Model generation device, water level prediction device, model generation method, model generation program, water level prediction method, water level prediction program, water level prediction system, and learned model

Abstract

To provide a model generation device, a method, a program, and a water level prediction device for predicting a water level with high accuracy by using a water detector having a simple configuration compared with a water level meter.SOLUTION: In a water level prediction system, a model generation device 3 comprises: a teacher data creation unit 34 for creating a teacher data set including time difference information for learning (ΔTL) about time differences between the points in time at which changes in the presence or absence of water at a plurality of predetermined water levels are detected by water detectors 2 installed at a plurality of specific locations, rain amount information for learning (RL) about the amount of rain at a specific location when detecting a change in the presence or absence of water at a predetermined water level and water level difference information for learning (ΔHL) about a water level difference between a plurality of predetermined water levels; and a model learning unit 35, using the set, with ΔTL and RL as input of a prediction model 4 and ΔHL as an output of the prediction model, for performing machine learning so that an input-output relation of the prediction model approaches an input-output relation of the teacher data set to generate a learned model.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present disclosure relates to a model generation device, a water level prediction device, a model generation method, a model generation program, a water level prediction method, a water level prediction program, a water level prediction system, and a trained model. [Background technology]

[0002] Climate change is increasing the frequency of intense rainfall, raising concerns that more rainwater will run onto roads in a shorter period of time than before. In areas where urbanization has increased the amount of land surface that makes it difficult for rainwater to penetrate underground, this will also have an impact. Road sections that cannot adequately drain the inflowing rainwater will be flooded and traffic restrictions will be required. For road users, residents living near sections where road flooding occurs, and road managers, predicting the onset and lifting of traffic restrictions due to road flooding is useful for the safety and efficiency of life and traffic. For example, based on the forecast information, road users can determine safe routes and road managers can operate roads efficiently.

[0003] As technologies related to such flood prediction, for example, Patent Documents 1 and 2 disclose a method of measuring time-series water level data using water level gauges at multiple points on the river whose water level is to be predicted, and predicting the water level of the river using the measured water level data.

[0004] On the other hand, Patent Document 3 discloses a configuration in which a water detector is installed at a detection target location to detect when a predetermined water level has been reached, for example to detect flooding of roads. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Publication No. 09-256338 [Patent Document 2] JP 2020-134300 A [Patent Document 3] JP 2023-66488 A Summary of the Invention [Problem to be solved by the invention]

[0006] In conventional water level prediction methods such as those described in Patent Documents 1 and 2, the prediction target is a river, so the installation location of the water level meter is a measurement environment where water flows constantly. For this reason, using a water level meter to constantly measure the water level and obtain time series data can be said to be effective in improving the accuracy of water level prediction. However, since the water level meter used in conventional water level prediction methods measures and transmits the water level at regular time intervals, it requires constant power, and requires expenses and installation space. For this reason, water level meter is generally not installed in places where water does not exist in normal times, such as roads and residential areas. In addition, even if a water level meter is installed in a place where water does not exist in normal times, the water level cannot be measured in most hours, so there is a high possibility that most of the measured time series data will be wasted. In addition, since the water level meter is operated constantly, wasteful power consumption may occur during hours when valid data cannot be measured.

[0007] On the other hand, conventional flood detection methods such as those described in Patent Document 3 use a water detector to detect that a certain water level has been reached at a detection location, and then recognize the occurrence of flooding at that location. Water detectors do not transmit data to external devices during times when there is no change in the presence or absence of water, so they have the advantage of reducing power consumption and being simpler in structure than water level gauges. However, conventional flood detection methods that use water detectors cannot predict in advance the occurrence of flooding at a detection location.

[0008] Moreover, in the water level prediction methods described in Patent Documents 1 and 2, time series data on the water level of the river to be predicted at the same time intervals is acquired and used for water level prediction. Meanwhile, the water detector in Patent Document 3 is a device for detecting changes in the presence or absence of water at a specified water level at the installation location. Therefore, even if multiple water detectors are installed at the same installation location so that they can detect different water levels, although information on the time when the presence or absence of water at the specified water level changed can be acquired in each water detector, it is not possible to constantly measure the water level before and after at a fixed time interval to acquire time series data. Therefore, in the water level prediction methods described in Patent Documents 1 and 2, the water detector in Patent Document 3 cannot be applied instead of a water level meter.

