KR102631539B1 - Method for predicting precipitation based on deep learning and apparatus implementing the same method - Google Patents

Abstract

A method performed by a computing device according to an embodiment of the present invention includes generating training data consisting of a series of consecutive frames using rainfall data and numerical meteorological model data, and sending the training data to an encoder-decoder. -decoder) structure, generating a precipitation prediction model, and inputting prediction target data into the precipitation prediction model to output precipitation prediction information.

Description

Deep learning-based precipitation prediction method and device for implementing the same {METHOD FOR PREDICTING PRECIPITATION BASED ON DEEP LEARNING AND APPARATUS IMPLEMENTING THE SAME METHOD}

The present invention relates to a deep learning-based precipitation prediction method, and a device for implementing the same. More specifically, it relates to a deep learning-based precipitation prediction method for predicting precipitation using a deep learning algorithm, and a device for implementing the same. will be.

Therefore, in order to predict ultra-short-term precipitation, technology that can increase the accuracy of prediction based on artificial intelligence is required. Additionally, when using various weather data in addition to rainfall radar data as learning data for precipitation prediction, a technology that can unify the format of data with different structures is required.

Japanese Patent Publication No. 2020-91171 (published on June 11, 2020)

The technical problem to be solved by the present invention is to provide a deep learning-based precipitation prediction method that can predict precipitation in a short period of time by applying learning data consisting of continuous frames to a deep learning model, and a device for implementing the same.

Another technical problem that the present invention aims to solve is to provide a deep learning-based precipitation prediction method that can increase the accuracy of prediction by using a multi-variable input method for precipitation prediction, and a device for implementing the same.

Another technical problem that the present invention aims to solve is a deep learning-based precipitation prediction method that can obtain more sophisticated prediction results by applying the prediction results by the encoder-decoder neural network to a generative adversarial network (GAN) in the precipitation prediction process; And it provides a device to implement this.

The technical problems of the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned can be clearly understood by those skilled in the art from the description below.

A method performed by a computing device according to an embodiment of the present invention for solving the above technical problem includes generating learning data consisting of a series of consecutive frames using rainfall data and numerical meteorological model data, It includes generating a precipitation prediction model by applying learning data to a neural network with an encoder-decoder structure, and inputting prediction target data into the precipitation prediction model to output precipitation prediction information.

As an example, the rainfall data may be a rainfall radar composite image.

As an embodiment, the step of generating learning data consisting of a series of consecutive frames using the rainfall data and numerical meteorological model data includes different spatial grids and temporal resolutions between the rainfall radar composite image and the numerical meteorological model data. A matching step may be included.

In one embodiment, the step of matching different spatial grids and temporal resolutions between the rainfall radar composite image and the meteorological numerical model data includes converting the rainfall radar composite image into hourly precipitation data through time integration, and the meteorological meteorological model data. It may include interpolating the numerical model data into the same spatial grid structure as the rainfall radar composite image.

As an embodiment, the step of generating learning data consisting of a series of consecutive frames using the rainfall data and numerical meteorological model data includes generating a multidimensional array consisting of consecutive frames of a first period and forming the learning data. The multidimensional array may be generated based on the number of frames, the number of spatial grids, and the number of input variables.

As an embodiment, the step of generating a multidimensional array composed of consecutive frames of the first period and configuring it as learning data includes using as learning data frames in which the number of spatial grids with rainfall values greater than a predetermined ratio among the total number of spatial grids is greater than a predetermined ratio. It may include steps.

As an embodiment, the step of generating a multidimensional array composed of consecutive frames of the first period and configuring it as learning data includes using the frames of the second period corresponding to the first half of the consecutive frames of the first period as an input sequence. It may include the step of configuring and configuring frames of the third period corresponding to the latter half as an output sequence.

As an embodiment, the step of generating a precipitation prediction model by applying the learning data to a neural network with an encoder-decoder structure includes extracting features by compressing the input sequence through the encoder. and restoring the compressed features to detailed information corresponding to the output sequence through the decoder.

As an embodiment, the step of generating a precipitation prediction model by applying the learning data to a neural network with an encoder-decoder structure includes MSE (Mean Squared) as a loss function in the process of learning the neural network. Error), Mean Absolute Error (MAE), Balanced MSE (Mean Squared Error), Balanced MAE (Mean Absolute Error), and SSIM (Scale Structural Similarity).

As an embodiment, the step of generating a precipitation prediction model by applying the learning data to a neural network with an encoder-decoder structure constitutes a neural network with a branched structure depending on the type of data included in the learning data. It may include steps.

As an embodiment, the step of generating a precipitation prediction model by applying the learning data to a neural network with an encoder-decoder structure includes a generative adversarial network (GAN) consisting of a generator and a discriminator. It may include adding a Generative Adversarial Network) to the neural network of the encoder-decoder structure.

As an embodiment, the step of adding a generative adversarial network (GAN) consisting of the generator and the discriminator to the neural network of the encoder-decoder structure includes the prediction information output from the decoder. It may include using as output information of the generator, and inputting the output information of the generator and actual precipitation information into the separator to determine the accuracy of the output information.

In order to solve the above technical problem, a computer-readable non-transitory recording medium according to an embodiment of the present invention stores a computer program that causes a computer to perform the method.

