Weather nowcasting method, device, computer equipment and storage medium

Abstract

The present application relates to a meteorological nowcasting method, device, computer equipment and storage medium, wherein the computer equipment obtains a radar echo map of a target area; then, the radar echo map is input into a nowcasting model to obtain meteorological nowcasting data of the target area; wherein the nowcasting model is obtained by performing migration training on a base model according to a regional sample set of the target area; the base model is obtained by training using a meta-learning mechanism and a general sample set; the regional sample set and the general sample set both include historical radar echo maps and nowcasting results corresponding to the historical radar echo maps. The above method can improve the accuracy of meteorological nowcasting and the rapid learning ability of the nowcasting model.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a weather nowcasting method, apparatus, computer equipment and storage medium. Background Art

[0002] With the development of deep learning technology, artificial intelligence (AI) weather nowcasting has become a hot topic of research. However, meteorological observation data in different regions are mainly collected through scattered radars in various regions. Due to the huge differences in climate among regions, the distribution of radar echo patterns in different regions is completely different, which makes it very challenging to use artificial intelligence for weather nowcasting.

[0003] In the traditional method, computer equipment conducts model training on historical radar echo images of different regions, obtains the nowcast model of the region, and performs meteorological nowcasting for the region based on the model.

[0004] However, due to the lack of radar echo data or low data quality in some areas, the obtained nowcast model is not accurate enough. Summary of the invention

[0005] Based on this, it is necessary to provide a weather nowcasting method, device, computer equipment and storage medium that can improve the accuracy of weather nowcasting in response to the above technical problems.

[0006] A meteorological nowcasting method, the method comprising:

[0007] Obtain radar echo map of target area;

[0008] The radar echo map is input into the nowcasting model to obtain the meteorological nowcasting data of the target area; the nowcasting model is obtained by transfer training the base model according to the regional sample set of the target area; the base model is obtained by training with a meta-learning mechanism and a general sample set; both the regional sample set and the general sample set include historical radar echo maps and the nowcasting results corresponding to the historical radar echo maps.

[0009] In one embodiment, the method further comprises:

[0010] Obtaining universal sample sets and regional sample sets;

[0011] The historical radar echo images in the general sample set are used as the input of the initial deep learning model, and the nowcast results corresponding to the historical radar echo images in the general sample set are used as the reference output of the initial model. The initial deep learning model is trained using a meta-learning mechanism to obtain a base model.

[0012] The historical radar echo images in the regional sample set are taken as the input of the base model, and the nowcast results corresponding to the historical radar echo images in the regional sample set are taken as the reference output of the base model. The base model is trained to obtain the nowcast model.

[0013] In one embodiment, the initial deep learning model is trained to obtain a base model, including:

[0014] The initial deep learning model is trained based on a first loss function to obtain a base model; the first loss function includes a second-order derivative of the model parameters.

[0015] In one embodiment, the initial deep learning model is trained based on the first loss function to obtain a base model, including:

[0016] dividing the universal sample set into a first data set and a second data set;

[0017] Based on the first data set and the first loss function, train an initial deep learning model to obtain an initial base model;

[0018] Based on the second data set and the first loss function, a first preset number of parameter iterations are performed on the initial base model to obtain a base model.

[0019] In one embodiment, the first preset number of times is 5 times.

[0020] In one embodiment, the first loss function is:

[0021]

[0022] Among them, θ is the model parameter, D T is the Tth sample data, and η is the learning rate.

[0023] In one embodiment, the training of the base model to obtain the nowcast model includes:

[0024] The base model is trained based on the second loss function to obtain the nowcasting model; the second loss function includes the first-order derivatives of the model parameters.

[0025] In one embodiment, the training of the base model based on the second loss function to obtain the nowcasting model includes:

[0026] Based on the second loss function, the base model is iterated for a second preset number of parameters to obtain a nowcasting model.

[0027] In one embodiment, the second preset number is greater than or equal to 5 and less than or equal to 10.

[0028] In one embodiment, the second loss function is:

[0029]

[0030] Among them, θ * are the parameters of the base model, is the sample data of the sample base in the Tth region, and η is the learning rate.

[0031] In one embodiment, the above-mentioned nowcasting model is a long short-term memory (LSTM) network model.

[0032] A weather nowcasting device, comprising:

[0033] An acquisition module is used to acquire a radar echo map of a target area;

[0034] The input module is used to input the radar echo map into the nowcasting model to obtain the meteorological nowcasting data of the target area; the nowcasting model is obtained by transferring the base model according to the regional sample set of the target area; the base model is obtained by training with the meta-learning mechanism and the general sample set; the regional sample set and the general sample set both include historical radar echo maps and the nowcasting results corresponding to the historical radar echo maps.