[0009] An object of the present disclosure is to enable highly accurate water level prediction using a water detector that has a simpler configuration than a water level meter. [Means for solving the problem]

[0010] A model generation device according to one aspect of an embodiment of the present invention is a model generation device for generating a model for predicting the water level at a specific location, and comprises a teacher data creation unit that creates a teacher dataset including learning time difference information regarding the time difference between the times when a change in the presence or absence of water at a plurality of specified water levels is detected by a water detector installed at the specific location and detecting a change in the presence or absence of water at the specified water level, learning rainfall information regarding the rainfall at the specific location when the change in the presence or absence of water at the specified water level is detected, and learning water level difference information regarding the water level difference at the plurality of specified water levels; and a model learning unit that uses the teacher dataset created by the teacher data creation unit to perform machine learning using the learning time difference information and the learning rainfall information as inputs to the model and the learning water level difference information as an output of the model, so that the input / output relationship of the model approaches the input / output relationship of the teacher dataset, thereby generating a trained model.

[0011] Similarly, a water level prediction device according to one aspect of an embodiment of the present invention is a water level prediction device for predicting the water level at a specific location, and calculates a correspondence between learning explanatory variables including learning time difference information relating to the time difference between the times when a change in the presence or absence of water at a plurality of specified water levels is detected by a water detector that is installed at the specific location and detects a change in the presence or absence of water at the specified water level, and learning rainfall information relating to the rainfall at the specific location when the change in the presence or absence of water at the specified water level is detected, and learning objective variables including learning water level difference information relating to the water level difference between the plurality of specified water levels. The system comprises a trained model obtained by machine learning, a prediction unit that inputs explanatory variables into the trained model, the explanatory variables including prediction time difference information regarding the time difference from the implementation time to the prediction time at which the water level is to be predicted, and prediction rainfall information regarding the predicted rainfall for the period from the implementation time to the prediction time, and outputs information on an objective variable related to the information of the explanatory variables, and predicts the water level difference for the period from the implementation time to the prediction time based on the objective variable, and a predicted water level calculation unit that calculates the water level at the prediction time based on the water level difference predicted by the prediction unit. Effect of the Invention

[0012] According to the present disclosure, water levels can be predicted with high accuracy using a water detector that has a simpler configuration than a water level meter. [Brief description of the drawings]

[0013] [Figure 1] FIG. 1 is a diagram showing the overall configuration of a water level prediction system according to a first embodiment. [Diagram 2] Functional block diagram of a model generating device according to a first embodiment [Diagram 3] Functional block diagram of a water level prediction device according to a first embodiment [Figure 4] Schematic diagram showing the relationship between model learning and each parameter of water level prediction [Diagram 5] Hardware configuration diagram of the model generation device and water level prediction device [Figure 6] Flowchart of model generation processing according to the first embodiment [Figure 7] FIG. 7 is a diagram for explaining a method for calculating the total rainfall amount in step S110 of FIG. [Figure 8] Flowchart of water level prediction process according to the first embodiment [Figure 9] FIG. 1 is a diagram showing experimental results using the water level prediction system according to the first embodiment. [Figure 10] Functional block diagram of a model generating device according to a second embodiment [Figure 11] Functional block diagram of a water level prediction device according to a second embodiment [Figure 12] FIG. 13 is a diagram showing the overall configuration of a water level prediction system according to a third embodiment. [Figure 13] Functional block diagram of a model generating device according to a third embodiment [Figure 14] Functional block diagram of a water level prediction device according to a third embodiment DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] Hereinafter, an embodiment will be described with reference to the accompanying drawings. In order to facilitate understanding of the description, the same components in each drawing are denoted by the same reference numerals as much as possible, and duplicated description will be omitted.

[0015] [First embodiment] The first embodiment will be described with reference to Figures 1 to 9. Figure 1 is a diagram showing the overall configuration of a water level prediction system 1 according to the first embodiment.

[0016] The water level prediction system 1 is a system for predicting the water level Le at a desired future prediction time Te in specific locations P1 to Pn where a water detector 2 is installed. The specific locations P1 to Pn as targets for predicting the water level Le in this system include locations where no water exists during normal times, such as roads and residential areas, and particularly locations where there is a relatively high possibility of flooding during rainfall, such as lowlands and underpasses, and locations where there is a relatively high possibility of flooding, such as basements, subway stations, underground shopping malls, underground parking lots, and electrical rooms, as facilities in urban areas. The specific locations P1 to Pn may also include locations where water exists during normal times and where there is a relatively high possibility of overflow during rainfall, such as rivers, lakes, reservoirs, and their banks. The specific locations to be predicted by the water level prediction system 1 are n locations, P1 to Pn, in FIG. 1, but may be a single location or any multiple locations.

