Learning device, learning method, and prediction device

Abstract

The present invention provides a learning device, a learning method, and a prediction device that can reduce the costs required for prediction. [Solution] The learning device 10 includes a learning data acquisition unit 12 that acquires a set of learning data consisting of whether or not there is flooding in the river at a certain first time point and classification information that categorizes the rainfall situation in the river's catchment area at at least one second time point prior to the first time point, and a learning processing unit 13 that has a learner and uses the learning data to train the learner. The learning processing unit 13 takes the classification information at at least one third time point as input and trains the learner to output whether or not there is flooding in the river at a fourth time point in the future of the third time point.

Description

[Technical Field]

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

[0002] In river construction projects, it is crucial for safety management to quickly determine whether or not flooding will occur. If only human evacuation is considered, it is sufficient to know whether or not flooding will occur a few hours in advance to ensure a safe evacuation. However, to ensure the safe evacuation of heavy machinery and equipment within the construction area, it is necessary to know whether or not flooding will occur approximately 24 hours in advance.

[0003] Generally, predictions of flooding use water level and flow rate prediction methods such as direct models and physical models. An example of a direct model is a regression model. A regression model calculates a regression equation between the water level at an observation station upstream of the prediction point and the water level at the prediction point, and then predicts the water level at the prediction point from the regression equation. An example of a physical model is a tank model or runoff model. A tank model represents the runoff and inflow due to rainfall in a watershed using tanks divided into surface, intermediate, and groundwater layers. A runoff model analyzes the flow of rainfall in a watershed, taking into account land use and roughness.

[0004] Furthermore, in recent years, data-driven modeling methods are sometimes used. Examples of data-driven models include machine learning methods such as neural networks and random forests, which are methods of AI (Artificial Intelligence) (hereinafter referred to as "AI models"). In predicting water levels and flow rates using AI models, for example, the relationship between hydrological information such as rainfall and temperature upstream of the prediction point and the water level and flow rate at the prediction point is learned. While location information such as observation values ​​at observation stations is often used for hydrological information, there is also a technique that uses machine learning to predict water levels 1 to 24 hours in advance, for example, by using the centroid coordinates of the distribution obtained from rain cloud images, the distance to the centroid coordinates, the average precipitation, and the most recent water level (see Patent Document 1). [Prior art documents] [Patent Documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2022-065522 (Figure 12) [Overview of the Initiative] [Problems that the invention aims to solve]

[0006] Conventional research on methods for determining water levels and flow rates has primarily focused on improving prediction accuracy, and has not adequately considered the burden on workers and computational load. For example, each of the prediction models mentioned above has the following problems. Direct models (one example being regression models) have the drawback of requiring time to construct the model equation and tune the parameters.

[0007] Physical models (such as tank models and outflow models) offer higher prediction accuracy than direct models, but they have the drawback of requiring more time for model equation construction and parameter tuning. Furthermore, physical models face the challenge of becoming computationally intensive and difficult to predict in real time if the model is made more precise or input values ​​are increased to improve accuracy.

[0008] Data-driven models have advantages over direct or physical models, such as not requiring the construction of model equations or parameter tuning. However, a certain amount of training data is necessary for accurate predictions. Furthermore, increasing the input data too much can lead to challenges such as longer training times, failure of weight convergence, and overfitting.

[0009] From this perspective, the present invention provides a learning device, a learning method, and a prediction device that can reduce the costs required for prediction (including the costs related to preparing for prediction). [Means for solving the problem]

[0010] Conventional models for predicting flooding use, for example, water levels (measured values) and rainfall (measured and forecasted) upstream of a prediction point as input to calculate the water level and flow rate at the prediction point. If the calculation results exceed predetermined water levels or flow rates, a flood is determined. The accuracy of the flood prediction depends on the accuracy of the calculation results for water level and flow rate at the prediction point; therefore, to improve the accuracy of the prediction, it is necessary to improve the accuracy of these calculation results. The inventors of this invention believed that by directly learning and predicting the presence or absence of flooding without involving the calculation of water level and flow rate, the costs required for prediction (including the costs related to preparing for the prediction) could be reduced.

[0011] The learning device according to the present invention comprises a learning data acquisition unit and a learning processing unit. The learning data acquisition unit acquires, as learning data, a pair of information consisting of whether or not there is flooding in the river at a certain first point in time, and classification information that categorizes the rainfall conditions in the river's catchment area at at least one second point in time prior to the first point in time. The learning processing unit has a learning device, and uses the learning data to train the learning device. The learning processing unit inputs the classification information at least one third time point in time, and trains the learning device to output whether or not the river floods at a fourth time point in the future, which is later than the third time point.

[0012] In the learning device according to the present invention, classification information categorizing rainfall conditions is used as input, and the presence or absence of flooding is directly learned through machine learning. In other words, the presence or absence of flooding is directly learned without the need to calculate water level or flow rate. Therefore, various costs required for preparing predictions (for example, the burden on operators and computational load) can be reduced. For example, by using machine learning, the construction of model equations and parameter tuning become unnecessary. Also, by using classification information as input, the features can be simplified, making learning less complex (learning becomes simpler). Therefore, the computational load can be reduced. Furthermore, the input data is easy to organize, and learning is relatively easy even when the amount of input data is increased.

[0013] The classification information may be obtained based on rainfall distribution information that shows the intensity of rainfall in two dimensions. In that case, the system may further include a data processing unit that categorizes multiple rainfall distribution information sets by cluster analysis.

[0014] In this way, for example, actual rainfall data, analyzed rainfall data, and rain cloud images distributed by the Japan Meteorological Agency can be used for training. Since this information covers the entire country of Japan and is easily available, it is possible to easily learn whether or not there is flooding at any point throughout Japan.

[0015] The data processing unit may include an information reduction processing unit that creates reduced rainfall distribution information by dividing the rainfall distribution information into m rows × n columns (where m and n are natural numbers) and calculating representative values ​​for each divided region, and a classification processing unit that classifies the reduced rainfall distribution information by cluster analysis.