In order to solve the above technical problem, a method performed by a computing device according to an embodiment of the present invention includes the operation of receiving information about a precipitation prediction model from the external device, and obtaining prediction target data by user input. An operation, inputting the prediction target data into the precipitation prediction model, and outputting precipitation prediction information from the precipitation prediction model, wherein the precipitation prediction model is a series of data using rainfall data and numerical meteorological model data. It was created by learning a neural network with an encoder-decoder structure using training data consisting of consecutive frames.

In one embodiment, the learning data is configured by generating a multidimensional array consisting of consecutive frames of a first period, and the multidimensional array can be generated based on the number of frames, the number of spatial grids, and the number of input variables. there is.

In one embodiment, the precipitation prediction model is generated by learning by adding a generative adversarial network (GAN) consisting of a generator and a discriminator to the neural network of the encoder-decoder structure. You can.

In order to solve the above technical problem, a rainfall prediction device according to an embodiment of the present invention uses a communication unit that communicates with an external server, rainfall data collected from the external server, and numerical meteorological model data to form a series of continuous frames. A learning unit that generates configured learning data and applies the learning data to a neural network with an encoder-decoder structure to generate a precipitation prediction model, and inputs prediction target data into the precipitation prediction model to generate precipitation prediction information. It includes a prediction unit that outputs.

In order to solve the above technical problem, a precipitation prediction device according to an embodiment of the present invention includes one or more processors, a communication interface for communicating with an external device, a memory for loading a computer program executed by the processor, and It includes storage for storing the computer program, wherein the computer program includes an operation of receiving information about a precipitation prediction model from the external device, an operation of acquiring prediction target data by a user input, and predicting precipitation by using the prediction target data. Includes instructions for performing an operation of inputting to a model and an operation of outputting precipitation prediction information from the precipitation prediction model, wherein the precipitation prediction model is a series of continuous operations using rainfall data and numerical meteorological model data. It was created by learning a neural network with an encoder-decoder structure using training data composed of frames.

1 is a conceptual diagram according to an embodiment of the present invention.
Figure 2 is a block diagram showing the configuration of a precipitation prediction device according to an embodiment of the present invention.
Figure 3 is a block diagram showing the configuration of a precipitation prediction device according to another embodiment of the present invention.
4 to 7 are flowcharts for explaining a precipitation prediction method according to another embodiment of the present invention.
Figure 8 is an example showing the structure of learning data according to some embodiments of the present invention.
Figure 9 is an example illustrating an encoder-decoder neural network structure for learning according to some embodiments of the present invention.
Figure 10 is an example of a loss function used in a learning process according to some embodiments of the present invention.
Figure 11 is an example of applying a branched structure according to the type of input data in the encoder according to some embodiments of the present invention.
Figure 12 is an example of applying a generative adversarial network (GAN) as an additional configuration to the encoder-decoder neural network according to some embodiments of the present invention.
Figure 13 is an example of using statistical indicators for learning verification according to some embodiments of the present invention.
Figure 14 is a hardware configuration diagram of an example computing device capable of implementing methods according to some embodiments of the present invention.

Hereinafter, preferred embodiments of the present disclosure will be described in detail with reference to the attached drawings. The advantages and features of the present disclosure and methods for achieving them will become clear by referring to the embodiments described in detail below along with the accompanying drawings. However, the technical idea of the present disclosure is not limited to the following embodiments and may be implemented in various different forms. The following examples are merely intended to complete the technical idea of the present disclosure and to be used in the technical field to which the present disclosure belongs. It is provided to fully inform those skilled in the art of the scope of the present disclosure, and the technical idea of the present disclosure is only defined by the scope of the claims.

When adding reference numerals to components in each drawing, it should be noted that identical components are given the same reference numerals as much as possible even if they are shown in different drawings. Additionally, in describing the present disclosure, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present disclosure, the detailed description will be omitted.

Unless otherwise defined, all terms (including technical and scientific terms) used in this specification may be used with meanings that can be commonly understood by those skilled in the art to which this disclosure pertains. Additionally, terms defined in commonly used dictionaries are not interpreted ideally or excessively unless clearly specifically defined. The terminology used herein is for the purpose of describing embodiments and is not intended to limit the disclosure. As used herein, singular forms also include plural forms, unless specifically stated otherwise in the context.

Additionally, in describing the components of the present disclosure, terms such as first, second, A, B, (a), and (b) may be used. These terms are only used to distinguish the component from other components, and the nature, sequence, or order of the component is not limited by the term. When a component is described as being “connected,” “coupled,” or “connected” to another component, that component may be directly connected or connected to that other component, but there is another component between each component. It will be understood that elements may be “connected,” “combined,” or “connected.”

As used in the specification, “comprises” and/or “comprising” refers to the presence of one or more other components, steps, operations and/or elements. or does not rule out addition.

Hereinafter, several embodiments of the present disclosure will be described in detail with reference to the attached drawings.

1 is a conceptual diagram according to an embodiment of the present invention. Referring to FIG. 1, the precipitation prediction device 1 according to an embodiment of the present invention predicts precipitation using a precipitation prediction model 31 generated by performing deep learning using learning data 2. Prediction information (5) can be output.

In the example shown, the learning data (2) consists of rainfall data and numerical meteorological model data, and the learning data (2) is used to learn a neural network (3) with an auto-encoder structure, a deep learning algorithm. The precipitation prediction model 31 can be created by. At this time, the neural network 3 with an auto-encoder structure may be composed of an encoder and a decoder.

The rainfall data constituting the learning data (2) can be prepared as a rainfall radar composite image provided by the Korea Meteorological Administration.

