KR101969716B1 - Method and system for predicting characteristic value of space - Google Patents

Abstract

Disclosed is an apparatus for predicting a space feature value. According to the present invention, the apparatus comprises: an input unit to which one or more first frames including an actual measured feature value of predetermined space and one or more second frames including a predicted feature value of the predetermined space are inputted at a predetermined period; and a processor increasing the calculation accuracy of a deep learning model using the actual measured feature value included in the first frames and the predicted feature value included in the second frames before a specific point (T) and predicting a feature value included in a frame after the specific point (T) using the deep learning model. Accordingly, the feature value of the space may be easily predicted.

Description

Prediction method of spatial characteristic value and system applying it {METHOD AND SYSTEM FOR PREDICTING CHARACTERISTIC VALUE OF SPACE}

The present invention relates to a method and system for predicting a characteristic value collected in a predetermined space, and more particularly, to a method and system for predicting a characteristic value of a predetermined space using a deep learning algorithm.

Even when using a super computer, a large amount of collected data and a large variation with time may occur when the accuracy of the predicted data does not follow the measured data.

Weather information is actually measured by real satellite systems, radar systems, and automatic weather stations (AWS), but also by various weather models (eg, Weather Research and Forecasting Model). However, the accuracy of the prediction is not always guaranteed.

Therefore, there is a need for a method for predicting more accurate weather information. Furthermore, a method for improving the accuracy of prediction information in various fields as well as forecasting weather information will be required.

On the other hand, the above information is only presented as background information to help the understanding of the present invention. No determination is made as to whether any of the above is applicable as the prior art concerning the present invention, and no claims are made.

Patent Publication No. 10-2004-0092403 (Published: 2004.11.3)

The present invention has been made to solve the above-described problem, an embodiment of the present invention proposes a method for more accurately predicting a characteristic value of a predetermined space and a system applying the same.

The technical problems to be achieved in the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned above will be clearly understood by those skilled in the art from the following description. Could be.

According to an embodiment of the present invention, an apparatus for predicting a feature of space may include one or more first frames including measured feature values of a predetermined space and at least one second frame including predicted feature values of the predetermined space at a predetermined time period. An input unit to be input; And before a specific point in time T, by using the measured characteristic value included in the first frame and the prediction characteristic value included in the second frame, to improve the calculation accuracy of the deep learning model, and after the specific point in time T. In the case of using the deep learning model to predict the characteristic value included in the frame after the specific time point (T).

More specifically, the processor is further configured to perform a first frame of the first time point T + 1 on the first frame of the specific time point T by a calculation of the deep learning model for a predetermined time interval after the specific time point T. Outputting the first frame of the first future view T + 1 and outputting the first frame of the second future view T + 2 by the operation of the deep learning model. Repeating for a predetermined time interval, outputting a second frame corresponding to the first frame output during the predetermined time interval by calculation of the deep learning model, and outputting a characteristic value of the first frame corresponding to the predetermined time interval; The correction may be based on the prediction characteristic value included in the second frame.

More specifically, the deep learning model includes a first sub-network model, a second sub-network model, and a third sub-network model, wherein the processor includes a measured characteristic value before the specific time point T. The first frame may be used as an input frame to perform an operation of the first sub-network model to output an output frame, and the output frame may be compared with the input frame to improve the calculation accuracy of the first sub-network model.

More specifically, the processor is configured to perform a calculation of the first sub-network model on a first frame including the measured characteristic value of the specific time point T to generate a first frame of a first future time point T + 1. Outputs the first frame at the first future time point T + 1 and outputs the first frame at the second future time point T + 2 by the operation of the second sub-network model, and outputs the second frame. The first frame of the future time point T + 2 is output to the first frame of the third future time point T + 3 by the operation of the second sub-network model, and the outputted third future time point T + 3 is output. The outputting of the first frame of as a first frame of the fourth future time point T + 4 by the calculation of the second sub-network model may be repeated during the predetermined time period.

More specifically, the processor is configured to predict a second frame by calculating a third sub-network model using a second frame corresponding to the first frame output during the predetermined time interval as the input of the third sub-network model. The characteristic value can be corrected.