[0016] This approach allows for learning that includes differences in rainfall distribution (for example, whether it's a uniform distribution or a concentrated downpour) while keeping the computational load down. Therefore, using a learner trained on this data, it's possible to predict the presence or absence of flooding with greater accuracy.

[0017] The categorization processing unit repeatedly performs a process of classifying the reduced rainfall distribution information into one of several groups, for example, to determine the final class. In this case, the categorization processing unit statistically processes the representative value of the divided region on a group basis to find the centroid of the group. The categorization processing unit also calculates the distance from the centroid to the reduced rainfall distribution information for all combinations of the group and the reduced rainfall distribution information, and designates the group with the closest distance from the centroid to the reduced rainfall distribution information as the new group for the reduced rainfall distribution information. The categorization processing unit then repeatedly calculates the centroid of the group and classifies it into the new group.

[0018] The learning data may further include the classification information at the first point in time. In that case, it is preferable that the learning processing unit trains the learner to output the presence or absence of river flooding at the fourth point in time by inputting the classification information at the third point in time and the fourth point in time.

[0019] The learning data may be selected to be within a predetermined period based on the time of flooding. In this way, the sensitivity to flooding can be increased, and the accuracy of flood prediction can be improved.

[0020] A learning method according to the present invention includes a learning data acquisition step and a learning processing step. In the learning data acquisition step, a set of the presence or absence of river flooding at a certain first point in time, and classification information obtained by typifying the rainfall situation in the catchment area of the river at at least one second point in time earlier than the first point in time, is acquired as learning data. The learning processing step includes a learner, and causes the learner to perform machine learning using the learning data. In the learning processing step, the learner is trained to output the presence or absence of river flooding at a fourth point in time that is later than the third point in time, by inputting the classification information at at least one third point in time.

[0021] In the learning method according to the present invention, the presence or absence of flooding is directly learned by machine learning using classification information obtained by typifying rainfall situations as an input value. That is, the presence or absence of flooding is directly learned without going through calculation of water level or flow rate. Therefore, various costs required for preparation for prediction (for example, the burden on workers and calculation load) can be reduced. For example, the use of machine learning eliminates the need for construction of model formulas and parameter tuning. In addition, by using the classification information as an input value, the feature quantity can be simplified, so that the learning is less likely to be complicated (the learning becomes simple). Therefore, the calculation load can be suppressed. Furthermore, the input data can be easily organized, and learning is relatively easy even when the input data is increased.

[0022] The prediction device according to the present invention includes a trained learner trained by the learning method described above, and inputs the classification information at the third time point into the trained learner to predict the presence or absence of river flooding at the fourth time point.

[0023] In the prediction device according to the present invention, classification information obtained by categorizing rainfall conditions is used as an input value, and a trained learner directly predicts the presence or absence of flooding. That is, the presence or absence of flooding is directly predicted without going through calculations of water level or flow rate. Therefore, various costs required for prediction can be reduced. Furthermore, by using classification information derived from weather forecast information (for example, rain cloud images distributed by the Japan Meteorological Agency), it is possible to predict the presence or absence of flooding far in advance. This makes it possible to evacuate and protect heavy machinery, materials and equipment in the construction yard with sufficient time to spare. Effects of the Invention

[0024] According to the present invention, the cost required for prediction can be reduced. Brief Description of the Drawings

[0025] [Figure 1] This is a functional configuration diagram of a learning device and a prediction device according to an embodiment of the present invention. [Figure 2] This is a schematic diagram of a river and a catchment area. [Figure 3] This is a diagram for explaining the processing of the image processing unit, where (a) is an example of rainfall distribution information, (b) is an example of a catchment area image obtained by extracting the rainfall distribution information within the catchment area, and (c) is an example of a grayscale image obtained by converting the catchment area image from RGB values to luminance values. [Figure 4] This is a diagram for explaining the processing of the information reduction processing unit, where (a) is a diagram in which divided regions are set on a grayscale image, (b) is an enlarged image of (a), and (c) is an example of reduced rainfall distribution information (only the range corresponding to (b)) with a reduced information amount. [Figure 5] This is a diagram for explaining the processing of the categorization processing unit. [Figure 6]This is a diagram illustrating the processing of the classification and processing unit. [Figure 7] This is an example of the results of classifying rainfall distribution information into four classes. [Figure 8] This is an example of a training dataset according to an embodiment of the present invention. [Figure 9] This is an example of a flowchart illustrating a learning method and a prediction method according to an embodiment of the present invention. [Figure 10] This is a flowchart showing the procedure for conducting a predictive test. [Figure 11] This is an example of classifying the presence or absence of flooding based on water levels at observation stations. (a) is a diagram showing the water level changes and the presence or absence of flooding in a certain river in Japan in "2019". (b) is a diagram showing the water level changes and the presence or absence of flooding in the same river on "September 10, 2019". (c) is a diagram showing the water level changes and the presence or absence of flooding in the same river on "October 13, 2019". [Figure 12] This figure shows the relationship between training data and prediction data in cross-validation. [Figure 13] Here are some examples of prediction test results: (a) is the prediction result for the number of classes "3", (b) is the prediction result for the number of classes "6", and (c) is the prediction result for the number of classes "10". [Figure 14] Here are some examples of prediction test results: (a) is the prediction result for the number of classes "3", (b) is the prediction result for the number of classes "6", and (c) is the prediction result for the number of classes "10". [Figure 15] Here are some examples of prediction test results: (a) is the prediction result for the number of classes "3", (b) is the prediction result for the number of classes "6", and (c) is the prediction result for the number of classes "10". [Figure 16] Here are some examples of prediction test results: (a) is the prediction result for the number of classes "3", (b) is the prediction result for the number of classes "6", and (c) is the prediction result for the number of classes "10". [Modes for carrying out the invention]

[0026] Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the drawings as appropriate. Each figure is only a schematic representation to the extent that the present invention can be fully understood. Therefore, the present invention is not limited to the illustrated examples. In each figure, common or similar components are denoted by the same reference numerals, and their redundant descriptions may be omitted.

