WO2021080107A1

WO2021080107A1 - Learning method and testing method for generating high-resolution weather and climate data, and testing method and testing apparatus using same

Info

Publication number: WO2021080107A1
Application number: PCT/KR2020/006489
Authority: WO
Inventors: 임효혁; 하지훈
Original assignee: (주)한국해양기상기술
Priority date: 2019-10-22
Filing date: 2020-05-18
Publication date: 2021-04-29
Also published as: KR102063358B1

Abstract

According to the present invention, provided is a learning method for generating high-resolution weather and climate data, the learning method comprising the steps of: (a) for each of specific grid points from among grid points corresponding to respective points at which a plurality of first axes parallel to a first direction and a plurality of second axes parallel to a second direction intersect, when first past forecast data as a value corresponding to weather and climate at each of t-mth past time points is acquired by a predetermined first forecast model, performing learning on a first deep learning module; (b) for each of the specific grid points, receiving at least a part of first specific past forecast data corresponding to a specific past time point and specific future forecast data corresponding to a specific future time point of at least a part of future time points so as to, as a result of performing a first deep learning operation, output, as second deep learning input forecast data, a value corresponding to specific weather and climate for each of the specific grid points and for each of at least a part of the specific past time point and the specific future time point; and (c) performing learning on a second deep learning module while performing a predetermined pre-processing on second specific deep learning input forecast data and generating reduced data reduced to a predetermined first magnification.

Description

Learning method and test method for generating high-resolution weather and climate data, test method and test apparatus using the same

The present invention relates to a learning method and a test method for generating high-resolution meteorological climate data, a learning device and a testing device using the same.

Along with the recent development of computing performance, technology that extracts and utilizes useful information through big data analysis is in the spotlight. One of the fields in which such big data analysis can be useful is the meteorological and climate forecasting field, and many studies are being conducted to effectively perform meteorological and climate forecasting by analyzing various and vast meteorological and climate data. The development and performance improvement of several numerical forecast models for forecasting are continuing.

However, since these numerical forecast models require high-performance computing capabilities, predictions are performed on a wide range of grid intervals, depending on the performance of the computing device, and thus respond to the demand for weather forecast information in a detailed area above the grid interval. There is a limit to this. In addition, there is a problem in that it takes a lot of cost and time to improve the performance so that the existing numerical forecast model can perform a forecast for a more detailed range, or to develop a new numerical forecast model.

According to the article posted on the Internet blog (“How to Increase the Accuracy of Weather Forecasting”, JW, https://brunch.co.kr/@bestbell/99), the numerical forecast model divides the geographic area into a certain grid. As a model for performing a computing operation on the properties of air in a grid region, it is described that the grid spacing needs to be small in order to increase accuracy, but there is a problem in that the amount of computing calculation and the calculation time increase accordingly.

Accordingly, there is a need for a method of obtaining weather and climate forecast data at more precise intervals by using forecast data that can be obtained by an existing forecast model.

It is an object of the present invention to solve all of the above-described problems.

In addition, another object of the present invention is to provide a method for generating high-resolution forecast data, thereby satisfying the demand for forecasting meteorological climate for an empty spot where the weather forecast data was not provided.

In addition, another object of the present invention is to provide a method of generating high-resolution forecast data using a deep learning algorithm to provide objective and high-resolution forecast data with high accuracy.

A characteristic configuration of the present invention for achieving the object of the present invention as described above and realizing the characteristic effects of the present invention to be described later is as follows.

According to an aspect of the present invention, as a learning method for generating high-resolution weather and weather data, (a) a point where a plurality of first axes parallel to a first direction and a plurality of second axes parallel to a second direction intersect For each of the grid points corresponding to each of the specific grid points included in the predetermined forecast region, the t-1, t-2, ..., tm th as a value corresponding to the meteorological climate of each past time point. 1 When the past forecast data is acquired by a predetermined first forecast model, the learning server (i) causes the first deep learning module to generate specific first past forecast data corresponding to at least some of the past time points. As a result of receiving input as first learning data and performing a predetermined first deep learning operation, predicting a value corresponding to at least one of the meteorological climates for each of the specific grid points and each of the specific past time points. And, (ii) the value corresponding to the predicted specific meteorological climate, the first GT (Ground Truth) re-analysis field data-The re-analysis field data, for each of the specific grid points and the specific past time point, 2 It is data that is corrected by referring to actual observation data collected separately from the second historical forecast data as a value corresponding to the specific weather climate acquired by the forecast model-Compared with, learning about the first deep learning module Allowing to perform; (b) In a state in which learning about the first deep learning module is completed, for each of the specific grid points, (i) each of the t-1, t-2, ..., tm-th past views The first past forecast data as a value corresponding to the meteorological climate and (ii) the future forecast as a value corresponding to the meteorological climate at each of the t+1, t+2, ..., t+nth future times When at least some of the data is acquired by the first forecast model, the learning server causes the first deep learning module to cause the specific first past forecast data and As a result of performing the first deep learning operation by receiving at least some of the specific future forecast data corresponding to the specific future time of at least some of the future time points, (i) each of the specific grid points and (ii) the specific past time point And outputting a value corresponding to the specific weather climate for at least some of the specific future time points as second deep learning input forecast data. And (c) performing a predetermined pre-processing on at least a part of the specific second deep learning input forecast data among the output second deep learning input forecast data to generate reduced data reduced by a predetermined first magnification, The learning server, (i) the second deep learning module receives the reduced data as second learning data, and as a result of applying a predetermined second deep learning operation to the reduced data, the expansion is enlarged by a second ratio. To calculate data, and (ii) to perform learning on the second deep learning module by comparing the enlarged data with the specific second deep learning input forecast data as a second GT (Ground Truth) corresponding thereto. step; Including, a learning method is provided.

As an example, in the step (a), the learning server causes the first deep learning module to additionally receive at least one of ground feature data and altitude feature data corresponding to each of the specific grid points to perform learning. However, in the case where the specific past time point is a plurality of time points, the server is equally as the number of the specific past time points so that at least one of the ground characteristic data and the altitude data corresponds to each of the specific past time points. At least one of the repeated first deep learning input ground feature data and the first deep learning input advanced feature data are additionally generated, and the first deep learning module receives the additional input to perform learning. To do, a learning method is provided.

As an example, the ground feature data has a preset category value for each ground feature corresponding to each of the specific grid points, and in the learning process of the first deep learning module, each of the category values is each of the specific grid points. A learning method is provided, characterized in that it is converted into a predetermined dimensional vector value corresponding to and used for learning of the first deep learning module.

As an example, the re-analysis field data is, for each of the specific grid points (i) for each of the first partial specific grid points in which the actual observation data exists within a predetermined range, referring to the actual observation data, the first 1 A value corresponding to the specific weather climate of each specific grid point is determined and corrected, and (ii) for each of the second partial specific grid points in which the actual observation data does not exist within a predetermined range, the second forecast model With reference to the value corresponding to the specific meteorological climate predicted by and at least one actual observation data measured at a position close to the second partial specific grid point, the specific weather for each of the second partial specific grid points A learning method is provided, characterized in that a value corresponding to the climate is determined and corrected.

As an example, in the step (c), the preprocessing process for generating the reduced data used as the learning data of the second deep learning module, the learning server, with respect to the specific second deep learning input forecast data A learning method is provided, characterized in that it is performed by additionally performing at least one of a blur process and noise addition.

As an example, in the step (c), the preprocessing process for generating the reduced data used as the learning data of the second deep learning module, the learning server, with respect to the specific second deep learning input forecast data A learning method is provided, characterized in that a process of deleting or adding a part of a specific row or a specific column is additionally performed so that the number of horizontal or vertical pixels becomes an integer value corresponding to a multiple of the second magnification.

As an example, a learning method is provided, characterized in that in step (c), the server generates the reduced data using a predetermined interpolation method.