[0035] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0036] Obtain radar echo map of target area;

[0037] The radar echo map is input into the nowcasting model to obtain the meteorological nowcasting data of the target area; the nowcasting model is obtained by transfer training the base model according to the regional sample set of the target area; the base model is obtained by training with a meta-learning mechanism and a general sample set; both the regional sample set and the general sample set include historical radar echo maps and the nowcasting results corresponding to the historical radar echo maps.

[0038] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0039] Obtain radar echo map of target area;

[0040] The radar echo map is input into the nowcasting model to obtain the meteorological nowcasting data of the target area; the nowcasting model is obtained by transfer training the base model according to the regional sample set of the target area; the base model is obtained by training with a meta-learning mechanism and a general sample set; both the regional sample set and the general sample set include historical radar echo maps and the nowcasting results corresponding to the historical radar echo maps.

[0041] The above-mentioned meteorological nowcasting method, device, computer equipment and storage medium, the computer equipment obtains the radar echo map of the target area; then, the radar echo map is input into the nowcasting model to obtain the meteorological nowcasting data of the target area; wherein the above-mentioned nowcasting model is obtained by transfer training the base model according to the regional sample set of the target area; the above-mentioned base model is obtained by adopting the meta-learning mechanism and the general sample set training; the above-mentioned regional sample set and the general sample set both include the historical radar echo map and the nowcasting results corresponding to the historical radar echo map. Since the above-mentioned nowcasting model is obtained by fast transfer training of the base model, the number of samples of the regional sample set required in the transfer training is small, and the accuracy of the nowcasting model can be improved for areas with a small number of samples or low quality of sample data; further, since the above-mentioned base model is obtained by adopting the meta-learning mechanism, the nowcasting model of the target area can be obtained by fast transfer training of the above-mentioned base model, thereby improving the fast learning ability of the nowcasting model. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A diagram showing an application environment of a weather nowcasting method in one embodiment;

[0043] Figure 2 A schematic diagram of a flow chart of a weather nowcasting method in one embodiment;

[0044] Figure 3 A schematic diagram of a weather nowcasting method in one embodiment;

[0045] Figure 4 A schematic flow chart of a weather nowcasting method in another embodiment;

[0046] Figure 5 A schematic flow chart of a weather nowcasting method in another embodiment;

[0047] Figure 6 A structural block diagram of a weather nowcasting device in one embodiment;

[0048] Figure 7 It is a structural block diagram of a weather nowcasting device in another embodiment;

[0049] Figure 8 It is a structural block diagram of a weather nowcasting device in another embodiment;

[0050] Fig. 9 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0052] The weather nowcasting method provided in this application can be applied to Figure 1 In the application environment shown, the computer device 100 is connected to the weather radar 200, and the current forecast result is obtained according to the radar echo map collected by the weather radar 200. The computer device 100 can be implemented by an independent server or a server cluster composed of multiple servers, and the weather radar 200 can be a conventional pulse radar or a Doppler radar, which is not limited here.

[0053] In one embodiment, Figure 2 As shown, a meteorological nowcasting method is provided, which is applied to Figure 1 The computer device 100 in FIG. 1 is taken as an example for explanation, including:

[0054] S101, obtaining a radar echo map of a target area.

[0055] Among them, the above-mentioned radar echo map can be the data collected by the meteorological radar in the target area to monitor the weather. After the meteorological radar transmits the pulse electromagnetic wave, the above-mentioned pulse electromagnetic wave can be scattered or absorbed by the target object (such as clouds, rain, precipitation particles, etc.); then, the meteorological radar can monitor the echo pulse to obtain the radar echo map, so that the computer equipment can make a weather forecast based on the radar echo map. The above-mentioned meteorological radar can be a conventional pulse radar or a Doppler radar, which is not limited here. The above-mentioned radar echo map can be a radar echo comprehensive map combining the map of the target area and the echo intensity of the meteorological radar, or it can be an echo intensity map in the coordinate system of the meteorological radar. The type of the above-mentioned radar echo map is not limited here.

[0056] The target area may be an area divided according to the distribution characteristics of meteorological radars, an area divided according to longitude and latitude, or an area divided according to cities. For example, the target area may be city A. The method of dividing the target area is not limited herein.