[0017] The water level prediction system 1 acquires various information when a water level reaches a plurality of predetermined levels using water detectors 2 installed at the specific locations P1 to Pn, and acquires the correspondence between the amount of rainfall and the water level at the specific locations P1 to Pn by machine learning based on the acquired information. The water level prediction system 1 can use the learning results to predict the water level Le at the specific locations P1 to Pn at a desired future prediction time Te.

[0018] As shown in FIG. 1, the water level prediction system 1 includes a water detector 2, a model generating device 3, a prediction model 4, and a water level prediction device 5.

[0019] The water detector 2 is a device that is installed in each of the specific locations P1 to Pn that are targets for water level prediction, and detects a change in the presence or absence of water at a predetermined water level. In the example of Fig. 1, the water detector 2 has detection units 21, 22, and 23 that detect the presence or absence of water at three different heights H1, H2, and H3 vertically above a reference plane G such as the ground, and each of the detection units 21, 22, and 23 detects the presence or absence of water, thereby detecting a change in the presence or absence of water at the three different predetermined water levels H1, H2, and H3 in the specific locations.

[0020] The number of predetermined water levels detected by the water detector 2 is not limited to three types, H1 to H3, and may be at least two or more. In short, it is sufficient that the number is such that the deviation (water level difference) of the water level at the specific location can be obtained based on the detection information by the water detector 2. However, the more predetermined water levels (for example, three or more types) that the water detector 2 can detect, the more teacher data used for learning the prediction model 4 described later can be created from data detected in the same time period, which is advantageous for improving the estimation accuracy of the trained model. In addition, in the example of FIG. 1, a single water detector 2 has multiple detection units 21, 22, and 23 and detects the presence or absence of water at multiple predetermined water levels H1, H2, and H3. However, multiple water detectors 2 capable of detecting different water levels may be installed in the same specific location, and a change in the presence or absence of water at multiple predetermined water levels may be detected based on the detection results of the multiple water detectors 2.

[0021] Here, the water detector 2 can be defined as a device that does not transmit during times when there is no change in the presence or absence of water, and transmits information about the occurrence of a change to an external device via wireless or wired communication only when there is a change in the presence or absence of water at a certain height. On the other hand, as a comparative example of the water detector 2, there is a device generally called a water level gauge. A water level gauge is a device that acquires the water level at regular intervals. In the above-mentioned Patent Documents 1 and 2, such a water level gauge is used to measure time series data on the water level of a river at the same time interval and transmit it to an external device. In order to operate the water level gauge, a commercial power source or solar power generation and a large-capacity storage battery are required, so construction work including the power source and a certain amount of space are required. In contrast, the water detector 2 can limit the value of the water level to be detected and does not need to transmit information related to the water level to an external device at regular intervals, so it can reduce power consumption more than the water level gauge, and can operate for several years with dry batteries, for example. In addition, since the water detector is smaller and cheaper than the water level gauge, it is highly likely to be installed at points where road flooding is a concern and where it is difficult to install a water level gauge. If there are surveillance cameras at the specific locations P1 to Pn that are the target of water level prediction, it is possible to use the surveillance cameras as water detectors by determining changes in water relative to a threshold value. In other words, the water detector 2 of this embodiment also includes detectors that can detect changes in the presence or absence of water using a mechanism different from the configuration in which the presence or absence of water is directly detected by the detection units 21, 22, and 23 as in the example of FIG.

[0022] The model generating device 3 is a device for generating, by using the prediction model 4, a model (trained model 4A) for predicting the water levels at the specific locations P1 to Pn.

[0023] The prediction model 4 is a model on which machine learning is performed by the model generating device 3, and after the machine learning is performed, the model is used to predict the water level by the water level prediction device 5. In the following explanation, the prediction model 4 in which a specific input / output relationship has been acquired after the machine learning is completed is also referred to as the "trained model 4A."

[0024] The water level prediction device 5 is a device for predicting the water level Le at each of the specific locations P1 to Pn by using the trained model 4A generated by the model generation device 3.

[0025] The model generating device 3 and the water level prediction device 5 according to the first embodiment will be described in detail with reference to Fig. 2 to Fig. 4. Fig. 2 is a functional block diagram of the model generating device 3 according to the first embodiment. Fig. 3 is a functional block diagram of the water level prediction device 5 according to the first embodiment. Fig. 4 is a schematic diagram showing the relationship between model learning and each parameter of water level prediction.

[0026] First, the model generating device 3 will be described in detail with reference to Fig. 2. As shown in Fig. 2, the model generating device 3 has a rainfall data acquiring unit 31, a detection information storing unit 32, a rainfall data storing unit 33, a teacher data creating unit 34, and a model learning unit 35.

