Learning device, learning method and prediction device

The learning device uses machine learning on historical water level and meteorological data to predict river water level trends, addressing the limitations of conventional systems and enhancing safety management in river construction.

JP2024170278APending Publication Date: 2024-12-06TAISEI CORP
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
JP2023193243
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-05-26
Filing Date
2023-11-13
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

Conventional flood warning systems struggle to accurately predict river water level trends up to several hours in advance, which is crucial for safety management and process control in river construction due to limitations in existing prediction methods.

Method used

A learning device and method that utilizes machine learning to analyze historical water level and meteorological data, including representative point information from weather distribution images, to predict future water level trends by performing machine learning on datasets that include meteorological item value information.

Benefits of technology

Enables accurate prediction of future water level trends, allowing for timely evacuation and restoration decisions in river construction, reducing the risk of damage to workers and equipment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024170278000001_ABST
    Figure 2024170278000001_ABST
Patent Text Reader

Abstract

To provide a learning device, a learning method and a prediction device, capable of grasping a trend (ascent / descent) of a water level in the future.SOLUTION: A learning device 10 includes: a learning data acquisition unit 12 that acquires learning data including a water level of a river at a first time point in a snow melting period in which snow exists in at least a part of the river, meteorological item value information at a plurality of second time points in a specific period including the previous period before the first time point, and a water level of the river at a plurality of time points in the future after the first time point; and a learning processing unit 13 that uses a learning data set obtained by collecting the learning data to allow a learning unit to undergo machine learning.SELECTED DRAWING: Figure 1
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to a learning device, a learning method, and a prediction device. [Background technology]

[0002] In recent years, the effects of global warming have led to an increase in heavy rain disasters, and damage caused by external and internal flooding of rivers is on the rise. In river construction work, rising water levels can cause damage (washing out, sinking, etc.) to workers and construction equipment (heavy machinery and materials). For this reason, it is important to predict water levels in advance and inform construction personnel in order to protect workers and construction equipment (heavy machinery and materials). Flood warning systems have been applied to river construction sites for some time now. An example of the configuration of a conventional flood warning system is shown in FIG. 49. The flood warning system shown in FIG. 49 obtains information about rain from the Japan Meteorological Agency and information about rivers from the Ministry of Land, Infrastructure, Transport and Tourism, and predicts the water level of the river by analyzing this information. The predicted river water level and a warning based on the prediction are notified to people involved in the construction site via the Internet, mobile email, etc. In conventional flood warning systems, the river water level is predicted using a physical model (water level prediction model) based on hydraulic formulas, and physical models such as a "numerical model," a "regression model," a "cumulative rainfall model," and a "conservation law model" are used (see, for example, Patent Documents 1 and 2).

[0003] The "numerical model" is a physical model that uses distributed runoff analysis to determine the water level at the construction site, with the water level upstream of the prediction point, rainfall distribution, land use, and elevation as input values. The "regression model" is a physical model that determines the regression equation between the water level at an observation station upstream of the prediction point and the water level at the prediction point, and predicts the water level at the prediction point from the regression equation. The "accumulated rainfall model" is a physical model that determines whether or not flooding will occur based on the relationship between the rainfall in the watershed upstream of the prediction point and the water level at the prediction point. The "conservation law model" is a physical model that predicts the water level at the prediction point from the relationship between the rainfall in the watershed upstream of the prediction point and the water level at the prediction point. Furthermore, regarding prediction of river water levels due to snowmelt, there is a technology described in Patent Document 3. The technology described in Patent Document 3 predicts the maximum water level for the next day, for example, by using a weather distribution image and the maximum water level from "0:00 to 23:00". [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2019-045290 A [Patent Document 2] JP 2008-015916 A [Patent Document 3] Patent Publication No. 2023-067705 Summary of the Invention [Problem to be solved by the invention]

[0005] In river construction work, grasping the water level trend (rise / fall) for the next few hours (for example, 24 hours) can be important for safety management and process management. However, the technology described in Patent Document 3 predicts the daily maximum water level (one point per day), making it difficult to grasp the water level trend for the next few hours. From this perspective, the present invention provides a learning device, a learning method, and a prediction device that are capable of grasping future water level trends (rise and fall). [Means for solving the problem]

[0006] The learning device according to the present invention includes a learning data acquisition unit and a learning processing unit. The learning data acquisition unit acquires learning data consisting of the water level of the river at a first point in time during a snowmelt period when snow is present in at least a portion of the river's basin, meteorological item value information at a plurality of second point in time during a specific period including a period in the past prior to the first point in time, and the water level of the river at a plurality of third point in time in the future prior to the first point in time. The learning processing unit trains a learning device by machine learning using a learning data set that is a compilation of the learning data. The learning processing unit trains the learning device to output the water level of the river at a plurality of the third time points by inputting the water level of the river at the first time point and meteorological item value information at a plurality of the second time points.

[0007] The learning data acquisition unit acquires learning data including at least one of the meteorological item value information, for example, average temperature per hour, accumulated precipitation per hour, average wind speed per hour, snow depth per hour, and sunshine hours per hour. By using a trained learning device trained by the learning device of the present invention, it is possible to predict future water level trends (rise and fall). This makes it possible to determine the timing of evacuation and restoration, which is useful for safety management and process management of river construction work.

[0008] The system may further include an image processing unit that calculates representative point information and meteorological item value information regarding representative points reflecting the intensity distribution in the watershed from a weather distribution image showing the intensity distribution of meteorological items by color changes. In this case, the learning data acquisition unit acquires learning data having the representative point information and the meteorological item value information calculated from the weather distribution image at the second time point. The image processing unit converts the RGB values ​​of each pixel constituting the weather distribution image at the second time point into a luminance value, calculates the center of gravity of the watershed as the representative point when the luminance value is regarded as weight, obtains the position of the center of gravity and the distance from a predetermined reference point to the center of gravity as the representative point information, and obtains a statistical value calculated from the luminance value of each pixel constituting the weather distribution image as the weather item value information.

[0009] Weather distribution images published by the Japan Meteorological Agency cover the entire country of Japan and are easy to obtain (for example, they can be purchased from the Japan Meteorological Agency). Therefore, by using the weather distribution images, it is easy to predict water levels at any point in Japan. In addition, by performing learning using representative point information rather than the weather distribution image itself, the calculation load can be reduced compared to inputting all data (all pixel values) of the weather distribution image.

[0010] The meteorological item value information may be AMeDAS observation values ​​at multiple points in the basin at the second time point. In this way, it is possible to predict the water level even if a meteorological distribution image is not available. Therefore, it is possible to grasp the water level trend (rise / fall). The learning processing unit may weight the meteorological item value information according to the meteorological items and perform learning using the weighted meteorological item value information. In this way, the contribution of the meteorological items to the water level prediction result is reflected in the learning data, making it possible to realize even more accurate prediction. The data set may be thinned based on flood periods, which reduces the computational load compared to inputting all data during snowmelt periods.

[0011] The learning method according to the present invention includes a learning data acquiring step and a learning processing step. The learning data acquisition process acquires learning data consisting of the water level of the river at a first point in time during a snowmelt period when snow is present in at least a portion of the river basin, meteorological item value information at a plurality of second point in time during a specific period including a period in the past prior to the first point in time, and the water level of the river at a plurality of third point in time in the future prior to the first point in time. In the learning process, a learning device is trained by machine learning using a learning data set that is a compilation of the learning data. In the learning process, the learning device is trained to output the water level of the river at a plurality of the third time points by inputting the water level of the river at the first time point and meteorological item value information at a plurality of the second time points. By using a trained learning device trained by the learning method of the present invention, it is possible to predict future water level trends (rise and fall). This makes it possible to determine the timing of evacuation and restoration, which is useful for safety management and schedule management of river construction work.