[0057] Specifically, the computer device can obtain the radar echo map sent by the weather radar by connecting to the weather radar; it can also call the radar echo map of the target area in the radar echo map database. The method for obtaining the above radar echo map is not limited here.

[0058] S102, inputting the radar echo map into the nowcasting model to obtain the meteorological nowcasting data of the target area; the nowcasting model is obtained by performing transfer training on the base model according to the regional sample set of the target area; the base model is obtained by training with a meta-learning mechanism and a general sample set; the regional sample set and the general sample set both include historical radar echo maps and the nowcasting results corresponding to the historical radar echo maps.

[0059] Among them, the above-mentioned nowcasting model is obtained by performing migration training on the base model according to the regional sample set of the target area. The above-mentioned regional sample set may include the historical radar echo map collected in the above-mentioned target area, and the nowcasting result obtained based on the historical radar echo map. The historical radar echo map in the above-mentioned regional sample set may be the radar echo map collected in a time period in the target area, such as the radar echo map collected by City A from April to June; in addition, the historical radar echo map in the above-mentioned regional sample set may also be the radar echo map corresponding to some weather types in the target area, such as the radar echo map corresponding to cloudy or less cloudy weather when the meteorological nowcast result is; in addition, it may also be the historical radar echo map obtained after filtering and screening the radar echo map collected in the target area; it is not limited here.

[0060] The above base model is obtained by training with a meta-learning mechanism and a general sample set. Among them, the meta-learning mechanism refers to the ability to learn through training so that the initial deep learning model can achieve the effect of rapid learning. The computer device can train the initial deep learning model to obtain a base model through the meta-learning mechanism to achieve parameter sharing; then, the current forecast model corresponding to the target area can be obtained by migration training based on the base model, which greatly saves the model training time. In other words, the base model is used to more quickly approximate the current forecast model of each region. There are many ways to implement the above meta-learning mechanism. The method of predicting gradients can be used to make the gradient descent of the neural network in the initial deep learning model more accurate to improve the learning speed; a parameter update mechanism can also be constructed so that the initial deep learning model can directly output new network parameters according to the current network parameters to achieve rapid learning; the implementation method of the above meta-learning mechanism is not limited here.

[0061] The above-mentioned general sample set may include historical radar echo maps of multiple regions, or historical radar echo maps of a region with a large number of samples, or historical radar echo maps of a region with high sample quality. The scope of the above-mentioned general sample set is not limited here. For example, the above-mentioned general sample set may include multiple historical radar echo maps of 10 cities such as City B, and the computer device may obtain the base model through the historical radar echo maps of multiple regions; or, the weather radar of City B has been established for a long time, corresponding to a large number of historical radar echo maps, and the computer device may also be trained based on the historical radar echo maps of City B to obtain the base model; or, the data quality collected by the weather radar of City C is high, such as small radar noise, high data integrity, etc., and the computer device may use the meta-learning mechanism and the historical radar echo map of City C to obtain the base model.

[0062] After the computer device obtains the base model, it can perform migration training on the base model according to the regional sample set of the target area to obtain the nowcast model of the target area. Specifically, the computer device can train the base model according to the regional sample set of the target area based on the parameters of the base model to obtain the nowcast model of the target area. Figure 3 Take the example to illustrate the above transfer training process. A curve function can be used to simply identify a nowcast model. The regional sample set of region A is used as the training sample to obtain the curve A corresponding to the model parameter θ1. The regional sample set of region B is used as the training sample to obtain the curve B corresponding to the model parameter θ2. The regional sample set of region 3 is used as the training sample to obtain the curve C corresponding to the model parameter θ3. For the model parameter θ1, the nowcast result corresponding to the radar echo map can be obtained based on the historical radar echo map of region 1. However, the above method needs to use different regional sample sets to train the model respectively, which requires a large number of samples and a long training time. Using the meta-learning mechanism, the model parameter θ of the base model can be obtained according to the general sample set, corresponding to the curve M. The above curve M can be a generalized function curve corresponding to multiple curves such as curve A and curve B, which can quickly approximate the function curve corresponding to each region. Based on the base model corresponding to the curve M, the computer device can use the regional sample set of region A to perform transfer training on the base function, and quickly obtain the model parameter θ1 of the nowcast model of region A. Similarly, the model parameters of the nowcast model corresponding to other regions can be obtained.