In addition, numerical meteorological model data is prepared as data used in a numerical forecast model, which is a type of computer program created to predict future weather using meteorological observation data. In particular, variables known to play an important role in the occurrence of precipitation are used. It can be targeted. By way of example, numerical meteorological model data may include dynamical and thermodynamic variables related to precipitation occurrence. As an example, numerical meteorological model data may be generated through a global numerical meteorological model or a local numerical meteorological model.

For precipitation prediction, when the prediction target data as input information (4) is input to the precipitation prediction model (31), a prediction result is generated using the precipitation prediction model (31), and precipitation prediction information (5) is generated as a prediction result. can be printed. Here, the prediction target data, which is the input information 4, may include a rainfall radar composite image at a reference point, dynamical and thermodynamic variables of a numerical meteorological model, etc.

According to the embodiment of the present invention as described above, the degree of future precipitation can be predicted from past rainfall status using deep learning technology. In particular, according to an embodiment of the present invention, the degree of ultra-short-term precipitation occurrence within, for example, 6 hours from a reference point can be predicted.

Figure 2 is a block diagram showing the configuration of a precipitation prediction device according to an embodiment of the present invention. Referring to FIG. 2, the precipitation prediction device 1 according to an embodiment of the present invention includes a learning unit 11 and a prediction unit 12, an external server 20, and a user terminal 10 and a network. can be connected through The precipitation prediction device 1 processes a request for creation and analysis of a deep learning model for precipitation prediction received from the user terminal 10 and provides the results.

The external server 20 may be implemented as each server that provides rainfall radar synthetic images and numerical meteorological model data, and the data provided therefrom may be collected to form learning data 2. At this time, the numerical meteorological model data included in the learning data 2 may be generated through a global numerical meteorological model or a local numerical meteorological model, and may include dynamical variables and thermodynamic variables related to the occurrence of precipitation.

The learning unit 11 is composed of a collection module 111, a preprocessing module 112, and a learning module 113, of which the collection module 111 is a rainfall radar composite image and weather provided from an external server 20. Numerical model data is collected to form learning data (2) and stored in the database.

The preprocessing module 112 reads the learning data 2 stored in the database by the collection module 111 and converts the learning data 2 into a form capable of deep learning.

As an example, the preprocessing module 112 may match different spatial grids and temporal resolutions between the rainfall radar synthetic image and the numerical meteorological model data. To this end, the preprocessing module 112 can convert the rainfall radar composite image into hourly rainfall data through time integration and perform the process of interpolating the meteorological numerical model data into the same spatial grid structure as the rainfall radar composite image. there is.

In addition, the preprocessing module 112 generates a multi-order sequence consisting of consecutive frames during the first period, and configures the frames during the second period corresponding to the first half of the consecutive frames of the first period as an input sequence, Frames during the third period corresponding to the second half can be configured as an output sequence. At this time, the multi-order array can be created based on the number of frames, the number of spatial grids in the north-south direction, the number of spatial grids in the east-west direction, and the number of input variables. For example, in consecutive frames at 1-hour intervals for 12 hours, the frames for the first 6 hours are configured as an input sequence, and the frames for the next 6 hours are configured as an output sequence to generate training data (2), It is possible to predict the occurrence of precipitation within, for example, 6 hours from the time point.

As an embodiment, in order to resolve the asymmetry of numerical data related to precipitation in the process of configuring the learning data 2, the preprocessing module 112 determines that the number of spatial grids in which the rainfall value is greater than the reference value within one frame is greater than a predetermined ratio. You can construct learning data by sampling frames.

The learning module 113 performs deep learning using the learning data 2 that has been preprocessed in the preprocessing module 112 and generates a precipitation prediction model 31 therefrom. At this time, as a deep learning algorithm used for learning, for example, an auto-encoder neural network consisting of an encoder and a decoder may be used. An autoencoder is a neural network that learns feature learning in the form of unsupervised learning to reduce dimensionality. The encoder receives input data, reduces the dimensionality and converts it into feature values, and the decoder converts the feature values into output data.

The prediction unit 12 is composed of an input module 121, a preprocessing module 122, and a prediction module 123, and results in precipitation prediction using prediction target data input from the user terminal 10 or provided from an external device. outputs.

In the input module 121, prediction target data provided from the user terminal 10 or an external device is input. At this time, the prediction target data may include rainfall status information such as a rainfall radar composite image at a reference point and dynamic and thermodynamic variables of a numerical meteorological model.

The preprocessing module 122 converts the prediction target data input from the input module 121 into a form that can be predicted based on a deep learning model. At this time, the conversion process of the prediction target data is the same as the process performed in the preprocessing module 112 of the learning unit 11.

The prediction module 123 loads the precipitation prediction model 31 generated in the learning module 113, inputs preprocessed prediction target data into the loaded precipitation prediction model 31, and generates a precipitation prediction result.

Additionally, the prediction module 123 loads the performance figures of the precipitation prediction model 31 generated in the learning module 113. The performance value of the precipitation prediction model 31 may include the value of a loss function calculated in a deep learning process. Here, the loss function is calculated using at least one of, for example, Mean Squared Error (MSE), Mean Absolute Error (MAE), Balanced Mean Squared Error (MSE), Balanced Mean Absolute Error (MAE), and Scale Structural Similarity (SSIM). It can be.