[0027] The rainfall data acquisition unit 31 acquires time-series data (rainfall data) DA of the rainfall at the specific locations P1 to Pn for predicting the water level. The rainfall data acquisition unit 31 can be configured to acquire the rainfall data DA from outside the water level prediction system 1, for example, by collecting the rainfall data DA at the specific locations P1 to Pn via an Internet line from observation data of a real-time rainfall observation system (XRAIN) using a high-performance weather radar operated by the Ministry of Land, Infrastructure, Transport and Tourism. Alternatively, the water level prediction system 1 can be equipped with rain gauges installed at the specific locations P1 to Pn, and the rainfall data acquisition unit 31 can acquire the rainfall data DA using elements inside the water level prediction system 1, for example, by acquiring the rainfall data DA measured by the rain gauge.

[0028] Graph A is shown in the upper right of Fig. 4 as a bar graph showing an example of time-series data of rainfall for each predetermined time (e.g., 10 minutes). The horizontal axis of graph A indicates time (minutes), and the vertical axis indicates rainfall (mm) for each time period. If the implementation time Tn for water level prediction, which will be described later, is the current time, the value of each bar graph to the left of the horizontal axis of time Tn in graph A corresponds to the rainfall data DA of this embodiment.

[0029] The detection information storage unit 32 stores and holds detection information including any of the predetermined water levels H1, H2, H3 (simply indicated as "H" in Figure 2) when water is observed by the water detector 2 or when the observed water disappears, and the times T1, T2, T3 (simply indicated as "T" in Figure 2) when these water level changes were detected.

[0030] The rainfall data storage unit 33 stores and holds the rainfall data DA acquired by the rainfall data acquisition unit 31.

[0031] The teacher data creation unit 34 creates a teacher data set including learning time difference information, learning rainfall information, and learning water level difference information.

[0032] The learning time difference information is information relating to the time difference between times T1, T2, and T3 at which the water detector 2 detects a change in the presence or absence of water at a plurality of predetermined water levels H1, H2, and H3. In the following explanation, as shown on the left side of Fig. 4, the time difference between times T1 and T2 at which the change in the presence or absence of water is detected at the predetermined water levels H1 and H2 is represented as ΔT12, the time difference between times T2 and T3 at which the change in the presence or absence of water is detected at the predetermined water levels H2 and H3 is represented as ΔT23, and the time difference between times T1 and T3 at which the change in the presence or absence of water is detected at the predetermined water levels H1 and H3 is represented as ΔT13. In the following explanation, these pieces of time difference information may be collectively referred to as learning time difference information ΔTL.

[0033] The learning rainfall information is information on the amount of rainfall at the specific locations P1 to Pn where the water detector 2 that performed the detection is installed when a change in the presence or absence of water at the predetermined water levels H1, H2, and H3 is detected. In the following description, as shown on the left side of FIG. 4, the amount of rainfall during the period when the change in the water level from the predetermined water level H1 to the predetermined water level H2 is detected is represented as R12, the amount of rainfall during the period when the change in the water level from the predetermined water level H2 to the predetermined water level H3 is detected is represented as R23, and the amount of rainfall during the period when the change in the water level from the predetermined water level H1 to the predetermined water level H3 is detected is represented as R13. In the following description, these pieces of rainfall information may be collectively represented as learning rainfall information RL. In this embodiment, when calculating the rainfall amounts R12, R23, and R13, the time difference (delay time Td) from rainfall to the rise in the water level is taken into consideration for the period when the change in the water level from the predetermined water level H1 to the predetermined water level H2 is detected. A method for calculating the rainfall amounts R12, R23, and R13 taking into account the delay time Td will be described later with reference to FIG.

[0034] The learning water level difference information is information relating to the water level difference between a plurality of predetermined water levels H1, H2, and H3 detected by the water detector 2. In the following explanation, as shown on the left side of Fig. 4, the water level difference between the predetermined water level H1 and the predetermined water level H2 is represented as ΔH12, the water level difference between the predetermined water level H2 and the predetermined water level H3 is represented as ΔH23, and the water level difference between the predetermined water level H1 and the predetermined water level H3 is represented as ΔH13. In the following explanation, these pieces of water level information may be collectively represented as learning water level difference information ΔHL.

[0035] The teacher data creation unit 34 calculates learning time difference information ΔT12, ΔT23, ΔT13 (ΔTL) based on the time difference between times T1, T2, T3 when the predetermined water level transitions between H1, H2, H3, for example, among the detection information stored in the detection information storage unit 32. In addition, based on the information of the detected predetermined water levels H1, H2, H3 among the detection information stored in the detection information storage unit 32, the teacher data creation unit 34 calculates the transition of the predetermined water level related to the time period of the calculated learning time difference information ΔTL as learning water level difference information ΔH12, ΔH23, ΔH13 (ΔHL). Furthermore, based on the rainfall data DA stored in the rainfall data storage unit 33, the teacher data creation unit 34 calculates the total rainfall related to the time period of the calculated learning time difference information ΔTL as learning rainfall information R12, R23, R13 (RL).