[0063] The above-mentioned regional sample set and general sample set both contain the nowcast results corresponding to the historical radar echo map. The above-mentioned nowcast results can be cloudy weather, less cloudy weather or cloudless weather, and can also be precipitation and air humidity, etc. The types of the above-mentioned nowcast results are not limited here. Accordingly, the computer device can determine the label of the historical radar echo map based on the nowcast result, and train to obtain the nowcast model, so that the nowcast model can obtain the nowcast data of the target area according to the radar echo map of the target area. The above-mentioned nowcast data is related to the label of the historical radar echo map, which can be cloudy, less cloudy or cloudless, etc.

[0064] The above-mentioned nowcasting model can be composed of a three-layer network or a network with other layers. The network structure of the nowcasting model is not limited here. Taking a specific model network as an example, the nowcasting model can include 4 downsampling layers, each of which can be composed of a convolution layer with a step size of 2, a batch normalization layer and a ReLU layer; after the 4 downsampling layers, it can also include a 1×1 convolution layer, a global average pooling layer and 5 upsampling layers, each of which can be composed of a deconvolution, a BN layer and a ReLU layer. The above-mentioned convolution layer can be used for dimensionality reduction and channel fusion of feature maps, and the channel dimension of the above-mentioned convolution layer can be 128. Optionally, the above-mentioned nowcasting model can be a long short-term memory (LSTM) network model. The above-mentioned LSTM network model can learn long-term dependency information, and the model parameter update method of the LSTM network model is similar to the update method of gradient descent, which can enable the model to have the ability to learn quickly.

[0065] Specifically, when the computer device inputs the radar echo map into the nowcast model, the radar echo map can be directly input into the nowcast model, or the radar echo map can be processed first and then input into the nowcast model. The above input method is not limited here. When the computer device processes the radar echo map, the radar echo map can be subjected to noise reduction processing, or adjacent radar echo maps can be subjected to frame insertion processing to obtain more radar echo features. The above processing method is not limited here.

[0066] In the above-mentioned meteorological nowcasting method, a computer device obtains a radar echo map of a target area; then, the radar echo map is input into a nowcasting model to obtain meteorological nowcasting data of the target area; wherein, the above-mentioned nowcasting model is obtained by transfer training a base model according to a regional sample set of the target area; the above-mentioned base model is obtained by training with a meta-learning mechanism and a general sample set; the above-mentioned regional sample set and the general sample set both include historical radar echo maps and nowcasting results corresponding to the historical radar echo maps. Since the above-mentioned nowcasting model is obtained by rapid transfer training of the base model, the number of samples of the regional sample set required in the transfer training is relatively small, and the accuracy of the nowcasting model can be improved for areas with a small number of samples or low quality of sample data; further, since the above-mentioned base model is obtained by adopting a meta-learning mechanism, the nowcasting model of the target area can be obtained by rapid transfer training of the above-mentioned base model, thereby improving the rapid learning ability of the nowcasting model.

[0067] Figure 4 FIG. 1 is a flow chart of a meteorological nowcasting method in an embodiment. This embodiment relates to a process of a computer device acquiring a nowcasting model. Based on the above embodiment, Figure 4 As shown, the above method includes:

[0068] S201: Obtain a general sample set and a regional sample set.

[0069] S202. Use the historical radar echo images in the general sample set as the input of the initial deep learning model, use the current forecast results corresponding to the historical radar echo images in the general sample set as the reference output of the initial deep learning model, and use the meta-learning mechanism to train the initial deep learning model to obtain a base model.

[0070] The computer device can use the historical radar echo map in the general sample set as the input of the initial deep learning model, and use the meta-learning mechanism based on the preset loss function to train the initial deep learning model to obtain a base model. When training the initial deep learning model, the computer device can use the nowcast result corresponding to the historical radar echo map in the general sample set as a reference, and through the preset loss function, the nowcast reference result output by the base model is not the nowcast result corresponding to the historical radar echo map, but a value similar to the nowcast result; through training with a large number of samples in the general sample set, a base model for approximating the nowcast model of each region can be obtained.

[0071] The preset loss function may be in different function forms, as long as a base model for approximating the nowcast model of each region can be obtained. Optionally, the preset loss function may be a first loss function, and the first loss function includes a second-order derivative of a model parameter. The computer device may train the initial deep learning model based on the first loss function to obtain a base model. For example, the first loss function may be:

[0072]

[0073] Among them, θ is the model parameter, D T is the Tth sample data, and η is the learning rate.

[0074] Specifically, the computer device can use the same training method for all samples in the general sample set to obtain the base model, or divide the general sample set into different sets and train the initial deep learning model in stages. The above training method is not limited here.