[0012] A prediction device according to the present invention includes a learning unit that has been trained by the above-described learning method, and predicts the water level of the river at a fourth time point during a snowmelt period and at a plurality of sixth time points in the future from the fourth time point by inputting meteorological item value information at a plurality of fifth time points during a specific period including a period in the past from the fourth time point. The prediction device according to the present invention can predict future water level trends (rise and fall), making it possible to determine the timing of evacuation and restoration, which is useful for safety management and process management of river construction work. Effect of the Invention

[0013] According to the present invention, it is possible to grasp the future trend (rise / fall) of water level. [Brief description of the drawings]

[0014] [Figure 1] FIG. 2 is a functional configuration diagram of a learning device and a prediction device according to an embodiment of the present invention. [Diagram 2] This is an image of a river and its basin. [Diagram 3] 11A and 11B are diagrams for explaining the processing of the image processing unit, in which (a) is an example of a weather distribution image, (b) is an example of a watershed image from which a weather distribution image within a watershed has been extracted, and (c) is an example of a representative point. [Figure 4] These are images of the calculation of the center of gravity G for each weather distribution image Fa, where (a) shows the calculation result for the hourly average temperature image F1a, (b) shows the calculation result for the hourly accumulated precipitation image F2a, and (c) shows the calculation result for the hourly average wind speed image F3a. [Figure 5A] This is an example of a training dataset. [Figure 5B] This is an enlarged view of data (1) from the training dataset. [Figure 5C] This is an enlarged view of data (2) from the training dataset. [Figure 5D] FIG. 1 is an enlarged view of the data (3) of the training dataset and the output data portion. [Figure 6] 10A and 10B are diagrams illustrating an example of thinning a learning dataset, where (a) is a diagram illustrating a method for selecting a maximum water level, and (b) is a graph showing the fluctuation of water level at an observation station. [Figure 7] 13 is a flow chart showing a prediction procedure in a simulation carried out to verify the effectiveness. [Figure 8] This is a diagram showing the location of the observatory within the Tsukigata Observatory basin. [Figure 9] 13 is a list of weight values ​​used in the simulation. [Figure 10] These are water level forecast results for fiscal year 2018 when using the Case 01 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 11] These are water level forecast results for fiscal year 2019 when using the Case 01 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 12] These are water level forecast results for fiscal year 2020 when using the Case 01 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 13] These are water level forecast results for fiscal year 2021 when using the Case 01 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 14] These are water level forecast results for fiscal year 2018 when using the Case 02 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 15] These are water level forecast results for fiscal year 2019 when using the Case 02 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 16]These are water level forecast results for fiscal year 2020 when using the Case 02 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 17] These are water level forecast results for fiscal year 2021 when using the Case 02 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 18] These are water level forecast results for fiscal year 2018 when using the Case 03 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 19] These are water level forecast results for fiscal year 2019 when using the Case 03 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 20] These are water level forecast results for fiscal year 2020 when using the Case 03 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 21] These are water level forecast results for fiscal year 2021 when using the Case 03 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 22] These are water level forecast results for fiscal year 2018 when using the Case 04 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 23] These are water level forecast results for fiscal year 2019 when using the Case 04 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 24] These are water level forecast results for fiscal year 2020 when using the Case 04 dataset, with (a) to (c) showing the forecast results for March to May. [Diagram 25] These are water level forecast results for fiscal year 2021 when using the Case 04 dataset, with (a) to (c) showing the forecast results for March to May. [Figure 26] This is a comparison of the predicted values ​​and the actual measured values ​​when using the data sets of Case01 and Case02. (a)~(d) are the results after 6, 12, and 18.24 hours for Case01, and (e)~(h) are the results after 6, 12, and 18.24 hours for Case02. [Figure 27]This is a comparison of the predicted values ​​and the actual measured values ​​when using the data sets of Case03 and Case04. (a)~(d) are the results after 6, 12, and 18.24 hours for Case03, and (e)~(h) are the results after 6, 12, and 18.24 hours for Case04. [Figure 28] The figures show the results of a simulation. (a) compares the slope of the correlation equation for each of Case01 to Case04, and (b) compares the correlation coefficient for each of Case01 to Case04. [Figure 29] These are water level forecast results for fiscal years 2018 and 2019 when using the Case 05 dataset. (a)-(b) are the forecast results for March-April 2018, and (c)-(d) are the forecast results for March-April 2019. [Diagram 30] These are water level forecast results for fiscal years 2020 and 2021 when using the Case 05 dataset, where (a)-(b) are the forecast results for March-April 2020, and (c)-(d) are the forecast results for March-April 2021. [Diagram 31] These are water level forecast results for fiscal years 2018 and 2019 when using the Case 06 dataset. (a)-(b) are the forecast results for March-April 2018, and (c)-(d) are the forecast results for March-April 2019. [Diagram 32] These are water level forecast results for fiscal years 2020 and 2021 when using the Case 06 dataset, where (a)-(b) are the forecast results for March-April 2020, and (c)-(d) are the forecast results for March-April 2021. [Diagram 33] These are water level forecast results for fiscal years 2018 and 2019 when using the Case 07 dataset. (a)-(b) are the forecast results for March-April 2018, and (c)-(d) are the forecast results for March-April 2019. [Diagram 34] These are water level forecast results for fiscal years 2020 and 2021 when using the Case 07 dataset, where (a)-(b) are the forecast results for March-April 2020, and (c)-(d) are the forecast results for March-April 2021. [Diagram 35]These are water level forecast results for fiscal years 2020 and 2021 when using the Case 08 dataset. (a)-(b) are the forecast results for March-April 2020, and (c)-(d) are the forecast results for March-April 2021. [Diagram 36] These are water level forecast results for fiscal years 2018 and 2019 when using the Case09 dataset. (a)-(b) are the forecast results for March-April 2018, and (c)-(d) are the forecast results for March-April 2019. [Figure 37] These are water level forecast results for fiscal years 2020 and 2021 when using the Case 09 dataset, where (a)-(b) are the forecast results for March-April 2020, and (c)-(d) are the forecast results for March-April 2021. [Figure 38] These are water level forecast results for fiscal years 2018 and 2019 when using the Case 10 dataset. (a)-(b) are the forecast results for March-April 2018, and (c)-(d) are the forecast results for March-April 2019. [Figure 39] These are water level forecast results for fiscal years 2020 and 2021 when using the Case 10 dataset, where (a)-(b) are the forecast results for March-April 2020, and (c)-(d) are the forecast results for March-April 2021. [Diagram 40] These are water level forecast results for fiscal years 2018 and 2019 when using the Case 11 dataset. (a)-(b) are the forecast results for March-April 2018, and (c)-(d) are the forecast results for March-April 2019. [Diagram 41] These are water level forecast results for fiscal years 2020 and 2021 when using the Case 11 dataset. (a)-(b) are the forecast results for March-April 2020, and (c)-(d) are the forecast results for March-April 2021. [Diagram 42] These are water level forecast results for fiscal years 2018 and 2019 when using the Case 12 dataset. (a)-(b) are the forecast results for March-April 2018, and (c)-(d) are the forecast results for March-April 2019. [Diagram 43]These are water level forecast results for fiscal years 2020 and 2021 when using the Case 12 dataset. (a)-(b) are the forecast results for March-April 2020, and (c)-(d) are the forecast results for March-April 2021. [Diagram 44] This is a comparison of the predicted values ​​and the actual measured values ​​when using the datasets for Case 05 and Case 06. (a) to (d) are the results for Case 05 after 6, 12, and 18.24 hours, and (e) to (h) are the results for Case 06 after 6, 12, and 18.24 hours. [Diagram 45] This is a comparison of the predicted values ​​and the actual measured values ​​when using the datasets for Case 07 and Case 08. (a) to (d) are the results for Case 07 after 6, 12, and 18.24 hours, and (e) to (h) are the results for Case 08 after 6, 12, and 18.24 hours. [Diagram 46] These are the results of a comparison between the predicted values ​​and the actual measured values ​​when using the data sets for Case 09 and Case 10. (a) to (d) are the results for Case 09 after 6, 12, and 18.24 hours, and (e) to (h) are the results for Case 10 after 6, 12, and 18.24 hours. [Figure 47] These are the results of a comparison between the predicted values ​​and the actual measured values ​​when using the data sets for Case 11 and Case 12. (a) to (d) are the results for Case 11 after 6, 12, and 18.24 hours, and (e) to (h) are the results for Case 12 after 6, 12, and 18.24 hours. [Figure 48] The figures show the results of a simulation. (a) compares the slope of the correlation equation for each of Case 01 to Case 12, and (b) compares the correlation coefficient for each of Case 01 to Case 12. [Figure 49] 1 shows an example of the configuration of a conventional flood warning system. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0015] Hereinafter, the embodiment of the present invention will be described in detail with reference to the drawings. Each drawing is merely a schematic illustration to allow a sufficient understanding of the present invention. Therefore, the present invention is not limited to the illustrated examples. In each drawing, common or similar components are denoted by the same reference numerals, and duplicated explanations thereof will be omitted.