[0027] <Configuration of the learning device and prediction device according to the embodiment> Referring to Figure 1, the learning device 10 and prediction device 20 according to this embodiment will be described. Figure 1 is a functional configuration diagram of the learning device 10 and prediction device 20 according to this embodiment. Note that the learning device 10 and prediction device 20 can also be configured as a single device.

[0028] The learning device 10 is a device that learns the relationship between classification information that categorizes rainfall conditions in a river's catchment area (also called a "watershed") and whether or not flooding occurs at a predicted point in the river. The classification information is obtained, for example, by processing rainfall distribution information F. Rainfall distribution information F represents the intensity of rainfall in two dimensions. The prediction device 20 is a device that predicts whether or not flooding will occur at a prediction point in a river after a predetermined time, based on classification information that categorizes rainfall conditions in the river's catchment area.

[0029] The rainfall distribution information F includes, for example, (1) image data related to rainfall, and (2) numerical information that forms the basis of said image data. Image data related to rainfall includes, for example, rain cloud images. Rain cloud images are RGB images that show the distribution of rainfall intensity through changes in color, and may be, for example, those published by the Japan Meteorological Agency. The numerical information that forms the basis of the image data is, for example, the analyzed rainfall (actual rainfall) with a resolution of "1 km" distributed by the Japan Meteorological Agency.

[0030] The rainfall distribution information F shows the rainfall amounts at multiple points in the river catchment area used for learning (or prediction). The rainfall amounts that make up the rainfall distribution information F may be either measured values ​​or predicted values. It is desirable that the rainfall distribution information F shows the rainfall distribution for the entire river catchment area, but it is not necessarily required that the entire area be shown (i.e., it may show only a part of the river catchment area). It is desirable that the rainfall distribution information F includes information indicating the time of day (e.g., date and time) for which the rainfall distribution is represented. In the following, the rainfall distribution information F used in the learning stage will be specifically referred to as "rainfall distribution information F1," and the rainfall distribution information F used in the prediction stage will be specifically referred to as "rainfall distribution information F2."

[0031] (Configuration of the learning device) As shown in Figure 1, the learning device 10 comprises a data processing unit 11, a learning data acquisition unit 12, and a learning processing unit 13. The data processing unit 11, the learning data acquisition unit 12, and the learning processing unit 13 are implemented, for example, by program execution processing by a CPU (Central Processing Unit). When these functions are implemented by program execution processing, the program for implementing these functions is stored in a storage unit (not shown).

[0032] The data processing unit 11 receives rainfall distribution information F1 as input. There is no particular limit to the number of rainfall distribution information F1 input to the data processing unit 11; for example, hourly rainfall distribution information F1 may be input. Based on the rainfall distribution information F1, the data processing unit 11 obtains classification information for the rainfall distribution information F1. The classification information is information that categorizes the rainfall conditions in the river's catchment area. The classification information is obtained, for example, by cluster analysis and output to the learning data acquisition unit 12. The data processing unit 11 includes, for example, an image processing unit 11a, an information reduction processing unit 11b, and a categorization processing unit 11c. If the rainfall distribution information F1 is not image data (i.e., it is numerical information such as analyzed rainfall), the image processing unit 11a is not necessary, and the data processing unit 11 only needs to have an information reduction processing unit 11b and a categorization processing unit 11c.

[0033] The processing of the image processing unit 11a will be explained with reference to Figures 2 and 3 (and Figure 1 as appropriate). Figure 2 is an image diagram of a river and a watershed. Figure 3 is a diagram for explaining the processing of the image processing unit 11a, where (a) is an example of rainfall distribution information F1, (b) is an example of a watershed image E from which rainfall distribution information F1 within the watershed has been extracted, and (c) is an example of a grayscale image D obtained by converting the watershed image E from RGB values ​​to luminance values.

[0034] In Figure 2, the river is represented by the symbol K1, and the catchment area is represented by the symbol K2. The catchment area K2 indicates the region where rainwater flows into the river K1, and the rainwater that falls in catchment area K2 flows into the sea K3 via the river K1. The catchment area K2 may be based on, for example, a watershed map published by the Ministry of Land, Infrastructure, Transport and Tourism. In addition, an observation station K4 is set up downstream of the river K1, where the water level of the river K1 can be measured. The time it takes for rainwater that falls upstream of the river K1 to affect the water level change at observation station K4 is determined by the characteristics of the river K1, for example, by the length of the river and the elevation difference.

[0035] The image processing unit 11a extracts the catchment area image E (see Figure 3(b)), which is an image of the catchment area K2, from the rainfall distribution information F1 (see Figure 3(a)), and converts the RGB values ​​of each pixel constituting the catchment area image E into luminance values. The purpose of this conversion from RGB values ​​to luminance values ​​is to assign a representative value to each pixel. The image processing unit 11a performs this conversion process when the rainfall distribution information F1 is an RGB image, but can omit this conversion process if the rainfall distribution information F1 is numerical information. In that case, the numerical information (for example, analyzed rainfall) is directly input to the information reduction processing unit 11b. Luminance values ​​are also called gray values.

[0036] As a result, the image processing unit 11a creates a grayscale image D (see Figure 3(c)) by converting the luminance of the catchment area image E (see Figure 3(b)). Since the value (luminance value) of each pixel that makes up the grayscale image D corresponds to the amount of rainfall, the grayscale image D is an example of rainfall distribution information. In other words, the image processing unit 11a creates new rainfall distribution information (in this case, the grayscale image D) by image processing the rainfall distribution information F1.

[0037] There are several formulas for converting to grayscale, such as "CIE XYZ" and "ITU-R Rec BT.601," but a formula that increases the conversion value when rainfall is heavy should be used. Conversion to grayscale is performed, for example, using formula (1) shown below. • Brightness value = 255 - (0.587 * r value + 0.114 * g value + 0.299 * b value) ··Equation (1) In rain cloud images, light blue indicates light rain and red indicates heavy rain. Therefore, a brightness conversion formula is constructed so that the brightness value of light blue is close to "0 (zero)" and the brightness value of red is close to "255". In other words, the red, blue, and green values ​​of the RGB color are weighted according to the intensity of rainfall in the rain cloud image.