[0075] S203, taking the historical radar echo images in the regional sample set as the input of the base model, taking the nowcast results corresponding to the historical radar echo images in the regional sample set as the reference output of the base model, training the base model, and obtaining the nowcast model.

[0076] Furthermore, the computer device can perform migration training on the base model based on the regional sample set after obtaining the above base model. The computer device can use the historical radar echo map in the regional sample set as the input of the base model, and use the nowcast results corresponding to the historical radar echo map in the regional sample set as the reference output of the base model to train the base model and obtain the nowcast model.

[0077] Similar to the input method of the radar echo map of the target area in S102, before the computer device inputs the historical radar echo map in the general sample set into the initial deep learning model, and before the computer device inputs the historical radar echo map in the regional sample set into the base model, it can also perform noise reduction or interpolation on the above historical radar echo map.

[0078] In the above-mentioned meteorological nowcasting method, the computer equipment adopts a meta-learning mechanism to train the initial deep learning model based on a general sample set to obtain a base model, so that the base model can be quickly migrated and trained using a regional sample set based on the model parameters in the base model, thereby improving the rapid learning ability of the nowcasting model.

[0079] Figure 5 FIG. 1 is a flow chart of a weather nowcasting method in an embodiment. This embodiment relates to a method for a computer device to obtain a base model. Based on the above embodiment, Figure 5 As shown, the above S202 includes:

[0080] S301: Divide a general sample set into a first data set and a second data set.

[0081] The computer device may divide the general sample set into a first data set and a second data set, and train the initial deep learning model in stages. Specifically, the computer device may randomly classify the historical radar echograms in the general sample set to obtain the first data set and the second data set; or, according to training requirements, select some samples from the general sample set to form the second data set, and the above classification method is not limited here.

[0082] When the computer device obtains the nowcast model of the target area, the historical radar echo map and its corresponding nowcast results similar to the weather conditions of the target area in the general sample set can be used as the second data set. For example, the general sample set is the historical radar echo map of area A and its corresponding nowcast results, and the weather conditions of the target area B are similar to those of area A from April to June, both of which are cloudy and rainy; then the computer device can use the historical radar echo map of area A from April to June and its corresponding nowcast results as the second data set, and then use other historical radar echo maps of area A and their corresponding nowcast results as the first data set. The above-mentioned first data set can be a support data set, and the above-mentioned second data set can be a query data set.

[0083] S302: Based on the first data set and the first loss function, train an initial deep learning model to obtain an initial base model.

[0084] Furthermore, the computer device can train the initial deep learning model based on the first data and the above-mentioned first loss function to obtain an initial base model.

[0085] S303: Based on the second data set and the first loss function, perform a first preset number of parameter iterations on the initial base model to obtain a base model.

[0086] On the basis of obtaining the initial base model, the computer device uses the historical radar echo map in the second data set as the input of the initial base model, uses the nowcast result corresponding to the historical radar echo map in the second data set as the reference output, and uses the first loss function to perform a first preset number of parameter iterations on the initial base model to obtain the base model. Optionally, the first preset number is 5 times.

[0087] In the above-mentioned meteorological nowcasting method, the computer device divides the general sample set into a first data set and a second data set, and obtains an initial base model based on the first data set, so that on the basis of the initial base model, the second data set and the first loss function can be used to perform a first preset number of parameter iterations on the initial base model. By limiting the number of parameter iterations, the training time process caused by too many iterations can be avoided, thereby improving the training speed of the base model.

[0088] In one embodiment, when the computer device obtains the nowcast model based on the base model, it can train the base model based on the second loss function to obtain the nowcast model of the target area; wherein the second loss function includes the first-order derivative of the model parameter. The second loss function can be a logarithmic loss function including the first-order derivative, or an absolute value loss function including the first-order derivative, which is not limited here. Optionally, the second loss function is:

[0089]

[0090] Among them, θ * are the parameters of the base model, is the sample data of the sample base in the Tth region, and η is the learning rate.

[0091] The learning rate in the above second loss function may be the same as or different from the learning rate of the first loss function, which is not limited here.

[0092] The computer device performs migration training on the base model based on the second loss function and the regional sample set, and can quickly migrate the parameters of the base model to the parameters of the short-term prediction model corresponding to the target area, so that the short-term prediction model can more accurately obtain the short-term prediction data corresponding to the radar echo map.

[0093] The computer device may perform a second preset number of parameter iterations on the base model based on the second loss function to obtain a nowcast model. The second preset number of iterations may be the same as or different from the first preset number of iterations, which is not limited here. Optionally, the second preset number of iterations may be greater than or equal to 5 and less than or equal to 10.