[0016] <Configuration of the learning device and prediction device according to the embodiment> A learning device 10 and a prediction device 20 according to an embodiment will be described with reference to Fig. 1. Fig. 1 is a functional configuration diagram of the learning device 10 and the prediction device 20 according to an embodiment. Note that the learning device 10 and the prediction device 20 can also be configured as a single device. The learning device 10 is a device that learns the relationship between the state of snowmelt in a river basin (also called a "catchment basin") and the water level of the river. The learning device 10 performs machine learning using meteorological item value information related to the water level of the river and the basin. The meteorological item value information may be, for example, information obtained by processing a weather distribution image F related to the basin, or an AMeDAS observation value observed at a point in the basin. The prediction device 20 is a device that predicts the water level of a river at a predetermined time from the state of snowmelt in the river basin. The prediction device 20 predicts the water level using a trained learning device (learning model) that has been trained using the learning device 10.

[0017] (Learning device configuration) 1, the learning device 10 includes an image processing unit 11, a learning data acquisition unit 12, and a learning processing unit 13. The image processing unit 11, the learning data acquisition unit 12, and the learning processing unit 13 are realized by program execution processing by a CPU (Central Processing Unit), a dedicated circuit, etc. When these functions are realized by program execution processing, a program for realizing the function is stored in a storage unit (not shown).

[0018] The image processing unit 11 has a function of obtaining meteorological item value information used for learning from the weather distribution image Fa. The meteorological item value information is information on meteorological item values ​​in a river basin (e.g., temperature, precipitation, wind speed, snow depth, sunshine hours, etc.). The weather distribution image Fa is input to the image processing unit 11. In addition to the meteorological item value information, the image processing unit 11 obtains representative point information from the weather distribution image Fa. The representative point here reflects the intensity distribution of meteorological items in the river basin, and is obtained by image processing the weather distribution image Fa. The representative point is, for example, the center of gravity of the basin when the intensity of the meteorological item value is considered as weight. The representative point information is, for example, the position of the representative point, and the distance from a predetermined reference point to the representative point.

[0019] The processing of the image processing unit 11 will be described with reference to Fig. 2 and Fig. 3. Fig. 2 is an image diagram of a river and a drainage basin. Fig. 3 is a diagram for explaining the processing of the image processing unit 11, in which (a) is an example of a weather distribution image Fa, (b) is an example of a drainage basin image E obtained by extracting the weather distribution image Fa within the drainage basin, and (c) is an example of a representative point. In FIG. 2, the river is represented by the symbol K1, and the basin is represented by the symbol K2. The basin K2 indicates an area where rain flows into the river K1, and the rain that falls in the basin K2 flows out into the sea K3 via the river K1. The basin K2 may be based on a basin map published by the Ministry of Land, Infrastructure, Transport and Tourism, for example. An observation station K4 is provided downstream of the river K1, and the water level of the river K1 can be measured. In this embodiment, the water level at the observation station K4 is predicted, and attention is paid to the basin K2A upstream of the observation station K4 (i.e., the melting of snow accumulated in the upstream basin K2A). It is preferable to select a portion of the basin K2 to be focused on depending on the location of the water level to be predicted. For example, when predicting the water level near the mouth of the river K1, it is preferable to focus on the entire basin K2, including the upstream basin K2A and the downstream basin K2B. It is preferable that the location for predicting water levels be a place that is not affected by tides, and predictions at a location away from the river mouth, where the effects of tidal fluctuations are less, are more accurate than predictions near the river mouth. The time it takes for water from melting snow that has accumulated in the upstream basin K2A to affect changes in the water level at observation station K4 is determined by the characteristics of river K1, such as its length and elevation difference. The river K1 shown in Figure 2 is the Ishikari River.

[0020] The image processing unit 11 extracts a watershed image E (see FIG. 3(b)), which is an image of the watershed K2A portion, from the weather distribution image Fa (see FIG. 3(a)), and converts the RGB values ​​of each pixel constituting the watershed image E into a luminance value. The luminance value is also called a gray value. The conversion from the RGB value to the luminance value is performed, for example, using equation (1). Brightness value = 255-(0.587 * r value + 0.114 * g value + 0.299 * b value) Equation (1) For example, if the intensity of meteorological item values ​​increases in the order of white, light blue, light blue, dark blue, yellow, orange, red, and brown, a luminance conversion equation is constructed so that the luminance value of white becomes closer to "0 (zero)" and the luminance value of brown becomes closer to "255" by luminance conversion. In other words, the RGB values ​​of red, blue, and green are weighted according to the intensity of the meteorological items in the intensity distribution. In this way, the image processing unit 11 creates a gray image D (see FIG. 3(c)) by luminance converting the watershed image E (see FIG. 3(b)).

[0021] The image processing unit 11 also calculates the center of gravity G of the watershed K2A as a representative point when the brightness value of each pixel constituting the gray image D (see FIG. 3(c)) is regarded as a weight. The calculation of the center of gravity of the watershed K2A can be found from the distance (e.g., coordinate value) from a reference point (here, origin O) set in the weather distribution image Fa (see FIG. 3(a)) to each pixel and the magnitude of the brightness value, and is performed using, for example, equation (2). Here, the origin O is the lower left vertex of the weather distribution image Fa, and the X-axis and Y-axis are set on the sides constituting the weather distribution image Fa, and the X-coordinate value and the Y-coordinate value correspond, for example, to the position (array) of the pixel. · X coordinate of center of gravity G G = (X 1 M 1 + X 2 M 2 + … + X n M n ) / (M 1 + M 2 + … + M n ) ···Formula (2) Here, "X 1 ,X 2 , … ,X n " is the distance from the origin O to each pixel in the X-axis direction (X coordinate value), and "M 1 ,M 2 , … ,M n " is the brightness value of each pixel. The Y coordinate value of the center of gravity G is calculated in the same way, and the center of gravity coordinate (X G , Y G ) is obtained. Then, the image processing unit 11 calculates the position of the center of gravity G (center of gravity coordinates (X G , Y G ) and the distance L from a reference point (here, origin O) to the center of gravity G are regarded as representative value information. Furthermore, the image processing unit 11 calculates a statistical value R of the brightness values ​​of each pixel constituting the gray image D (see FIG. 3(c)), and regards the calculated statistical value R as meteorological item value information. The statistical value R of the brightness values ​​may be, for example, the average value of the brightness values ​​of each pixel constituting the gray image D.