The precipitation prediction device 1 according to the embodiment of the present invention as described above can be implemented as a device that performs both a learning process and a prediction process for precipitation prediction. Accordingly, the precipitation prediction device 1 can predict precipitation within a short period of time by applying learning data consisting of continuous frames to a deep learning model. Additionally, the accuracy of prediction can be increased by using a multi-variable input method for precipitation prediction.

Figure 3 is a block diagram showing the configuration of a precipitation prediction device according to another embodiment of the present invention. Referring to FIG. 3, the precipitation prediction device 7 according to an embodiment of the present invention includes a prediction unit 12 and is connected to a server 6 connected through a network. The server 6 includes a learning unit 11 and can be connected to the external server 20 through a network.

In the illustrated example, the server 6 includes a learning unit 11 that generates a precipitation prediction model 31 through deep learning, and the precipitation prediction device 7 predicts the precipitation generated by the server 6. It includes a prediction unit 12 that generates a prediction result for input prediction target data using the model 31. At this time, the learning unit 11 included in the server 6 and the prediction unit 12 included in the precipitation prediction device 7 are the learning unit 11 included in the precipitation prediction device 1 shown in FIG. Since they are components corresponding to the and prediction units 12, detailed descriptions of the operations performed by each component will be omitted.

The server 6 performs an operation to create a model for precipitation prediction. The server 6 collects the rainfall radar synthetic image provided from the external server 20 and the numerical meteorological model data to form learning data 2, and uses this to perform deep learning to create the rainfall prediction model 31. creates .

The precipitation prediction device 7 processes an analysis request for precipitation prediction input from the user and provides the results on the screen.

As an embodiment, the precipitation prediction device 7 may transmit a request for creating a model for precipitation prediction to the server 6 and receive information about the precipitation prediction model 31 generated by the server 6. . The precipitation prediction device 7 may use the precipitation prediction model 31 provided from the server 6 to generate a prediction result of precipitation occurrence for prediction target data input from the user and display the result on the screen. At this time, the prediction target data input from the user may include rainfall status information such as a rainfall radar composite image at a reference point and dynamic and thermodynamic variables of a numerical meteorological model.

The precipitation prediction device 7 according to the embodiment of the present invention as described above only performs a prediction process for precipitation prediction, and the learning process for generating a model for prediction may be performed through a separate server 6. Accordingly, since the learning process and prediction process for precipitation prediction are performed in different devices, prediction results can be provided with high performance without delay in predicting the degree of precipitation occurrence.

4 to 6 are flowcharts for explaining a precipitation prediction method according to another embodiment of the present invention.

The precipitation prediction method according to this embodiment may be executed by the computing device 100, for example, by the precipitation prediction device 1. The computing device 100 that executes the method according to this embodiment may be a computing device equipped with an application execution environment. Note that description of the subject performing some operations included in the method according to this embodiment may be omitted, and in such case, the subject is the computing device 100.

Referring to FIG. 4, first, in operation S41, learning data 2 consisting of a series of continuous frames is generated using rainfall data and numerical meteorological model data. Here, rainfall data can be prepared as a rainfall radar composite image provided by the Korea Meteorological Administration.

Referring to FIG. 5, operation S41 is an operation S411 that matches different spatial grids and temporal resolutions between the rainfall radar synthetic image and the numerical meteorological model data, and generates a multidimensional array consisting of continuous frames consisting of an input sequence and an output sequence. Thus, an operation S412 of configuring learning data may be included.

As an embodiment, operation S411 includes converting a rainfall radar composite image into hourly rainfall data through time integration, and interpolating numerical meteorological model data into the same spatial grid structure as the rainfall radar composite image. It can be included.

As an example, operation S412 configures frames during a second period corresponding to the first half as an input sequence among consecutive frames at a predetermined interval during the first period, and frames during a third period corresponding to the second half of the sequence. It can include actions that consist of an output sequence.

As an example, a multidimensional array consisting of 12 hours of continuous frames is configured as learning data, but the multidimensional array can be created to have a size of 12 * n _y * n _x * n _{var (} n _y: space in the north-south direction Number of grids _, n _x: Number of spatial grids in the east-west direction, n _var: Number of species of input data). At this time, among the 12-hour continuous frames, the first 6-hour frame is composed of an input sequence, and the next 6-hour frame is composed of an output sequence. In the learning process, the output sequence is used as the model's It is provided as a learning target, and in the prediction process, the input sequence can be input into the learned model to derive the output sequence.

As an embodiment, operation S412 generates learning data consisting of consecutive frames at predetermined intervals during the first period, such that the number of spatial grids with a rainfall value greater than the reference value among the total number of spatial grids within one frame is greater than or equal to a predetermined ratio. Frames can be included as samples in training data.

For example, within each frame of consecutive frames for 12 hours, frames in which the number of grids with a rainfall value of 10 mm/h or more is 10% or more of n _y * n _x may be included as samples of the learning data. Accordingly, data with meaningful precipitation distribution can be used for learning.

Next, in operation S42, the precipitation prediction model 31 is generated by applying the learning data 2 to the neural network 3 with an encoder-decoder structure.

Referring to FIG. 6, operation S42 includes operation S421 of learning the neural network 3 of the encoder-decoder structure using learning data 2, and a generative adversarial neural network (Generator) and Discriminator. Operation S422 of adding a Generative Adversarial Network (GAN) to the neural network (3) of the encoder-decoder structure, and using the prediction information output from the decoder as output information of the generator, and inputting the output information of the generator and actual precipitation information into the separator. Thus, an operation S423 of determining the accuracy of the output information may be included.