[0094] In the above-mentioned meteorological nowcasting method, the computer device uses the second loss function to perform migration training on the base model, which can improve the accuracy of the nowcasting model; further, the computer device performs a second preset number of parameter iterations on the base model based on the second loss function, which can avoid the training time process caused by too many iterations by limiting the number of parameter iterations, thereby improving the migration training speed of the nowcasting model.

[0095] It should be understood that although Figure 2-5The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 2-5 At least part of the steps may include multiple steps or multiple stages. These steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed in turn or alternately with other steps or at least part of the steps or stages in other steps.

[0096] In one embodiment, Figure 6 As shown, a weather nowcasting device is provided, comprising: an acquisition module 10 and an input module 20, wherein:

[0097] An acquisition module 10 is used to acquire a radar echo map of a target area;

[0098] The input module 20 is used to input the radar echo map into the nowcasting model to obtain the meteorological nowcasting data of the target area; the nowcasting model is obtained by performing migration training on the base model according to the regional sample set of the target area; the base model is obtained by training with a meta-learning mechanism and a general sample set; the regional sample set and the general sample set both include historical radar echo maps and the nowcasting results corresponding to the historical radar echo maps.

[0099] In one embodiment, based on the above embodiment, Figure 7 As shown, the above device also includes a training module 30, and the training module 30 includes:

[0100] An acquisition unit 301 is used to acquire a general sample set and a regional sample set;

[0101] The first training unit 302 is used to use the historical radar echo images in the general sample set as the input of the initial deep learning model, use the nowcast results corresponding to the historical radar echo images in the general sample set as the reference output of the initial deep learning model, and use the meta-learning mechanism to train the initial deep learning model to obtain a base model;

[0102] The second training unit 303 is used to use the historical radar echo images in the regional sample set as the input of the base model, use the nowcast results corresponding to the historical radar echo images in the regional sample set as the reference output of the base model, train the base model, and obtain the nowcast model.

[0103] In one embodiment, based on the above embodiment, the above first training unit 302 is specifically used to: train the initial deep learning model based on a first loss function to obtain a base model; the first loss function includes the second-order derivative of the model parameters.

[0104] In one embodiment, based on the above embodiment, Figure 8 As shown, the first training unit 302 includes:

[0105] The classification subunit 3021 is used to divide the general sample set into a first data set and a second data set;

[0106] A training subunit 3022 is used to train an initial deep learning model based on the first data set and the first loss function to obtain an initial base model;

[0107] The iterative subunit 3023 is used to perform a first preset number of parameter iterations on the initial base model based on the second data set and the first loss function to obtain a base model.

[0108] In one embodiment, based on the above embodiment, the first preset number of times is 5 times.

[0109] In one embodiment, based on the above embodiment, the first loss function is: Among them, θ is the model parameter, D T is the Tth sample data, and η is the learning rate.

[0110] In one embodiment, based on the above embodiment, the second training unit 303 is specifically used to: train the base model based on a second loss function to obtain a nowcasting model; the second loss function includes a first-order derivative of a model parameter.

[0111] In one embodiment, based on the above embodiment, the second training unit 303 is specifically used to: perform a second preset number of parameter iterations on the base model based on a second loss function to obtain a nowcasting model.

[0112] In one embodiment, based on the above embodiment, the second preset number is greater than or equal to 5 and less than or equal to 10.

[0113] In one embodiment, based on the above embodiment, the second loss function is: Among them, θ * are the parameters of the base model, is the sample data of the sample base in the Tth region, and η is the learning rate.

[0114] In one embodiment, based on the above embodiment, the nowcasting model is a long short-term memory (LSTM) network model.

[0115] The weather nowcasting device provided above can execute the weather nowcasting method embodiment described above, and its implementation principle and technical effect are similar, which will not be described in detail here.

[0116] The specific definition of the meteorological nowcasting device can be found in the above definition of the meteorological nowcasting method, which will not be repeated here. Each module in the above-mentioned meteorological nowcasting device can be implemented in whole or in part by software, hardware and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0117] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Fig. 9 As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store meteorological nowcasting data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a meteorological nowcasting method is implemented.

[0118] Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0119] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0120] Obtain radar echo map of target area;

[0121] The radar echo map is input into the nowcasting model to obtain the meteorological nowcasting data of the target area; the nowcasting model is obtained by transfer training the base model according to the regional sample set of the target area; the base model is obtained by training with a meta-learning mechanism and a general sample set; both the regional sample set and the general sample set include historical radar echo maps and the nowcasting results corresponding to the historical radar echo maps.