[0022] With reference to Fig. 4, a process will be described in which three types of weather distribution images F, an hourly average temperature image F1a, an hourly accumulated precipitation image F2a, and an hourly average wind speed image F3a, are used. Fig. 4 shows an image of calculation of the center of gravity G in each weather distribution image Fa, with Fig. 4(a) showing the calculation result in the hourly average temperature image F1a, Fig. 4(b) showing the calculation result in the hourly accumulated precipitation image F2a, and Fig. 4(c) showing the calculation result in the hourly average wind speed image F3a. As shown in Fig. 4(a), the center of gravity coordinates (X1, Y1), distance L1 to the center of gravity G, and brightness statistical value R1 are obtained from the hourly average temperature image F1a. The hourly average temperature (°C) is the average temperature in an hourly unit (for example, between 0 and 59 minutes). The average of observation values ​​measured for a specific few minutes between 0 and 59 minutes (for example, 10 minutes from 50 to 59 minutes) may be used as the hourly average temperature, and in this embodiment, the value calculated based on the observation values ​​for 10 minutes is used as the hourly average temperature. 4(b), the barycenter coordinates (X2, Y2), the distance L2 to the barycenter G, and the luminance statistic R2 are obtained from the time-integrated precipitation image F2a. The time-integrated precipitation is the cumulative value of the precipitation in a unit of time (e.g., between 0 and 59 minutes). The time-integrated precipitation may be obtained based on the observation values ​​measured for a specific few minutes between 0 and 59 minutes (e.g., 10 minutes from 50 to 59 minutes), and in this embodiment, the time-integrated precipitation is calculated based on the observation values ​​for 10 minutes. 4(c), the center of gravity coordinates (X3, Y3), distance L3 to the center of gravity G, and luminance statistical value R3 are obtained from the hourly average wind speed image F3a. The hourly average wind speed is the average wind speed in an hourly unit (for example, between 0 and 59 minutes). The average value of observation values ​​measured for a specific few minutes between 0 and 59 minutes (for example, 10 minutes from 50 to 59 minutes) may be used as the hourly average wind speed, and in this embodiment, the value calculated based on the observation values ​​for 10 minutes is used as the hourly average wind speed. In this way, by using an hourly weather distribution image F that reflects the cumulative and average observation values, it is expected that the accuracy of water level predictions can be improved. However, it is also possible to use a weather distribution image F that reflects representative values ​​for each hour (for example, observation values ​​measured at a specific time between 0 and 59 minutes).

[0023] The learning data acquisition unit 12 shown in FIG. 1 has a function of organizing learning data. Information on the water level of the river, information on meteorological item values ​​of the basin, information on the representative value of the basin, etc. are input to the learning data acquisition unit 12. The learning data acquisition unit 12 organizes the learning data according to instructions from a person who predicts the water level, for example. The information on the water level of the river, information on meteorological item values ​​of the basin, and information on the representative value of the basin may be registered in advance in the storage unit, and the learning data acquisition unit 12 may acquire the registered information as learning data. The storage unit may be included in a device connected via a communication line. Hereinafter, the organized collection of learning data may be referred to as a "learning data set."

[0024] Fig. 5A shows an example of a learning data set organized by the learning data acquiring unit 12. Fig. 5A shows an example of a learning data set. The learning data set shown in Fig. 5A is composed of multiple pieces of learning data having different reference times (see reference numeral 51 in Fig. 5A). The learning dataset shown in FIG. 5A includes information for 2018 to 2021, and the information for each year corresponds to the hourly water level at the river observation station from March to May. Each learning data is composed of an input data portion (see reference numeral 52 in FIG. 5A) and an output data portion (see reference numeral 53 in FIG. 5A). The input data portion is information that is input when predicting the water level, and the output data portion is information that is output when predicting the water level. March to May is an example of a snowmelt period when snow is present in the watershed.

[0025] As shown in FIG. 5A, the input data portion is mainly composed of data (1) (see reference numeral 54 in FIG. 5A), data (2) (see reference numeral 55 in FIG. 5A), and data (3) (see reference numeral 56 in FIG. 5A). Figure 5B is an enlarged view of data (1) of the training data set, Figure 5C is an enlarged view of data (2) of the training data set, and Figure 5D is an enlarged view of data (3) of the training data set and a portion of the output data.

[0026] As shown in FIG. 5B, data (1) is the “center of gravity coordinate (X G ,Y G ) "distance to the center of gravity L" and "average value within the region R". As shown in FIG. 5C, data (2) is composed of AMeDAS observation values ​​of snow depth and sunshine duration from t-24h to t+24h, where t is the reference time. As shown in Figure 5D, data (3) consists of the water level at the observation station at time t. As shown in FIG. 5D, the output data portion is made up of water levels at the observation station from t+1h to t+24h, where t is a reference time. That is, when making a prediction, data (1) to (3) are input to a trained learning device, and the water level for each hour up to 24 hours ahead is output as shown in data (4). Note that time t in the training data is an example of a "first time point," times t-24h to t+24h are an example of a "second time point," and times t+1h to t+24h are an example of a "third time point."

[0027] Data (1) shown in FIG. 5B is obtained, for example, from the weather distribution image F. Data (1) assuming that time t is "2018 / 3 / 1 1:00" will be described. In this case, data (1) has a total of 25 center of gravity coordinates (X1, Y1), distance L1 to the center of gravity G, and average value R1 within the watershed, obtained from the hourly average temperature image F1a (see FIG. 4(a)) at time t-24h (2018 / 2 / 28 1:00), time t-23h (2018 / 2 / 28 2:00), ··, time t+23h (2018 / 3 / 2 0:00), and time t+24h (2018 / 3 / 2 1:00). In addition, it has a total of 25 center of gravity coordinates (X2, Y2), distance L2 to the center of gravity G, and average value R2 within the basin, which are obtained from the time-integrated precipitation image F2a (see Figure 4(b)) at time t-24h (2018 / 2 / 28 1:00), time t-23h (2018 / 2 / 28 2:00), ··, time t+23h (2018 / 3 / 2 0:00), and time t+24h (2018 / 3 / 2 1:00). In addition, it has a total of 25 center of gravity coordinates (X3, Y3), distance L3 to the center of gravity G, and average value R3 within the watershed, which are obtained from the hourly average wind speed image F3a (see Figure 4(c)) at time t-24h (2018 / 2 / 28 1:00), time t-23h (2018 / 2 / 28 2:00), ··, time t+23h (2018 / 3 / 2 0:00), and time t+24h (2018 / 3 / 2 1:00).

