KR102218179B1 - System and method for emulating radiative transfer parameterization of korea local analysis and prediction system - Google Patents

Abstract

The present invention relates to a system and a method for emulating radiative physical parameterization of the Korea local analysis and prediction system (KLAPS) of the Meteorological Administration. The present invention is able to divide a figure prediction data set of the KLAPS into an independent meteorological element information and a subsidiary meteorological element information, build an artificial neural network, receive input of meteorological element information for figure prediction, calculate parameterized meteorological element information, and improve the accuracy of the calculation results and the calculation speed. To this end, according to the present invention, the system for emulating radiative physical parameterization of the KLAPS can comprise: a prior learning data classification module, a storage module, a neural network module, a control module, and an input and output module.

Description

System and method to emulate copy physics parameterization of ultra-short forecast model {SYSTEM AND METHOD FOR EMULATING RADIATIVE TRANSFER PARAMETERIZATION OF KOREA LOCAL ANALYSIS AND PREDICTION SYSTEM}

The present invention relates to a system and method for emulating the radiative physics parameterization technique of the Korea Local Analysis and Prediction System (KLAPS) of the Korea Meteorological Administration.

In more detail, the artificial neural network of the emulation system is constructed using radiative physical information such as vertical air pressure and vertical temperature included in the data set of the ultra-short-term forecast model, and radiative physical information such as the vertical heating rate and heat radiation parameterized accordingly. And, the present invention relates to an emulation system and method that outputs parameterized copy physics information when inputting copy physics information for numerical forecasting, and improves computation speed and accuracy of the output copy physics information.

Numerical forecasting is a series of processes that analyze current atmospheric conditions and quantitatively predict future atmospheric conditions by continuously numerically integrating governing equations for the dynamics and physical principles of atmospheric phenomena.

The numerical forecast model is for realizing the numerical forecast, which is a meteorological modeling of the Earth's meteorological system using the dynamics and physical equations governing the state and motion of the atmosphere.

Since the meteorological system is a continuum in time and space, calculating the meteorological system mathematically directly requires a lot of resources. Therefore, the numerical forecast model can be configured to divide the earth into a grid and calculate an equation for the state and motion of the atmosphere at each grid point.

The horizontal resolution used at this time may be defined as the distance between grid points in the horizontal direction, and the horizontal resolution may be hundreds of m to hundreds of km. In addition, the vertical resolution can be defined according to the number of layers by viewing the space between the grid points in the vertical direction as one layer.

Numerical forecast models can be classified into various types according to the purpose of operation and the scope of operation.

For example, Global Data Assimilation and Prediction System (GDAPS) for global mid-term prediction, Local Data Assimilation and Prediction System (LDAPS) for short-term prediction of the Korean Peninsula, and ultra-short forecast for ultra-short prediction of the Korean Peninsula. Model (Korea Local Analysis and Prediction System, KLAPS), wave height model for global ocean wave height prediction (Global Wave Watch-3, GWW3), regional tide/Storm Surge Model (RTSM) for Asian storm tsunami prediction , And is operating numerical forecasting models such as the Asian Dust Aerosol Model 2 (ADAM2) for predicting the transport of yellow dust.

Among them, KLAPS produces 3D prediction data for the Korean peninsula, and has a horizontal resolution of 5 km and a vertical number of 40 stories. In addition, the number of daily operations of the very short-term forecast model (KLAPS) is 144 times, and the forecasting time is 12 hours.

Since the number of daily operations of the very short-term forecast model (KLAPS) is large, it is necessary to calculate the numerical forecast model at high speed while maintaining the accuracy of the weather forecast. And in order to prevent social and economic losses due to bad weather, the need for a high-speed numerical forecast model is constantly being raised.

It is the radiative physics process that describes the energy transfer process between the sun and the earth that accounts for the largest amount of computation in the numerical forecast model.

The process including heat transfer by radiation may be, for example, emission, absorption, reflection, etc. of radiant energy for each wavelength by different surfaces (grounds, clouds, etc.) and components of the atmosphere (ozone, water vapor, etc.) . Parameterizing these processes is very complex and requires a lot of computation time.

In addition, since computing resources for operating the Rapid Radiative Transfer Model for GCM (general circulation model Korea, RRTMG-K) of KLAPS are limited, the numerical value depends only on the performance improvement of the hardware. There is a limit to improving the computational speed of the forecast model.

Therefore, there is a need for a system and method that improves the computational speed and maintains the accuracy of numerical forecasting while emulating the radiative physics parameterization technique (RRTMG-K) of the very short-term forecast model (KLAPS).

An object of the present invention is to solve the above problems, and to provide a system and method for emulating a radiation physics parameterization technique (RRTMG-K) of an ultra-short forecast model (KLAPS).

In addition, it is to provide an emulation system and method that improves the emulation speed of the radiative physics parameterization technique (RRTMG-K) of the very short-term forecast model (KLAPS) and maintains the accuracy of the numerical forecast.

In order to achieve the above object, the present invention receives a numerical forecast data set of an ultra-short forecast model (KLAPS), extracts independent meteorological element information for each row from the numerical forecast data set, and processes the pre-input data (PREV_IN). And a pre-learning data classification module for extracting dependent meteorological element information and processing reference output data COMP_OUT; A storage module including a memory for storing at least one of the pre-input data PREV_IN, the reference output data COMP_OUT, weight data, bias data, and pre-output data PREV_OUT, respectively; A neural network module for calculating the pre-output data PREV_OUT including the dependent meteorological element information after the pre-input data PREV_IN, the weight data, and the bias data are mapped; And a control module for comparing the pre-output data PREV_OUT and the reference output data COMP_OUT, and then updating the weight data and the bias data and storing them in the storage module, wherein the control module comprises: the neural network module The pre-input data (PREV_IN), the weight data, and the bias data are mapped to calculate the pre-output data (PREV_OUT), and the pre-output data (PREV_OUT) and the reference output data (COMP_OUT) are compared. And updating the weight data and the bias data and storing them in the storage module for a first operation time.