[0036] The teacher data creation unit 34 creates a set of teacher data sets by associating each of the calculated learning time difference information ΔTL, learning water level difference information ΔHL, and learning rainfall information RL. The teacher data creation unit 34 can create multiple sets of teacher data sets by performing the above process multiple times. In the example of FIG. 4, three types of learning time difference information ΔT12, ΔT23, and ΔT13, learning water level difference information ΔH12, ΔH23, and ΔH13, and learning rainfall information R12, R23, and R13 are associated with each other to create three types of teacher data sets (ΔT12, ΔH12, R12), (ΔT23, ΔH23, R23), and (ΔT13, ΔH13, R13).

[0037] The model learning unit 35 uses the teacher data set created by the teacher data creation unit 34 to learn the prediction model 4 and generate the trained model 4A. The model learning unit 35 uses the learning time difference information ΔTL and the learning rainfall information RL of the teacher data as the input of the prediction model 4, and the learning water level difference information ΔHL as the output of the prediction model 4, and performs machine learning so that the input / output relationship of the prediction model 4 approaches the input / output relationship of the teacher data set, thereby generating the trained model 4A. In this embodiment, it is preferable to use multiple regression analysis as the machine learning method, and it is preferable that the prediction model 4 and the trained model 4A are regression models by multiple regression analysis. In addition, it is preferable that the learning time difference information ΔTL and the learning rainfall information RL are standardized before being used as input information of the prediction model 4, and the learning water level difference information ΔHL is also standardized before being used as output information of the prediction model 4.

[0038] The water level prediction device 5 will be described in detail with reference to Fig. 3. As shown in Fig. 3, the water level prediction device 5 has a prediction time input unit 51, a predicted rainfall data acquisition unit 52, a predicted rainfall data storage unit 53, a prediction time difference calculation unit 54, a predicted rainfall calculation unit 55, a prediction unit 56, a predicted water level calculation unit 57, a prediction result table 58, and a determination unit 59. The water level prediction device 5 can also be expressed as having a trained model 4A.

[0039] The trained model 4A is a model that obtains the correspondence between the learning explanatory variables including the above-mentioned learning time difference information ΔTL and learning rainfall information RL and the learning objective variables including the learning water level difference information ΔHL by machine learning. Based on the input of the explanatory variables including the forecast time difference information ΔT and forecast rainfall information ΔR, the trained model 4A outputs the objective variable (forecast water level difference ΔL) related to the input explanatory variables.

[0040] The predicted time input unit 51 accepts input of information on a future predicted time Te at which the water level Le at any specific location P1 to Pn is to be predicted. The information on the predicted time Te is acquired, for example, by the predicted time input unit 51 itself or other elements in the water level prediction device 5 automatically and periodically updating the information. Alternatively, the information on the predicted time Te may be acquired by a user of the water level prediction device 5 inputting a desired time using an input device such as a keyboard.

[0041] Furthermore, the predicted time input unit 51 acquires, for example, time information when input of the predicted time Te information is accepted as an implementation time Tn, which is a reference time for water level prediction, and outputs it together with the predicted time Te to the predicted time difference calculation unit 54. Note that the implementation time Tn is not necessarily limited to the current time when input of the predicted time Te information is accepted, and the user may input any time information other than the current time.

[0042] The predicted rainfall data acquisition unit 52 acquires time series data (predicted rainfall data) DB of the predicted rainfall for a time period ahead of the current time at the specific locations P1 to Pn for which the water level is predicted. The predicted rainfall data acquisition unit 52 can be configured to acquire the predicted rainfall data DB from outside the water level prediction system 1, for example, by using a service such as high-resolution precipitation nowcast provided by the Japan Meteorological Agency to collect the predicted rainfall data DB for the specific locations P1 to Pn via an Internet line or the like.

[0043] Graph A is shown in the upper right of Fig. 4 as a bar graph showing an example of time series data of rainfall for each predetermined time (e.g., 10 minutes). The horizontal axis of graph A indicates time (minutes), and the vertical axis indicates rainfall (mm) for each time period. If the implementation time Tn for water level prediction, which will be described later, is the current time, the values ​​of each bar graph to the right of the horizontal axis of time Tn in graph A correspond to the predicted rainfall data DB of this embodiment.