[0028] Data (2) shown in FIG. 5C is, for example, information measured at specific points (for example, a total of 20 points) in the region. The information of data (2) may be, for example, AMeDAS observation values. Data (2) will be described assuming that time t is "2018 / 3 / 1 1:00". In this case, data (2) has a total of 500 pieces of snow depth information measured at points 1 to 20 at time t-24h (2018 / 2 / 28 1:00), time t-23h (2018 / 2 / 28 2:00), ··, time t+23h (2018 / 3 / 2 0:00), and time t+24h (2018 / 3 / 2 1:00). In addition, the system has information on a total of 500 sunshine hours measured at locations 1 to 20 at times t-24h (2018 / 2 / 28 1:00), t-23h (2018 / 2 / 28 2:00), ··, t+23h (2018 / 3 / 2 0:00), and t+24h (2018 / 3 / 2 1:00).

[0029] Data (3) shown in FIG. 5D is information measured at a specific point (e.g., an observation station) on a river. When time t is assumed to be “2018 / 3 / 1 1:00”, data (3) is the water level at time t (2018 / 3 / 1 1:00). Data (4) shown in Fig. 5D is information measured at a specific point (e.g., an observation station) on a river. When time t is assumed to be "2018 / 3 / 1 1:00", data (4) is the water level at time t+1h (2018 / 3 / 1 2:00), ··, time t+23h (2018 / 3 / 2 0:00), and time t+24h (2018 / 3 / 2 1:00).

[0030] In this embodiment, the learning data is assumed to have data (1) and (2), but the learning data may have only data (1) or (2). In other words, the learning data may have a configuration of data (1), (3), and (4) or a configuration of data (2), (3), and (4). Also, the configurations of data (1) and (2) shown in FIGS. 5B and 5C are merely examples, and the types of meteorological items included in data (1) and (2) are not limited to those shown in FIGS. 5B and 5C. For example, it is also possible to include items other than the meteorological items shown in FIGS. 5B and 5C in data (1) and (2), or to exclude some of the meteorological items from data (1) and (2). In the present embodiment, meteorological items (temperature, precipitation, wind speed) for which it is easy to obtain a meteorological distribution image are used as data (1), and meteorological items (snow depth, sunshine duration) for which it is difficult to obtain a meteorological distribution image are used as data (2). Also, data (1) and (2) can also be composed only of information for the period from time t - 24h to t. That is, data (1) and (2) may not include information for the period from time t + 1h to t + 24h.

[0031] The training dataset may be thinned out based on the flooding period. Referring to FIG. 6, an example of thinning out the dataset will be described. FIG. 6(a) is a diagram for explaining a method of selecting the maximum water level, and (b) is a graph showing the fluctuation of the water level at the observation point. Here, it is assumed that thinning is performed based on the flooding when the water level at the observation point exceeds "7m" (see FIG. 6(b)). The water level of "7m" at the observation point is the water level at the time of flooding that occurs several times a year. Regarding the water level WL at a certain time t, a time t that satisfies the conditions of "WL(t - 1) < WL(t)", "WL(t - 2) < WL(t)", "WL(t + 1) < WL(t)", and "WL(t + 2) < WL(t)" (see FIG. 6(a)) and has a water level exceeding "7m" is selected as the flooding time, and a flooding period T is set based on the flooding time t. Since the water level at the flooding time t is higher than the water levels in the preceding and succeeding periods, it is called the maximum water level. The flooding period T is defined as t ± 72h of the flooding time t. If another maximum water level is included in the flooding period T, the flooding period T is from the time t - 72h of the maximum water level at the earliest timing to the time t + 72h of the maximum water level at the latest timing (see FIG. 6(b)).

[0032] Then, data to be included in the dataset is selected based on the flood period T. Examples of thinned datasets are shown below in Case 02 to Case 04. Note that Case 01 shows the state of the dataset before thinning, and for example, all data for the entire period (one example is the snowmelt period (March to May)) is included in the learning dataset. In Case 02, only data from the flood period T is included in the learning dataset. The amount of data in Case 02 is "N." In other words, Case 02 is obtained by thinning out the data from Case 01 for the period T' outside the flood period. In Case 03, data was randomly selected from the non-flood period T' as much as the number of data N in flood period T (N x 1), and the data from flood period T was combined with the data selected from the non-flood period T'. The number of data in Case 03 is "N x 2". In Case 04, data that is twice the number of data N in flood period T (N x 2) is randomly selected from non-flood period T', and the data in flood period T is combined with the data selected from non-flood period T'. The number of data in Case 04 is "N x 3".

[0033] The learning processing unit 13 shown in FIG. 1 has a learning device, and the learning device is trained by machine learning using a learning data set. The learning device is a learning system in machine learning, and can lead to good results by comparing the results of classification, prediction, judgment, etc. based on given data with the actual results that are correct answers, and adjusting various parameters. The learning device is, for example, a neural network, and the present embodiment will also be described assuming a neural network. A neural network is generally composed of an input layer, an intermediate layer, and an output layer. An explanatory variable is input to the input layer, and a target variable is output from the output layer. The intermediate layer plays a role of relating the two. There is no restriction on the number of intermediate layers and the number of neurons in each intermediate layer, and they can be set arbitrarily.

[0034] The method of training the learning device is not particularly limited, and it is preferable to train the learning device by a method suitable for the type of the learning device. For example, (1) the "center of gravity coordinates (X G ,Y G ) “distance to the center of gravity L” “average value within the region R” (2) AMeDAS observation values ​​of snow depth and sunshine hours from time t-24h to t+24h, and (3) water level at the observation station at time t are input into the input layer, and the middle layer is adjusted based on the error between the result output from the output layer and (4) water level at the observation station from time t+1h to t+24h. In other words, the learning processing unit 13 inputs meteorological item value information about the river water level at a first point in time (one example is time t) during the snowmelt period, and at multiple second points in time (one example is times t-24h to t+24h) at specified times during a specific period including the first point in time, and trains the learning device to output the river water level at multiple third points in time (one example is times t+1h to t+24h), which are second points in time after the first point in time.

[0035] (Configuration of prediction device) 1, the prediction device 20 includes an image processing unit 21, a prediction data acquisition unit 22, and a prediction processing unit 23. The image processing unit 21, the prediction data acquisition unit 22, and the prediction processing unit 23 are realized by a program execution process by a CPU (Central Processing Unit), a dedicated circuit, etc. When these functions are realized by a program execution process, a program for realizing the function is stored in a storage unit (not shown).

[0036] The image processing unit 21 shown in FIG. 1 has a function of obtaining meteorological item value information used for prediction from the weather distribution image Fb. The meteorological item value information is information on meteorological item values ​​in the river basin (for example, temperature, precipitation, wind speed, snow depth, sunshine hours, etc.). The weather distribution image Fb is input to the image processing unit 21. In addition to the meteorological item value information, the image processing unit 21 obtains representative point information from the weather distribution image Fb. The processing of the image processing unit 21 is similar to the processing of the image processing unit 11 in the learning stage, so a detailed description will be omitted. The weather distribution image Fb may be of the past (for example, actually measured) or of the future (for example, forecast). In other words, the current or near future water level can be predicted from the past actual weather distribution image Fb, and the water level in the even more distant future can be predicted from the future weather distribution image Fb published as a forecast.

[0037] The forecast data acquisition unit 22 shown in Fig. 1 has a function of organizing forecast data. Information on the water level of the river (at a time before the prediction time), information on meteorological item values ​​of the basin, information on the representative value of the basin, etc. are input to the forecast data acquisition unit 22. The forecast data acquisition unit 22 organizes the forecast data according to instructions from a person who predicts the water level, for example. Note that the information on the water level of the river, information on meteorological item values ​​of the basin, and information on the representative value of the basin may be registered in advance in the storage unit, and the forecast data acquisition unit 22 may acquire the registered information as forecast data. The storage unit may be provided in a device connected via a communication line.