[0038] The processing of the information reduction processing unit 11b will be explained with reference to Figure 4 (and Figures 1 to 3 as appropriate). Figure 4 is a diagram for explaining the processing of the information reduction processing unit 11b, where (a) is a diagram in which a divided region 8 is set on a gray image D, (b) is an enlarged image of the range 9 in (a), and (c) is an example of reduced rainfall distribution information H (only the part corresponding to range 9) with reduced information volume.

[0039] The information reduction processing unit 11b divides the rainfall distribution information F1 (for example, numerical information that forms the basis of the image data) or the image-processed information of the rainfall distribution information F1 (for example, a grayscale image D) into a grid, and calculates statistical values ​​of the amount of rainfall for each divided region 8. The calculated statistical values ​​are representative values ​​of rainfall in the divided region 8. For example, the divided regions 8 are arranged in a "m rows × n columns (m and n are natural numbers)" pattern so as to have no gaps in the rainfall distribution information F1 or the image-processed information of the rainfall distribution information F1. The number of divided regions 8 (number of divisions N (N is a natural number greater than or equal to 2)) is best set by referring to the area upstream of the prediction point and the east-west and north-south distances.

[0040] Figure 4 shows an example where a divided region 8 is set for a grayscale image D, as shown in (a). The divided region 8 is set to contain multiple pixels. In Figure 4(b), the divided region 8 is set to contain "3 vertically x 3 horizontally (9 pixels in total)". In Figure 4(b), the rectangular regions labeled 7 represent individual pixels, and the numbers inside region 7 represent the brightness values.

[0041] The information reduction processing unit 11b calculates, for example, the average rainfall amount (in this case, the average of the brightness values) for each divided region 8, and uses the calculated average rainfall amount as a representative value for rainfall in the divided region 8. The set of representative values ​​for the divided region 8 (in this case, assumed to be the average rainfall amount) is called reduced rainfall distribution information H because it is obtained by reducing the amount of information (reducing the size) of the gray image D. In the example shown in Figure 4, the amount of information is reduced to "1 / 9" by creating reduced rainfall distribution information H from the gray image D.

[0042] Each representative value that makes up the reduced rainfall distribution information H corresponds to the amount of rainfall, so the reduced rainfall distribution information H is an example of rainfall distribution information. In other words, the information reduction processing unit 11b creates new rainfall distribution information with reduced information volume based on the rainfall distribution information F1 (for example, numerical information that forms the basis of image data) or information obtained by image processing of the rainfall distribution information F1 (for example, grayscale image D).

[0043] The categorization processing unit 11c shown in Figure 1 categorizes the reduced rainfall distribution information H (see Figure 4(c)) using machine learning cluster analysis and assigns classes to it. In this embodiment, we assume that the reduced rainfall distribution information H is categorized into classes using "k-means cluster analysis," which categorizes based on the distance between centroids. The processing of the information reduction processing unit 11b will be explained with reference to Figures 5 and 6 (and Figures 1 to 4 as appropriate). Figures 5 and 6 are diagrams illustrating the processing of the categorization processing unit 11c.

[0044] As shown in Figure 5, the classification processing unit 11c receives, for example, hourly reduced rainfall distribution information H for a certain river's catchment area. The reduced rainfall distribution information H is created based on the original gray image D. The reduced rainfall distribution information H shown in Figure 5 is divided into "11" vertically and horizontally, resulting in a total of "121" divided regions 8. The classification processing unit 11c clusters the reduced rainfall distribution information H as an N-dimensional (N is the number of divisions) value based on the divided regions 8.

[0045] Specifically, as shown in Figure 5, row numbers (in Figure 5, row numbers "00" to "10") and column numbers (in Figure 5, column numbers "00" to "10") are assigned to the reduced rainfall distribution information H. Next, the representative values ​​of the divided region 8 (here, the average of the brightness values ​​is assumed) are arranged in a single row based on the row and column numbers, so that one reduced rainfall distribution information H is represented as a 1xN matrix. After converting all the input reduced rainfall distribution information H into a 1xN matrix, the converted reduced rainfall distribution information H is arranged based on date and time to create information G1 that summarizes all the input reduced rainfall distribution information H.

[0046] Next, the classification processing unit 11c classifies the reduced rainfall distribution information by repeatedly calculating the centroid of the group and classifying it into a new group. The specific procedure is as follows, for example. (Preparation: "Setting Initial Information") As shown in Figure 6, in the preliminary "initial information setting" step, the reduced rainfall distribution information H is arbitrarily grouped. For example, the reduced rainfall distribution information H is grouped so that the number of groups is evenly distributed.

[0047] (First grouping process) In the first grouping, the centroid of each group is determined based on the grouping done in the preliminary preparation. Specifically, group centroid information G2 is created for each group by statistically processing the representative values ​​of the divided region 8 on a group basis (here, the mean of the representative values ​​is assumed). Each element that makes up the group centroid information G2 is the centroid value (here, the mean) in the divided region 8 at the same position (1st row, jth column (j=1 to N)) within the same group.

[0048] Next, we determine the distance L from the centroid of each group to the respective reduced rainfall distribution information H. The distance L to the centroid of each group is calculated using, for example, equation (2) shown below. L = (Σ(representative value of the divided region (j) - centroid value of the divided region (j)) 2 ) 0.5 ·Formula (2)

[0049] For example, the distance Ldt1(i-1) from the centroid of group (i-1) up to reduced rainfall distribution information Hdt1 at the date and time "October 12, 2019, 8:00 PM" is as follows: ·Distance Ldt1(i-1)= {(Representative value of partitioned region "00_00" - Centroid value of partitioned region "00_00" (i-1)) 2 + (Representative value of the divided region "01_00" - Centroid value of the divided region "01_00" (i-1)) 2 +··· (Representative value of the partitioned region "10_10" - Centroid value of the partitioned region "10_10" (i-1)) 2} 0.5

[0050] Furthermore, the distance Ldt1(i) from the centroid of group(i) up to the reduced rainfall distribution information Hdt1 at the time of date and time "October 12, 2019, 20:00" is as follows: ·Distance Ldt1(i)= {(Representative value of partitioned region "00_00" - Centroid value of partitioned region "00_00" (i)) 2 + (Representative value of the divided region "01_00" - Centroid value of the divided region "01_00" (i)) 2 +··· (Representative value of segmented region "10_10" - Centroid value (i) of segmented region "10_10") 2} 0.5

[0051] In this manner, information G3 is created that summarizes the distance L from each group for all the reduced rainfall distribution information H. Then, for each piece of the reduced rainfall distribution information H, the group with the centroid having the shortest distance is selected, and said group is set as a new group. For example, the reduced rainfall distribution information Hdt1 at the date and time "20:00 on October 12, 2019" was classified into group (i-1) in the pre-preparation grouping, but when the distance Ldt1(i) from the centroid of group (i) is the shortest, group (i) is set as the new group.