[0122] In one embodiment, when the processor executes the computer program, the following steps are also implemented: obtaining a general sample set and a regional sample set; using the historical radar echo map in the general sample set as the input of the initial deep learning model, using the short-term forecast results corresponding to the historical radar echo map in the general sample set as the reference output of the initial deep learning model, and using the meta-learning mechanism to train the initial deep learning model to obtain a base model; using the historical radar echo map in the regional sample set as the input of the base model, using the short-term forecast results corresponding to the historical radar echo map in the regional sample set as the reference output of the base model, training the base model, and obtaining a short-term forecast model.

[0123] In one embodiment, when the processor executes the computer program, the following steps are also implemented: training the initial deep learning model based on the first loss function to obtain a base model; the first loss function includes the second-order derivatives of the model parameters.

[0124] In one embodiment, when the processor executes the computer program, the following steps are also implemented: dividing the general sample set into a first data set and a second data set; training an initial deep learning model based on the first data set and a first loss function to obtain an initial base model; and performing a first preset number of parameter iterations on the initial base model based on the second data set and the first loss function to obtain a base model.

[0125] In one embodiment, the first preset number of times is 5 times.

[0126] In one embodiment, the first loss function is: Among them, θ is the model parameter, D T is the Tth sample data, and η is the learning rate.

[0127] In one embodiment, when the processor executes the computer program, the following steps are also implemented: training the base model based on the second loss function to obtain the nowcasting model; the second loss function includes the first-order derivatives of the model parameters.

[0128] In one embodiment, when the processor executes the computer program, the following steps are further implemented: based on the second loss function, performing a second preset number of parameter iterations on the base model to obtain a nowcasting model.

[0129] In one embodiment, the second preset number is greater than or equal to 5 and less than or equal to 10.

[0130] In one embodiment, the second loss function is: Among them, θ * are the parameters of the base model, is the sample data of the sample base in the Tth region, and η is the learning rate.

[0131] In one embodiment, the nowcasting model is a long short-term memory (LSTM) network model.

[0132] The implementation principle and technical effects of the computer device provided in this embodiment are similar to those of the above method embodiments, and will not be repeated here.

[0133] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0134] Obtain radar echo map of target area;

[0135] The radar echo map is input into the nowcasting model to obtain the meteorological nowcasting data of the target area; the nowcasting model is obtained by transfer training the base model according to the regional sample set of the target area; the base model is obtained by training with a meta-learning mechanism and a general sample set; both the regional sample set and the general sample set include historical radar echo maps and the nowcasting results corresponding to the historical radar echo maps.

[0136] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: obtaining a general sample set and a regional sample set; using the historical radar echo map in the general sample set as the input of the initial deep learning model, using the short-term forecast results corresponding to the historical radar echo map in the general sample set as the reference output of the initial deep learning model, and using the meta-learning mechanism to train the initial deep learning model to obtain a base model; using the historical radar echo map in the regional sample set as the input of the base model, using the short-term forecast results corresponding to the historical radar echo map in the regional sample set as the reference output of the base model, training the base model, and obtaining a short-term forecast model.

[0137] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented: training the initial deep learning model based on a first loss function to obtain a base model; the first loss function includes the second-order derivatives of the model parameters.

[0138] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented: dividing the general sample set into a first data set and a second data set; training an initial deep learning model based on the first data set and a first loss function to obtain an initial base model; and performing a first preset number of parameter iterations on the initial base model based on the second data set and the first loss function to obtain a base model.

[0139] In one embodiment, the first preset number of times is 5 times.

[0140] In one embodiment, the first loss function is: Among them, θ is the model parameter, D T is the Tth sample data, and η is the learning rate.

[0141] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented: training the base model based on a second loss function to obtain a nowcasting model; the second loss function includes the first-order derivatives of the model parameters.

[0142] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: based on the second loss function, performing a second preset number of parameter iterations on the base model to obtain a nowcasting model.

[0143] In one embodiment, the second preset number is greater than or equal to 5 and less than or equal to 10.

[0144] In one embodiment, the second loss function is: Among them, θ * are the parameters of the base model, is the sample data of the sample base in the Tth region, and η is the learning rate.

[0145] In one embodiment, the nowcasting model is a long short-term memory (LSTM) network model.

[0146] The computer storage medium provided in this embodiment has similar implementation principles and technical effects as those of the above method embodiments, and will not be described in detail here.