[0044] The forecast rainfall data storage unit 53 stores and holds the forecast rainfall data DB acquired by the forecast rainfall data acquisition unit 52.

[0045] The predicted time difference calculation unit 54 calculates the time difference (predicted time difference) ΔT between the predicted time Te input by the predicted time input unit 51 and the implementation time Tn at which the water level prediction is performed, and outputs information on the calculated time difference ΔT to the prediction unit 56 as prediction time difference information ΔT.

[0046] The forecast rainfall calculation unit 55 refers to the forecast rainfall data DB stored in the forecast rainfall data storage unit 53, calculates the total rainfall ΔR for the time period of the forecast time difference ΔT calculated by the forecast time difference calculation unit 54, and outputs information on the calculated total rainfall ΔR to the prediction unit 56 as forecast rainfall information ΔR.

[0047] The prediction unit 56 inputs explanatory variables including the prediction time difference information ΔT and the prediction rainfall information ΔR to the trained model 4A, and outputs a target variable related to the input explanatory variables. Here, the prediction time difference information ΔT is information on the time difference from the implementation time Tn at which the prediction is performed to the prediction time Te at which the water level is to be predicted, and is input from the prediction time difference calculation unit 54 to the prediction unit 56. The prediction rainfall information ΔR is information on the predicted rainfall during the period from the implementation time Tn to the prediction time Te, and is input from the predicted rainfall calculation unit 55 to the prediction unit 56. The prediction time difference information ΔT and the prediction rainfall information ΔR correspond to the learning time difference information ΔT12, ΔT23, ΔT13 (ΔTL) and the learning rainfall information R12, R23, R13 (RL), respectively, used as teacher data for machine learning. The prediction unit 56 predicts the water level difference ΔL during the period ΔT from the implementation time Tn to the prediction time Te, based on the objective variables output by the trained model 4A.

[0048] The predicted water level calculation unit 57 calculates the water level Le at the prediction time Te based on the water level difference ΔL predicted by the prediction unit 56. The predicted water level calculation unit 57 calculates, for example, a value obtained by adding any one (Hn) of the predetermined water levels H1, H2, and H3 detected by the water detector 2 at the implementation time Tn to the water level difference ΔL predicted by the prediction unit 56, as the predicted water level Le at the prediction time Te.

[0049] The prediction result table 58 stores information on the water level prediction results up to now. Every time the prediction unit 56 and the predicted water level calculation unit 57 calculate a new water level difference ΔL and a new predicted water level Le, the prediction result table 58 is updated by additionally storing the calculated information and related information such as the predicted time Te.

[0050] Based on information on the predicted water level Le at the prediction time Te calculated by the predicted water level calculation unit 57, the determination unit 59 determines whether the water level Le at the prediction time Te is a dangerous water level LVj (j is an integer).

[0051] The bottom right of Fig. 4 illustrates an example of graph B of a water level prediction at any one of the specific locations P1 to Pn. The horizontal axis of graph B indicates time (minutes), and the vertical axis indicates water level (m). The time on the horizontal axis of graph B is the same as that of graph A shown in the upper right of Fig. 4. The water level on the vertical axis of graph B corresponds to the height of each of the detection units 21, 22, and 23 of the water detector 2 shown on the left side of Fig. 4 relative to the reference plane G. Graph B illustrates, with dotted lines, the water levels corresponding to the predetermined heights H1, H2, and H3 detected by each of the detection units 21, 22, and 23, with the position of the reference plane G set as 0.

[0052] In the first embodiment, the range of the teacher data used for the machine learning of the model is within the range from the minimum value H1 to the maximum value H3 of the predetermined height that the water detector 2 can detect. For this reason, for example, as in the prediction graph C shown by the solid line in the graph B, when the water level Ln1 at the implementation time Tn is the same as the predetermined height H1, and the predicted water level Le1 at the prediction time Te is larger than the predetermined height H2 by the water level difference ΔL1, the prediction using the trained model 4A is an interpolation prediction, so it is considered that a highly accurate prediction is possible. On the other hand, for example, as in the prediction graph D shown by the dashed line in the graph B, when the water level Ln2 at the implementation time Tn is the same as the reference level G (i.e., a state in which there is no water) and the predicted water level Le2 at the prediction time Te is larger than the predetermined height H3 by the water level difference ΔL2, the prediction using the trained model 4A is an extrapolation prediction, but in this embodiment, the water level can be predicted with high accuracy even in this case.

[0053] FIG. 5 is a hardware configuration diagram of the model generating device 3 and the water level predicting device 5.