As an embodiment, the neural network 3 of the encoder-decoder structure may be composed of an encoder (see symbol '91' in FIG. 9) and a decoder (see symbol '92' in FIG. 9) having a structure of multiple layers. . The encoder 91 and decoder 92 can reduce or enlarge the dimension of data by performing convolution operations such as Conv2D and ConvLSTM2D, for example. For example, when the input sequence constituting the learning data 2 is input to the encoder 91, the dimensionality is reduced and key features having compressed information can be extracted. Additionally, when information compressed through the encoder 91 is input to the decoder 92, the dimension is expanded and detailed information matching the output sequence can be output.

As an embodiment, operation S421 is a loss function that calculates the loss of the error generated during the learning process of the neural network 3 of the encoder-decoder structure, such as MSE (Mean Squared Error) and MAE (Mean Absolute Error). , may include an operation using at least one of Balanced MSE (Mean Squared Error), Balanced MAE (Mean Absolute Error), and SSIM (Scale Structural Similarity).

As an embodiment, operation S421 may include an operation of configuring a neural network with a branched structure according to the type of data included in the learning data 2. For example, if the data provided as learning data 2 is of different data types, such as radar composite image and weather numerical model data, the input layer of the encoder 91 and the Conv2D layer connected thereto are branched according to the type of data. It can be composed of one structure.

As an embodiment, operation S422 may include an operation of configuring the encoder-decoder structured neural network 3 as a generator of a generative adversarial network (GAN).

In addition, in operation S423, the prediction result output from the decoder 92 is not output as the final precipitation prediction information, but is used as output information of the generator, and the actual precipitation information is used together with the output information through a generative adversarial neural network ( By inputting it into the discriminator of GAN, it can be determined whether the output information is accurately predicted information to match the actual precipitation information, or is inaccurately predicted fake information.

As an embodiment, operation S423 may further include an operation of feeding back a loss value generated from a discriminator of a generative adversarial network (GAN) to the generator. In other words, the loss value generated during the learning process of the classifier is fed back to the generator, and by repeatedly performing this feedback process, the generator can generate prediction information close to actual precipitation information.

Finally, in operation S43, the prediction target data 4 is input into the precipitation prediction model 31, and precipitation prediction information 5 is output. At this time, the prediction target data 4 may include a rainfall radar composite image at a reference point, dynamic and thermodynamic variables of a numerical meteorological model, etc. The precipitation prediction information 5 is a result of predicting the occurrence of ultra-short-term precipitation, such as after 6 hours from a reference point, and may include, for example, a rainfall radar composite image predicted by the precipitation prediction model 31.

According to the method according to the embodiment of the present invention as described above, the degree of precipitation occurrence within a short period of time can be predicted by applying learning data consisting of continuous frames to a deep learning model. Additionally, in the precipitation prediction process, more sophisticated prediction results can be obtained by applying the prediction results from the encoder-decoder neural network to a generative adversarial network (GAN).

Figure 7 is a flowchart for explaining a precipitation prediction method according to another embodiment of the present invention.

The precipitation prediction method according to this embodiment may be executed by the computing device 100, for example, by the precipitation prediction device 7. The computing device 100 that executes the method according to this embodiment may be a computing device equipped with an application execution environment. Note that description of the subject performing some operations included in the method according to this embodiment may be omitted, and in such case, the subject is the computing device 100.

Referring to FIG. 7, in operation S71, prediction target data 4 is input to the precipitation prediction model 31, and in operation S72, precipitation prediction information 5 is output from the precipitation prediction model 31. At this time, the precipitation prediction model 31 uses a neural network 3 with an encoder-decoder structure using learning data 2 consisting of a series of consecutive frames using rainfall data and numerical meteorological model data. It may have been created through learning.

Operations S71 and S72 according to this embodiment are performed by the precipitation prediction device 7 that performs a prediction process, and correspond to operation S43 described in the embodiment of FIG. 4, so detailed descriptions will be omitted.

Figure 8 is an example showing the structure of learning data according to some embodiments of the present invention. Referring to FIG. 8, it shows the configuration of learning data 80 used to generate the precipitation prediction model 31.

In the example shown, training data 80 may include samples comprised of consecutive frames spaced 1 hour apart for 12 hours. At this time, among the consecutive frames for 12 hours, a plurality of frames (T-5h, T-4h, T-3h, T-2h, T-1h, T) for the previous 6 hours are input sequences. ) (81), and multiple frames (T+1h, T+2h, T+3h, T+4h, T+5h, T+6h) for the next 6 hours are output sequence (82) ) can be used. At this time, the collection period and frame interval of the learning data 80 can be set variously in consideration of the data collection environment and prediction time.

As an example, among all samples (s000001, s000002, ??, sn-2, sn-1) of the learning data 80, only frames in which the number of spatial grids with a rainfall value greater than or equal to the reference value among the total number of spatial grids are greater than a predetermined ratio are selected. It can be included as a sample of learning data.

As an example, within each frame of 80 consecutive frames for 12 hours, the number of grids with a rainfall value of 10 mm/h or more is 10% or more of n _y * n _x so that the frames are included as samples of the learning data. (n _y: number of spatial grids in the north-south direction _, n _x: number of spatial grids in the east-west direction). Accordingly, only data with a meaningful precipitation distribution can be included as samples of the learning data 80 and used for learning.