In addition, according to another aspect of the present invention, as a test method for generating high-resolution weather and weather data, (a) by a learning server, (1) a plurality of first axes parallel to the first direction and parallel to the second direction. For each specific grid point included in a predetermined forecast region among grid points corresponding to each of the points where a plurality of second axis lines intersect, t-1, t-2, ..., tm past views When the first historical forecast data for learning as a value corresponding to each meteorological climate is acquired by a predetermined first forecast model, (i) the first deep learning module causes the first deep learning module to correspond to at least some of the specific past time points. As a result of receiving specific first past forecast data for learning as first learning data, and performing a predetermined first deep learning operation, a value corresponding to at least one specific meteorological climate among the meteorological climates is determined, respectively, and the specific grid points. Each of the specific past time points is predicted, and (ii) the value corresponding to the predicted specific weather climate is calculated as the first GT (Ground Truth) for learning re-analysis field data-The learning re-analysis field data is each of the specific grid points. And, for each of the specific past time points, the second historical forecast data for learning as a value corresponding to the specific weather climate obtained by the second forecast model is corrected with reference to actual observation data collected separately-and comparison with Thus, the learning for the first deep learning module is performed, and (2) in a state in which the learning for the first deep learning module is completed, for each of the specific grid points, (i) the t-1, The first past forecast data for learning as values corresponding to the weather climate at each of the t-2, ..., and tm-th past times, and (ii) the t+1, t+2, ..., and When at least some of the future forecast data for learning as a value corresponding to the weather climate at each of the t+n future points is acquired by the first forecast model, the first deep learning module causes the first deep learning module to specify the at least some of the past points. The learning specific first corresponding to the past point of view As a result of performing the first deep learning operation by receiving at least some of the past forecast data and the specific future forecast data for learning corresponding to at least some of the future time points and performing the first deep learning operation, (i) each of the specific grid points and (ii) ) Outputting the value corresponding to the specific weather climate for at least some of the specific past time point and the specific future time point as input forecast data for second deep learning learning, and (3) the outputted second input for deep learning learning In a state in which at least some of the forecast data is generated by performing a predetermined pre-processing on the input forecast data for specific second deep learning learning, and generating the reduced learning data reduced to a predetermined first magnification, (i) the second deep learning module It causes the learning reduction data to be input as second learning data, and to calculate the learning expansion data enlarged at a second ratio as a result of applying a predetermined second deep learning operation to the learning reduction data, and (ii) the learning expansion By comparing the data with the input prediction data for the specific second deep learning learning as a corresponding second GT (Ground Truth) to perform learning on the second deep learning module, the first deep learning module and the first 2 In a state in which learning for each of the deep learning modules is completed, the test server causes the first deep learning module to test as a value corresponding to the meteorological climate at each of the specific past points obtained by the first forecast model. As a result of performing the first deep learning operation by receiving at least some of the specific first past forecast data and the specific future forecast data for testing as a value corresponding to the weather climate at each of the specific future time points, (i) the specific grid Outputting a value corresponding to the specific weather climate for each of the points and (ii) at least some of the specific past time point and the specific future time point as input forecast data for a second deep learning test; And (b) the test server receives input forecast data for at least some of the input forecast data for the second deep learning test by the second deep learning module, and receives the second deep learning. Obtaining the output high-resolution forecast data by outputting the high-resolution forecast data expanded by the second magnification of the input forecast data for the specific second deep learning test as a result of performing the operation; Including, a test method is provided.

As an example, the test server may determine a value corresponding to the specific weather climate for at least some of the high resolution grid points corresponding to the high resolution forecast data, corresponding to the specific high resolution grid points among the actual observation data. A test method is provided, characterized in that it is compared with specific actual observation data measured at a location, and when the difference is greater than or equal to a threshold, it is determined that additional learning is necessary, and the additional learning is supported.

As an example, the test server inputs a first error value obtained by comparing the specific first historical forecast data with the reanalyzing field data and the first historical forecast data into the first deep learning module to provide the first deep learning. By allowing the module to perform the first deep learning operation, the first forecast model and the second error value obtained by comparing the outputted second deep learning test input forecast data with the reanalysis field data are respectively referenced. 1 A test method is provided, which supports to determine the performance of the deep learning module, but supports to determine a higher performance model as the error value decreases.

In addition, according to another aspect of the present invention, there is provided a learning server for generating high-resolution weather and weather data, comprising: at least one memory for storing instructions; And at least one processor configured to execute the instructions. Including, wherein the processor, (I) a predetermined forecast region among grid points corresponding to points at which a plurality of first axes parallel to a first direction and a plurality of second axes parallel to a second direction intersect each other. For each of the specific grid points included in, the first historical forecast data as a value corresponding to the meteorological climate at each of the t-1, t-2, ..., and tmth past points is transferred to a predetermined first forecast model. When obtained by, (i) the first deep learning module receives specific first past forecast data corresponding to at least some of the past time points as first learning data, and performs a predetermined first deep learning operation. As a result of the execution, a value corresponding to at least one specific weather climate among the meteorological climates is predicted for each of the specific grid points and each of the specific past time points, and (ii) a value corresponding to the predicted specific weather climate is calculated. , Re-analysis field data as a first GT (Ground Truth)-The re-analysis field data is a value corresponding to the specific weather climate obtained by a second forecast model for each of the specific grid points and each of the specific past time points. A process of performing learning on the first deep learning module by comparing the second historical forecast data with the corrected data with reference to actual observation data collected separately; (II) In a state in which learning about the first deep learning module is completed, for each of the specific grid points, (i) each of the t-1, t-2, ..., tm-th past views The first past forecast data as a value corresponding to the meteorological climate and (ii) the future forecast as a value corresponding to the meteorological climate at each of the t+1, t+2, ..., t+nth future times When at least some of the data is acquired by the first forecast model, the learning server causes the first deep learning module to cause the specific first past forecast data and As a result of performing the first deep learning operation by receiving at least some of the specific future forecast data corresponding to the specific future time of at least some of the future time points, (i) each of the specific grid points and (ii) the specific past time point And a process of outputting a value corresponding to the specific weather climate for at least some of the specific future time points as second deep learning input forecast data. And (III) performing a predetermined pre-processing on at least some of the output second deep learning input forecast data on at least a portion of the second deep learning input forecast data to generate reduced data reduced by a predetermined first magnification, The learning server, (i) the second deep learning module receives the reduced data as second learning data, and as a result of applying a predetermined second deep learning operation to the reduced data, the expansion is enlarged by a second ratio. To calculate data, and (ii) to perform learning on the second deep learning module by comparing the enlarged data with the specific second deep learning input forecast data as a second GT (Ground Truth) corresponding thereto. process; To perform, a learning server is provided.

As an example, in the process (I), the processor causes the first deep learning module to additionally receive at least one of ground feature data and altitude feature data corresponding to each of the specific grid points to perform learning. However, if the specific past time point is a plurality of time points, the processor repeats the same as the number of the specific past time points so that at least one of the ground characteristic data and the altitude data corresponds to each of the specific past time points. At least one of ground feature data for first deep learning input and advanced feature data for first deep learning input are additionally generated, and the first deep learning module receives additional input to perform learning. That is, a learning server is provided.

As an example, the ground feature data has a preset category value for each ground feature corresponding to each of the specific grid points, and in the learning process of the first deep learning module, each of the category values is each of the specific grid points. A learning server is provided, characterized in that it is converted into a predetermined dimensional vector value corresponding to and used for learning of the first deep learning module.

As an example, the re-analysis field data is, for each of the specific grid points (i) for each of the first partial specific grid points in which the actual observation data exists within a predetermined range, referring to the actual observation data, the first 1 A value corresponding to the specific weather climate of each specific grid point is determined and corrected, and (ii) for each of the second partial specific grid points in which the actual observation data does not exist within a predetermined range, the second forecast model With reference to the value corresponding to the specific meteorological climate predicted by and at least one actual observation data measured at a position close to the second partial specific grid point, the specific weather for each of the second partial specific grid points A learning server is provided, characterized in that a value corresponding to the climate is determined and corrected.

As an example, in the (III) process, the preprocessing process for generating the reduced data used as the learning data of the second deep learning module, the processor, blurs the specific second deep learning input forecast data. A learning server is provided, characterized in that it is performed by additionally performing at least one process of (blur) processing and noise addition.

As an example, in the (III) process, the pre-processing process for generating the reduced data used as the learning data of the second deep learning module, the processor, horizontally with respect to the specific second deep learning input forecast data. Alternatively, a learning server is provided, characterized in that a process of deleting or adding a part of a specific row or a specific column is additionally performed so that the number of vertical pixels becomes an integer value corresponding to a multiple of the second magnification.

As an example, in the process (III), a learning server is provided, characterized in that the processor generates the reduced data using a predetermined interpolation method.

