point to be tested. The embodiment of the present invention realizes the short-term warning and forecast of hail 0-3 hours, effectively improves the hail recognition and classification effect, and improves the accuracy of hail weather monitoring.

[0113] The embodiments of the present invention have at least the following characteristics:

[0114] 1. SMOTE can improve model performance. By solving the problem of class imbalance, SMOTE enables the model to better identify patterns and generalize, thereby improving overall performance. SMOTE helps reduce the bias caused by class imbalance and ensures that the model does not favor the majority class and ignore the minority class. SMOTE is algorithm-independent and compatible with various machine learning algorithms. The synthetic samples generated by the SMOTE algorithm interpolate between minority class samples, maintaining the distribution characteristics of the original data.

[0115] (ii) Temporal convolutional networks can efficiently process large amounts of time series data in the meteorological field, capture the dynamic characteristics of weather changes over time, and improve forecast accuracy. At the same time, they can capture the complex spatiotemporal correlations in the weather system, which helps to more accurately predict weather changes. Temporal convolutional networks rely on data-driven methods and make predictions through training with a large amount of historical meteorological data. They can automatically extract features and make predictions, reducing reliance on physical models. Compared with traditional numerical weather forecasting methods, temporal convolutional networks are expected to significantly reduce the time required for meteorological forecasting and improve the accuracy of forecasts, especially in long-term weather forecasting.

[0116] (III) The cross entropy loss function has good mathematical properties, such as convexity and differentiability, and can handle multi-category classification problems well. It can effectively evaluate model performance by measuring the classification error of each category. The cross entropy loss function conforms to the principle of maximum likelihood estimation, which makes it perform well in multi-category classification problems and can directly measure the difference between model predictions and true labels.

[0117] Based on the above embodiments, the present invention provides a short-term forecast and warning device for hail based on a time convolutional network. Figure 3 The schematic diagram of the structure of a short-term forecast and warning device for hail based on a temporal convolutional network is shown, and the device mainly includes the following parts:

[0118] The data processing module 302 is used to determine the label and multi-dimensional feature data corresponding to each grid point in the target area based on the historical radar observation data, the historical automatic station observation data and the historical forecast data corresponding to the target area, and the labels are divided into hail labels and non-hail labels;

[0119] The sample construction module 304 is used to construct a plurality of hail samples and a plurality of non-hail samples according to the label and multi-dimensional feature data corresponding to each grid point, wherein the hail samples include the label and multi-dimensional feature data corresponding to each grid point in the first area, and the non-hail samples include the label and multi-dimensional feature data corresponding to each grid point in the second area, wherein the first area is an area composed of grid points where hail events occur in the target area, and the second area is an area composed of grid points where hail events do not occur in the target area, and the size of the first area is consistent with the size of the second area;

[0120] The sample expansion module 306 is used to expand the hail samples so that the ratio between the total number of hail samples after the expansion process and the total number of non-hail samples meets a preset condition;

[0121] The network training module 308 is used to train the temporal convolutional network using the hail samples and non-hail samples after expansion processing. The trained temporal convolutional network is used to perform short-term forecast and warning of hail in the target area based on the radar observation data to be tested and the forecast data to be tested.

[0122] The device for short-term hail forecasting and warning based on a temporal convolutional network provided by an embodiment of the present invention integrates multi-source data such as historical radar observation data, historical automatic station observation data and historical forecast data to generate labels and multi-dimensional feature data corresponding to each grid point in the target area, so as to construct hail samples and non-hail samples, and use the hail samples and non-hail samples after expansion processing to train the temporal convolutional network, which can not only enable the temporal convolutional network to better learn the mapping relationship between labels and multi-dimensional feature data, but also reduce the problem of poor training effect of the temporal convolutional network due to uneven samples, thereby significantly improving the short-term hail forecasting and warning capabilities.

[0123] In one implementation, the data processing module 302 is specifically configured to:

[0124] The label corresponding to each grid point in the target area is determined by using the location information of the hail event described by the historical automatic station observation data corresponding to the target area;

[0125] Furthermore, using the occurrence location information of hail events described by the historical automatic station observation data corresponding to the target area, the historical radar observation data and historical forecast data corresponding to the target area are preprocessed to determine the multi-dimensional feature data corresponding to each grid point in the target area.

[0126] In one implementation, the data processing module 302 is specifically configured to:

[0127] Using the location information of the hail event described by the historical automatic station observation data corresponding to the target area, spatial matching is performed on the historical radar observation data and historical forecast data corresponding to the target area to determine the historical radar observation data and historical forecast data corresponding to each grid point in the target area;

[0128] Based on the historical radar observation data and historical forecast data corresponding to each grid point, the multi-dimensional feature data corresponding to each grid point is extracted.