More specifically, the first sub-network model, the second sub-network model, and the third sub-network model include a plurality of compression layers for compressing characteristic values included in the frame into abstraction information, and includes abstraction information. It may include a plurality of decompression (decompression) layer to be decompressed by the feature value included in the frame.

More specifically, the processor is further configured to determine a specific state value included in the compression layer of the first sub-network model of the second sub-network model and the third sub-network model corresponding to the compression layer of the first sub-network model. Can be provided as a decompression layer.

More specifically, the predetermined space is divided into a plurality of regions, and the processor collects the measured characteristic values of at least a portion of the divided regions in chronological order, and includes a divided region in which the measured characteristic values are collected; The predictive characteristic values for the partitioned region in which the measured characteristic values are not collected may be collected in chronological order.

More specifically, the measured characteristic value and the predicted characteristic value may be numerical values representing precipitation information of the predetermined space.

Meanwhile, the method for predicting a feature value of a space performed by a processor according to an embodiment of the present disclosure may include at least one first frame including an actual feature value of a predetermined space and at least one prediction feature value of the predetermined space. Inputting a second frame at a predetermined time period; Before a specific point in time, improving a computational accuracy of the deep learning model using the measured characteristic values included in the first frame and the predicted characteristic values included in the second frame; And after the specific time point T, predicting a characteristic value included in the frame after the specific time point T using the deep learning model.

On the other hand, in a non-transitory computer-readable recording medium that records a program for execution on a computer according to an embodiment of the present invention, when the program is executed by a processor, the program includes a measured characteristic value of a predetermined space An operation of inputting at least one second frame including at least a first frame and a predicted characteristic value of the predetermined space at a predetermined time period, and before a specific point in time, the measured characteristic value included in the first frame and the first The operation of improving the computational accuracy of the deep learning model using the prediction characteristic values included in the two frames, and after a specific time point T, is included in the frame after the specific time point T using the deep learning model. And an executable instruction to perform an operation of predicting a characteristic value.

By providing the method of predicting the spatial characteristic value and the system applying the same, the following effects are generated.

First, by learning the characteristic values for a predetermined space collected up to a specific time point, the accuracy of the characteristic value prediction for the predetermined space after the specific time point may be improved.

Second, when both the measured data and the predicted data are provided, the future predicted data can be derived more accurately.

The effects obtainable in the present invention are not limited to the above-mentioned effects, and other effects not mentioned above may be clearly understood by those skilled in the art from the following description. will be.

1 illustrates measured characteristic values of a predetermined space according to an embodiment of the present invention.
2 (a) and 2 (b) are diagrams for explaining a schematic driving of a spatial characteristic value prediction system according to an embodiment of the present invention.
3 is a block diagram illustrating a configuration of a spatial characteristic value prediction system according to an embodiment of the present invention.
4 to 7 are views for explaining the driving of the spatial characteristic value according to an embodiment of the present invention.
8 is a graph showing the performance of the spatial characteristic value prediction system according to an embodiment of the present invention.

Hereinafter, various embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In the present specification, in describing a method of accurately predicting measured information, precipitation information is included among weather information, but this is only an example, and various measured information may be predicted by the following description.

However, in describing the present invention, when it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the subject matter of the present invention, a detailed description thereof will be omitted.

1 illustrates measurement characteristic values of a predetermined space SP according to an embodiment of the present invention. Here, the predetermined space includes the national territory of the Republic of Korea and the nearby sea, and the measured characteristic value means the value of actually measuring the characteristic value. Can be set.

AWS (Automatic Weather Station) is a system that automatically performs weather observation, a system that can automatically collect temperature information, wind direction information, humidity information, air pressure information, precipitation information, solar radiation information, sunshine information, etc. In the following description, precipitation information collected by AWS will be described as an example. The precipitation information may be collected through satellite or radar in addition to AWS, and may also be collected through other measurement equipment.