In addition, according to another aspect of the present invention, a test method for generating high-resolution weather and weather data, comprising: at least one memory for storing instructions; And at least one processor configured to execute the instructions. Including, the processor, (I) by the learning server, (1) a plurality of first axis parallel to the first direction and a plurality of second axis parallel to the second direction corresponding to each of the intersection point The first past forecast for learning as a value corresponding to the meteorological climate of each of the t-1, t-2, ..., tm past points for each of the specific grid points included in the predetermined forecast area among the grid points When data is acquired by a predetermined first forecast model, (i) the first deep learning module inputs the first specific first forecast data for learning corresponding to at least some of the specific past time points as the first learning data. In response, as a result of performing a predetermined first deep learning operation, a value corresponding to at least one specific weather climate among the meteorological climates is predicted for each of the specific grid points and each of the specific past time points, and (ii) the The value corresponding to the predicted specific weather climate is calculated as the first GT (Ground Truth) for learning re-analysis field data-The training re-analysis field data is for each of the specific grid points and the specific past point in a second forecast model. It is data that is corrected by referring to actual observation data collected separately from the second historical forecast data for learning as a value corresponding to the specific meteorological climate acquired by-to perform learning on the first deep learning module by comparing with And (2) in a state in which the learning of the first deep learning module is completed, for each of the specific grid points, (i) each of the t-1, t-2, ..., and tm-th past views The first past forecast data for learning as a value corresponding to the meteorological climate of and (ii) a value corresponding to the meteorological climate at each of the t+1, t+2, ..., t+nth future times When at least a part of the future forecast data for learning is acquired by the first forecast model, the first deep learning module causes the learning specific first past forecast data corresponding to the specific past time and At least some of the above future time points As a result of performing the first deep learning operation by receiving at least some of the corresponding specific future forecast data for learning, (i) each of the specific grid points and (ii) at least some of the specific past time and the specific future time The value corresponding to the specific weather climate for is output as second deep learning input forecast data, and (3) at least some of the output second deep learning input forecast data for second deep learning learning input forecast data In a state in which the reduced data for learning reduced by a predetermined first magnification is generated by performing a predetermined preprocessing on the (i) the second deep learning module receives the reduced data for learning as the second learning data, and the As a result of applying a predetermined second deep learning operation to the reduced data for learning, the expanded data for learning expanded at a second ratio is calculated, and (ii) the expanded data for learning is referred to as the second GT (Ground Truth) corresponding thereto. In a state in which learning for each of the first deep learning module and the second deep learning module is completed by comparing it with the input forecast data for specific second deep learning learning, the test server A, the first deep learning module allows the specific first past forecast data for testing as a value corresponding to the meteorological climate of each of the specific past times acquired by the first forecast model and the weather of each of the specific future times. As a result of performing the first deep learning operation by receiving at least some of the specific future forecast data for a test as a value corresponding to the climate, (i) each of the specific grid points and (ii) the specific past time and the specific future time A process of outputting a value corresponding to the specific weather climate for each of at least some of the second deep learning test input forecast data; And (II) the test server receives input forecast data for at least some of the input forecast data for the second deep learning test by the second deep learning module, and receives the second deep learning data. A process of obtaining the output high-resolution forecast data by outputting the high-resolution forecast data expanded by the second magnification of the input forecast data for the specific second deep learning test as a result of performing the operation; To perform, a test server is provided.

As an example, the processor may determine a value corresponding to the specific weather climate for at least some of the high-resolution grid points corresponding to the high-resolution forecast data, and a position corresponding to the specific high-resolution grid point in the actual observation data. Compared with the specific actual observation data measured in, if the difference is greater than or equal to a threshold, it is determined that additional learning is necessary, and a test server is provided, characterized in that supporting the additional learning to be performed.

As an example, the processor inputs a first error value obtained by comparing the specific first past forecast data with the reanalysis field data, and the first past forecast data into the first deep learning module, and the first deep learning module The first prediction model and the first predicted model and the first predicted model and the first predicted model and the first predicted model and the first predicted model and the first predicted model and the first predicted model and the first A test server is provided, which supports to determine the performance of the deep learning module, but supports to determine a higher performance model as the error value decreases.

According to the present invention, the following effects are obtained.

The present invention provides a method of generating high-resolution forecast data based on the existing forecast data, thereby complementing the existing forecast model to perform efficient weather climate prediction.

1 is a diagram showing a schematic configuration of a learning server for generating high-resolution weather and climate data according to an embodiment of the present invention.

2 is a diagram schematically illustrating a flow of data used in a learning and testing process for generating high-resolution weather climate data according to an embodiment of the present invention.

3 is an exemplary diagram of data input and output to a first deep learning module in a learning and testing process for generating high-resolution weather climate data according to an embodiment of the present invention.

4A is an exemplary diagram of data input and output to a second deep learning module in a learning process for generating high-resolution weather climate data according to an embodiment of the present invention.

4B is an exemplary diagram of data input and output to a second deep learning module in a test process for generating high-resolution weather climate data according to an embodiment of the present invention.

5 is a flowchart schematically illustrating a procedure in which a learning server for generating high-resolution weather climate data performs learning according to an embodiment of the present invention.

6 is a flowchart schematically illustrating a procedure in which a test server for generating high-resolution weather climate data performs a test according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The detailed description of the present invention described below refers to the accompanying drawings, which illustrate specific embodiments in which the present invention may be practiced. These embodiments are described in detail sufficient to enable a person skilled in the art to practice the present invention. It should be understood that the various embodiments of the present invention are different from each other, but need not be mutually exclusive. For example, specific shapes, structures, and characteristics described herein may be implemented in other embodiments without departing from the spirit and scope of the present invention in relation to one embodiment. In addition, it should be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the present invention. Accordingly, the detailed description to be described below is not intended to be taken in a limiting sense, and the scope of the present invention, if appropriately described, is limited only by the appended claims, along with all ranges equivalent to those claimed by the claims. Like reference numerals in the drawings refer to the same or similar functions over several aspects.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings in order to enable those of ordinary skill in the art to easily implement the present invention.

Referring to FIG. 1, a learning server 100 for generating high-resolution weather climate data may include a memory 110 and a processor 120. At this time, the memory 110 may store instructions of the processor 120, specifically, the instructions are codes generated for the purpose of causing the learning server 100 to function in a specific manner, and a computer or other programmable It may be stored in computer usable or computer readable memory that may be directed to data processing equipment. Instructions may perform processes for performing the functions described in the specification of the present invention.

In addition, the processor 120 of the learning server 100 may include a hardware configuration such as a micro processing unit (MPU) or a central processing unit (CPU), a cache memory, and a data bus. In addition, it may further include an operating system and a software configuration of an application that performs a specific purpose.

In addition, the learning server 100 includes a first deep learning module 210 and a second deep learning module 220 and may be interlocked with each of them. Learning for each of the 1 deep learning module 210 and the second deep learning module 220 may be performed in advance.

And, when the learning of each of the first deep learning module 210 and the second deep learning module 220 is completed, a test process may be performed by a test server for generating high-resolution weather climate data. In this case, the test server May be the same server as the learning server 100, or may be a separate server including the first deep learning module 210 and the second deep learning module 220 in which the learning has been completed, which is subject to the implementation conditions of the invention. It can be done differently. In addition, if the test server is a separate server different from the learning server 100, the corresponding test server may be configured in the same manner as the learning server 100 as shown in FIG. 1.

Details of the learning process and the test process will be described again with reference to separate drawings below.

Next, the learning server 100 or the test server may be interlocked with a database (not shown) including data used in a learning and testing process for generating high-resolution weather climate data, where the database is a flash memory type (flash memory type), hard disk type, multimedia card micro type, card type memory (e.g. SD or XD memory), RAM (Random Access Memory, RAM), SRAM (Static Random Access Memory), ROM (ReadOnly Memory, ROM), EEPROM (Electrically Erasable Programmable ReadOnly Memory), PROM (Programmable ReadOnly Memory), magnetic memory, magnetic disk, can include at least one type of storage medium of optical disk However, the present invention is not limited thereto and may include any medium capable of storing data. In addition, the database may be installed separately from the learning server 100 or the test server, or alternatively, it may be installed inside the learning server 100 or the test server to transmit data or record received data. It may be implemented separately in two or more, which may vary depending on the operating conditions of the invention.