[0054] As shown in Figure 5, the model generating device 3 and water level prediction device 5 of the first embodiment can be physically configured as a computer system including a CPU (Central Processing Unit) 101, a GPU (Graphics Processing Unit) 108, RAM (Random Access Memory) 102 and ROM (Read Only Memory) 103 which are main memory devices, input devices 104 such as a keyboard and a mouse which are input devices, an output device 105 such as a display, a communication module 106 which is a data transmission / reception device such as a network card, an auxiliary memory device 107 such as a hard disk, and the like.

[0055] 2 are realized by loading predetermined computer software (model generation program) onto hardware such as the CPU 101 and RAM 102, operating the communication module 106, the input device 104, and the output device 105 under the control of the CPU 101, and reading and writing data from and to the RAM 102 and the auxiliary storage device 107. That is, by executing the model generation program of the first embodiment on a computer, the model generation device 3 functions as the rainfall data acquisition unit 31, the detection information storage unit 32, the rainfall data storage unit 33, the teacher data creation unit 34, the model learning unit 35, and the prediction model 4 of FIG.

[0056] Similarly, each function of the water level prediction device 5 shown in Fig. 3 is realized by loading a predetermined computer software (water level prediction program) onto hardware such as the CPU 101 and RAM 102, operating the communication module 106, the input device 104, and the output device 105 under the control of the CPU 101, and reading and writing data from and to the RAM 102 and the auxiliary storage device 107. That is, by executing the water level prediction program of the first embodiment on a computer, the water level prediction device 5 functions as the prediction time input unit 51, the predicted rainfall data acquisition unit 52, the predicted rainfall data storage unit 53, the prediction time difference calculation unit 54, the predicted rainfall calculation unit 55, the prediction unit 56, the predicted water level calculation unit 57, the prediction result table 58, the determination unit 59, and the trained model 4A of Fig. 3.

[0057] The model generation program and water level prediction program according to the first embodiment are stored, for example, in a storage device provided in a computer. Note that the model generation program and water level prediction program may be configured so that a part or all of them are transmitted via a transmission medium such as a communication line, and are received and recorded (including installed) by a communication module 106 provided in the computer. Also, the model generation program and water level prediction program may be configured so that a part or all of them are recorded (including installed) in the computer from a state in which they are stored in a portable storage medium such as a CD-ROM, DVD-ROM, or flash memory.

[0058] Similarly, the trained model 4A generated by the model generation device 3 may be stored alone in a storage medium and carried around, transmitted via a transmission medium, or recorded in a computer.

[0059] Next, a model generation method according to the first embodiment will be described with reference to Fig. 6 and Fig. 7. Fig. 6 is a flowchart of a model generation process according to the first embodiment. Each process in the flowchart shown in Fig. 6 is performed by the model generation device 3 according to the first embodiment.

[0060] First, in step S101, a process of acquiring data used to create teacher data is started, and the processes of steps S102 to S107 are repeated until it is determined in step S108 that acquisition of a predetermined number of pieces of data has been completed. The processes of steps S101 to S108 are performed by the detection information storage unit 32, and the acquired data corresponds to the above-mentioned "detection information."

[0061] In addition, Fig. 6 illustrates, as an example of data acquisition, a case in which a water detector 2 detects a continuous rise in water level from a predetermined height H1 to a predetermined height H3 at a specific location where the water detector 2 is installed, but data acquisition cases are not limited to the example of Fig. 6. For example, cases in which the water level continuously falls from a predetermined height H3 to a predetermined height H1, as opposed to the example of Fig. 6, cases in which the water level increases or decreases between the predetermined height H1 and the predetermined height H2, and cases in which the water level increases or decreases between the predetermined height H2 and the predetermined height H3 are also included.

[0062] In the case where the water level rises continuously from a predetermined height H1 to a predetermined height H3, the system waits until the water detector 2 detects the predetermined water level H1 (No in S102). Then, when the water level rises and the water detector 2 detects the predetermined water level H1 (Yes in S102), the time T1 at which the water level reaches the predetermined water level H1 and information on the reached water level H1 are linked and recorded as detection information in the detection information storage unit 32 (S103).

[0063] Next, the system waits until the water detector 2 detects the predetermined water level H2 (No in S104). Then, when the water level further rises from the predetermined water level H1 and the water detector 2 detects the predetermined water level H2 (Yes in S104), the time T2 when the water level reaches the predetermined water level H2 and information on the reached water level H2 are linked together and recorded as detection information in the detection information storage unit 32 (S105).

[0064] Next, the system waits until the water detector 2 detects the predetermined water level H3 (No in S106). Then, when the water level further rises from the predetermined water level H2 and the water detector 2 detects the predetermined water level H3 (Yes in S106), the time T3 when the water level reaches the predetermined water level H3 and information on the reached water level H3 are linked and recorded as detection information in the detection information storage unit 32 (S107).