In the present specification, information for predicting actual weather information uses WRF (Weather Research and Forecasting Model), and WRF is a local weather numerical model, which has initial conditions and boundary conditions, and information on the atmosphere, hydrosphere, ice field, life zone, and land zone. Can be predicted and has a relatively high space-time resolution.

Hereinafter, a schematic driving of the spatial characteristic prediction system 100 of FIG. 3 according to an exemplary embodiment of the present invention will be described with reference to FIGS. 2A and 2B. 2 (a) shows measured characteristic values, and FIG. 2 (b) shows predicted characteristic values.

2 (a) and 2 (b), assuming a specific time T, T + 1 is the time when one day has passed, T + 2 is the time when two days have passed, and T-1 Is a point in time a day ago, T-2 represents a point in time two days ago, but this is only one example, and according to an embodiment, +1 and -1 may be implemented to represent seconds, minutes, and hours.

The spatial characteristic value prediction system 100 predicts the measured characteristic value 15 of the T to T + 1 section by using the measured characteristic value of the T-1 to T section 11 of the measured characteristic value, and predicts the measured characteristic. Based on the prediction characteristic value 13 of T-T + 1, the measured value 15 can correct the measured characteristic value 15 of the T-T + 1 section.

Here, the measured characteristic value may be collected from one or more sources. For example, precipitation information may be collected together through AWS, satellites, and radar, and future data may be predicted by reflecting both data and prediction data from multiple sources.

Hereinafter, the configuration of the spatial characteristic value prediction system 100 will be described in detail with reference to FIG. 3. The spatial characteristic value prediction system 100 may predict future characteristic values for the predetermined space based on the measured characteristic values and the predicted characteristic values collected from the predetermined space.

Referring to FIG. 3, the spatial characteristic prediction system 100 includes an input unit 110, a display 120, a storage unit 130, and a processor 140. However, the above-described components are not necessarily required to describe the present invention, and therefore, the spatial characteristic prediction system 100 according to an exemplary embodiment of the present invention may include more or less components than those described above.

In detail, the input unit 110 may receive the measured characteristic value and the predicted characteristic value of the predetermined space SP in a predetermined time period. The input unit is a frame, and the frame is a size including all the characteristic values of the predetermined groin SP. The input unit may be divided into a plurality of regions, and the characteristic values may be included in each divided region.

The display 120 may visualize various information under the control of the processor 140. The display 120 is a display unit on which data is displayed. The display 120 is a liquid crystal display (LCD), a thin film transistor-liquid crystal display (TFT LCD), and an organic light emitting diode (organic light). emitting diodes (OLEDs), flexible displays, 3D displays, and e-ink displays.

The storage unit 130 is a module in which the collected data is stored. The storage unit 130 is a flash memory type, a hard disk type, a solid state disk type, or an SDD type. (Silicon Disk Drive type), multimedia card micro type, card type memory (e.g. SD or XD memory, etc.), random access memory (RAM), static random access memory (SRAM), It may include a storage medium of at least one type of read-only memory (ROM), electrically erasable programmable read-only memory (EPEROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, and optical disk. .

The processor 140 corresponds to a module for controlling the overall spatial characteristic prediction system 100. In the training phase, the processor 140 learns a deep learning algorithm using the measured characteristic value and the predicted value of the meteorological numerical model. If you want to use the learning algorithm to make predictions and reflect the trends of the latest data, you can relearn the latest data.

If the specific time point T is a time point that separates the training step from the prediction step, the processor 140 performs a deep learning operation by inputting first frames including the measured characteristic value before the specific time point T. The accuracy of the deep learning algorithm may be improved by comparing the output first frame with the actually measured first frame. In addition, the processor 140 performs a deep learning operation by inputting second frames including prediction characteristic values before a specific point in time T, and outputs the second frame and the first frame (Ground Truth) actually measured. The accuracy of the deep learning algorithm can be improved by comparing

Hereinafter, specific operations of the processor 140 will be described with reference to FIGS. 4 to 7. 4 illustrates an overall system diagram of a deep learning model according to an embodiment of the present invention, and FIGS. 5 to 7 illustrate partial system diagrams of the deep learning model. Here, the measured characteristic value may be data collected from AWS, or data measured from a satellite system or data measured from a radar system may be used depending on an embodiment, and data measured from a plurality of sources may be applied together.