In addition, the learning server 100 or the test server, among the data used in the learning and testing process for generating high-resolution meteorological and climate data, for data that can be obtained from a separate meteorological and climate data provider such as the Meteorological Agency, such as the Internet. Corresponding data may be transmitted and obtained from the separate meteorological and climate data provider through an external network, and this may also be configured differently according to the implementation conditions of the present invention.

Referring to FIG. 2, first, the learning of the first deep learning module 210 during the learning process for generating high-resolution weather and climate data is, among the first past forecast data 11 obtained from the first forecast model 10. At least a portion of the specific first past forecast data 11-1 corresponding to a specific past time point may be input to the first deep learning module 210 as first learning data. In this case, the first historical forecast data 11 is determined from among grid points corresponding to points where a plurality of first axis lines parallel to a first direction and a plurality of second axis lines parallel to a second direction cross each other. For each of the specific grid points included in the forecast area, it may be a value corresponding to the weather climate of each of the t-1, t-2, ..., tm past times, and the entire past time according to the implementation conditions of the invention. At least some of the past views may be selected as the specific past views. In addition, the meteorological climate may include a plurality of characteristics such as temperature, humidity, atmospheric pressure, wind vector, visibility, cumulative precipitation for a predetermined time, and cumulative snowfall for a predetermined time for each of the specific grid points, The present invention is not limited thereto, and characteristics different from the above characteristics may be selected and used according to the implementation conditions of the invention. In addition, the first axis and the second axis may have the same spacing, or the spacing may be variably determined according to the characteristics of the forecast area. For example, the neighborhood forecast model used by the Meteorological Agency may be the first forecast model 10, and the forecast data predicted from the corresponding neighborhood forecast model may be used as the first past forecast data 11, and the corresponding neighborhood forecast data May be a value corresponding to the meteorological climate of each specific grid point included in a predetermined forecast area among all grid points consisting of a first axis line at 5 km intervals parallel to the latitude line and a second axis line at 5 km intervals parallel to the longitude line. .

In addition, the reanalysis field data 22 together with the first past forecast data 11 may be input to the first deep learning module 210 as a first GT (Ground Truth). At this time, the re-analysis field data 22 is obtained by the second forecast model 20 when the second past forecast data as a value corresponding to a specific weather climate for each specific grid point and each specific past time point is obtained, a specific grid It may be data corrected according to a predetermined criterion with reference to actual observation data 30 corresponding to at least some of the points. In this case, the second forecast model 20 may be used by selecting a more accurate numerical forecast model than the first forecast model 10 according to the implementation conditions of the invention. ) Is evaluated to be more accurate than the light bulb model (UM) currently in use by the Meteorological Agency, and thus, ECMWF may be selected and used as the second forecast model 20.

In addition, the actual observation data 30 is data measured and collected from a plurality of meteorological and climate data measuring devices installed in a predetermined forecast area, and can correspond only to some of each of the specific grid points. Depending on whether the actual observation data 30 exists within the range, the reanalyzing field data 22 may be generated by differently correcting each value corresponding to a specific weather climate for each specific grid point. For example, the reanalysis field data 22 is for each specific grid point (i) for each of the first partial specific grid points in which the actual observation data 30 exists within a predetermined range, the actual observation data 30 With reference to, a value corresponding to a specific weather climate for each of the first specific grid points is determined and corrected, and (ii) for each of the second specific grid points in which the actual observation data 30 does not exist within a predetermined range , With reference to the value corresponding to the specific weather climate predicted by the second forecast model 20 and the at least one actual observation data measured at a position close to the second partial specific grid point, each of the second partial specific grid points A value corresponding to the specific meteorological climate for is determined and corrected. In this case, when there are a plurality of actual observation data referenced for correction for each of the first specific grid points or each of the second partial specific grid points, only one actual observation data closest to each of the corresponding grid points may be referenced. , Actual observation data corresponding to a predetermined number may be referenced together in close order, and this may be different according to the implementation conditions of the invention.

In addition, the specific meteorological climate is at least one specific characteristic of a plurality of characteristics corresponding to the meteorological climate, for example, may be one characteristic such as'temperature', or a plurality of'temperature' and'humidity'. The high-resolution forecast data 221 obtained as a result of implementing the present invention may be data on a value corresponding to the specific weather climate.

In addition, as an example of the invention, at least one of the ground feature data 41 and the elevation feature data 42 corresponding to each specific grid point may be additionally input to the first deep learning module to perform learning. At this time, each of the ground characteristic data 41 and the altitude characteristic data 42 is data having a constant value that is not affected by the passage of time, and in particular, the ground characteristic data 41 indicates the ground characteristic of each of the specific grid points. It is data expressed by a predetermined category value, and is data having different properties from a value corresponding to the meteorological climate. Therefore, by additionally inputting at least one of the ground characteristic data 41 and the altitude characteristic data 42 as separate input data together with the specific first past forecast data 11-1 to the first deep learning module, each At least one of ground characteristics and altitude characteristics along with the meteorological and climate characteristics by generating a result value of the same size by performing a predetermined first deep learning operation on and combining them and performing a predetermined first deep learning operation again. Learning of the first deep learning module using data to which the characteristics of is additionally reflected may be performed.

In this case, when a specific past time point is a plurality of time points, the learning server 100 is configured as many as the number of specific past time points so that at least one of the ground feature data 41 and the high feature data 42 corresponds to each particular past time point. At least one of ground feature data for the first deep learning input and altitude feature data for the first deep learning input that are repeated identically is additionally generated, and the first deep learning module 210 receives this additionally to perform learning. Can be done.

In addition, the ground characteristic data 41 has a preset category value for each ground characteristic corresponding to each specific grid point, but in the learning process of the first deep learning module 210, each of the category values is a specific grid point. It is converted into a predetermined dimensional vector value corresponding to each, and can be used for learning of the first deep learning module 210.

As described above, the data input for learning by the first deep learning module 210 will be described in detail with reference to FIG. 3 as follows.

Referring to FIG. 3, a specific grid point corresponding to a predetermined forecast area may be a grid point having a size of (149, 253), and for each of the corresponding grid points,'barometric pressure','humidity', and each of 28 specific past points. Data (310-1, 310-2, ...) of the size of (149,253, 28) including values corresponding to each of the meteorological and climate characteristics such as'temperature' and'wind vector' is a specific first past forecast Each may be acquired as data 11-1, and the learning server 100 refers to the acquired data 310-1, 310-2, ... ) Can be created and used as the first learning data.

In addition, in the case of the high characteristic data 42, it may be data of a size of (149,253) having a constant value according to a change in time, and the learning server 100 may repeat (149,253,28) constantly as many as the number of specific past points. ) Sized data can be generated as the first deep learning input altitude feature data 320 and used for learning of the first deep learning module 210.

And, in the case of the ground characteristic data 41, a value corresponding to each specific grid point is a predetermined category value, and the learning server 100 performs an embedding operation that converts each of the values into a predetermined dimension. 3, an embedding operation for converting each of the category values included in the ground characteristic data 41 of the size (149,253) into 15 dimensions is performed to first dip the data of the size (149,253,28x15). It can be generated as ground feature data 330 for learning input and used for learning of the first deep learning module 210. For example, as illustrated in FIG. 3, the embedding operation may be performed on the category value 41-1 of '1' to convert it into 15 dimensional vector values 330-1. At this time, the learning server 100 may include a separate module for the embedding operation, or allow a separate device to perform the embedding operation and obtain the result, which is subject to the implementation conditions of the invention. It can be done differently.

As described above, since data input for learning of the first deep learning module have different sizes and properties, a process of combining the data into one data may be additionally performed. A first deep learning module causes the inputted first learning data, the first deep learning input ground feature data 330, and the first deep learning input altitude feature data 320 to receive a predetermined first for each of at least one of the first learning data inputted by the first deep learning module. 1 Perform a deep learning operation to output data of the same size (340-1, 340-2, 340-3), and combine each of the output result values in a specific dimensional direction to obtain one data 350. The result data 360 may be output from the first deep learning module by generating and performing a predetermined first deep learning operation again. In this case, the result data 360 may be a predicted value corresponding to a specific weather climate for each specific grid point and each specific past time point.