[0129] In one embodiment, the multi-dimensional characteristic data includes one or more of maximum reflectivity factor, vertically integrated liquid water content, vertically integrated liquid water content density, echo top height, strong echo centroid height, zero-degree layer height, and convective effective potential energy.

[0130] In one implementation, the sample expansion module 306 is specifically configured to:

[0131] For each hail sample, multiple adjacent hail samples are determined according to the Euclidean distance between the hail sample and other hail samples, multiple target adjacent hail samples are randomly selected from the adjacent hail samples, and multiple new hail samples are generated based on the hail sample and the target adjacent hail samples to achieve expansion processing of the hail samples.

[0132] In one embodiment, it is specifically used for:

[0133] Determine a difference hail sample between the hail sample and a target adjacent hail sample;

[0134] The random number function is used to adjust the difference hail samples;

[0135] The adjusted difference hail samples are fused with the target adjacent hail samples to obtain new hail samples.

[0136] In one embodiment, a hail forecast and warning module is also included, which is used to:

[0137] Based on the radar observation data to be measured and the forecast data to be measured corresponding to the target area, the multi-dimensional feature data to be measured corresponding to each grid point in the target area is extracted;

[0138] Through the trained temporal convolutional network, short-term hail forecast and warning are carried out for the target area based on the multi-dimensional feature data to be tested, and the short-term hail forecast and warning results corresponding to the target area are obtained.

[0139] The device provided in the embodiment of the present invention has the same implementation principle and technical effects as those of the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the device embodiment, reference may be made to the corresponding contents in the aforementioned method embodiment.

[0140] An embodiment of the present invention provides an electronic device. Specifically, the electronic device includes a processor and a storage device. The storage device stores a computer program, and when the computer program is executed by the processor, it executes the method described in any one of the above-mentioned implementation methods.

[0141] Figure 4 A structural diagram of an electronic device provided in an embodiment of the present invention, the electronic device 100 includes: a processor 40, a memory 41, a bus 42 and a communication interface 43, wherein the processor 40, the communication interface 43 and the memory 41 are connected via the bus 42; the processor 40 is used to execute an executable module stored in the memory 41, such as a computer program.

[0142] The memory 41 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 43 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

[0143] The bus 42 may be an ISA bus, a PCI bus, or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 4 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0144] Among them, the memory 41 is used to store programs, and the processor 40 executes the program after receiving the execution instruction. The method executed by the device for flow process definition disclosed in any embodiment of the above-mentioned embodiment of the present invention can be applied to the processor 40 or implemented by the processor 40.

[0145] The processor 40 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the processor 40. The above processor 40 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present invention can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiment of the present invention can be directly embodied as a hardware decoding processor to be executed, or the hardware and software modules in the decoding processor can be executed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 41, and the processor 40 reads the information in the memory 41 and completes the steps of the above method in combination with its hardware.

[0146] The computer program product of the readable storage medium provided in the embodiment of the present invention includes a computer-readable storage medium storing program code, and the instructions included in the program code can be used to execute the methods described in the previous method embodiments. The specific implementation can be referred to the previous method embodiments, which will not be repeated here.

[0147] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0148] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A short-term hail forecasting and warning method based on a temporal convolutional network, characterized in that: include: Based on the historical radar observation data, historical automatic station observation data and historical forecast data corresponding to the target area, determine the label and multi-dimensional feature data corresponding to each grid point in the target area, the label is divided into a hail label and a non-hail label, and the multi-dimensional feature data includes one or more of the maximum reflectivity factor, the vertically integrated liquid water content, the vertically integrated liquid water content density, the echo top height, the strong echo centroid height, the zero-degree layer height and the convective effective potential energy; According to the label and the multi-dimensional feature data corresponding to each of the grid points, a plurality of hail samples and a plurality of non-hail samples are constructed, wherein the hail samples include the label and the multi-dimensional feature data corresponding to each grid point in a first area, and the non-hail samples include the label and the multi-dimensional feature data corresponding to each grid point in a second area, wherein the first area is an area composed of the grid points where hail events occur in the target area, and the second area is an area composed of the grid points where hail events do not occur in the target area, and the size of the first area is consistent with the size of the second area; Performing expansion processing on the hail samples so that the ratio between the total number of the hail samples after the expansion processing and the total number of the non-hail samples meets a preset condition; The time convolution network is trained by using the hail samples and the non-hail samples after the expansion process, and the trained time convolution network is used to perform short-term forecast and warning of hail in the target area based on the radar observation data to be measured and the forecast data to be measured; According to the label and the multi-dimensional feature data corresponding to each of the grid points, a plurality of hail samples and a plurality of non-hail samples are constructed, including: pre-configuring a slice size based on the occurrence range of the hail event, slicing the target area according to the slice size to obtain a plurality of the first areas and a plurality of the second areas, and combining the label and the multi-dimensional feature data corresponding to each of the grid points to construct a plurality of hail samples and a plurality of non-hail samples.