[0052] (Second and subsequent grouping processes) In the second grouping, the centroid of each group is determined based on the first grouping, and the reduced rainfall distribution information H is classified into new groups based on the determined group centroids. The third and subsequent groupings are similarly performed, and the creation of group centroid information G2 and the classification into new groups based on the group centroids are repeatedly executed until the groups no longer change (or the change rate falls below a threshold). When the iterative grouping process ends, the final group (class) of the reduced rainfall distribution information H (that is, the rainfall distribution information F1) is determined. FIG. 7 is an example of the result of categorizing the reduced rainfall distribution information H (that is, the rainfall distribution information F1) into four classes (class "0" to class "3").

[0053] The learning data acquisition unit 12 shown in FIG. 1 combines the presence or absence of river flooding and the class value (classification information) determined by the data processing unit 11 to organize the combined data as learning data. In the learning data, when a specific water level is exceeded, flooding may be estimated to be present, and the water level in that case may be, for example, one published by the Ministry of Land, Infrastructure, Transport and Tourism. The classification information is information obtained by categorizing the rainfall conditions in the river catchment area, and the class value is an example of the classification information.

[0054] The learning data acquisition unit 12 organizes the learning data according to the instructions of, for example, a person who predicts whether or not there will be flooding (hereinafter referred to as the "user"). Alternatively, the presence or absence of river flooding and the class value (classification information) may be pre-registered in the storage unit, and the learning data acquisition unit 12 may acquire the registered information as learning data. The storage unit may be provided by a device connected via a communication line (for example, a cloud system). Hereinafter, the organized collection of learning data may be referred to as the "learning dataset."

[0055] Figure 8 shows an example of a training dataset organized by the training data acquisition unit 12. The training dataset shown in Figure 8 consists of multiple training datasets with different reference times. The training dataset shown in Figure 8 includes information from the reference time "September 9, 2019, 0:00 to September 9, 2019, 23:00". Each training dataset consists of the "reference time", "class value" for each hour from 24 hours before to 24 hours after the reference time, and "presence or absence of flooding" for each hour from 1 hour to 24 hours after the reference time.

[0056] Figure 8 shows an example of creating a dataset when categorized into five classes (class "0" to class "4"). The "input section" on the left shows the class values ​​"0 to 4" for each hour relative to "time (reference time)" from "t-24h to t+24h". The "output section" on the right shows the values ​​for whether or not flooding occurred ("1" indicates flooding, "0" indicates no flooding) after "t+1h to t+24h" relative to "time (reference time)".

[0057] Note that (1) class values ​​for times in the future beyond the reference time, and (2) class values ​​for the reference time, do not necessarily have to be included in the training data. Including this information has the advantage of improving prediction accuracy. In other words, because future rainfall will change the rise and fall of water levels, relying solely on past information may not improve the accuracy of predicting water levels in the future, especially at times far from the present. In such cases, it is better to use class values ​​for times in the future beyond the reference time and for the reference time.

[0058] In this way, the learning data acquisition unit 12 acquires a pair of data as learning data: whether or not there is flooding at a certain first time point (a time between 1 hour and 24 hours after the reference time in Figure 8) and class values ​​(classification information) at at least one second time point prior to the first time point (a time between 24 hours before the reference time in Figure 8 and the reference time in Figure 8). Note that the class values ​​(classification information) at the first time point may also be included in the learning data.

[0059] The learning processing unit 13 shown in Figure 1 has a learner, and uses the learning data acquired by the learning data acquisition unit 12 to train the learner through machine learning. The learner compares the results of classification, prediction, and judgment based on the given data with the actual correct answer, and can derive better results by adjusting various parameters. The learner is, for example, a neural network, and in this embodiment, a neural network will be assumed and explained. The learner is also called a "learning model".

[0060] As is well known, neural networks are an information processing method developed with the aim of having computers mimic the workings of the human brain. Neural networks excel at complex nonlinear processing and have the characteristic of not requiring the formalization of the relationship between explanatory variables and target variables by using learning functions. A neural network generally consists of an input layer, hidden layers, and an output layer. Explanatory variables are input to the input layer, and the target variable is output from the output layer. The hidden layers play the role of relating the two. There are no constraints on the number of hidden layers or the number of neurons in each hidden layer; they can be set arbitrarily.

[0061] The method for training the learning model is not particularly limited; for example, the learning model can be trained using a method appropriate to the type of learning model. Specifically, the "class values" for each hour from 24 hours before to 24 hours after a reference time that constitute the learning data are input to the input layer, and the intermediate layer is adjusted based on the error between the result output from the output layer and the presence or absence of flooding for each hour from 1 hour to 24 hours after the reference time. In other words, the learning processing unit 13 learns the learning model by inputting class values ​​(classification information) at least one third time point in time so that it outputs whether or not there is flooding in the river at a fourth time point in the future from the third time point. The number of learning data to be used for training is not particularly limited; for example, training can be terminated when the expected accuracy is reached.

[0062] (Configuration of the prediction device) As shown in Figure 1, the prediction device 20 comprises a data processing unit 21, a prediction data acquisition unit 22, and a prediction processing unit 23. The data processing unit 21, the prediction data acquisition unit 22, and the prediction processing unit 23 are implemented, for example, by program execution processing by a CPU (Central Processing Unit). When these functions are implemented by program execution processing, the program for implementing these functions is stored in a storage unit (not shown).