In addition, as the data input to the first deep learning module are three-dimensional spatiotemporal data, the deep learning algorithm used in the first deep learning module may be a 3D-CNN algorithm, and in this case, the first deep learning module The first deep learning operation performed on input data may include a convolution operation using a predetermined filter, a feature map generation, and pooling, etc. This is not limited thereto, and various deep learning algorithms may be used according to the implementation conditions of the invention, and accordingly, details of the first deep learning operation and the learning method for the first deep learning module may be applied differently.

Next, when the first deep learning module 210 predicts a value corresponding to a specific weather climate for each of a specific grid point and each of a specific past time point, the learning server 100 compares it with the reanalyzed field data as the first GT. Thus, learning of the first deep learning module 210 may be performed. In this case, the first deep learning module 210 may acquire an optimal parameter by repeatedly adjusting the parameter so that the difference between the predicted value and the reanalyzed field data is minimized while repeatedly performing learning.

Then, for each of the specific grid points for the first deep learning module that has been learned through the above-described learning process, (i) t-1, t-2, ..., tm past views The first historical forecast data (11) and (ii) as values corresponding to each meteorological climate, as values corresponding to each of the t+1th, t+2, ..., t+nth future times When at least a part of the future forecast data 12 is acquired by the first forecast model 10, the learning server 100 causes the first deep learning module 210 to A predetermined first deep learning operation is performed by receiving at least a part of the corresponding specific first past forecast data 11-1 and the specific future forecast data 12-1 corresponding to at least some of the future time points. As a result of the execution, a value corresponding to the specific weather climate for each of (i) a specific grid point and (ii) a specific past time point and a specific future time point may be output as second deep learning input forecast data.

At this time, in the learning process of the first deep learning module, if at least one of the ground feature data 41 and the high feature data 42 is additionally input and the learning is performed, the second deep learning input forecast data is output. In the process, at least one of the corresponding ground characteristic data 41 and the altitude characteristic data 42 may be input together into the first deep learning module, and may be referred to and used for outputting the second deep learning input forecast data.

In addition, a predetermined pre-processing is performed on at least some of the second deep learning input forecast data 211 output from the first deep learning module 210. In a state in which the reduced data reduced by the magnification is generated, the learning server 100 (i) receives the reduced data as the second learning data by the second deep learning module, and performs a second deep learning operation on the reduced data. As a result of applying, the enlarged data enlarged at a second ratio is calculated, and (ii) the enlarged data is compared with the specific second deep learning input forecast data as a corresponding second GT (Ground Truth) to obtain a second deep It can be done to learn about the learning module.

At this time, as an example of the invention, the pre-processing process for generating reduced data used as the learning data of the second deep learning module, the learning server 100, the specific second deep learning input forecast data 211-1 For example, the number of horizontal or vertical pixels may be an integer value corresponding to a multiple of the second magnification.

In addition, as another example of the invention, the learning server 100 may generate the reduced data using a predetermined interpolation method.

As described above, data input for learning by the second deep learning module 220 will be described in detail with reference to FIG. 4A as follows.

4A is an exemplary diagram of data input and output to a second deep learning module in a learning process for generating high-resolution weather climate data according to an embodiment of the present invention. In FIG. 4A, each of the input and output data is illustrated as image data, but the present invention is not limited thereto, and each data may be data including only a numerical value, not an image.

Referring to FIG. 4A, with respect to at least a portion of the second deep learning input forecast data 211-1 of the second deep learning input forecast data 211 output from the first deep learning module, the learning server 100 The reduced data 430 may be generated through a predetermined pre-processing. In this case, the learning server 100 may generate the reduced data 430 by using a predetermined interpolation method such as cubic spline interpolation for the original specific second deep learning input forecast data 211-1.

In addition, the original specific second deep learning input forecast data 211-1 may be data of a size of (149, 253), as shown in FIG. 4A, and a predetermined first magnification for generating the reduced data 430 is In the case of 1/5 times, the reduced data must have a size of (149/5, 253/5), but since each of the reduced horizontal and vertical sizes is not an integer value, a specific second deep learning input to make it an integer value. A specific row or at least a part of a specific column of the forecast data 211-1 can be added or deleted. Each of the vertical size as the final number of rows and the horizontal size as the final number of columns is a predetermined second deep learning of the reduced data 430. It may be a multiple of the second magnification applied to the generation of the enlarged data 440 calculated through the operation. As shown in FIG. 4A, when the second magnification is 5 times, the learning server 100 includes two columns at the left and right edges and one upper edge row for the specific second deep learning input forecast data 211-1. And, by deleting the data corresponding to one lower edge row and reducing its size to (145,250) and reducing it to 1/5 times the first magnification, the horizontal and vertical size is reduced data of size (29,25) each of which is an integer value. 430 can be created.

Then, the learning server 100 causes the second deep learning module 220 to receive the reduced data 430 and perform a predetermined second deep learning operation to obtain the enlarged data 440 enlarged by a second magnification. The calculation may be performed, and the second deep learning module may compare this with the original specific second deep learning input forecast data 211-1 to perform learning on the second deep learning module.

At this time, as another example of the invention, the second deep learning module may borrow an image high-resolution deep learning algorithm such as SRCNN, and in this case, the learning server 100 inputs a specific second deep learning module to the second deep learning module. The deep learning input forecast data 211-1 may be converted according to a predetermined image conversion method, and the output image data may be converted again according to a predetermined data conversion method. At this time, the pre-processing process for generating the reduced data 430 used as the learning data of the second deep learning module, the learning server 100, the specific second deep learning input forecast data (211-1) A learning method characterized in that at least one of a blur process and noise addition is additionally performed on the converted image. In addition, in the second deep learning operation, a predetermined convolution operation performed on the reduced data 430 may be repeatedly performed to calculate each pixel value of a high-resolution image.

In addition, even when the specific second deep learning input forecast data 211-1 is not converted into an image and is input as it is, the blur processing and noise addition algorithm are performed on the pixel values of the image. It can be applied and applied to the deep learning input forecast data 211-1 itself.

In addition, the SRCNN algorithm that can be used in the second deep learning module may be borrowed from the paper (https://arxiv.org/abs/1501.00092) as a prior document, but is not limited thereto, and implementation of the invention. Various deep learning algorithms may be used according to conditions, and accordingly, details of the second deep learning operation and the learning method for the second deep learning module 220 may be applied differently.

When learning for the first deep learning module 210 and the second deep learning module 220 is completed through the learning process as described above, the first deep learning module 210 and the second deep learning module 220 The test process may be performed by the included test server.

Specifically, the test server causes the first deep learning module 210 on which the training has been completed to generate a specific first past for testing as a value corresponding to the meteorological climate at each specific past time point acquired by the first forecast model 10. As a result of performing the first deep learning operation by receiving at least some of the forecast data and the specific future forecast data for testing as a value corresponding to the weather climate at each specific future point, (i) each of the specific grid points and (ii) the specific A value corresponding to the specific weather climate for at least some of a past time point and a specific future time point may be output as input forecast data for the second deep learning test.

Then, the test server receives input forecast data for a specific second deep learning test by the second deep learning module 220 and performs a second deep learning operation. By outputting the high-resolution forecast data 221 enlarged at the second magnification, the output high-resolution forecast data 221 may be obtained.

At this time, each of the specific first past forecast data for test, specific future forecast data for test, and input forecast data for second deep learning test used in the test process is the learning server 100 as a prerequisite for performing the test process. In order to distinguish it from learning specific first past forecast data, specific future forecast data for learning, and input forecast data for second deep learning learning used for learning of the first deep learning module and the second deep learning module,'test' and'test The expression'dragon' is used, and each of them may be data having the same format and properties as the specific first past forecast data, specific future forecast data, and second deep learning input forecast data represented in FIG. 2.

In addition, data used in the process of the test server outputting the high-resolution forecast data by the second deep learning module will be described with reference to FIG. 4B as follows.

4B is an exemplary diagram of data input and output to a second deep learning module in a test process for generating high-resolution weather climate data according to an embodiment of the present invention. In FIG. 4B, each of the input and output data is illustrated as image data, but the present invention is not limited thereto, and each data may be data including only a numerical value, not an image.

Referring to FIG. 4B, when specific second deep learning test input data 410 is input to the second deep learning module 220, the test server causes the second deep learning module 220 to The running operation may be performed to output the high-resolution forecast data 221 enlarged at the second magnification. At this time, if the specific second deep learning test input data 410, which is the input data, has a size of (149,253), a second deep learning operation is performed on it, and the (745,1265) is enlarged to a second magnification of 5 times. High-resolution forecast data 221 of a size may be output.