Another embodiment of the present invention includes the steps of: (S1) receiving, by a pre-training data classification module, a numerical forecast data set of an ultra-short-term forecast model (KLAPS); (S2) The pre-learning data classification module extracts N independent meteorological element information from the numerical forecast data set to process the pre-input data (PREV_IN), and extracts M dependent meteorological element information to obtain reference output data (COMP_OUT). ) Processing, and transferring it to a storage module; (S3) storing, by the storage module, the pre-input data PREV_IN and the reference output data COMP_OUT; (S4) storing, by the storage module, randomly setting and storing first weight data, first bias data, second weight data, and second bias data; (S5) setting, by the control module, a first operation time; (S6) the control module mapping the first weight data and the first bias data, the second weight data and the second bias data to a neural network module; (S7) mapping the pre-input data (PREV_IN) to the neural network module by the control module; (S8) calculating, by the neural network module, pre-output data PREV_OUT according to Equation 1 below;

[Equation 1]

: 상기 사전 입력 데이터(PREV_IN) /

: 상기 제 1 가중치 데이터 /

: 상기 제 1 편향 데이터 /

: 상기 제 2 가중치 데이터 /

: 상기 제 2 편향 데이터 /

: 활성화 함수 /

: 상기 사전 출력 데이터(PREV_OUT) )(

: The above pre-input data (PREV_IN) /

: The first weight data /

: The first bias data /

: The second weight data /

: The second bias data /

: Activation function /

: The pre-output data (PREV_OUT))

(S9) After the control module compares the pre-output data PREV_OUT and the reference output data COMP_OUT, the first weight data and the first bias data, the second weight data and the second bias data Updating to; (S10) After determining whether the control module has reached the first operation time, it ends if the first operation time has been reached, and repeats steps (S6) to (S9) if the first operation time has not been reached. It provides a method for emulation of radiative physics parameterization comprising the steps of executing the method.

The radiative physical parameterization emulation system and method according to the present invention receives a numerical forecast data set of an ultra-short-term forecast model (KLAPS) of the Meteorological Administration, extracts independent meteorological element information and dependent meteorological element information, respectively, and provides pre-input data (PREV_IN) and The reference output data (COMP_OUT) is processed, and the pre-output data (PREV_OUT) calculated after mapping the pre-input data (PREV_IN), the weight data, and the bias data to the artificial neural network is compared with the reference output data (COMP_OUT), The bias data can be updated.

In addition, since the accuracy of weight data and bias data can be improved by repeatedly executing such a task for a set operation time, a numerical value having a high degree of similarity to the copy physics parameterization technique (RRTMG-K) of the ultra-short forecast model (KLAPS). Forecast results can be calculated, and the computational speed can be increased by eliminating the need to directly calculate the complex equations of the radiative physics parameterization technique (RRTMG-K).

FIG. 1A is a block diagram schematically illustrating a radiation physics parametric emulation system according to an embodiment of the present invention, FIG. 1B is a schematic diagram of a numerical forecast data set, and FIG. 1C is a schematic diagram of a neural network module.
FIG. 2A is a flow chart briefly showing the steps of building a neural network in the emulation method for radiative physics parameterization according to an embodiment of the present invention, FIG. 2B is a flowchart briefly showing steps for reinforcing a neural network, and FIG. 2C is a radiative physics parameterization It is a flow chart that outlines the steps to emulate the technique.
3A and 3B are graphs showing a long-wave radiation heating rate and a short-wave radiation heating rate, respectively, when using the radiation physics parameterization technique of the ultra-short-term forecast model and the radiation physics parameterization emulation system according to an embodiment of the present invention. .
4A to 4F and 5A to 5F show the amount of long-wave radiation and the amount of short-wave radiation, respectively, when using the radiation physics parameterization technique of the ultra-short forecast model and when using the radiation physics parameterization emulation system according to an embodiment of the present invention. It is a graph.
6 is a table comparing computation speed and similarity (coefficient of determination) when using a radiative physics parameterization technique of an ultra-short forecast model and when using a radiative physics parameterization emulation system according to an embodiment of the present invention.

The present invention may be implemented with various changes without departing from the spirit, and may have one or more embodiments. In the present invention, the embodiments described in “specific details for carrying out the invention” and “drawings” are examples for specifically describing the present invention, and do not limit or limit the scope of the present invention.

Therefore, what can be easily inferred by those of ordinary skill in the technical field to which the present invention belongs, from the "specific details for carrying out the invention" and "drawings" of the present invention, is interpreted as belonging to the scope of the present invention. can do.

In addition, the size and shape of each component shown in the drawings may be exaggerated for description of the embodiments, and do not limit the size and shape of the invention actually implemented.

Unless a term used in the specification of the present invention is specifically defined, it may have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 1A is a block diagram schematically illustrating a radiation physics parametric emulation system according to an embodiment of the present invention, FIG. 1B is a schematic diagram of a numerical forecast data set, and FIG. 1C is a schematic diagram of a neural network module.

As shown in FIG. 1A, the copy physics parametric emulation system 1 according to an embodiment of the present invention includes a pre-learning data classification module 100 and a storage module 200, a neural network module 300, and a control module. 400), and may further include an input module 500 and an output module 600.

The copy physical parameterization emulation system 1 according to an embodiment of the present invention can be applied without limitation, if it is a computing device capable of reading and processing binary data such as a server computer, a desktop computer, and a smart phone. I can.

The pre-learning data classification module 100 may receive a numerical forecast data set (KLAPS_DS) of an ultra-short-term forecast model (KLAPS) produced by the Meteorological Agency.

In the ultra-short-term forecast model (KLAPS), the grid spacing may be 5 km, the number of horizontal grids is 283 x 235, the number of vertical floors is 40, the central latitude may be 38.0 ° (N) north latitude, and the central longitude may be 126.0 ° east longitude (E). . In addition, the analysis period of the very short-term forecast model (KLAPS) may be in units of 1 hour.