[0063] The data processing unit 21 receives rainfall distribution information F2 as input. There is no particular limit to the number of rainfall distribution information F2 input to the data processing unit 21; for example, continuous rainfall distribution information F2 every hour may be input. Based on the rainfall distribution information F2, the data processing unit 21 obtains classification information for the rainfall distribution information F2. The classification information is information that categorizes the rainfall conditions in the river's catchment area. The classification information is obtained, for example, by cluster analysis and output to the prediction data acquisition unit 22.

[0064] The data processing unit 21 includes an image processing unit 21a, an information reduction processing unit 21b, and a categorization processing unit 21c. The processing of the data processing unit 21 (image processing unit 21a, information reduction processing unit 21b, and categorization processing unit 21c) is the same as the processing of the data processing unit 11 in the learning phase, so a detailed explanation is omitted. The rainfall distribution information F2 may be from the past (e.g., measured data) or from the future (e.g., forecast data). In other words, it is possible to predict whether or not there will be flooding a few hours later (e.g., now or in the near future) from past actual rainfall distribution information F2, and it is also possible to predict whether or not there will be flooding a few hours later from future rainfall distribution information F2 that has been published as a forecast.

[0065] The prediction data acquisition unit 22 shown in Figure 1 organizes the class values ​​(classification information) obtained by the data processing unit 21 as prediction data. The prediction data acquisition unit 22 organizes the prediction data according to the instructions of, for example, the person making the prediction of whether or not there will be flooding (hereinafter referred to as the "user"). Alternatively, the class values ​​(classification information) may be registered in a storage unit in advance, and the prediction data acquisition unit 22 may acquire the registered information as prediction data. The storage unit may be provided by a device connected via a communication line (for example, a cloud system).

[0066] In this manner, the prediction data acquisition unit 22 acquires classification information as prediction data at least one third point in time (in this embodiment, this may include a period from 24 hours before the reference time to the reference time and 24 hours after). It is desirable that the prediction data is calculated using rainfall distribution information F2 in the same format as rainfall distribution information F1.

[0067] The prediction processing unit 23 shown in Figure 1 has a trained learner. The trained learner was trained using machine learning by the learning device 10. In other words, the trained learner is trained to output whether or not there will be river flooding at a fourth time point in the future, by inputting class values ​​(classification information) at at least one third time point. The prediction processing unit 23 inputs the prediction data acquired by the prediction data acquisition unit 22 into the trained learner, predicts whether or not there will be river flooding at a fourth time point in the future, which is included in the prediction data, and outputs the predicted result.

[0068] <Regarding the learning method and prediction method according to the embodiment> Referring to Figure 9 (and Figures 1 to 8 as appropriate), the learning method and prediction method according to the embodiment will be described. Figure 9 is an example of a flowchart of the learning method and prediction method according to the embodiment.

[0069] As shown in Figure 9, the process of building a prediction model corresponding to the learning method according to the embodiment (step S10) mainly consists of a preparation step of input data (e.g., past data) (step S11), a parameter tuning step (step S12), and a prediction accuracy determination step (step S13). Furthermore, the flood prediction process corresponding to the prediction method according to the embodiment (step S20) mainly consists of a preparation step of input data (e.g., real-time data) (step S21), a calculation step using the prediction method (step S22), a confirmation step of the prediction results (step S23), and a provision step to construction personnel (step S24).

[0070] (Preparation process for input data (e.g., past data) (Step S11)) The user prepares rainfall distribution information F1 and river flooding information to be used in training the prediction model (learner). For example, the user prepares hourly rainfall distribution information F1 and river flooding information. Rainfall distribution information F1 can be, for example, information published by the Japan Meteorological Agency, and since the information published by the Japan Meteorological Agency covers all of Japan, it can be prepared without much effort or time. River flooding can be estimated by whether or not the river water level has exceeded the dangerous level.

[0071] The data processing unit 11 of the learning device 10 obtains class values ​​(classification information) that categorize rainfall conditions from the rainfall distribution information F1 using the method described above. The categorization processing unit 11c categorizes the reduced rainfall distribution information H (see Figure 4(c)) using, for example, machine learning cluster analysis, and obtains class values. In cluster analysis, for example, the rainfall distribution information F1 is categorized into classes based on the distance from the centroid.

[0072] The learning data acquisition unit 12 combines the presence or absence of river flooding with the class value (classification information) obtained by the data processing unit 11 and organizes it as learning data. The organized learning data is stored in a storage unit (not shown).

[0073] (Parameter tuning process (step S12)) The learning data acquisition unit 12 of the learning device 10 sequentially acquires learning data, and the learning processing unit 13 uses the acquired learning data to train the learner. Assuming a neural network as the learner, the hidden layers are adjusted according to the learning data. The learning processing unit 13 inputs class values ​​(classification information) at least one third time point, and trains the learner to output whether or not there is a river flood at a fourth time point that is in the future of the third time point.

[0074] (Prediction accuracy determination process (step S13)) In this step, the prediction accuracy of the trained model is determined, and the learning process is terminated if the prediction accuracy meets the criteria set by the user (if the prediction accuracy is "good"). On the other hand, if the prediction accuracy of the trained model does not meet the set criteria (if the prediction accuracy is "poor"), the parameter tuning process (step S12) is performed, and the learning process continues. Note that the parameter tuning process (step S12) and the prediction accuracy determination process (step S13) may be implemented by cross-validation. When learning is performed using cross-validation, for example, the data for the year to be predicted from the organized data is classified as test data, and the rest is classified as learning data to predict whether or not there will be a flood.

[0075] (Preparation process for input data (e.g., real-time) (step S21)) The user prepares rainfall distribution information F2 to be used to predict whether or not there will be river flooding. The user prepares, for example, continuously time-series rainfall distribution information F2 every hour. The rainfall distribution information F2 can be, for example, information published by the Japan Meteorological Agency. The data processing unit 21 of the prediction device 20 obtains class values ​​(classification information) from the rainfall distribution information F2 in the same manner as the preparation step S11. Then, the prediction data acquisition unit 22 organizes the class values ​​(classification information) obtained by the data processing unit 21 as prediction data. Note that when making predictions in real time for the present and beyond, the rainfall distribution information F2 at future points in time may be forecast values.