The deep learning model 400 is a model of a modified Recurrent Neural Network (RNN) algorithm that is easy to process data input in chronological order, and various deep learning algorithms may be applied according to embodiments. The deep learning model 400 may include a first sub network model 1SNM, a second sub network model 2SNM, and a third sub network model 3SNM.

The input of the deep learning model 400 may be an input frame, and the input frame may include a first frame including the measured characteristic value and a second frame including the prediction characteristic value. The output of the deep learning model may include a first frame including the measured characteristic value and a second frame including the predicted characteristic value. Each frame can be divided into a plurality of areas. For example, when expressing precipitation information of the Korean Peninsula, when a predetermined space has a resolution of 1Km in a 708 * 708 grid, an actual precipitation value or a predicted precipitation value may be expressed in each divided grid. .

The processor 140 may sequentially receive the input frames in chronological order, and the input receiving period may be 10 minutes to 6 hours, and may be input in various periods according to the embodiment. The deep learning operation for one input frame is completed, and the next input frame may sequentially perform the deep learning operation. Further, according to an embodiment, since there is no data dependency between input frames in the CNN layer (convolution cell and deconvolution cell to be described later), multiple input frames can be processed in parallel, and in the convolution LSTM to be described later, the previous frame is used. Numerical processing may be performed sequentially depending on the processing result.

The first sub network model 1SNM to the third sub network model 3SNM of the deep learning model 400 will be described in detail. The first sub-network model 1SNM may set the input frame and the output frame in the same manner to generate a compressed tensor while retaining the nature of data.

Referring to FIG. 4 together with FIG. 5, the first sub-network model 1SNM performs a deep learning operation on the first frames A _t-5 to A _t representing the measured characteristic values before a specific point in time T. The first frames A ' _{t -5} to A' _t can be output. Since the first frames A _t-5 to A _t are all measured data, a difference from the output first frames A ' _{t- 5} to A' _t can be accurately corrected.

The first sub-network model 1SNM includes operations of convolution cells C11 to C13 and deconvolution cells D11 to D13 and convolutional LSTMs LS11 to LS15.

The convolution cells C11 to C13 sequentially characterize the characteristic value information of the input frame into more abstracted information. The convolution cells C11 to C13 include a convolution process, a batch normalization process, and an active function (ReLU) operation process. do. The convolution cells C11 to C13 are not sensitive to data one by one and can be referred to as a process of abstracting a specific feature.

The deconvolution cells D11 to D13 are processes corresponding to the convolution cells C11 to C13 to detail the abstracted information into the characteristic value information of the output frame. The deconvolution process, the normalization process, It includes an active function calculation process. Specifically, the deconvolution cell D11 corresponds to the convolution cell C11, the deconvolution cell D12 corresponds to the convolution cell C12, and the deconvolution cell D13 is the convolution cell. Corresponding to C13, respectively.

Convolution LSTM (LS11 to LS15) is an upgraded version of LSTM that enables the convolutional input and output of LSTM input and output as vector as well as the function of reflecting the past data to the current operation. . The most abstracted information is stored in the LS3 convolution LSTM, and detailed information may be stored toward the LS2 and LS1. The convolutional LSTM (LS11 to LS15) has an input gate, an output gate, and an oblivion gate, and these gates control recursive operation and transfer of LSTM state values so that the characteristics of past data required can be reflected in the operation of the current data. have.

In the LS1 convolution LSTM (LS1), state data may be transmitted to LS10 and LS15 convolutional LSTMs of the second sub-network model 2SNM and the third sub-network model 3SNM. In this case, the LS10 and LS15 convolutional LSTMs receive the state data used to perform the abstraction, so that the data can be refined more effectively.

Similarly, the LS2 convolutional LSTM (LS2) may provide status data to LS9 and LS14 of the second subnetwork model 2SNM and the third subnetwork model 3SNM, and the LS3 convolutional LSTM (LS3) is the second. Deep learning operation efficiency may be more effective by providing state data to LS8 and LS13 of the sub network model 2SNM and the third sub network model 3SNM.