The pre-learning data classification module 100 may receive a numerical forecast data set (KLAPS_DS) through a computer network using the open API of the meteorological data open portal (https://data.kma.go.kr/). ,

Alternatively, the pre-learning data classification module 100 may directly receive a file storing the numerical forecast data set (KLAPS_DS). In this case, the file type may be NetCDF, but it is not limited to this and may be various types of files.

As shown in FIG. 1B, the numerical forecast data set KLAPS_DS may have a table structure in the form of a matrix.

In this case, each row can represent individual numerical forecast data measured in a specific grid unit and every time unit. In addition, for each row, information on meteorological factors such as air pressure, temperature, water vapor amount, longitude, latitude, surface albedo, all-electric long-wave vertical heating rate, all-short-wave vertical heating rate, and all-long-wave upward radiation at the top of the atmosphere, each row is assigned a unique number. And, a serial number for distinguishing the numerical forecast data stored in each row can be indicated. In this case, a unique number may be assigned to distinguish weather element information.

That is, the numerical forecast data set (KLAPS_DS) may be a set of data storing meteorological element information measured every hour for each specific grid. In the numerical forecast data set (KLAPS_DS), a cell formed by meeting a row and a column may indicate meteorological element information measured every hour for each specific grid.

The pre-learning data classification module 100 may extract pre-learning data from the numerical forecast data set (KLAPS_DS).

The pre-learning data classification module 100 may extract independent meteorological element information (IW) selected by the user from among the meteorological element information, and process it into a pre-input data set PREV_IN_DS.

The pre-input data set PREV_IN_DS may be a subset configured by selecting a column corresponding to the independent meteorological element information IW from the numerical forecast data set KLAPS_DS. The dictionary input data set PREV_IN_DS may include one or more rows, and each row may represent the dictionary input data PREV_IN. Accordingly, the pre-input data PREV_IN may be processed in a form including independent meteorological element information IW.

The pre-learning data classification module 100 may extract dependent meteorological element information DW selected by the user from the meteorological element information, and may process it into a reference output data set COMP_OUT_DS.

The reference output data set COMP_OUT_DS may be a subset configured by selecting a column corresponding to the dependent meteorological element information DW from the numerical forecast data set KLAPS_DS. The reference output data set COMP_OUT_DS may include one or more rows, and each row may represent the reference output data COMP_OUT. Accordingly, the reference output data COMP_OUT may be processed in a form including dependent meteorological element information DW.

The independent meteorological element information (IW) may include 322 meteorological elements, for example, as shown in Table 1 below.

number Independent Weather Factor Information (IW) 1 to 39 Vertical barometric pressure for each floor from 1st to 39th floor 40 ~ 78 Vertical temperature per floor from 1st to 39th floor 79 ~ 117 Vertical water vapor volume by floor 1 to 39 118 ~ 156 Vertical ozone amount by floor 1 to 39 157 ~ 195 Vertical cloud volume for each floor from 1st to 39th floor 196-234 Number of vertical clouds by floor 1 to 39 235-273 1st to 39th floor vertical cloud ice quantity 274-312 1st to 39th floor vertical snow and cold snow 313 Land/Marine classification 314 Hardness 315 Latitude 316 Date (month) 317 Prediction time 318 Surface temperature 319 Surface emissivity 320 Solar constant 321 Solar zenith angle 322 Surface albedo

The dependent meteorological element information DW may include 90 meteorological elements, for example, as shown in Table 2 below.

number Dependent Weather Factor Information (DW) 1 to 39 All-in-one long-wave vertical heating rate for each floor from 1st to 39th floor 40 ~ 78 All-in-one shortwave vertical heating rate for each floor from 1st to 39th floor 79 Total long-wave upward radiation at the top of the atmosphere 80 Blue sky long-wave upward radiation at the top of the atmosphere 81 Total long-wave upward radiation from the surface 82 Blue sky long-wave upward radiation from the surface 83 Total long-wave downward radiation from the surface 84 Blue-sky long-wave downward radiation from the surface 85 All shortwave upward radiation at the top of the atmosphere 86 Blue sky shortwave upward radiation at the top of the atmosphere 87 All shortwave upward radiation from the surface 88 Blue sky shortwave upward radiation from the surface 89 Total shortwave downward radiation from the surface 90 Clear shortwave downward radiation from the surface

The storage module 200 may store a pre-input data set (PREV_IN_DS) and a reference output data set (COMP_OUT_DS).

In addition, the storage module 200 may store one or more weight data and one or more bias data. The weight data may be divided into one or more first weight data W1 _ji and one or more second weight data W2 _qj. The bias data may be divided into one or more first bias data B1 _j and one or more second bias data B2 _q. Weight data and bias data can be initially set randomly.

In addition, the storage module 200 may store pre-output data PREV_OUT, forecast input data CAST_IN, and forecast output data CAST_OUT. The storage module 200 may include a memory for storing them.

The neural network module 300 may include an input layer 310, an output layer 320, and one or more hidden layers 330.

In FIG. 1C, the units are shown in a circle shape, and as shown, the input layer 310 may include one or more units, for example, N units. In this case, the N units may be the same as the number of independent meteorological element information IW, and each of the pre-input data PREV_IN and the forecast input data CAST_IN may include N independent meteorological element information IW.

The output layer 320 may include one or more units, for example, M units. At this time, the M units may be the same as the number of dependent weather element information (DW), and the pre-output data (PREV_OUT), reference output data (COMP_OUT), and forecast output data (CAST_OUT) are each M dependent weather element information (DW). ) Can be included.

The concealment layer 330 may include one or more units, for example, K units, and the concealment layer 330 may include as many as H units.

A relationship line may be formed between the N units of the input layer 310 and the K units of the hidden layer 330. In addition, a relationship line may be formed between the K units of the hidden layer 330 and the M units of the output layer 320.

The storage module 200 may include as many as N x K number of relationship lines between the unit of the input layer 310 and the unit of the hidden layer 330, and the first weight data W1 _ji Data B1 _j may be provided as many as K, which is the number of units of the hidden layer 330.