[0076] (Calculation process using prediction method (Step S22)) The prediction data acquisition unit 22 of the prediction device 20 acquires prediction data, and the prediction processing unit 23 inputs the prediction data into a trained learner to predict whether or not there will be flooding in the river.

[0077] (Confirmation of prediction results (Step S23) and provision of results to construction personnel (Step S24)) The user confirms the predicted presence or absence of flooding (prediction result) in step S22 and provides information on the presence or absence of flooding to construction personnel as needed. Based on the provided information on the presence or absence of flooding, the construction personnel evacuate heavy machinery and equipment in the construction yard, and take appropriate measures such as securing them.

[0078] <Effects of the learning device and prediction device according to the embodiment> As described above, the learning device 10 and prediction device 20 according to the embodiment directly learn the presence or absence of flooding using machine learning, with classification information categorizing rainfall conditions as input values. In other words, the presence or absence of flooding is learned directly without going through calculations of water level or flow rate. Therefore, various costs required for preparing for prediction (for example, the burden on operators and computational load) can be reduced. For example, by using machine learning, the construction of model equations and parameter tuning become unnecessary. In addition, since the features can be simplified by using classification information as input values, learning does not become complex (learning becomes simpler). Therefore, the computational load can be reduced. Furthermore, it is easy to organize the input data, and learning is relatively easy even when the amount of input data is increased.

[0079] Furthermore, the learning device 10 and prediction device 20 according to this embodiment obtain classification information based on rainfall distribution information F, which shows the intensity of rainfall in two dimensions. In this way, for example, actual rainfall, analyzed rainfall, and rain cloud images distributed by the Japan Meteorological Agency can be used for learning and prediction. Since this information covers the entire country of Japan and is easily available, it is possible to easily learn and predict whether or not there will be flooding at any point throughout Japan. In addition, by using classification information obtained from weather forecast information (for example, rain cloud images), it is possible to predict whether or not there will be flooding long in advance. Therefore, it is possible to evacuate and protect heavy machinery and equipment in the construction yard with sufficient time.

[0080] Furthermore, in the learning device 10 and prediction device 20 according to this embodiment, the rainfall distribution information F is divided into m rows × n columns (where m and n are natural numbers), and representative values ​​are calculated for each divided region 8 to create reduced rainfall distribution information H with a reduced amount of information. Then, the reduced rainfall distribution information H is categorized using cluster analysis. In this way, the learning can be performed while keeping the computational load down, and differences in rainfall distribution (for example, whether it is a uniform distribution or a concentrated downpour) can be included in the learning. As a result, the presence or absence of flooding can be predicted with greater accuracy.

[0081] To verify the effectiveness of the learning device 10 and prediction device 20 according to the embodiment, a prediction test using actual data was conducted and will be described below. In the prediction test, the presence or absence of flooding was predicted using a domestic observation station as an example. The target river is the one indicated by the symbol K1 in Figure 2, and the observation station that serves as the prediction point is located at the position indicated by the symbol K4 in Figure 2. As shown in Figure 2, the observation station is located in the lower reaches of the river.

[0082] The procedure for conducting the prediction test is shown in Figure 10. Figure 10 is a flowchart of the procedure for conducting the prediction test. In the prediction test, hourly rain cloud images from June to November for the period from 2007 to 2023 were prepared as rainfall distribution information F (step S31). In order to focus on the presence or absence of flooding due to rainfall, periods other than June to November were excluded from the target to avoid the influence of snowmelt.

[0083] Next, data within the catchment area (watershed) is extracted from the rainfall distribution information F, and the RGB values ​​of each pixel in the extracted area are converted to brightness values ​​from "0 to 255" (step S32). Subsequently, the grayscale image D is divided into m rows × n columns (m and n are natural numbers), and representative values ​​are calculated for each of the 8 divided areas to create reduced rainfall distribution information H with a reduced amount of information (step S33). Subsequently, the reduced rainfall distribution information H is categorized into rainfall distribution information F by cluster analysis (step S34).

[0084] In addition, hourly water level data at the prediction points for the same period as the rainfall distribution information F is prepared (step S35). In the prediction test, water level data at the prediction points was obtained from the hydrological water quality database managed by the Ministry of Land, Infrastructure, Transport and Tourism. Next, the water level data is judged and classified into "flood present" and "no flood present" (step S36). In the prediction test, floods that occur a few times a year were set to "flood present" when the water level at the observation station exceeded "9m" and to "no flood present" when it was less than "9m". Figure 11 shows the relationship between the water level at the observation station and the presence or absence of floods in the prediction test, with (a) showing the water level changes and the presence or absence of floods in "2019", (b) showing the water level changes and the presence or absence of floods on "September 10, 2019", and (c) showing the water level changes and the presence or absence of floods on "October 13, 2019".

[0085] Next, create the training dataset (step S37). The training dataset for the prediction test is a combination of the input and output values ​​shown below, with the reference time being "t=0". • Input value: Class value of reduced rainfall distribution information H is "t-24h~t+24h". • Output value: Whether or not water is being discharged ("1" or "0") in "t+1 to +24h" A portion of the training dataset created is shown in Figure 8. In the prediction test, multiple datasets with different numbers of classes were created, and the presence or absence of flooding was predicted for each number of classes.

[0086] Next, the presence or absence of flooding is predicted using machine learning with the aforementioned training dataset (Step S38). In the prediction test, prediction was performed using a regression-type neural network. In addition, as shown in Figure 12, cross-validation was performed year by year in the prediction test. For example, when predicting whether or not there would be flooding in "2019", the model was trained using the dataset from "2007-2023" excluding the "2019" dataset, and the "2019" dataset was input during the prediction.