LS1 does not compress much of the characteristics of the data, but it can retain some of the local characteristics of the objects in the original data. LS2 compresses the characteristics of the data meaningfully, losing some of the local characteristics of the objects in the original data, but it can abstract the overall appearance to some extent. While the LS3 compresses the features of the data as much as possible, it loses the local characteristics of the objects in the original data, but can abstract it while preserving the overall shape of the objects in the frame.

Referring to FIG. 4 together with FIG. 6, the second sub-network model 2SNM includes operations of convolution cells C21 to C23, deconvolution cells D21 to D23, and convolutional LSTMs LS6 to LS10. .

The second sub-network model 2SNM receives the state information of LS1, LS2, and LS3 from the first sub-network model 1SNM as the state initialization value of the LSTM, and receives compressed input frames of the second sub-network model 2SNM. You can help to refine.

The processor 140 may predict a measured characteristic value of a predetermined space of a predetermined time interval t + 1 to t + 6 after a specific time point T. In detail, the processor 140 may predict A ′ _{t +1 based} on the calculation of the first sub-network model 1SNM of the first frame of the specific time point T.

When the processor 140 performs the operation of the second sub-network model 2SNM with the A ' _{t + 1} frame as an input frame, the A' _{t + 2} frame is output and the A ' _{t + 2} frame is input as it is. When the operation of the second sub network model 2SNM is performed as a frame, the A ' _{t + 3} frame is output and the operation of the second sub network model 2SNM is performed using the A' _{t + 3} frame as an input frame. When A ' _{t + 4} frames are output and A' _{t + 4} frames are used as input frames, the operation of the second sub-network model 2SNM is performed, and A ' _{t + 5} frames are output and A' _{t + When} the operation of the second sub-network model 2SNM is performed using _five frames as input frames, A ′ _{t + 6} frames may be output.

As illustrated, the LS8 may be used as a guide for restoring the overall shape of the prediction frame based on the most abstracted information of the objects present in the frame by using the compressed tensor generated in the first sub-network model 1SNM. The LS9 may be used as a guide for restoring the overall shape based on the partial shape of the object existing in the frame by using the compressed tensor generated in LS2 of the first sub network model 1SNM. The LS10 can generate the predicted data by restoring the local characteristics of the objects in the original data and controlling the amount of change over time by using the generated compressed tensor.

In this manner, a prediction characteristic value closer to the actual measured characteristic value may be derived.

Referring to FIG. 4 together with FIG. 7, the third sub-network model 3SNM includes operations of convolution cells C31 to C33, deconvolution cells D31 to D33, and convolutional LSTMs LS11 to LS15. .

The third sub-network model 3SNM receives the state information of LS1, LS2, and LS3 from the first sub-network model 1SNM as the state initialization value of the LSTM, and receives compressed input frames of the third sub-network model 3SNM. You can help to refine.

Since the third sub-network model 3SNM initially receives an input frame as a prediction characteristic value of a predetermined space, W _{t + 1} to W _{t + 6} can be predicted at a specific time T, and corresponding W _{t + 1} to The W _{t + 6} frames may correspond to the output frames of the second sub network model 2SNM, respectively.

The third sub-Performing the operations on the network model (3SNM) W _{'t +1} to W' _{+ t} and the predicted characteristic value of ₆ can be obtained, W _{'t +1} to W' _t W _{+ 6} frame _{t +} The prediction error of ₁ to W _{t + 6} frames is corrected through the third sub-network model 3SNM.

The third sub-network model 3SNM finally performs convolution operations on A ' _{t +1} to A' _{t + 6} frames and W ' _{t +1} to W' _{t + 6} frames to perform W ' _{t +1} to F _{t + 1} to F _{t + 6} frames in which the characteristic values of the W ′ _{t + 6} frames are reflected in the A ′ _{t +1} to A ′ _{t + 6} frames may be finally calculated.

Hereinafter, the driving performance of the spatial characteristic value prediction system 100 will be described with reference to FIG. 8.