In addition, the storage module 200 may include the second weight data W2 _qj as much as K x M, which is the number of relationship lines between the unit of the output layer 320 and the unit of the hidden layer 330, and the second bias data (B2 _q ) may be provided as many as M, which is the number of units of the output layer 320.

The units included in the input layer 310, the output layer 320, and the hidden layer 330 may have a relationship as shown in Equation 1 below. When Equation 1 is applied, one hidden layer 330 may be provided.

는 사전 입력 데이터(PREV_IN) 또는 예보 입력 데이터(CAST_IN),

는 제 1 가중치 데이터,

는 제 1 편향 데이터,

는 제 2 가중치 데이터,

는 제 2 편향 데이터,

는 활성화 함수,

는 사전 출력 데이터(PREV_OUT) 또는 예보 출력 데이터(CAST_OUT)일 수 있다. i 는 입력층(310)의 개별 유닛을 나타낼 수 있고, j 는 은닉층(330)의 개별 유닛을 나타낼 수 있으며, q 는 출력층(320)의 개별 유닛을 나타낼 수 있다.In Equation 1

Is pre-input data (PREV_IN) or forecast input data (CAST_IN),

Is the first weight data,

Is the first biased data,

Is the second weight data,

Is the second biased data,

Is the activation function,

May be pre-output data PREV_OUT or forecast output data CAST_OUT. i may represent an individual unit of the input layer 310, j may represent an individual unit of the hidden layer 330, and q may represent an individual unit of the output layer 320.

The activation function may be one of hyperbolic tangent (tanh), sigmoid, ReLU, and softmax, but is not limited thereto and may be another function.

Accordingly, after the weight data and bias data are mapped, the neural network module 300 may calculate the pre-output data PREV_OUT when the pre-input data PREV_IN is mapped, and output the forecast when the forecast input data CAST_IN is mapped. Data (CAST_OUT) can be calculated.

The control module 400 uses the first weight data W1 _ji , the first bias data B1 _j , the second weight data W2 _qj , and the second bias data B2 _q stored in the storage module 200 into a neural network. It can be mapped to the module 300.

For example, the control module 400 includes N x K first weight data W1 _ji stored in the storage module 200, and N input layer 310 units and hidden layer 330 units of the neural network module 300 Each of the K relationship lines can be mapped to N x K lines. In addition, the control module 400 may _{map the K pieces of first bias data B1 j} to K units of the hidden layer 330, respectively. The control module 400 may _{map K x M second weight data W2 qj} to K x M relationship lines between K hidden layer 330 units and M output layer 320 units, respectively. In addition, the control module 400 may _{map M pieces of second deflection data B2 q} to M units of the output layer 320, respectively.

The control module 400 may map individual pre-input data PREV_IN corresponding to a row of the pre-input data set PREV_IN_DS stored in the storage module 200 to the input layer 310 unit of the neural network module 300. .

For example, the control module 400 may map N independent meteorological element information IW constituting the individual pre-input data PREV_IN to N units of the input layer 310, respectively. In addition, the neural network module 300 may calculate the pre-output data PREV_OUT for each M unit of the output layer 320 by performing an operation according to Equation 1.

The control module 400 compares the pre-output data PREV_OUT with the individual reference output data COMP_OUT stored in the storage module 200, and then compares the first weight data W1 _{ji stored in the storage module 200 and the first weight data.} The bias data B1 _j , the second weight data W2 _qj , and the second bias data B2 _q may be updated and stored again in the storage module 200.

The input module 500 may receive the forecast input data CAST_IN from the outside of the copy physical parameterization emulation system 1 and transmit it to the storage module 200.

The forecast input data CAST_IN may include N independent weather element information IW, and the storage module 200 may additionally store the forecast input data CAST_IN.

The control module 400 may map the forecast input data CAST_IN stored in the storage module 200 to the input layer 310 unit of the neural network module 300. The neural network module 300 may perform an operation according to Equation 1 to calculate a result displayed by M units of the output layer 320 as forecast output data CAST_OUT.

The forecast output data CAST_OUT may include M dependent weather element information DW. The control module 400 may transmit the forecast output data CAST_OUT to the storage module 200, and the storage module 200 may additionally store the forecast output data CAST_OUT.

The output module 600 may receive the forecast output data CAST_OUT from the storage module 200 and output it to the outside of the copy physical parameterization emulation system 1.

The forecast output data (CAST_OUT) can be output in the form of a database or a file, and can be transmitted to the outside through a computer network.

FIG. 2A is a flow chart briefly showing the steps of building a neural network in the emulation method for radiative physics parameterization according to an embodiment of the present invention, FIG. 2B is a flowchart briefly showing steps for reinforcing a neural network, and FIG. 2C is a radiative physics parameterization It is a flow chart that outlines the steps to emulate the technique.

The radiative physics parameterization emulation method can be divided into steps of building a neural network (S1 to S10), reinforcing a neural network (Q1 to Q7), and a step of emulating a radiative physics parameterization technique (R1 to R6).

In the steps of constructing a neural network (S1 to S10), weight data and bias data are repeatedly modeled to model the relationship between independent meteorological element information (IW) and dependent meteorological element information (DW) of the numerical forecast data set (KLAPS_DS). It is to obtain and study in advance.

The first step (S1) of building a neural network (S1 to S10) is a step in which the pre-learning data classification module 100 receives a numerical forecast data set (KLAPS_DS) of an ultra-short-term forecast model (KLAPS).

The numerical forecast data set KLAPS_DS may have the same matrix structure as described in FIG. 1B.

In the second step (S2), the pre-learning data classification module 100 extracts the independent meteorological element information DW from the numerical forecast data set KLAPS_DS, processes the pre-input data set PREV_IN_DS, and processes the dependent meteorological element information ( IW) is extracted, the reference output data set (COMP_OUT_DS) is processed, and then transferred to the storage module 200, respectively.

Individual (row-by-row) pre-input data PREV_IN included in the pre-input data set PREV_IN_DS may include N independent meteorological element information IW. In addition, the individual (row-by-row) reference output data COMP_OUT included in the reference output data set COMP_OUT_DS may include M dependent weather element information DW.