Therefore, when comparing each of the values included in the high-resolution forecast data 221 with points corresponding to each of the values included in the original specific second deep learning test input forecast data 410, the horizontal 5 times for the same forecast space. And 5 times the length of data, and as a result, it can be used as forecast data corresponding to the fine grid spacing of 5 times. That is, if the original specific input forecast data 410 for the second deep learning test is data with a resolution of 5 km, the high-resolution forecast data 221 is data with a resolution of 1 km and includes forecast data for a finer grid interval. It may be used, and detailed contents of the present invention may be adjusted and implemented to obtain higher resolution data according to the implementation conditions of the present invention.

In addition, as another example of the invention, the second deep learning module 220 may borrow an image high-resolution deep learning algorithm such as SRCNN, and in this case, the test server is input to the second deep learning module. The running test input forecast data 410 may be converted according to a predetermined image conversion method, and the output image data may be converted again according to a predetermined data conversion method. In this case, the second deep learning operation may be performed by repeatedly performing a predetermined convolution operation on the input forecast data 410 for a specific second deep learning test.

In addition, even when the input forecast data 410 for a specific second deep learning test is not converted into an image and a real value is reflected as it is, the blur processing and noise addition algorithm are performed on the pixel values of the image. 2 It may be applied to the input forecast data 410 for deep learning test itself and performed.

And, the SRCNN algorithm that can be used in the second deep learning module as described above is not limited thereto, and various deep learning algorithms can be used according to the implementation conditions of the invention, and accordingly, the second deep learning operation and the second deep learning algorithm Details of the test method for the learning module 220 may be applied differently.

As another example of the invention, the test server determines a value corresponding to the specific weather climate for at least some of the high resolution grid points corresponding to the high resolution forecast data 221, the actual observation data 30 Among them, it is compared with specific actual observation data measured at a location corresponding to a specific high-resolution grid point, and if the difference is greater than or equal to a threshold, it is determined that additional learning is necessary, and the additional learning can be supported. For example, if there is a specific actual observation data measured at a location equal to or within a predetermined range of a specific high resolution grid point of the high resolution forecast data 221, the test server If the value corresponding to'temperature' is 30 degrees, the actual observation data corresponding to it is 25 degrees, the difference is 5, and the predetermined threshold is 3 degrees, the test server determines that additional learning is necessary, and related information It is possible to support re-learning by providing a user terminal or allowing a separate user terminal to provide it.

In addition, as another example of the invention, the test server compares the specific first historical forecast data 11-1 with the reanalyzing field data 22 and the first error value and the specific first historical forecast data 11-1. ) To the first deep learning module 210 to cause the first deep learning module 210 to perform the first deep learning operation, and the output second deep learning test input forecast data 221 is reanalyzed field data. With reference to each of the second error values compared to (22), the performance of the first forecast model 10 and the first deep learning module 210 may be compared and determined. For example, by comparing the specific first past forecast data 11-1 obtained from the first forecast model 10 with the reanalyzing field data 22 used as the first GT, each specific grid point and each specific past time point A second deep learning test that is output as a result of calculating each of the first error values for, and the first deep learning module 210 receives specific first past forecast data 11-1 and performs the first deep learning operation. Comparing the input forecast data 410 for the reanalysis field data 22 to calculate the second error values for each of a specific grid point and each of a specific past time point, a value obtained by summing each of the first error values If this is 10 and the sum of each of the second error values is 8, it may be determined that the performance of the first deep learning module 210 is better than the first forecast model 10, but is not limited thereto. A calculation method and a performance criterion applied to each of the error values may be determined differently according to the implementation conditions of the invention.

Referring to FIG. 5, in the learning process for generating high-resolution weather and climate data 221, first past forecast data 11 for each specific grid point included in a predetermined forecast area are converted to a predetermined first forecast model. It starts from what is acquired (S501).

Then, the learning server 100 may input (S502) specific first past forecast data 11-1 corresponding to a specific past time point into the first deep learning module as first learning data.

Next, the learning server 100 may cause the first deep learning module 210 to predict a value corresponding to a specific weather climate for each specific grid point and each specific past time point (S503).

Then, the learning server 100 causes the first deep learning module 210 to compare the predicted value corresponding to the specific weather climate with the reanalysis field data 22 as the first GT, and the first deep learning module learns ( S504).

In a state in which the learning of the first deep learning module 210 is completed through the processes S501 to S504, at least a part of the first past forecast data 11 and the future forecast data 12 for each specific grid point is the first It may be acquired (S505) by the forecast model.

Then, the learning server 100 may input at least some of the acquired first past forecast data 11 and future forecast data 12 to the first deep learning module 210 (S506), and the first The deep learning module 210 outputs a value corresponding to a specific weather climate for each specific grid point and at least some of a specific past time point and a specific future time point as the second deep learning input forecast data 211 (S507). can do.

Next, the learning server 100 performs a predetermined pre-processing on at least some of the second deep learning input forecast data 211-1 of the second deep learning input forecast data 211 output from the first deep learning module. Thus, the reduced data 430 reduced to the first magnification can be generated (S508), and the generated reduced data 430 can be input (S509) to the second deep learning module 220 as second learning data. have.

Then, the learning server 100 may cause the second deep learning module 220 to calculate the enlarged data 440 enlarged at a second ratio (S510), and the calculated enlarged data 440 corresponding thereto. The second deep learning module 220 may be trained (S511) by comparing it with the specific second deep learning input forecast data 211-1 as the second GT (Ground Truth).

Referring to FIG. 6, the process of performing a test by the test server for generating high-resolution weather climate data is, in a state in which learning of the first deep learning module 210 and the second deep learning module 220 is completed, the first It starts from obtaining (S601) specific first past forecast data for testing and second specific past forecast data for testing by the forecast model 10.

Then, the test server inputs at least a part of the acquired specific first past forecast data for testing and the specific second past forecast data for testing into the first deep learning module 210 (S602), so that the second deep learning The module 220 outputs a value corresponding to a specific weather climate for each of a specific grid point and at least some of a specific past time point and a specific future time point as the input forecast data 410 for the second deep learning test (S603). can do.

Then, the test server inputs the outputted second deep learning test input forecast data 410 to the second deep learning model 220 (S604), thereby causing the second deep learning model 220 to generate a second magnification. By outputting the high-resolution forecast data 221 enlarged to (S605), the high-resolution forecast data 221 may be finally obtained.

And, as an example of the invention, through the implementation of the present invention, a value corresponding to a specific meteorological climate for a blank area, which is an area inside the grid point of the past forecast data predicted by a weather forecast information provider such as the Meteorological Administration It can be predicted, and based on this, it is possible to develop and support new business models such as the meteorological appraisal industry that enables the use of values corresponding to specific weather conditions that are estimated to have affected events that occurred in the blank areas at a specific past time. I can.

The embodiments according to the present invention described above may be implemented in the form of program instructions that can be executed through various computer components and recorded in a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, data structures, and the like alone or in combination. The program instructions recorded in the computer-readable recording medium may be specially designed and configured for the present invention, or may be known and usable to those skilled in the computer software field. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magnetooptical media such as floptical disks. , And a hardware device specially configured to store and execute program instructions such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules to perform the processing according to the present invention, and vice versa.

In the above, the present invention has been described by specific matters such as specific elements and limited embodiments and drawings, but this is provided only to help a more general understanding of the present invention, and the present invention is not limited to the above embodiments. , Anyone having ordinary knowledge in the technical field to which the present invention pertains can make various modifications and variations from these descriptions.

Therefore, the spirit of the present invention is limited to the above-described embodiments and should not be defined, and all modifications that are equally or equivalent to the claims as well as the claims to be described later fall within the scope of the spirit of the present invention. I would say.