[0038] The prediction data is composed of information corresponding to the input data portion of the learning data (see reference numeral 52 in FIG. 5A). In other words, the prediction data is, for example, the following: (1) The "center of gravity coordinates (X G ,Y G), "distance to the center of gravity L", "average value within the region R", (2) AMeDAS observation values ​​of snow depth and sunshine hours from time t-24h to t+24h, and (3) water level at the observation station at time t. Based on this prediction data, (4) water level at the observation station from time t+1h to t+24h can be predicted. If it is assumed that the reference time t is the current time and the water level is predicted for each hour from the present time up to 24 hours ahead, the time t+1h to t+24h portion of data (1) and (2) included in the prediction data may be forecast values ​​or information obtained by separately performing numerical analysis. If data (1) and (2) of the learning data are composed only of information for the period from time t-24h to t (i.e., if they do not include information for the period from time t+1h to t+24h), data (1) and (2) of the prediction data can also be composed only of information for the period from time t-24h to t. In this way, water levels can be predicted further into the future without using forecast values, etc. Note that time t in the prediction data and the prediction results is an example of a "fourth time point," times t-24h to t+24h are an example of a "fifth time point," and times t+1h to t+24h are an example of a "sixth time point."

[0039] The prediction processing unit 23 shown in FIG. 1 has a trained learning device. The trained learning device is a learning device that has been machine-learned by the learning device 10. In other words, the trained learning device is trained to output the water level of the river at a fourth time point (one example is time t) during a snowmelt period, and at a plurality of fifth time points (one example is times t-24h to t+24h) at predetermined times during a specific period including the fourth time point, by inputting weather item value information at the fifth time points (one example is times t+1h to t+24h) at predetermined times during a specific period including the fourth time point. The prediction processing unit 23, for example, outputs the "center of gravity coordinates (X G ,Y G ) ``distance to the center of gravity L'' ``average value within the region R'' (2) AMeDAS observation values ​​of snow depth and sunshine hours from time t-24h to t+24h, and (3) water level at the observation station at time t are input into the learning machine, and (4) the water level at the observation station from time t+1h to t+24h is predicted.

[0040] (Regarding weighting) The contribution of meteorological items related to snowmelt may differ depending on the region and time of year. Therefore, the contribution of meteorological items to water level prediction may be set as a weight, and the learning data and prediction data may be weighted (e.g., multiplied by the weight) to predict river water levels. For example, weighting may be performed using multiple regression analysis. Multiple regression analysis is a method of modeling the relationship between the objective variable and explanatory variables as in equation (3), and constructing an equation to find the objective variable (y in equation (3)) using the explanatory variables (xi in equation (3)). The wi that is multiplied by the explanatory variables in equation (3) corresponds to the weight. ·y=w0+w1x1+w2x2+w3x3+···+wnxn ···Formula (3) In multiple regression analysis, multiple combinations of xi and y are used to find the value of wi closest to y using a mathematical method (such as Lagrange's undetermined multiplier method). In this embodiment, the explanatory variables xi correspond to meteorological item values, the objective variable y corresponds to the river water level, and the weight of the meteorological items corresponds to wi.

[0041] The river water level and meteorological item values ​​used in the multiple regression analysis (i.e., information used to calculate the weighting value) are preferably those within the basin of the river for which the water level is predicted, and are preferably observed values ​​at the prediction point where the river water level is predicted or at a point upstream of the prediction point. The multiple regression analysis may be performed for each of a plurality of points, and the results obtained by performing statistical calculations on the results may be used as the weighting value. The river water level and meteorological item values ​​used in the multiple regression analysis are preferably those for a period related to the learning data used to learn the river water level and the prediction data used for prediction, and are preferably those for a time point corresponding to the learning data and the prediction data. The types and number of meteorological items used in the multiple regression analysis are not particularly limited, but are preferably meteorological items corresponding to the learning data and the prediction data. The weighting value may be set using a method other than the multiple regression analysis. The weighting value may be set not only based on mathematical calculations, but may be based on experience, for example.

[0042] During learning and prediction, the representative point information and / or weather item value information are weighted using a weight value, and learning or prediction is performed using the weighted information. For example, the learning processing unit 13 multiplies the representative point information and / or weather item value information included in the learning data by a weight value, and performs learning using the information after multiplication by the weight value. Furthermore, the prediction processing unit 23 multiplies the representative point information and / or weather item value information included in the prediction data by a weight value, and performs prediction using the information after multiplication by the weight value. Note that the representative point information and / or weather item value information may be weighted directly, or the representative point information and / or weather item value information may be weighted indirectly by weighting the weather distribution image F.

[0043] <Effects of the learning device and the prediction device according to the embodiment> As described above, by using a trained learning device trained by the learning device 10 according to the embodiment, the prediction device 20 can predict the future trend (rise / fall) of water levels. Therefore, it is possible to determine the timing of evacuation / restoration, which is useful for safety management and process management of river construction work. Furthermore, by learning using weighted meteorological item value information, the contribution of meteorological items to water level prediction results is reflected in the learning data, making it possible to achieve even more accurate predictions. In addition, by learning using a learning data set that has been thinned out based on the flood period T, the calculation load can be reduced compared to inputting all data during the snowmelt period.

[0044] In order to verify the effects of the learning device 10 and the prediction device 20 according to the embodiment, a simulation was performed using actual data, which will now be described. In the simulation, in order to compare the effect of thinning out the learning data set, learning was performed using data sets of Case 01 to Case 04, and the prediction results were compared. In addition, in order to compare the effect of weighting, learning was performed with and without weighting, and the prediction results were compared. The flow of the simulation is shown in FIG. 7. FIG. 7 is a prediction procedure flow in the simulation. The prediction procedure mainly includes the steps of selecting a water level prediction point (S1), preparing data (S2), calculating center of gravity information of the image, etc. (S3), creating thinned data (S4), weighting meteorological items (S5), and learning and predicting the water level (S6).

[0045] (Selection of water level prediction point "S1") In the simulation, the Tsukigata water level observation station on the Ishikari River, which flows through Hokkaido, was selected as the water level prediction point, and the water level at that observation station was predicted (see Figure 2). The Ishikari River is the river indicated by the symbol K1 in Figure 2, and the Tsukigata observation station is located at the position indicated by the symbol K4 in Figure 2. The drainage area of ​​the Ishikari River is approximately 9,400 km2. 2 As shown in Figure 2, Tsukigata Observatory is located downstream of the Ishikari River.

[0046] (Data preparation "S2") For the simulation, hourly weather distribution images and AMeDAS observation values ​​from March to May (snowmelt period) from 2018 to 2021 were prepared. The meteorological items for which weather distribution images were prepared were temperature, precipitation, and wind speed, and images were prepared within the watershed. The types of images were "estimated weather distribution" for temperature, "analyzed rainfall" for precipitation, and "hourly atmospheric analysis" for wind speed. These images were provided by the Japan Meteorological Agency and can be obtained. AMeDAS observation values ​​were used for snow depth and sunshine hours. The observation points were 20 observation stations within the Tsukigata Observatory's basin (diamond-shaped locations indicated by reference number 80 in Figure 8), and the observation values ​​were obtained from the Japan Meteorological Agency's website. Figure 8 shows the locations of the observation stations within the Tsukigata Observatory's basin. Initially, we considered using weather distribution images for all meteorological items (temperature, precipitation, wind speed, snow depth, and sunshine hours), but there was insufficient provision of weather distribution images for snow depth and sunshine hours, so we prioritized increasing the amount of learning data and used AMeDAS observation values.