The independent meteorological element information IW and the dependent meteorological element information DW may include, for example, one or more meteorological element information shown in Tables 1 and 2, respectively.

The third step S3 is a step in which the storage module 200 stores the pre-input data set PREV_IN_DS and the reference output data set COMP_OUT_DS.

Accordingly, the storage module 200 may store individual pre-input data PREV_IN and individual reference output data COMP_OUT.

In the fourth step (S4), the storage module 200 randomizes _{the first weight data W1 ji} and the first bias data B1 _j , the second weight data W2 _qj and the second bias data B2 _q. This is the step to set and save.

This can be seen as a step of initializing weight data and bias data.

The fifth step S5 is a step in which the control module 400 sets the first operation time T1.

The copy physics parameterization emulation system 1 calculates the pre-output data (PREV_OUT) during the set first operation time (T1) and compares it with the reference output data (COMP_OUT) to _{bias the weight data (W1 ji} , W2 _qj ). You can repeat the process of updating data (B1 _j , B2 _{q ).}

Accordingly, the weight data and the bias data can be gradually updated so as to be similar to the result calculated by the copy physics parameterization technique (RRTMG-K) of the numerical forecast data set (KLAPS_DS).

The first operation time T1 may be, for example, 6 hours, but is not limited thereto and may exceed or be less than 6 hours.

In the sixth step (S6), the control module 400 includes first weight data W1 _ji , first bias data B1 _j , second weight data W2 _qj , and second bias stored in the storage module 200. This is a step of mapping the data B2 _q to the neural network module 300.

The seventh step S7 is a step in which the control module 400 maps the pre-input data PREV_IN stored in the storage module 200 to the neural network module 300.

After reflecting the weight data and bias data updated during the first calculation time T1 to the neural network module 300, the pre-input data PREV_IN may be input to the neural network module 300 to calculate output data.

는 사전 입력 데이터(PREV_IN)일 수 있고,

는 사전 출력 데이터(PREV_OUT)일 수 있다.In the eighth step S8, the neural network module 300 calculates the pre-output data PREV_OUT according to Equation 1. At this time, in Equation 1

May be pre-input data (PREV_IN),

May be pre-output data PREV_OUT.

)는 하이퍼볼릭 탄젠트(hyperbolic tangent, tanh), 시그모이드(sigmoid), 렐루(ReLU), 소프트맥스(softmax) 중 하나일 수 있으나, 이에 한정하지 않고 다른 함수일 수 있음은 동일하다.In Equation 1, the activation function (

) May be one of hyperbolic tangent (tanh), sigmoid, ReLU, and softmax, but is not limited thereto and may be another function.

In the ninth step (S9), after the control module 400 compares the pre-output data PREV_OUT and the reference output data COMP_OUT, the first weight data W1 _ji and the first bias data B1 _j 2 This is a step of updating the weight data W2 _qj and the second bias data B2 _q.

In the tenth step (S10), after determining whether the control module 400 has reached the first calculation time (T1), if the first calculation time (T1) is reached, the pre-learning is terminated, and the first calculation time (T1) is reached. If not, it is a step of repeatedly executing the sixth to ninth steps (S6 to S9).

In an embodiment of the present invention, after calculating the pre-output data PREV_OUT by inputting a plurality of pre-input data PREV_IN as pre-learning data, comparing the difference with the reference output data COMP_OUT, and then comparing the difference with the weight data Data can be updated.

By repeatedly executing these steps, weight data and bias data that can minimize the difference between the pre-output data PREV_OUT and the reference output data COMP_OUT can be obtained.

In the step of reinforcing the neural network (Q1 to Q7), after inputting pre-input data (PREV_IN) to the constructed neural network module 300, pre-output data (PREV_OUT) is calculated, and compared with reference output data (COMP_OUT). In the steps of strengthening the neural network (Q1 to Q7), a first step (Q1) is a step in which the control module 400 sets a second calculation time (T2).

The second operation time T2 may be, for example, 6 hours, but is not limited thereto and may exceed or be less than 6 hours.

_{The second step (Q2) is the first weight data (W1 ji} ), the first bias data (B1 _j ), and the second weight data, updated in the steps of building a neural network (S1 to S10) and stored in the storage module 200. This is a step in which the control module 400 maps (W2 _qj ) and the second bias data B2 _{q to the neural network module 300.}

The weight data and the bias data may be data obtained immediately before completing the steps of constructing a neural network (S1 to S10).

The third step Q3 is a step in which the control module 400 maps the pre-input data PREV_IN stored in the storage module 200 to the neural network module 300.

는 사전 입력 데이터(PREV_IN)일 수 있고,

는 사전 출력 데이터(PREV_OUT)일 수 있다. 그리고 활성화 함수(

)는 신경망을 구축하는 단계(S1 ~ S10)와 동일할 수 있으나, 이에 한정하지 않고 다른 함수일 수도 있다.The fourth step Q4 is a step in which the neural network module 300 calculates the pre-output data PREV_OUT according to Equation 1. At this time, in Equation 1

May be pre-input data (PREV_IN),

May be pre-output data PREV_OUT. And the activation function (

) May be the same as the steps of building a neural network (S1 to S10), but is not limited thereto and may be another function.

The fifth step (Q5) is a step in which the control module 400 compares the pre-output data PREV_OUT and the reference output data COMP_OUT to obtain a similarity of a calculation result indicating accuracy, and calculates an operation speed.

At this time, the similarity of the calculation result is the similarity between the reference output data COMP_OUT extracted from the numerical forecast data set (KLAPS_DS) and the pre-output data PREV_OUT calculated through the neural network module 300 constructed in an embodiment of the present invention. Can be seen as.

In the sixth step (Q6), it is determined whether the control module 400 has reached the second calculation time (T2), and if the second calculation time (T2) is reached, the step is terminated, and the second calculation time (T2) is reached. If not, the second to fifth steps (Q2 to Q5) are repeatedly executed.

In the seventh step (Q7), the control module 400 determines whether the similarity and the calculation speed of the calculation result obtained in the fifth step (Q5) are within the range set as the target, and then ends the step when the result is within the target range. If it is out of the target range, it is the step of rebuilding the neural network.