Figure 9 is an example illustrating an encoder-decoder neural network structure for learning according to some embodiments of the present invention. Referring to FIG. 9, in order to perform deep learning using the training data 80, a neural network with an auto-encoder structure including an encoder 91 and a decoder 92 may be used.

In the example shown, the encoder 91 and decoder 92 may be composed of three layers including Conv2D and ConvLSTM2D. Conv2D and ConvLSTM2D have different convolution operation methods and can reduce or enlarge the dimension of input data.

As an embodiment, the encoder 91 receives the radar composite image 93 corresponding to the input sequence 81 and undergoes a compression process to reduce the dimension, thereby extracting main features. The decoder 92 goes through a restoration process that expands the dimension of the compressed information through the encoder 91, and through this, the detailed information matching the output sequence 82 can be reconstructed to output the predictive radar composite image 94. there is.

As an embodiment, in the process of learning an auto-encoder neural network consisting of the encoder 91 and the decoder 92, a loss function (loss) that calculates the loss of the error, which is the difference between the predicted value and the actual value, is used to find the optimal weight. function) can be used. In particular, since the precipitation data used for precipitation prediction has an asymmetric distribution, a loss function that takes this into account must be used.

As an example, referring to Figure 10, the loss functions include Mean Squared Error (MSE), Mean Absolute Error (MAE), Balanced Mean Squared Error (BMSE) (1010), Balanced Mean Absolute Error (BMAE) (1020), and At least one of SSIM (Scale Structural Similarity) 1030 may be used. Additionally, as a loss function, a combination 1040 of BMSE (1010) and BMAE (1020) may be used, or BALL (1050), which is a combination of BMSE (1010), BMAE (1020), and SSIM (1030), may be used. . In this way, when several types of loss functions are used in combination, the ratio of each term can be adjusted through empirical experiments.

As an embodiment, in the process of learning an auto-encoder neural network composed of the encoder 91 and the decoder 92, a neural network with a branched structure can be constructed according to the type of data included in the learning data 2.

As an example, referring to FIG. 11, the data provided as learning data 2 is of different types such as radar composite image 1110, global weather numerical model data 1120, and local weather numerical model data 1130. In this case, a neural network can be constructed in a structure in which the input layer of the encoder 91 and the Conv2D layer connected thereto are branched according to each type of data (1110, 1120, 1130).

Figure 12 is an example of applying a generative adversarial network (GAN) as an additional configuration to the encoder-decoder neural network according to some embodiments of the present invention. Referring to FIG. 12, a generative adversarial network (GAN) consisting of a generator 1210 and a discriminator 1220 is applied as an additional configuration to an auto-encoder neural network consisting of an encoder 1211 and a decoder 1212. Thus, the precipitation prediction model 31 can be created.

In the illustrated example, the generator 1210 includes an encoder 1211 and a decoder 1212, and the prediction result output from the decoder 1212 can be used as output information 1213 of the generator 1210.

When the output information 1213 of the generator 1210 and the actual precipitation information 1221 are input, the separator 1220 determines whether the output information 1213 is accurately predicted information to match the actual precipitation information 1221, or is incorrectly predicted. It is possible to determine whether it is predicted fake information.

As an embodiment, the separator 1220 feeds back the loss 1222 generated during the learning process to the generator 1210, and by repeatedly performing this feedback process, the generator 1210 generates actual precipitation information ( Output information 1213 close to 1221) can be generated.

According to the embodiment of the present invention as described above, in generating a precipitation prediction model, by adding a generative adversarial network (GAN) to the neural network of the encoder-decoder structure, in the case of using only the neural network of the encoder-decoder structure, More sophisticated predictions may become possible.

Figure 13 is an example of using statistical indicators for learning verification according to some embodiments of the present invention. Referring to FIG. 13, when configuring learning data 2, in addition to sample data for learning, test data for learning verification is separately prepared, and the predicted result using the input sequence is actually used. The validity of learning can be judged by comparing it with observed results.

In the example shown, to verify the learning of a neural network with an encoder-decoder structure, prediction is performed using verification data, and in the process, the prediction results of neural networks learned differently according to the loss function applied to the neural network are graphed. In the above, predictions can be provided by time slot (F01, F02, F03, F04, F05, F06). At this time, as the loss function, at least one of MSE (Mean Squared Error), MAE (Mean Absolute Error), BMSE (1010), BMAE (1020), and SSIM (1030) may be used.

In this way, by providing prediction results for each prediction time period on a graph, the prediction performance of the neural network according to the configuration of the loss function for each prediction time period can be determined. At this time, statistical indicators such as CSI, POD, FAR, BIAS, RMSE, etc. can be presented as prediction results to provide a judgment standard for learning verification.

Figure 14 is a hardware configuration diagram of an example computing device capable of implementing methods according to some embodiments of the present invention. As shown in FIG. 14, the computing device 100 loads one or more processors 101, a bus 107, a network interface 102, and a computer program 105 performed by the processor 101. It may include a memory 103 that stores a computer program 105 and a storage 104 that stores a computer program 105. However, only components related to the embodiment of the present invention are shown in Figure 14. Accordingly, anyone skilled in the art to which the present invention pertains can recognize that other general-purpose components may be included in addition to the components shown in FIG. 14.

The processor 101 controls the overall operation of each component of the computing device 100. The processor 101 includes at least one of a Central Processing Unit (CPU), Micro Processor Unit (MPU), Micro Controller Unit (MCU), Graphic Processing Unit (GPU), or any type of processor well known in the art of the present invention. It can be configured to include. Additionally, the processor 101 may perform operations on at least one application or program to execute methods/operations according to various embodiments of the present invention. Computing device 100 may include one or more processors.