[0047] (Calculation of image center of gravity information, etc. "S3") The temperature, precipitation, and wind speed using the weather distribution image are calculated by dividing the image into representative values ​​by the "center of gravity coordinates (X G ,Y G )," "distance to the center of gravity L," and "brightness statistical value R" were calculated (see Figure 4). If all data from the weather distribution image (for example, information from all pixels) is used, problems such as deep learning taking a long time or the deep learning weights not converging can occur, so to prevent this, information from representative values ​​is used.

[0048] (Creating thinned data "S4") The Case 01 dataset used in the simulation contains information from 2018 to 2021, and the information for each year corresponds to the hourly water level at the river observation station from March to May (see Figure 5A). This results in a total of approximately 8,800 data sets (92 days / year x 4 years from 2018 to 2021 x 24 hours) (approximately 1,500 data / year). Each learning data is the "center of gravity coordinate (X G ,Y G ) "distance to the center of gravity L" and "average value within the watershed R", (2) AMeDAS observation values ​​of snow depth and sunshine hours from time t - 24h to t + 24h, (3) water level at Tsukigata Observatory at time t, and (4) water level at Tsukigata Observatory from time t + 1h to t + 24h.

[0049] Moreover, the data sets of Case 02 to Case 04 used in the simulation were created by the thinning method described with reference to FIG. In Case 02, only data from the flood period T is included in the learning dataset. The amount of data in Case 02 is "N." In other words, Case 02 is obtained by thinning out the data from Case 01 for the period T' outside the flood period. In Case 03, data was randomly selected from the non-flood period T' as much as the number of data N in flood period T (N x 1), and the data from flood period T was combined with the data selected from the non-flood period T'. The number of data in Case 03 is "N x 2". In Case 04, data that is twice the number of data N in flood period T (N x 2) is randomly selected from non-flood period T', and the data in flood period T is combined with the data selected from non-flood period T'. The number of data in Case 04 is "N x 3".

[0050] (Weather item weighting "S5") A multiple regression equation (Equation (4)) was obtained between hourly AMeDAS observations from March to May 2022 and the water level at Tsukigata Observatory, and the regression coefficients were used as weighting values. ·y=w1x1+w2x2+w3x3+w4x4+w5x5+b ·Formula (4) Here, y is the water level at Tsukigata Observatory, w1 is the weight of snow cover, w2 is the weight of temperature, w3 is the weight of precipitation, w4 is the weight of sunshine hours, w5 is the weight of wind speed, and b is a coefficient. The observation points used to calculate the weights were Furano (observation station indicated by reference number 81 in Figure 8), Asahikawa (observation station indicated by reference number 82 in Figure 8), and Shumarinai (observation station indicated by reference number 83 in Figure 8), located near the center of the basin. Multiple regression analysis was performed for each observation station, and a multiple regression equation was derived to calculate the water level at Tsukigata observation station from the 24-hour cumulative AMeDAS observation values. The multiple regression coefficients obtained are shown in the table in Figure 9, and in this simulation, the average value of the three stations was used as the weight value. Figure 9 is a list of the weight values ​​used in the simulation.

[0051] (Water level learning and prediction "S6") Water levels were predicted for each year using the data sets of Case 01 to Case 04. Learning was performed excluding the data for the year to be predicted from Case 01 to Case 04, and when making predictions, data (1) to (3) for the year to be predicted were input. The water level prediction results from the simulation are shown in Figures 10 to 25. Figures 10 to 13 show the prediction results when the data set for Case01 was used, Figures 14 to 17 show the prediction results when the data set for Case02 was used, Figures 18 to 21 show the prediction results when the data set for Case03 was used, and Figures 22 to 25 show the prediction results when the data set for Case04 was used. Moreover, the comparison results between the predicted values ​​and the actual measured values ​​after 6, 12, 18, and 24 hours are shown in Figures 26 and 27. Figure 26 shows the comparison results when the data sets of Case01 and Case02 are used, and Figure 27 shows the comparison results when the data sets of Case03 and Case04 are used.

[0052] From the time series graphs shown in Figures 10 to 25, we can see that the prediction accuracy improved in each case by applying weighting, and that the results of Case 01 without thinning were similar to those of Case 03 and Case 04 with thinning. Also, from FIG. 26 and FIG. 27, it can be seen that Case 03 and Case 04 have roughly the same accuracy as Case 01 without thinning.

[0053] Figure 28 shows the results of comparing the slope of the correlation equation and the correlation coefficient for each case. Figure 28(a) compares the slope of the correlation equation for Cases 01 to 04, and (b) compares the correlation coefficient for Cases 01 to 04. As shown in Figure 28, it was confirmed that Cases 03 and 04 had roughly similar values ​​to Case 01. These results show that thinning out the learning data does not significantly reduce prediction accuracy, and that water levels can be predicted accurately by using the same number of non-flood period data as the flood period T or more.

[0054] In addition, a simulation was performed by changing the method of thinning the learning data set, which will be explained below. In the additional simulation, learning was performed using the data sets from Case 05 to Case 12, and the prediction results were compared. Other than the thinning method, the simulation was the same as the simulation described above. In Case 02 to Case 04, the flood period T was set to "t ± 72 h", but in Case 05 to Case 09, the flood period was changed to "t ± 36 h" or "t ± 144 h". In addition, in Case 02 to Case 04, data was arbitrarily selected from the period T' other than the flood period, and the data in the flood period T and the data selected from the period T' other than the flood period were combined. In Case 10 to Case 12, data for a continuous period starting from 01:00 on April 1, 2019 was selected as the data for the period T' other than the flood period, and combined with the data for the flood period T. Details are as follows.

[0055] In Case 05, only the data during the flood period T (t ± 36 h) was included in the learning dataset. The number of data in Case 05 is “N”. In Case 06, data was randomly selected from the period T' outside the flood period as much as the number of data N in the flood period T (t ± 36 h) (N × 1 times), and the data from the flood period T (t ± 36 h) was combined with the data selected from the period T' outside the flood period. The number of data in Case 06 is "N × 2". In Case 07, data that is twice the number of data N (N x 2) during flood period T (t ± 36 h) is randomly selected from the period T' outside the flood period, and the data during flood period T (t ± 36 h) is combined with the data selected from the period T' outside the flood period. The number of data in Case 07 is "N x 3".

[0056] In Case 08, only the data during the flood period T (t ± 144 h) was included in the learning dataset. The number of data in Case 08 is “N”. In Case 09, data was randomly selected from the period T' outside the flood period as the number of data N in the flood period T (t ± 144 h) (N × 1 times), and the data from the flood period T (t ± 144 h) was combined with the data selected from the period T' outside the flood period. The number of data in Case 09 is "N × 2".

[0057] In Case 10, "N x 1" data items are selected every hour from the starting point set in the period T' outside the flood period (here, 01:00 on April 1, 2019), and the "N x 1" data items in the flood period T (t ± 36 h) are combined with the data selected from the period T' outside the flood period. The number of data items in Case 10 is "N x 2". In Case 11, "N x 1" data items are selected every hour from the starting point set in the period T' outside the flood period (here, 01:00 on April 1, 2019), and the "N x 1" data items in the flood period T (t ± 72 h) are combined with the data selected from the period T' outside the flood period. The number of data items in Case 11 is "N x 2". In Case 12, "N x 1" data items are selected every hour from the starting point set in the period T' outside the flood period (here, 01:00 on April 1, 2019), and the "N x 1" data items in the flood period T (t ± 144 h) are combined with the data selected from the period T' outside the flood period. The amount of data in Case 12 is "N x 2".