Claims

As a learning method for generating high-resolution weather and climate data,

(a) A specific grid point included in a predetermined forecast area among grid points corresponding to points where a plurality of first axis lines parallel to the first direction and a plurality of second axis lines parallel to the second direction cross each other. On the other hand, when the first historical forecast data as a value corresponding to the meteorological climate of each of the t-1, t-2, ..., tm past time points is obtained by a predetermined first forecast model, the learning server, (i) As a result of the first deep learning module receiving specific first past forecast data corresponding to at least some of the past time points as first learning data, and performing a predetermined first deep learning operation. A value corresponding to at least one of the meteorological climates is predicted for each of the specific grid points and each of the specific past time points, and (ii) a value corresponding to the predicted specific weather climate is calculated as the first GT Re-analysis field data as (Ground Truth)-The re-analysis field data is a second past forecast as a value corresponding to the specific weather climate obtained by a second forecast model for each of the specific grid points and each of the specific past time points. Comparing the data with the corrected data with reference to actual observation data collected separately, and performing learning on the first deep learning module;

(b) In a state in which learning about the first deep learning module is completed, for each of the specific grid points, (i) each of the t-1, t-2, ..., tm-th past views The first past forecast data as a value corresponding to the meteorological climate and (ii) the future forecast as a value corresponding to the meteorological climate at each of the t+1, t+2, ..., t+nth future times When at least some of the data is acquired by the first forecast model, the learning server causes the first deep learning module to cause the specific first past forecast data and As a result of performing the first deep learning operation by receiving at least some of the specific future forecast data corresponding to the specific future time of at least some of the future time points, (i) each of the specific grid points and (ii) the specific past time point And outputting a value corresponding to the specific weather climate for at least some of the specific future time points as second deep learning input forecast data. And

(c) In a state in which reduced data reduced by a predetermined first magnification is generated by performing a predetermined pre-processing on at least a part of specific second deep learning input forecast data among the output second deep learning input forecast data, the The learning server, (i) the second deep learning module receives the reduced data as second learning data, and as a result of applying a predetermined second deep learning operation to the reduced data, the enlarged data enlarged at a second ratio And (ii) comparing the enlarged data with the specific second deep learning input forecast data as a second GT (Ground Truth) corresponding thereto to perform learning on the second deep learning module ;

Containing, learning method.
The method of claim 1,

In step (a),

The learning server, characterized in that for causing the first deep learning module to perform learning by additionally receiving at least one of ground characteristic data and altitude characteristic data corresponding to each of the specific grid points,

When the specific past time point is a plurality of time points, the first deep learning that the server repeats the same as the number of the specific past time points so that at least one of the ground characteristic data and the altitude data corresponds to each of the specific past time points. A learning method, characterized in that at least one of ground feature data for input and advanced feature data for first deep learning input is additionally generated, and the first deep learning module receives the additional input to perform learning.
The method of claim 2,

The ground characteristic data has a preset category value for each ground characteristic corresponding to each of the specific grid points,

In the learning process of the first deep learning module, each of the category values is converted into a predetermined dimensional vector value corresponding to each of the specific grid points, and is used for learning of the first deep learning module. Way.
The method of claim 1,

The re-analysis field data is, for each of the specific grid points (i), for each of the first partial specific grid points in which the actual observation data exists within a predetermined range, the first specific grid is referred to the actual observation data. The value corresponding to the specific weather climate of each point is determined and corrected, and (ii) for each of the second partial specific grid points in which the actual observation data does not exist within a predetermined range, predicted by the second forecast model. With reference to the value corresponding to the specified weather climate and at least one actual observation data measured at a position close to the second part specific grid point, corresponding to the specific weather climate for each of the second part specific grid points. The learning method, characterized in that the value is determined and corrected.
The method of claim 1,

In step (c),

In the pre-processing process for generating the reduced data used as the learning data of the second deep learning module, the learning server performs blur processing and noise on the specific second deep learning input forecast data. The learning method, characterized in that made by additionally performing at least one process of addition.
The method of claim 1,

In step (c),

In the pre-processing process for generating the reduced data used as the learning data of the second deep learning module, the learning server includes, for the specific second deep learning input forecast data, the number of horizontal or vertical pixels is equal to the second magnification. A learning method, characterized in that it is formed by additionally performing a process of deleting or adding a part of a specific row or a specific column so that it can be an integer value corresponding to a multiple.
The method of claim 1,

In step (c),

The learning method, characterized in that the server generates the reduced data using a predetermined interpolation method.
As a test method for generating high-resolution weather and climate data,

(a) By the learning server, (1) a predetermined forecast area among grid points corresponding to points where a plurality of first axes parallel to a first direction and a plurality of second axes parallel to a second direction intersect. For each specific grid point included in, the first historical forecast data for learning as a value corresponding to the meteorological climate of each of the t-1, t-2, ..., and tm past points is a predetermined first forecast model. When obtained by (i) the first deep learning module receives specific first past forecast data for learning corresponding to at least some of the past time points as first learning data, and a predetermined first deep learning As a result of performing the calculation, a value corresponding to at least one specific meteorological climate among the meteorological climates is predicted for each of the specific grid points and each of the specific past time points, and (ii) a value corresponding to the predicted specific weather climate Value, the learning re-analysis field data as a first GT (Ground Truth)-The learning re-analysis field data is, for each of the specific grid points and each of the specific past time points, in the specific weather climate obtained by the second forecast model. This data is corrected by referring to actual observation data collected separately from the second historical forecast data for learning as a corresponding value.-Compared with, to perform learning on the first deep learning module, and (2) the first In a state in which learning about the deep learning module is completed, for each of the specific grid points, (i) a value corresponding to the weather climate at each of the t-1, t-2, ..., and tm-th past points At least a portion of the first past forecast data for learning as and (ii) the future forecast data for learning as values corresponding to the weather climate at each of the t+1, t+2, ..., and t+nth future times When is obtained by the first forecast model, the first deep learning module causes the learning specific first past forecast data to correspond to the specific past time point and at least a part of the future time point. Specific future forecast for learning corresponding to the future point of view As a result of performing the first deep learning operation by receiving at least a portion of the data, (i) each of the specific grid points and (ii) the specific weather climate for at least some of the specific past time point and the specific future time point The corresponding value is output as second deep learning learning input forecast data, and (3) a predetermined preprocessing is performed on at least some of the output second deep learning input forecast data. In a state in which the reduced data for learning reduced to a predetermined first magnification is generated, (i) the second deep learning module receives the reduced data for learning as second learning data, As a result of applying the second deep learning operation, the expanded data for learning expanded at a second ratio is calculated, and (ii) the expanded data for learning is used as the corresponding second GT (Ground Truth) for the specific second deep learning learning. Compared with the input forecast data to perform learning on the second deep learning module, and in a state in which learning on each of the first deep learning module and the second deep learning module is completed, the test server The learning module allows the specific first past forecast data for testing as a value corresponding to the meteorological climate at each of the specific past time points acquired by the first forecast model and as a value corresponding to the weather climate at each of the specific future time points. As a result of performing the first deep learning operation by receiving at least some of the specific future forecast data for test, (i) each of the specific grid points, and (ii) each of the specific past time points and at least some of the specific future time points. Outputting a value corresponding to the specific weather climate as input forecast data for a second deep learning test; And

(b) The test server receives input forecast data for at least a portion of the second deep learning test input forecast data from the second deep learning module and performs the second deep learning operation. As a result of performing the specific second deep learning test input forecast data to output high-resolution forecast data enlarged by the second magnification, thereby obtaining the output high-resolution forecast data;

Containing, the test method.
The method of claim 8,

The test server measures a value corresponding to the specific weather climate for at least some of the high-resolution grid points corresponding to the high-resolution forecast data at a location corresponding to the specific high-resolution grid point in the actual observation data. Compared with the specific actual observation data, if the difference is greater than or equal to a threshold, it is determined that additional learning is necessary, and the additional learning is supported so that the additional learning can be performed.
The method of claim 8,

The test server inputs a first error value obtained by comparing the specific first past forecast data with the reanalysis field data, and the first past forecast data into the first deep learning module to cause the first deep learning module to The first prediction model and the first deep learning by referring to a second error value obtained by comparing the input forecast data for the second deep learning test output by performing the first deep learning operation with the data of the reanalysis field. It supports to determine the performance of the module, characterized in that support to determine the higher performance model as the error value is smaller, the test method.
As a learning server to generate high-resolution weather and climate data,