The memory 103 stores various data, instructions and/or information. The memory 103 may load one or more programs 105 from the storage 104 to execute methods/operations according to various embodiments of the present invention. For example, when the computer program 105 is loaded into memory 103, logic (or modules) may be implemented on memory 103. An example of the memory 103 may be RAM, but is not limited thereto.

Bus 107 provides communication functionality between components of computing device 100. The bus 107 may be implemented as various types of buses, such as an address bus, a data bus, and a control bus.

The network interface 102 supports wired and wireless Internet communication of the computing device 100. The network interface 102 may support various communication methods other than Internet communication. To this end, the network interface 102 may be configured to include a communication module well known in the art of the present invention.

Storage 104 may non-transitory store one or more computer programs 105. The storage 104 may include non-volatile memory such as flash memory, a hard disk, a removable disk, or any type of computer-readable recording medium well known in the art to which the present invention pertains.

The computer program 105 may include one or more instructions implementing methods/operations according to various embodiments of the present invention. When the computer program 105 is loaded into the memory 103, the processor 101 can perform methods/operations according to various embodiments of the present invention by executing the one or more instructions.

For example, the computer program 105 includes an operation of generating learning data consisting of a series of consecutive frames using rainfall data and numerical meteorological model data collected from an external server, and an encoder-decoder operation of the learning data. ) may include instructions for generating a precipitation prediction model by applying it to a neural network, and inputting prediction target data into the precipitation prediction model to output precipitation prediction information.

So far, various embodiments of the present invention and effects according to the embodiments have been mentioned with reference to FIGS. 1 to 14. The effects according to the technical idea of the present invention are not limited to the effects mentioned above, and other effects not mentioned can be clearly understood by those skilled in the art from the description below.

The technical idea of the present invention described so far can be implemented as computer-readable code on a computer-readable medium. The computer-readable recording medium may be, for example, a removable recording medium (CD, DVD, Blu-ray disk, USB storage device, removable hard disk) or a fixed recording medium (ROM, RAM, computer-equipped hard disk). You can. The computer program recorded on the computer-readable recording medium can be transmitted to another computing device through a network such as the Internet, installed on the other computing device, and thus used on the other computing device.

In the above, even though all the components constituting the embodiments of the present invention have been described as being combined or operated in combination, the technical idea of the present invention is not necessarily limited to these embodiments. That is, as long as it is within the scope of the purpose of the present invention, all of the components may be operated by selectively combining one or more of them.

Although operations are shown in the drawings in a specific order, it should not be understood that the operations must be performed in the specific order shown or sequential order or that all illustrated operations must be performed to obtain the desired results. In certain situations, multitasking and parallel processing may be advantageous. Moreover, the separation of the various components in the embodiments described above should not be construed as necessarily requiring such separation, and the program components and systems described may generally be integrated together into a single software product or packaged into multiple software products. You must understand that it exists.

Although embodiments of the present invention have been described above with reference to the attached drawings, those skilled in the art will understand that the present invention can be implemented in other specific forms without changing the technical idea or essential features. I can understand that there is. Therefore, the embodiments described above should be understood in all respects as illustrative and not restrictive. The scope of protection of the present invention shall be interpreted in accordance with the claims below, and all technical ideas within the equivalent scope shall be construed as being included in the scope of rights of the technical ideas defined by the present invention.

Claims (18)