[0147] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0148] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0149] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A weather nowcasting method, characterized in that: The method comprises: Obtain radar echo map of target area; Performing noise reduction processing or frame insertion processing on the radar echo image to obtain a processed radar echo image; Inputting the processed radar echo map into a nowcasting model to obtain meteorological nowcasting data of the target area; the nowcasting model is obtained by performing transfer training on a base model according to a regional sample set of the target area; the base model is obtained by adopting a meta-learning mechanism and training with a general sample set; the regional sample set and the general sample set both include historical radar echo maps and nowcasting results corresponding to the historical radar echo maps; The method further comprises: Obtaining universal sample sets and regional sample sets; The historical radar echo map in the general sample set is used as the input of the initial deep learning model, and the nowcast result corresponding to the historical radar echo map in the general sample set is used as the reference output of the initial deep learning model. The general sample set is divided into a first data set and a second data set by adopting a meta-learning mechanism; based on the first data set and a first loss function, the initial deep learning model is trained to obtain an initial base model; based on the second data set and the first loss function, the initial base model is iterated for a first preset number of parameters to obtain the base model; the first loss function contains the second-order derivatives of the model parameters; Taking the historical radar echo map in the regional sample set as the input of the base model, taking the nowcast result corresponding to the historical radar echo map in the regional sample set as the reference output of the base model, training the base model, and obtaining the nowcast model; Among them, the general sample set includes historical radar echo maps of multiple areas; the second data set includes historical radar echo maps in the general sample set that are similar to the weather conditions of the target area and their corresponding nowcast results; the first data set includes historical radar echo maps in the general sample set that are not similar to the weather conditions of the target area and their corresponding nowcast results.

2. The weather nowcasting method according to claim 1, characterized in that: The first loss function is: Among them, θ is the model parameter, D T is the Tth sample data, and η is the learning rate.

3. The weather nowcasting method according to claim 1, characterized in that: The step of training the base model to obtain the nowcast model includes: The base model is trained based on a second loss function to obtain the nowcasting model; the second loss function includes the first-order derivatives of the model parameters.

4. The weather nowcasting method according to claim 3, characterized in that: The step of training the base model based on the second loss function to obtain the nowcasting model includes: Based on the second loss function, a second preset number of parameter iterations are performed on the base model to obtain the nowcasting model.

5. The weather nowcasting method according to claim 3 or 4, characterized in that: The second loss function is: Among them, θ * are the parameters of the base model, is the sample data of the sample base in the Tth region, and η is the learning rate.

6. A weather nowcasting device, characterized in that: The device comprises: An acquisition module is used to acquire a radar echo map of a target area; Performing noise reduction processing or frame insertion processing on the radar echo image to obtain a processed radar echo image; An input module is used to input the processed radar echo map into a nowcasting model to obtain the meteorological nowcasting data of the target area; the nowcasting model is obtained by performing migration training on a base model according to a regional sample set of the target area; the base model is obtained by adopting a meta-learning mechanism and training with a general sample set; the regional sample set and the general sample set both include historical radar echo maps and the nowcasting results corresponding to the historical radar echo maps; A training module, used to obtain a general sample set and a regional sample set; use the historical radar echo map in the general sample set as the input of the initial deep learning model, use the current forecast results corresponding to the historical radar echo map in the general sample set as the reference output of the initial deep learning model, and adopt a meta-learning mechanism to divide the general sample set into a first data set and a second data set; based on the first data set and a first loss function, train the initial deep learning model to obtain an initial base model; based on the second data set and the first loss function, perform a first preset number of parameter iterations on the initial base model to obtain the base model; the first loss function contains the second-order derivatives of the model parameters; Taking the historical radar echo map in the regional sample set as the input of the base model, taking the nowcast result corresponding to the historical radar echo map in the regional sample set as the reference output of the base model, training the base model, and obtaining the nowcast model; Among them, the general sample set includes historical radar echo maps of multiple areas; the second data set includes historical radar echo maps in the general sample set that are similar to the weather conditions of the target area and their corresponding nowcast results; the first data set includes historical radar echo maps in the general sample set that are not similar to the weather conditions of the target area and their corresponding nowcast results.

7. The device according to claim 6, characterized in that The first loss function is: Among them, θ is the model parameter, D T is the Tth sample data, and η is the learning rate.

8. The device according to claim 6, characterized in that The training module is also used to: The base model is trained based on the second loss function to obtain the nowcasting model; the second loss function includes the first-order derivatives of the model parameters.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images