2. The method for short-term forecasting and early warning of hail based on a temporal convolutional network according to claim 1, characterized in that: Based on the historical radar observation data, historical automatic station observation data and historical forecast data corresponding to the target area, the label and multi-dimensional feature data corresponding to each grid point in the target area are determined, including: Using the occurrence location information of the hail event described by the historical automatic station observation data corresponding to the target area, determine the label corresponding to each grid point in the target area; Furthermore, using the occurrence location information of hail events described by the historical automatic station observation data corresponding to the target area, the historical radar observation data and the historical forecast data corresponding to the target area are preprocessed to determine the multi-dimensional feature data corresponding to each grid point in the target area.

3. The method for short-term forecasting and early warning of hail based on a temporal convolutional network according to claim 2, characterized in that: Using the occurrence location information of the hail event described by the historical automatic station observation data corresponding to the target area, the historical radar observation data and the historical forecast data corresponding to the target area are preprocessed to determine the multi-dimensional feature data corresponding to each grid point in the target area, including: Using the occurrence location information of the hail event described by the historical automatic station observation data corresponding to the target area, spatial matching is performed on the historical radar observation data and the historical forecast data corresponding to the target area to determine the historical radar observation data and the historical forecast data corresponding to each grid point in the target area; Based on the historical radar observation data and the historical forecast data corresponding to each grid point, multi-dimensional feature data corresponding to each grid point is extracted.

4. The method for short-term forecasting and early warning of hail based on a temporal convolutional network according to claim 1, characterized in that: The hail sample is expanded and processed, including: For each of the hail samples, a plurality of adjacent hail samples are determined according to the Euclidean distance between the hail sample and other hail samples, a plurality of target adjacent hail samples are randomly selected from the adjacent hail samples, and a plurality of new hail samples are generated based on the hail sample and the target adjacent hail samples, so as to achieve expansion processing of the hail samples.

5. The method for short-term forecasting and early warning of hail based on a temporal convolutional network according to claim 4 is characterized in that: Generating a plurality of new hail samples based on the hail sample and the target adjacent hail samples comprises: determining a difference hail sample between the hail sample and the target adjacent hail sample; Using a random number function to adjust the difference hail samples; The adjusted difference hail samples are fused with the target adjacent hail samples to obtain new hail samples.

6. The method for short-term forecasting and early warning of hail based on a temporal convolutional network according to claim 1, characterized in that: The method further comprises: Extracting the multi-dimensional feature data to be measured corresponding to each of the grid points in the target area based on the radar observation data to be measured and the forecast data to be measured corresponding to the target area; Through the trained temporal convolutional network, a short-term forecast and warning of hail is performed on the target area based on the multi-dimensional feature data to be tested, and a short-term forecast and warning result of hail corresponding to the target area is obtained.

7. A short-term hail forecast and warning device based on a temporal convolutional network, characterized in that: include: A data processing module, for determining a label and multi-dimensional feature data corresponding to each grid point in the target area based on historical radar observation data, historical automatic station observation data and historical forecast data corresponding to the target area, wherein the label is divided into a hail label and a non-hail label, and the multi-dimensional feature data includes one or more of a maximum reflectivity factor, a vertically integrated liquid water content, a vertically integrated liquid water content density, an echo top height, a strong echo centroid height, a zero-degree layer height and a convective effective potential energy; a sample construction module, configured to construct a plurality of hail samples and a plurality of non-hail samples according to the label and the multi-dimensional feature data corresponding to each grid point, wherein the hail samples include the label and the multi-dimensional feature data corresponding to each grid point in a first area, and the non-hail samples include the label and the multi-dimensional feature data corresponding to each grid point in a second area, wherein the first area is an area composed of the grid points where hail events occur in the target area, and the second area is an area composed of the grid points where hail events do not occur in the target area, and the size of the first area is consistent with the size of the second area; A sample expansion module, used for expanding the hail samples so that the ratio between the total number of the hail samples after the expansion process and the total number of the non-hail samples meets a preset condition; A network training module, used for training a temporal convolutional network using the hail samples and the non-hail samples after the expansion process, wherein the trained temporal convolutional network is used for performing short-term forecast and early warning of hail in the target area based on the radar observation data to be measured and the forecast data to be measured; The sample construction module pre-configures a slice size based on the occurrence range of the hail event, slices the target area according to the slice size to obtain a plurality of the first areas and a plurality of the second areas, and combines the label corresponding to each of the grid points and the multi-dimensional feature data to construct a plurality of hail samples and a plurality of non-hail samples.

8. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the method according to any one of claims 1 to 6.

Images