In a method performed by a computing device,
Generating learning data consisting of a series of continuous frames using radar synthetic image data, global numerical meteorological model data, and local numerical meteorological model data;
Generating a precipitation prediction model by applying the learning data to a neural network with an encoder-decoder structure; and
Including the step of inputting prediction target data into the precipitation prediction model and outputting precipitation prediction information,
The step of generating the precipitation prediction model is,
Comprising the step of branching and configuring an input layer of the encoder and a convolution layer connected to the input layer according to the type of data included in the learning data,
The input layer is branched into a first input layer corresponding to the radar composite image data, a second input layer corresponding to the global numerical meteorological model data, and a third input layer corresponding to the local numerical meteorological model data,
The convolutional layer is branched into a first convolutional layer connected to the first input layer, a second convolutional layer connected to the second input layer, and a third convolutional layer connected to the third input layer,
How to predict precipitation. delete According to claim 1,
The step of generating learning data consisting of a series of consecutive frames using the radar composite image data, global weather numerical model data, and local weather numerical model data,
Comprising the step of matching different spatial grids and temporal resolutions between the radar composite image data, global numerical meteorological model data, and local numerical meteorological model data,
How to predict precipitation. According to clause 3,
The step of matching different spatial grids and temporal resolutions between the radar composite image data, global numerical meteorological model data, and local numerical meteorological model data,
Converting the radar composite image data into hourly precipitation data through time integration; and
Comprising the step of interpolating the global weather numerical model data and the local weather numerical model data into the same spatial grid structure as the radar composite image data,
How to predict precipitation. According to claim 1,
The step of generating learning data consisting of a series of consecutive frames using the radar composite image data, global weather numerical model data, and local weather numerical model data,
Generating a multidimensional array consisting of consecutive frames of a first period and constructing it as training data,
The multidimensional array is generated based on the number of frames, the number of spatial grids, and the number of input variables,
How to predict precipitation. According to clause 5,
The step of generating a multidimensional array composed of consecutive frames of the first period and configuring it as learning data includes:
Including the step of using as learning data frames in which the number of spatial grids in which the rainfall value is greater than or equal to the reference value is more than a predetermined ratio among the total number of spatial grids,
How to predict precipitation. According to clause 5,
The step of generating a multidimensional array composed of consecutive frames of the first period and configuring it as learning data includes:
Comprising the step of configuring frames of a second period corresponding to the first half of the consecutive frames of the first period as an input sequence, and configuring frames of a third period corresponding to the second half of the continuous frames as an output sequence,
How to predict precipitation. According to clause 7,
The step of generating a precipitation prediction model by applying the learning data to a neural network with an encoder-decoder structure,
extracting features by compressing the input sequence through the encoder; and
Recovering the compressed features through the decoder into detailed information consistent with the output sequence,
How to predict precipitation. According to claim 1,
The step of generating a precipitation prediction model by applying the learning data to a neural network with an encoder-decoder structure,
In the process of learning the neural network, the loss functions include MSE (Mean Squared Error), MAE (Mean Absolute Error), Balanced MSE (Mean Squared Error), Balanced MAE (Mean Absolute Error), and SSIM (Scale Structural Similarity). ) Comprising the step of using at least one of,
How to predict precipitation. delete According to claim 1,
The step of generating a precipitation prediction model by applying the learning data to a neural network with an encoder-decoder structure,
Comprising the step of adding a generative adversarial network (GAN) consisting of a generator and a discriminator to the neural network of the encoder-decoder structure,
How to predict precipitation. According to claim 11,
The step of adding a generative adversarial network (GAN) consisting of the generator and the discriminator to the neural network of the encoder-decoder structure,
Using prediction information output from the decoder as output information of the generator; and
Including the step of inputting the output information of the generator and actual precipitation information into the separator, and determining the accuracy of the output information,
How to predict precipitation. A computer program is stored that causes a computer to perform the precipitation prediction method of any one of claims 1, 3 to 9, 11, and 12,
A non-transitory computer-readable recording medium. In a method performed by a computing device,
Receiving information about a precipitation prediction model from a server;
Obtaining prediction target data by user input;
Inputting the prediction target data into the precipitation prediction model; and
Comprising the step of outputting precipitation prediction information from the precipitation prediction model,
The precipitation prediction model learns a neural network with an encoder-decoder structure using training data consisting of a series of consecutive frames using radar composite image data, global numerical weather model data, and local numerical weather model data. It was created by
The precipitation prediction model is constructed by branching the input layer of the encoder and the convolution layer connected to the input layer according to the type of data included in the learning data,
The input layer is branched into a first input layer corresponding to the radar composite image data, a second input layer corresponding to the global numerical meteorological model data, and a third input layer corresponding to the local numerical meteorological model data,
The convolutional layer is branched into a first convolutional layer connected to the first input layer, a second convolutional layer connected to the second input layer, and a third convolutional layer connected to the third input layer,
How to predict precipitation. According to claim 14,
The training data is constructed by generating a multidimensional array consisting of consecutive frames of a first period,
The multidimensional array is generated based on the number of frames, the number of spatial grids, and the number of input variables,
How to predict precipitation. According to claim 14,
The precipitation prediction model is generated by learning by adding a generative adversarial network (GAN) consisting of a generator and a discriminator to the neural network of the encoder-decoder structure,
How to predict precipitation. A communication unit that communicates with an external server;
Using the radar composite image data, global weather numerical model data, and local weather numerical model data collected from the external server, learning data consisting of a series of continuous frames is generated, and the learning data is sent to an encoder-decoder. ) A learning unit that generates a precipitation prediction model by applying it to a neural network of the structure; and
A prediction unit that inputs prediction target data into the precipitation prediction model and outputs precipitation prediction information,
In generating the precipitation prediction model, the learning unit branches and configures an input layer of the encoder and a convolutional layer connected to the input layer according to the type of data included in the learning data,
The input layer is branched into a first input layer corresponding to the radar composite image data, a second input layer corresponding to the global numerical meteorological model data, and a third input layer corresponding to the local numerical meteorological model data,
The convolutional layer is branched into a first convolutional layer connected to the first input layer, a second convolutional layer connected to the second input layer, and a third convolutional layer connected to the third input layer,
Precipitation prediction device. One or more processors;
A communication interface for communicating with external devices;
a memory that loads a computer program executed by the processor; and
Including storage for storing the computer program,
The computer program is,
An operation of receiving information about a precipitation prediction model from the external device,
An operation of acquiring prediction target data by user input,
An operation of inputting the prediction target data into the precipitation prediction model, and
Includes instructions for performing an operation of outputting precipitation prediction information from the precipitation prediction model,
The precipitation prediction model learns a neural network with an encoder-decoder structure using training data consisting of a series of consecutive frames using radar composite image data, global numerical weather model data, and local numerical weather model data. It was created by
The precipitation prediction model is constructed by branching the input layer of the encoder and the convolution layer connected to the input layer according to the type of data included in the learning data,
The input layer is branched into a first input layer corresponding to the radar composite image data, a second input layer corresponding to the global numerical meteorological model data, and a third input layer corresponding to the local numerical meteorological model data,
The convolutional layer is branched into a first convolutional layer connected to the first input layer, a second convolutional layer connected to the second input layer, and a third convolutional layer connected to the third input layer,
Precipitation prediction device.