[0058] Water levels were predicted for each year using the data sets from Case 05 to Case 12. Learning was performed excluding the data for the year to be predicted from Case 05 to Case 12, and when making predictions, data (1) to (3) for the year to be predicted were input. The water level prediction results from the simulation are shown in Figures 29 to 43. Figures 29 and 30 show the prediction results when the data set for Case05 is used, Figures 31 and 32 show the prediction results when the data set for Case06 is used, and Figures 33 and 34 show the prediction results when the data set for Case07 is used. FIG. 35 shows the prediction results when the dataset for Case08 is used, and FIGS. 36 and 37 show the prediction results when the dataset for Case09 is used. 38 and 39 show the prediction results when the Case 10 dataset is used, FIGS. 40 and 41 show the prediction results when the Case 11 dataset is used, and FIGS. 42 and 43 show the prediction results when the Case 12 dataset is used.

[0059] Moreover, the comparison results between the predicted values ​​and the actual measured values ​​after 6, 12, 18, and 24 hours are shown in Figures 44 to 47. Figure 44 shows the comparison results when the datasets of Case 05 and Case 06 are used, and Figure 45 shows the comparison results when the datasets of Case 07 and Case 08 are used. Figure 46 shows the comparison results when the datasets of Case 09 and Case 10 are used, and Figure 47 shows the comparison results when the datasets of Case 11 and Case 12 are used. The results of comparing the slope of the correlation equation and the correlation coefficient for each case are shown in Figure 48. Figure 48(a) compares the slope of the correlation equation for each of Case 01 to Case 12, and (b) compares the correlation coefficient for each of Case 01 to Case 12.

[0060] The combination of data from the flood period T and data from the non-flood period T' will be examined based on the results of Case02-09. Note that in Case02-09, data from the non-flood period T' is arbitrarily extracted. As shown in Figure 48, the prediction accuracy decreased in Case 02, which only had data for the flood period T (T = t ± 72 h), and in Case 05, which only had data for the flood period T (T = t ± 36 h). Water level predictions were made regardless of whether flooding occurred or not, but in Cases 02 and 05, it is believed that the reason for the decrease in water level prediction accuracy for the period T' outside of flood periods was that there was little data for the period T' outside of flood periods in the learning data. In Case 08, which only had data for the flood period T (T = t ± 144 h), it is believed that the prediction accuracy did not decrease because the learning data included a lot of data for the flood period T and the period T' outside of flood periods.

[0061] Data thinning will be considered based on the results of Cases 10 to 12. In Cases 10 to 12, data from the period T' outside the flood period is selected in continuous time. For Case 10 with flood period T (T = t ± 36 h), the prediction accuracy decreased, as shown in Figure 48. This is thought to be because the data selected after April 1, 2019 contained a small amount of data for flood period T and non-flood period T'. For Case 11 with flood period T (T = t ± 72 h) and Case 12 with flood period T (T = t ± 144 h), the prediction accuracy did not decrease because the data selected after April 1, 2019 contained sufficient data for flood period T and non-flood period T'.

[0062] From the simulation results using the Ishikari River as an example, the following was found: [1] In the method described in this embodiment, the training data can be thinned out. [2] The optimal thinning method is to divide the training data into flood periods T and non-flood periods T', and to select the training data by determining period T' based on the number of data points in flood periods T. [3] There are two methods for selecting data based on the period T' other than the flood period. The first is to select data arbitrarily, and the second is to select data continuously from a certain time. [4] Assume that data is selected arbitrarily from a period T' outside the flood period. If the flood period T is shorter than "t ± 72 h" relative to the time t of the peak water level (maximum water level) during flooding, it is best to select the same number of data points from the period T' outside the flood period as the number of data points for the flood period T (see the results of Case 06) or to select more than that (see the results of Case 07) for learning. Also, if the flood period T is "t ± 144 h" or longer, it is possible to use only the data from the flood period T as learning data. [5] Assume that data for a continuous period is selected from a starting point set in the non-flood period T'. In this case, the number of data selected from the non-flood period T' should be the same as the number of data in the flood period T, and the flood period T should be selected to be at least "t±72 hours" from the time t of the peak water level (maximum water level) during flooding (see the results of Case 11 and Case 12). [Explanation of symbols]

[0063] 10 Learning Device 11 Image processing section 12 Learning data acquisition section 13 Learning processing unit 20 Prediction Device 21 Image processing section 22 Prediction Data Acquisition Section 23 Prediction processing unit F, Fa, Fb Weather distribution image

Claims

1. a learning data acquisition unit that acquires learning data consisting of a water level of the river at a first time point during a snowmelt period when snow is present in at least a part of the river basin, meteorological item value information at a plurality of second time points during a specific period including a period in the past of the first time point, and water levels of the river at a plurality of third time points in the future of the first time point; A learning processing unit that performs machine learning on a learning device using a learning data set that is a compilation of the learning data, the learning processing unit inputs the water level of the river at the first time point and meteorological item value information at a plurality of the second time points, thereby causing the learning unit to learn so as to output the water level of the river at a plurality of the third time points; A learning device characterized by:

2. the learning data acquisition unit acquires learning data including at least one of meteorological item value information of an hourly average temperature, an hourly accumulated precipitation, an hourly average wind speed, an hourly snow depth, and an hourly sunshine duration; 2. The learning device according to claim 1 .

3. An image processing unit is further provided for calculating representative point information and meteorological item value information relating to representative points reflecting the intensity distribution in the watershed from a meteorological distribution image showing the intensity distribution of meteorological items by color changes, the learning data acquisition unit acquires learning data having the representative point information and the weather item value information calculated from the weather distribution image at the second time point; 2. The learning device according to claim 1 .

4. The image processing unit includes: converting the RGB values ​​of each pixel constituting the weather distribution image at the second time point into a luminance value, calculating the center of gravity of the watershed as the representative point when the luminance value is regarded as a weight, and determining the position of the center of gravity and the distance from a predetermined reference point to the center of gravity as the representative point information; a statistical value calculated from the brightness value of each pixel constituting the weather distribution image as the weather item value information; 4. The learning device according to claim 3.

5. 2. The learning device according to claim 1, wherein the meteorological item value information is AMeDAS observation values ​​at a plurality of points within the watershed at the second time point.

6. the learning processing unit weights the weather item value information according to the weather items, and performs learning using the weighted weather item value information.

2. The learning device according to claim 1 .

7. 2. The learning device according to claim 1, wherein the data set is thinned out based on a flood period.

8. a learning data acquisition step of acquiring learning data consisting of a water level of the river at a first time point during a snowmelt period when snow is present in at least a part of the river basin, meteorological item value information at a plurality of second time points during a specific period including a period in the past of the first time point, and water levels of the river at a plurality of third time points in the future of the first time point; A learning process step of performing machine learning on a learning device using a learning data set that summarizes the learning data, In the learning process, the learning device is trained to output water levels of the river at a plurality of third time points by inputting the water level of the river at the first time point and meteorological item value information at a plurality of second time points. A learning method comprising:

9. A learning device that has been trained by the learning method according to claim 8 is provided, and the water level of the river at a fourth time point during a snowmelt period and meteorological item value information at a plurality of fifth time points during a specific period including a period in the past before the fourth time point are input to predict the water level of the river at a plurality of sixth time points in the future than the fourth time point. A prediction device comprising:

Citation Information

Patent Citations

  • Monitoring system

    JP2008015916A

  • River water level prediction system

    JP2019045290A

  • Learning device, learning method, and prediction device

    JP2023067705A