[0087] As an example of the results of the prediction test, Figures 13 to 15 show the prediction results when the number of classes in the cluster analysis is set to "3, 6, and 10". Figures 13 to 15 are examples of the results of the prediction test, where (a) is the prediction result when the number of classes is "3", (b) is the prediction result when the number of classes is "6", and (c) is the prediction result when the number of classes is "10". In the prediction results shown in Figures 13 to 15, the actual water level at the time of flooding and whether or not there was actual flooding are shown together, as well as the prediction results for "3, 6, 12, and 24 hours" ahead. Note that in Figures 13 to 15, the prediction results for "3, 6, 12, and 24 hours" ahead are plotted at a time obtained by adding the respective prediction time to the reference time corresponding to each prediction result for "3, 6, 12, and 24 hours" ahead.

[0088] To predict whether or not there will be flooding, it is necessary to make predictions during or before the flooding period. Here, the flooding period refers to the period from when the actual water level begins to exceed the designated danger level until it starts to fall. Observing Figures 13 to 15, it can be seen that the prediction results for "3, 6, 12, and 24 hours" ahead of the flooding period are generally judged to be the presence of flooding, so it can be said that the prediction of whether or not there will be flooding is possible. For example, looking at the "24-hour" prediction for class number "6" shown in Figure 13(b), it is predicted that flooding will occur within 24 hours from the time of the actual flooding (July 15, 2014, 16:00) (see symbol 61). Also, looking at the "24-hour" prediction result for class number "6" shown in Figure 14(b), it is predicted that flooding will occur at the same time as the actual flooding (October 6, 2014, 18:00) (see symbol 71). This can be said to predict flooding 24 hours before the actual time of flooding. Furthermore, focusing on the prediction result for class number "6" in Figure 14(b), which is "6h" ahead, it can be said that it predicts flooding 1 hour before the actual time of flooding, and 7 hours before the actual time of flooding.

[0089] Here, we can see that the prediction results differ depending on the number of classes in the cluster analysis. In particular, when the number of classes is "10", the prediction results sometimes fail to predict flooding even at the time when flooding actually occurs (see, for example, Figure 13(c)). This means that increasing the number of classes in the cluster analysis does not necessarily improve prediction accuracy, but rather that there is a number of classes that yields the highest prediction accuracy. From the standpoint of prediction accuracy, 2 to 8 classes are preferable, and 3 to 6 classes are even preferable.

[0090] Furthermore, in the prediction test, a method was used to improve prediction accuracy by selecting training data and enhancing the quality of the learning process. Specifically, by focusing on flood events and extracting information for the period from 48 hours before the start of a flood to 48 hours after the end of the flood, only for the number of flood events, and using this information for training, the sensitivity to floods was increased. Figure 16 shows the results of predicting floods using this method. Comparing this with Figure 15, where all data was used as training data, it can be seen that the prediction accuracy for floods has improved, and therefore the frequency of flood predictions has increased. Thus, it can be confirmed that prediction accuracy increases by selecting training data and increasing sensitivity to floods. [Explanation of Symbols]

[0091] 10 Learning device 11 Data Processing Unit 12. Training Data Acquisition Unit 11a, 21a Image Processing Unit 11b, 21b Information reduction processing unit 11c,21c Classification processing unit 13 Learning Processing Unit 20 Prediction device 21 Data Processing Unit 22 Prediction Data Acquisition Unit 23 Prediction Processing Unit 8 divided areas F,F1,F2 Rainfall distribution information E. Watershed image D Grayscale image H Reduced rainfall distribution information

Claims

1. A learning data acquisition unit acquires, as learning data, a pair of data consisting of whether or not there is flooding in a river at a certain first point in time, and classification information that categorizes the rainfall conditions in the river's catchment area at at least one second point in time prior to the first point in time. It comprises a learning unit having a learning device and a learning processing unit that uses the learning data to perform machine learning on the learning device, The learning processing unit, by inputting the classification information at least one or more third time points, causes the learning device to train to output whether or not the river floods at a fourth time point that is later than the third time point. A learning device characterized by the following features.

2. The aforementioned classification information was obtained based on rainfall distribution information that shows the intensity of rainfall in two dimensions. The system further includes a data processing unit that categorizes multiple rainfall distribution information sets by cluster analysis. The learning device according to feature 1.

3. The aforementioned data processing unit An information reduction processing unit creates reduced rainfall distribution information by dividing the aforementioned rainfall distribution information into m rows × n columns (where m and n are natural numbers) and calculating representative values ​​for each divided region, thereby reducing the amount of information; The system includes a classification processing unit that classifies the reduced rainfall distribution information by cluster analysis. The learning device according to feature 2.

4. The classification processing unit repeatedly processes the reduced rainfall distribution information into one of several groups to determine the final class. By statistically processing the representative values ​​of the divided region on a group basis, the centroid of the group is determined. For all combinations of the aforementioned group and the reduced rainfall distribution information, the distance from the centroid to the reduced rainfall distribution information is calculated, and the group with the closest distance from the centroid to the reduced rainfall distribution information is designated as the new group for the reduced rainfall distribution information. The calculation of the centroid of the group and the classification into the new group are repeated. The learning device according to feature 3.

5. The aforementioned training data further includes the classification information at the first time point, The learning processing unit, by inputting classification information at the third and fourth time points, causes the learning device to learn to output whether or not there is flooding in the river at the fourth time point. The learning device according to feature 1.

6. The aforementioned training data is selected from data collected within a predetermined period based on the time of flooding. The learning device according to feature 1.

7. A learning data acquisition step involves acquiring a pair of data as learning data: the presence or absence of flooding in a river at a certain first point in time, and classification information that categorizes the rainfall conditions in the river's catchment area at at least one second point in time prior to the first point in time. The system includes a learning device, and a learning process that uses the learning data to train the learning device, In the learning process, the learning device is trained to output whether or not the river floods at a fourth time point in the future, by inputting the classification information at least one or more third time points. A learning method characterized by the following:

8. The system includes a trained learning device that has been trained by the learning method described in claim 7, and by inputting the classification information at the third time point into the trained learning device, it predicts whether or not the river will flood at the fourth time point. A prediction device characterized by the following features.

Images