Referring to FIG. 8, the X axis represents the learning repetition frequency and the Y axis represents the RMSE. The number of iterations was about 800 repetitions. RMSE is an observation index that is expressed as the square root of the value divided by the total number after sequential summation of squared values obtained by subtracting the measured value. RMSE measured by the precipitation prediction information model WRF. While the error of (W-rmse) is constant, RMSE (D_rmse) using the spatial characteristic value prediction system according to an embodiment of the present invention can be observed that the deviation between the measured value and the predicted value gradually decreases as the number of repetitions increases.

Meanwhile, in the method for predicting a feature value of a space performed by a processor, at least one first frame including an actual feature value of a predetermined space and at least one second frame including a predicted feature value of the predetermined space are provided at a predetermined time period. Step input,

In the case before a specific point in time T, improving the calculation accuracy of the deep learning model using the measured characteristic value included in the first frame and the predicted characteristic value included in the second frame, and the specific point in time T Subsequently, the method may include predicting a characteristic value included in a frame after the specific time point T using the deep learning model.

In addition, in a non-transitory computer readable recording medium having recorded thereon a program for executing on a computer according to an embodiment of the present invention, when the program is executed by a processor, the program includes a measured characteristic value of a predetermined space. An operation of inputting at least one second frame including at least one first frame and a predicted characteristic value of the predetermined space at a predetermined time period, and before a specific point in time, the measured characteristic value included in the first frame and the Improving operation accuracy of the deep learning model by using prediction characteristic values included in a second frame, and in a case after the specific time point T, using the deep learning model in a frame after the specific time point T And an executable instruction to perform an operation of predicting an included characteristic value.

Embodiments of the present invention can be implemented in the form of program instructions that can be executed by various computer means and recorded on a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the media may be those specially designed and constructed for the present invention, or they may be of the kind well-known and available to those having skill in the computer software arts. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tape, optical media such as CD-ROMs, DVDs, and magnetic disks, such as floppy disks. Magneto-optical media, and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine code generated by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.

Although the present invention has been described in detail with reference to preferred embodiments, the present invention is not limited to the above-described embodiments, and the technical field to which the present invention belongs without departing from the gist of the present invention as claimed in the following claims. Anyone skilled in the art will have the technical idea of the present invention to the extent that various modifications or changes are possible.

Claims (11)