When the similarity of the calculation result and the calculation speed are out of the target range, the pre-input data set (PREV_IN_DS) and the reference output data set (COMP_OUT_DS) are changed, and the number of units included in the hidden layer 330 of the neural network module 300 ( After K) is changed, the neural network module 300 may be newly constructed by executing the steps (S1 to S10) of building the neural network again.

In an embodiment of the present invention, the neural network module 300 may be newly constructed until the similarity of the calculation result and the calculation speed reach a target range.

In the step of emulating the radiation physics parameterization technique (R1 to R6), after inputting independent meteorological element information (IW) to the constructed neural network module 300, dependent meteorological element information (DW) is calculated to obtain a numerical forecast result. For.

The first step (R1) of emulating the copy physical parameterization technique (R1 to R6) is a step in which the input module 500 receives the forecast input data CAST_IN and transmits the received forecast input data CAST_IN to the storage module 200.

The input module 500 may receive forecast input data CAST_IN through a computer network in the form of, for example, a database or a file.

The storage module 200 may store the forecast input data CAST_IN in the same way as the individual pre-input data CAST_IN.

In the second step (R2), the control module 400 includes first weight data W1 _ji , first bias data B1 _j , second weight data W2 _qj , and second bias stored in the storage module 200. This is a step of mapping the data B2 _q to the neural network module 300.

The weight data and bias data may be data updated through the steps of constructing a neural network (S1 to S10) and steps of reinforcing the neural network (Q1 to Q7).

The third step R3 is a step in which the control module 400 maps the forecast input data CAST_IN stored in the storage module 200 to the neural network module 300.

The fourth step R4 is a step in which the neural network module 300 calculates the forecast output data CAST_OUT according to Equation 1.

는 예보 입력 데이터(CAST_IN)일 수 있고,

는 예보 출력 데이터(CAST_OUT)일 수 있다. 활성화 함수(

)는 신경망을 구축하는 단계(S1 ~ S10)와 동일할 수 있으나, 이에 한정하지 않고 다른 함수일 수도 있다.At this time, in Equation 1

May be forecast input data (CAST_IN),

May be forecast output data (CAST_OUT). Activation function (

) May be the same as the steps of building a neural network (S1 to S10), but is not limited thereto and may be another function.

The fifth step R5 is a step in which the control module 200 transmits the forecast output data CAST_OUT to the storage module 200 and the storage module 200 stores it.

The sixth step R6 is a step in which the output module 600 outputs the forecast output data CAST_OUT stored in the storage module 200 to the outside of the copy physical parameterization emulation system.

The forecast output data (CAST_OUT) can be output in the form of a database or a file, and can be transmitted to the outside through a computer network.

3A and 3B are graphs showing a long-wave radiation heating rate and a short-wave radiation heating rate, respectively, when using the radiation physics parameterization technique of the ultra-short-term forecast model and the radiation physics parameterization emulation system according to an embodiment of the present invention. .

In the graph, the horizontal axis represents the time (hours), the vertical axis represents the height (km), and the color represents the average atmospheric radiant heating rate (K/day). And the graph shows the average atmospheric radiative heating rate from the surface to a height of 15 km over 6 hours. Such a graph can be viewed as representing the average atmospheric radiative heating rate at a specific time and at a specific height.

As shown in FIG. 3A, a graph (A) calculated by a radiation physics parameterization technique of an ultra-short-term forecast model and a graph (B) calculated by a radiation physics parameterization emulation system according to an embodiment of the present invention are the average atmospheric long wave It can be seen that there is little difference in the distribution of radiant heating rates.

That is, the radiative physics parameterization emulation system according to an embodiment of the present invention can calculate a long-wave radiant heating rate almost similar to the radiative physics parameterization technique of an ultra-short forecast model.

In addition, as shown in FIG. 3B, a graph (C) calculated by the radiative physics parameterization technique of the ultra-short-term forecast model and a graph (D) calculated by the radiative physics parameterization emulation system according to an embodiment of the present invention are also shown. It can be seen that the distribution of shortwave radiant heating rates is similar.

Accordingly, the radiative physics parameterization emulation system according to an embodiment of the present invention can calculate a short-wave radiant heating rate similar to the radiative physics parameterization technique of the ultra-short forecast model.

4A to 4F and 5A to 5F show the amount of long-wave radiation and the amount of short-wave radiation, respectively, when using the radiation physics parameterization technique of the ultra-short forecast model and when using the radiation physics parameterization emulation system according to an embodiment of the present invention. It is a graph.

In the graph, the horizontal axis represents time (hours), the vertical axis represents the horizontal direction (Horizontal) (km), and the color represents the amount of radiation, the flux (W/m2). And the graph shows the amount of radiation from the reference point (0 km) to a distance (location) of 25 km in the east direction and 25 km in the west direction for 6 hours. In this case, the reference altitude may be at the top of the atmosphere or near the surface. Such a graph can be viewed as representing the amount of radiation at a specific time and at a specific horizontal position.

As shown in Fig. 4A, comparing the graph of the amount of radiation from the top of the atmosphere in the upward direction of the whole sky, when using the radiation physics parameterization technique of the very short-term forecast model (E-1), and an embodiment of the present invention. When using the radiation physics parameterization emulation system according to (E-2), it can be seen that the radiation amount distribution is similar.

Clear sky upward direction (F-1, F-2) from the upper atmosphere of FIG. 4B, all stream upward direction (G-1, G-2) near the surface of FIG. 4C, and clear sky upward from the surface of FIG. 4D Directions (H-1, H-2), all-cheon downward direction (I-1, I-2) near the surface of Fig. 4e, and blue-cheon downward radiation amount graph (J-1, J-2) near the surface of Fig. 4f In comparison, it can be seen that there is no significant difference in the radiation amount distribution when using the radiative physics parameterization technique of the ultra-short-term forecast model and when the radiative physics parameterization emulation system according to an embodiment of the present invention is used.