At least one memory for storing instructions; And

At least one processor configured to execute the instructions; Including,

The processor,

(I) A specific grid point included in a predetermined forecast area among grid points corresponding to points where a plurality of first axis lines parallel to the first direction and a plurality of second axis lines parallel to the second direction cross each other. On the other hand, if the first historical forecast data as a value corresponding to the meteorological climate at each of the t-1, t-2, ..., tm past time points is obtained by a predetermined first forecast model, (i) the first 1 The deep learning module receives specific first past forecast data corresponding to at least some of the past time points as first learning data, and performs a predetermined first deep learning operation. A value corresponding to at least one specific weather climate is predicted for each of the specific grid points and each of the specific past time point, and (ii) a value corresponding to the predicted specific weather climate is determined as a first GT (Ground Truth) Re-analysis field data as-The re-analysis field data separately collects second past forecast data as a value corresponding to the specific weather climate obtained by a second forecast model for each of the specific grid points and each of the specific past time points. A process of performing learning on the first deep learning module by comparing the data with the corrected data with reference to the actual observed data; (II) In a state in which learning about the first deep learning module is completed, for each of the specific grid points, (i) each of the t-1, t-2, ..., tm-th past views The first past forecast data as a value corresponding to the meteorological climate and (ii) the future forecast as a value corresponding to the meteorological climate at each of the t+1, t+2, ..., t+nth future times When at least some of the data is acquired by the first forecast model, the learning server causes the first deep learning module to cause the specific first past forecast data and As a result of performing the first deep learning operation by receiving at least some of the specific future forecast data corresponding to the specific future time of at least some of the future time points, (i) each of the specific grid points and (ii) the specific past time point And a process of outputting a value corresponding to the specific weather climate for at least some of the specific future time points as second deep learning input forecast data. And (III) performing a predetermined pre-processing on at least some of the output second deep learning input forecast data on at least a portion of the second deep learning input forecast data to generate reduced data reduced by a predetermined first magnification, The learning server, (i) the second deep learning module receives the reduced data as second learning data, and as a result of applying a predetermined second deep learning operation to the reduced data, the expansion is enlarged by a second ratio. To calculate data, and (ii) to perform learning on the second deep learning module by comparing the enlarged data with the specific second deep learning input forecast data as a second GT (Ground Truth) corresponding thereto. process; To do, learning server.
The method of claim 11,

In the above (I) process,

The processor is characterized in that for causing the first deep learning module to perform learning by additionally receiving at least one of ground feature data and altitude feature data corresponding to each of the specific grid points,

When the specific past time point is a plurality of time points, the processor is a first deep learning that repeats the same as the number of the specific past time points so that at least one of the ground characteristic data and the altitude data corresponds to each of the specific past time points. A learning server, characterized in that by additionally generating at least one of ground feature data for input and advanced feature data for first deep learning input, and allowing the first deep learning module to receive additional input and perform learning.
The method of claim 12,

The ground characteristic data has a preset category value for each ground characteristic corresponding to each of the specific grid points,

In the learning process of the first deep learning module, each of the category values is converted into a predetermined dimensional vector value corresponding to each of the specific grid points, and is used for learning of the first deep learning module. server.
The method of claim 11,

The re-analysis field data is, for each of the specific grid points (i), for each of the first partial specific grid points in which the actual observation data exists within a predetermined range, the first specific grid is referred to the actual observation data. The value corresponding to the specific weather climate of each point is determined and corrected, and (ii) for each of the second partial specific grid points in which the actual observation data does not exist within a predetermined range, predicted by the second forecast model. With reference to the value corresponding to the specified weather climate and at least one actual observation data measured at a position close to the second part specific grid point, corresponding to the specific weather climate for each of the second part specific grid points. Learning server, characterized in that the value is determined and corrected.
The method of claim 11,

In the above (III) process,

In the preprocessing process for generating the reduced data used as the training data of the second deep learning module, the processor processes a blur and adds noise to the specific second deep learning input forecast data. Learning server, characterized in that formed by additionally performing at least one of the processes.
The method of claim 11,

In the above (III) process,

In the pre-processing for generating the reduced data used as the training data of the second deep learning module, the processor includes the number of horizontal or vertical pixels for the specific second deep learning input forecast data, which is a multiple of the second magnification. A learning server, characterized in that it is formed by additionally performing a process of deleting or adding a part of a specific row or a specific column to become an integer value corresponding to.
The method of claim 11,

In the above (III) process,

The learning server, characterized in that the processor generates the reduced data using a predetermined interpolation method.
As a test method for generating high-resolution weather and climate data,

At least one memory for storing instructions; And

At least one processor configured to execute the instructions; Including,

The processor,

(I) By the learning server, (1) a predetermined forecast area among grid points corresponding to points where a plurality of first axes parallel to a first direction and a plurality of second axes parallel to a second direction intersect. For each specific grid point included in, the first historical forecast data for learning as a value corresponding to the meteorological climate of each of the t-1, t-2, ..., and tm past points is a predetermined first forecast model. When obtained by (i) the first deep learning module receives specific first past forecast data for learning corresponding to at least some of the past time points as first learning data, and a predetermined first deep learning As a result of performing the calculation, a value corresponding to at least one specific meteorological climate among the meteorological climates is predicted for each of the specific grid points and each of the specific past time points, and (ii) a value corresponding to the predicted specific weather climate Value, the learning re-analysis field data as a first GT (Ground Truth)-The learning re-analysis field data is, for each of the specific grid points and each of the specific past time points, in the specific weather climate obtained by the second forecast model. This data is corrected by referring to actual observation data collected separately from the second historical forecast data for learning as a corresponding value.-Compared with, to perform learning on the first deep learning module, and (2) the first In a state in which learning about the deep learning module is completed, for each of the specific grid points, (i) a value corresponding to the weather climate at each of the t-1, t-2, ..., and tm-th past points At least a portion of the first past forecast data for learning as and (ii) the future forecast data for learning as values corresponding to the weather climate at each of the t+1, t+2, ..., and t+nth future times When is obtained by the first forecast model, the first deep learning module causes the learning specific first past forecast data to correspond to the specific past time point and at least a part of the future time point. Specific future forecast for learning corresponding to the future point of view As a result of performing the first deep learning operation by receiving at least a portion of the data, (i) each of the specific grid points and (ii) the specific weather climate for at least some of the specific past time point and the specific future time point The corresponding value is output as second deep learning learning input forecast data, and (3) a predetermined pre-processing is performed on at least some of the output second deep learning input forecast data. In a state in which the reduced data for learning reduced to a predetermined first magnification is generated, (i) the second deep learning module receives the reduced data for learning as second learning data, As a result of applying the second deep learning operation, the expanded data for learning expanded at a second ratio is calculated, and (ii) the expanded data for learning is used for the specific second deep learning learning as a corresponding second GT (Ground Truth). Compared with the input forecast data to perform learning on the second deep learning module, and in a state in which learning on each of the first deep learning module and the second deep learning module is completed, the test server The running module allows the specific first past forecast data for testing as a value corresponding to the meteorological climate at each of the specific past time points acquired by the first forecast model, and a value corresponding to the weather climate at each of the specific future time points. As a result of performing the first deep learning operation by receiving at least some of the specific future forecast data for test, (i) each of the specific grid points, and (ii) each of the specific past time points and at least some of the specific future time points. A process of outputting a value corresponding to the specific weather climate as input forecast data for a second deep learning test; And (II) the test server receives input forecast data for at least some of the input forecast data for the second deep learning test by the second deep learning module, and receives the second deep learning data. A process of obtaining the output high-resolution forecast data by outputting the high-resolution forecast data expanded by the second magnification of the input forecast data for the specific second deep learning test as a result of performing the operation; To do, the test server.
The method of claim 18,

The processor measures a value corresponding to the specific weather climate for at least some of the high-resolution grid points corresponding to the high-resolution forecast data at a location corresponding to the specific high-resolution grid point in the actual observation data. Compared with specific actual observation data, if the difference is greater than or equal to a threshold, it is determined that additional learning is necessary, and the test server is characterized in that supporting the additional learning to be performed.
The method of claim 18,

The processor inputs a first error value obtained by comparing the specific first past forecast data with the reanalysis field data, and the first past forecast data into the first deep learning module, causing the first deep learning module to cause the The first prediction model and the first deep learning module by referring to a second error value obtained by comparing the input forecast data for the second deep learning test output by performing a first deep learning operation with the data of the reanalysis field. Support to determine the performance of, but the smaller the error value, the test server, characterized in that it supports to determine the higher performance model.