One or more first frames including measured characteristic values of a predetermined space collected from a first device and one or more second frames including predicted characteristic values of the predetermined space collected from a second device different from the first device are predetermined. An input unit input in a time period of; And
Before a specific point in time, the calculation accuracy of the deep learning model is improved by using the measured characteristic values included in the first frame and the predicted characteristic values included in the second frame.
In the case after a specific time point T, the processor includes a processor that predicts a feature value included in a frame after the specific time point T using the deep learning model.
The processor,
The first frame of the specific time point T is input and output to the first frame of the first future time point T + 1 by performing the deep learning model operation.
Outputting the first frame of the first future time point T + 1 as an input to output the first frame of the second future time point T + 2 by performing the operation of the deep learning model during a predetermined time period. Repeat until N future time points (T + N),
The first future time point T + 1 to the Nth future time point T + N by performing the operation of the deep learning model by inputting a second frame corresponding to each time point of each of the first frames output during the predetermined time period. Output the second frame of,
And correct the characteristic value of the first frame corresponding to the predetermined time interval based on the predicted characteristic value included in the output second frame. delete The method of claim 1,
The deep learning model,
A first sub-network model, a second sub-network model and a third sub-network model that are provided with information about the operation from the first sub-network model,
The processor,
Receives one or more first frames including the measured characteristic value before the specific time point T as an input frame, performs an operation of the first sub-network model, and outputs an output frame, and outputs the output frame and the input frame. And a comparison property of the first sub-network model to improve the calculation accuracy. The method of claim 3,
The processor,
The first frame including the measured characteristic value of the specific time point T is calculated by performing the operation of the first sub-network model to output the first frame of the first future time point T + 1, and the outputted first future time. The first frame of the time point T + 1 is output to the first frame of the second future time point T + 2 by the operation of the second sub-network model, and the output of the second future time point T + 2 is output. The first frame is output to the first frame of the third future time point T + 3 by the operation of the second sub-network model, and the outputted first frame of the third future time point T + 3 is output to the second frame. And outputting the first frame at the fourth future time point (T + 4) by the calculation of the sub network model to the Nth future time point (T + N) during the predetermined time period. The method of claim 4, wherein
The processor,
A second frame corresponding to each of the first frames output from the first future time point T + 1 to the Nth future time point T + N during the predetermined time period is used as an input of the third sub network model. And calculating a third sub-network model and correcting the characteristic value of the first frame based on the output characteristic value of the second frame. The method of claim 5,
The first sub-network model, the second sub-network model and the third sub-network model include a plurality of compression layers for compressing characteristic values included in the frame into abstraction information, and the abstraction information is included in the frame. A spatial characteristic value prediction apparatus comprising a plurality of decompression layers decompressed with characteristic values. The method of claim 6,
The processor,
Providing a specific state value included in the compression layer of the first sub network model to the decompression layers of the second sub network model and the third sub network model corresponding to the compression layer of the first sub network model. Characteristic prediction device of the. The method of claim 1,
The predetermined space is divided into a plurality of regions,
The processor,
Collects measured characteristic values of at least a portion of the divided regions in chronological order, and predicts characteristic values of the divided region in which the measured characteristic values are collected and the divided regions in which the measured characteristic values are not collected in chronological order. Collecting, the characteristic value prediction apparatus of space. The method of claim 1,
And the measured characteristic value and the predicted characteristic value are numerical values representing precipitation information of the predetermined space. A method of predicting a characteristic value of a space performed by a processor.
One or more first frames including measured characteristic values of a predetermined space collected from a first device and one or more second frames including predicted characteristic values of the predetermined space collected from a second device different from the first device are predetermined. Inputting at a time period of;
Before a specific point in time, improving a computational accuracy of the deep learning model using the measured characteristic values included in the first frame and the predicted characteristic values included in the second frame; And
After the specific time point T, using the deep learning model, predicting a characteristic value included in a frame after the specific time point T;
The processor,
The first frame of the specific time point T is input and output to the first frame of the first future time point T + 1 by performing the deep learning model operation.
Outputting the first frame of the first future time point T + 1 as an input to output the first frame of the second future time point T + 2 by performing the operation of the deep learning model during a predetermined time period. Repeat until N future time points (T + N),
The first future time point T + 1 to the Nth future time point T + N by performing the operation of the deep learning model by inputting a second frame corresponding to each time point of each of the first frames output during the predetermined time period. Output the second frame of,
And correcting the characteristic value of the first frame corresponding to the predetermined time interval based on the predicted characteristic value included in the outputted second frame. In a non-transitory computer readable recording medium having recorded thereon a program for execution on a computer, the program, when executed by a processor,
The processor may include one or more first frames including measured characteristic values of a predetermined space collected from a first apparatus and one or more first predicted characteristic values of the predetermined space collected from a second apparatus different from the first apparatus. The operation accuracy of the deep learning model is calculated using an operation in which two frames are input at a predetermined time period, and a measurement characteristic value included in the first frame and a prediction characteristic value included in the second frame before a specific point in time. And an executable instruction for performing an operation of improving and predicting a characteristic value included in a frame after the specific time point T using the deep learning model.
The processor,
The first frame of the specific time point T is input and output to the first frame of the first future time point T + 1 by performing the deep learning model operation.
Outputting the first frame of the first future time point T + 1 as an input to output the first frame of the second future time point T + 2 by performing the operation of the deep learning model during a predetermined time period. Repeat until N future time points (T + N),
The first future time point T + 1 to the Nth future time point T + N by performing the operation of the deep learning model by inputting a second frame corresponding to each time point of each of the first frames output during the predetermined time period. Output the second frame of,
And correct the characteristic value of the first frame corresponding to the predetermined time interval based on the predicted characteristic value included in the output second frame.

Images