That is, the radiation physics parameterization emulation system according to an embodiment of the present invention can calculate a long-wave radiation amount almost similar to the radiation physics parameterization technique of an ultra-short forecast model.

In addition, as shown in Figs. 5A to 5F, all streams are upward from the top of the atmosphere (K-1, K-2), clear streams are upward from the top of the atmosphere (L-1, L-2), and all streams are upward from the surface. Direction (M-1, M-2), upward direction of the clear stream near the surface (N-1, N-2), downward direction of the entire stream near the surface (O-1, O-2), the amount of radiation in the downward direction of the clear stream near the surface Comparing the graphs (P-1, P-2), there is no significant difference in the distribution of radiative quantities when using the radiative physics parameterization technique of the ultra-short forecast model and when using the radiative physics parameterization emulation system according to an embodiment of the present invention. Can be seen.

Accordingly, the radiation physics parameterization emulation system according to an embodiment of the present invention can calculate a short-wave radiation amount almost similar to the radiation physics parameterization technique of an ultra-short forecast model.

6 is a table comparing computation speed and similarity when using a radiative physics parameterization technique of an ultra-short forecast model and when using a radiative physics parameterization emulation system according to an embodiment of the present invention.

The table shows the relative computation speed of the radiative physics parameterization emulation system according to an embodiment of the present invention and the similarity of the calculation results, as compared with the radiative physics parameterization technique of the ultra-short-term forecast model.

It can be seen that as the number of units included in the hidden layer 330 decreases from 300 to 56, the computational speed significantly increases. (20.64 times → 100.36 times)

In addition, the long-wave and short-wave radiation heating rates, and the long-wave and short-wave radiation amounts calculated by the radiation physics parameterization emulation system according to an embodiment of the present invention are compared with the calculation results according to the radiation physics parameterization technique of the ultra-short-term forecast model, and the similarity is 86. You can see that it is more than %.

In addition, even if the number of units included in the hidden layer 330 is reduced, it can be seen that the similarity between the long-wave radiation amount and the short-wave radiation amount is almost the same.

As described above, the present invention can emulate the radiative physics parameterization technique by calculating a result similar to the radiative physics parameterization technique of the ultra-short forecast model. In addition, it is not necessary to directly calculate the equations executed in the radiative physics parameterization technique of the ultra-short forecast model, and since numerical forecast data can be calculated using an artificial neural network, the calculation speed can be increased.

Although the embodiments of the present invention have been described above, the present invention is not limited to the above embodiments, and various within the scope of the detailed description of the invention and the accompanying drawings as long as the effect is not impaired without departing from the spirit of the present invention. It can be changed and implemented. In addition, it is natural that such an embodiment falls within the scope of the present invention.

1: Radiative physics parameterization emulation system
100: pre-learning data classification module
200: storage module
300: neural network module
400: control module
500: input module
600: output module
S1 ~ S10: Steps to build a neural network
Q1 ~ Q7: Steps to strengthen the neural network
R1 ~ R6: Steps to emulate the radiative physics parameterization technique

Claims (14)

Receives the numerical forecast data set of the KLAPS, extracts independent meteorological element information for each row from the numerical forecast data set, processes pre-input data (PREV_IN), and extracts dependent meteorological element information for reference output data. A pre-learning data classification module processing (COMP_OUT);
A memory for storing at least one of the pre-input data PREV_IN and the reference output data COMP_OUT, first weight data, second weight data, first bias data, second bias data, and pre-output data PREV_OUT, respectively, A storage module including;
After the pre-input data PREV_IN and the first weight data, the second weight data, the first bias data, and the second bias data are mapped, the pre-output data PREV_OUT including the dependent meteorological element information A neural network module that calculates;
After comparing the pre-output data PREV_OUT and the reference output data COMP_OUT, the first weight data, the second weight data, the first bias data, and the second bias data are updated and stored in the storage module. Including a control module to store,
The control module maps the pre-input data PREV_IN, the first weight data, the second weight data, the first bias data, and the second bias data to the neural network module to map the pre-output data PREV_OUT. And, after comparing the pre-output data PREV_OUT and the reference output data COMP_OUT, the first weight data, the second weight data, the first bias data, and the second bias data are updated. A copy physics parameterization emulation system that repeats the operation to be stored in the storage module for a first operation time.
The method of claim 1,
The neural network module includes an input layer having N units, H hidden layers having K units, and an output layer having M units.
The method of claim 2,
After the first calculation time has elapsed, the control module sends the pre-input data PREV_IN, the first weight data, the second weight data, the first bias data, and the second bias data to the neural network module. After mapping the pre-output data (PREV_OUT) to be calculated, the similarity is calculated by comparing the pre-output data (PREV_OUT) and the reference output data (COMP_OUT), and the operation of calculating the operation speed is repeated for the second operation time. ,
After the second calculation time has elapsed, the control module determines whether the similarity and the calculation speed are included in a target range, and if the target range is out of the target range, the pre-input data PREV_IN and the reference output data COMP_OUT And, after changing the number of K units in the hidden layer, and rebuilding the neural network module.
The method of claim 3,
An input module that receives forecast input data (CAST_IN) including the independent meteorological element information from the outside of the copy physical parameterization emulation system and transmits it to the storage module,
The storage module further comprises an output module for receiving the forecast output data (CAST_OUT) including the dependent meteorological element information and outputting it to the outside of the copy physical parameterization emulation system,
The storage module stores the forecast input data (CAST_IN) and the forecast output data (CAST_OUT),
The neural network module calculates the forecast output data CAST_OUT after the forecast input data CAST_IN and the first weight data, the second weight data, the first bias data, and the second bias data are mapped. Radiative physics parameterization emulation system.
The method of claim 4,
The numerical forecast data set includes at least one cell partitioned into a row representing individual numerical forecast data and a column representing meteorological element information to which a unique number is assigned.
The method of claim 5,
In the storage module,
The pre-input data (PREV_IN) and the forecast input data (CAST_IN) each include N pieces of independent meteorological element information,
The pre-output data PREV_OUT, the reference output data COMP_OUT, and the forecast output data CAST_OUT each include M pieces of dependent weather element information,
The first weight data includes N x K, the second weight data includes K x M,
The first bias data includes K pieces, and the second bias data includes M pieces.
The method of claim 6,
The number of N is 322,
The independent meteorological element information included in the pre-input data (PREV_IN) and the forecast input data (CAST_IN),
Vertical atmospheric pressure (IW1 to IW39), vertical temperature (IW40 to IW78), vertical water vapor amount (IW79 to IW117), vertical ozone amount (IW118 to IW156), vertical cloud volume (IW157 to IW195), vertical of the first to 39th layers, respectively Cloud water quantity (IW196~IW234), vertical cloud ice quantity (IW235~IW273), vertical snow and cold weather (IW274~IW312),
Land and ocean classification (IW313), longitude (IW314), latitude (IW315), date (IW316), predicted time (IW317), surface temperature (IW318), surface emissivity (IW319), solar constant (IW320), solar zenith angle ( IW321), radiative physics parameterization emulation system, which is ground albedo (IW322).
The method of claim 7,
The number of M is 90,
The dependent weather element information included in the pre-output data PREV_OUT, the reference output data COMP_OUT, and the forecast output data CAST_OUT,
The all-electric long-wave vertical heating rate (DW1-DW39), the all-electric short-wave vertical heating rate (DW40-DW78) of the first to 39th layers, respectively, and
Total long-wave upward radiation from the top of the atmosphere (DW79), Blue sky long-wave upward radiation from the top of the atmosphere (DW80), All-sky long-wave upward radiation from the surface (DW81), Blue sky long-wave upward radiation from the surface (DW82), all the sunlight from the surface Longwave Downward Radiation (DW83), Clear Long Wave Downward Radiation from the Surface (DW84), All-Short Wave Upward Radiation at the Top of the Air (DW85), Clear Short-wave Upward Radiation from the Top of the Air (DW86), All-Short Upward Radiation from the Surface ( DW87), blue shortwave upward radiation from the surface (DW88), total shortwave downward radiation from the surface (DW89), and blue shortwave downward radiation from the surface (DW90).
The method of claim 8,
The number of H is 1,
The neural network module is a copy physics parameterization emulation system that calculates the pre-output data PREV_OUT or the forecast output data CAST_OUT according to Equation 1 below.
[Equation 1]

(

: The prior input data (PREV_IN) or the forecast input data (CAST_IN) /

: The first weight data /

: The first bias data /

: The second weight data /

: The second bias data /

: Activation function /

: The pre-output data (PREV_OUT) or the forecast output data (CAST_OUT) ).
The method of claim 9,
The activation function (

) Is the hyperbolic tangent function (tanh), a radiative physics parametric emulation system.
The method of claim 10,
The pre-learning data classification module is a copy physics parameterization emulation system that receives the numerical forecast data set through an open API of a weather data open portal or in the form of a NetCDF file.
(S1) receiving, by the pre-learning data classification module, a numerical forecast data set of an ultra-short-term forecast model (KLAPS);
(S2) The pre-learning data classification module extracts N independent meteorological element information from the numerical forecast data set to process the pre-input data (PREV_IN), and extracts M dependent meteorological element information to obtain reference output data (COMP_OUT). ) Processing, and transferring it to a storage module;
(S3) storing, by the storage module, the pre-input data PREV_IN and the reference output data COMP_OUT;
(S4) storing, by the storage module, randomly setting and storing first weight data, first bias data, second weight data, and second bias data;
(S5) setting, by the control module, a first operation time;
(S6) the control module mapping the first weight data and the first bias data, the second weight data and the second bias data to a neural network module;
(S7) mapping the pre-input data (PREV_IN) to the neural network module by the control module;
(S8) calculating, by the neural network module, pre-output data PREV_OUT according to Equation 1 below;
[Equation 1]

(

: The above pre-input data (PREV_IN) /

: The first weight data /

: The first bias data /

: The second weight data /

: The second bias data /

: Activation function /

: The pre-output data (PREV_OUT))
(S9) After the control module compares the pre-output data PREV_OUT and the reference output data COMP_OUT, the first weight data and the first bias data, the second weight data and the second bias data Updating to;
(S10) After determining whether the control module has reached the first operation time, it ends if the first operation time has been reached, and repeats the steps (S6) to (S9) if the first operation time has not been reached. Steps to run by
Radiative physics parameterization emulation method comprising a.
The method of claim 12,
(Q1) setting, by the control module, a second operation time;
(Q2) the control module mapping the first weight data, the first bias data, the second weight data and the second bias data to the neural network module;
(Q3) mapping the pre-input data (PREV_IN) to the neural network module by the control module;
(Q4) calculating, by the neural network module, the pre-output data PREV_OUT according to Equation 1;
(Q5) calculating, by the control module, a degree of similarity by comparing the pre-output data PREV_OUT and the reference output data COMP_OUT, and calculating an operation speed;
(Q6) determining whether the control module has reached the second operation time, and if the second operation time has not been reached, repeating the steps (Q2) to (Q5);
(Q7) When the second calculation time is reached, the control module determines whether the similarity and the calculation speed are included in the target range, and ends when the similarity and the calculation speed are included in the target range. After changing (PREV_IN) and the reference output data (COMP_OUT), changing the number of units of the hidden layer of the neural network module, and executing steps (S1) to (S10) again, a copy physics parameterization emulation method.
The method of claim 13,
(R1) receiving, by an input module, forecast input data (CAST_IN), and transmitting it to the storage module;
(R2) the control module mapping the first weight data and the first bias data, the second weight data and the second bias data to the neural network module;
(R3) mapping the forecast input data (CAST_IN) to the neural network module by the control module;
(R4) the neural network module in Equation 1

Input as the forecast input data (CAST_IN),

Calculating as forecast output data (CAST_OUT);
(R5) the control module transferring the forecast output data (CAST_OUT) to the storage module;
(R6) an output module outputting the forecast output data CAST_OUT stored in the storage module
Radiative physics parameterization emulation method further comprising.

Images