[0065] In this case, the times T1, T2, T3 at which the specified water levels H1, H2, H3 were reached through a series of water level rises, and information on each reached water level H1, H2, H3 are stored in the detection information storage unit 32 as a data set of detection information.

[0066] The processes of steps S102 to S107 are repeated, and when the number of data sets stored in the detection information storage unit 32 reaches a predetermined number, the process proceeds to step S109 and subsequent steps.

[0067] In step S109, the teacher data creation unit 34 calculates the water level difference (learning water level difference information ΔHL) and the time difference (learning time difference information ΔTL) for each acquired data. For each data set of detection information acquired in the processing of steps S101 to S108, the teacher data creation unit 34 calculates the time difference ΔT12 from the difference T2-T1 between the arrival time T1 of the predetermined water level H1 and the arrival time T2 of the predetermined water level H2, calculates the water level difference ΔH12 from the difference H2-H1 between the predetermined water level H1 and the predetermined water level H2, and links the time difference ΔT12 and the water level difference ΔH12 to form a set of data sets. Similarly, a time difference ΔT23 is calculated from the difference T3-T2 between the arrival time T2 of the predetermined water level H2 and the arrival time T3 of the predetermined water level H3, a water level difference ΔH23 is calculated from the difference H3-H2 between the predetermined water level H2 and the predetermined water level H3, and the time difference ΔT23 and the water level difference ΔH23 are linked to form a set of data. Similarly, a time difference ΔT13 is calculated from the difference T3-T1 between the arrival time T1 of the predetermined water level H1 and the arrival time T3 of the predetermined water level H3, and the water level difference ΔH13 is calculated from the difference H3-H1 between the predetermined water level H1 and the predetermined water level H3, and the time difference ΔT13 and the water level difference ΔH13 are linked to form a set of data.

[0068] In step S110, the teacher data creation unit 34 calculates the total rainfall (learning rainfall information RL) for the time period corresponding to the time difference (learning time difference information ΔTL) calculated in step S109, from the rainfall data DA stored in the rainfall data storage unit 33. The teacher data creation unit 34 extracts time series information of the rainfall in the time periods related to each of the time differences ΔT12, ΔT23, and ΔT13 calculated in step S109 from the rainfall data DA, and calculates the total rainfall corresponding to each of the time differences ΔT12, ΔT23, and ΔT13.

[0069] Fig. 7 is a diagram for explaining a method for calculating the total rainfall in step S110 in Fig. 6. Fig. 7 illustrates an example in which the water level rises from H1 to H2 due to rainfall. The horizontal axis in Fig. 7 is the time axis. One vertical axis in Fig. 7 indicates the water level, and the positions of the predetermined water levels H1 and H2 are shown by dotted lines in the graph. The other vertical axis in Fig. 7 indicates the amount of rainfall, and Fig. 7 illustrates the amount of rainfall for each predetermined time period with four bar graphs.

[0070] Here, the median time (delay time) required for water to rise after flowing into the location targeted for water level prediction from the time of rainfall observation is Td. Td can also be expressed as the time difference until the water level at a specific location actually starts to rise due to the influence of rainfall. In the example of FIG. 7, at the specific location targeted for water level prediction, the water level first rises from a waterless state due to rainfall, and at time T1, the water detector 2 detects that the water level has reached a predetermined water level H1. After that, the water level continues to rise due to rainfall, and at time T2, the water detector 2 detects that the water level has reached a predetermined water level H2. The water level difference obtained at this time is ΔH12=H2-H1, and the time difference is ΔT12=T2-T1.

[0071] In this embodiment, the total rainfall R12 associated with the water level difference ΔH12 and the time difference ΔT12 is calculated taking into consideration the delay time Td described above. In the example of FIG. 7, the rainfall observed in the time period of the time difference ΔT12 is the sum of the rainfall amounts indicated by the three bar graphs Rb, Rc, and Rd. Here, when the time until the rain on the ground affects the water level is the delay time Td described above, the delay time Td affects both the rise in the water level of the predetermined water level H1 and the rise in the water level of the predetermined water level H2. Therefore, in this embodiment, the rainfall amount in the time period from the time T1 back by the delay time Td to the time T2 back by the delay time Td is set as the total rainfall R12. In the example of FIG. 7, the rainfall amount observed in the time period back by the delay time Td from the time T1 is the rainfall amount indicated by one bar graph Ra, so this rainfall amount Ra is added to the total rainfall R12. Furthermore, going back from time T2 by the delay time Td, the rainfall indicated by the bar graph Rd for the time period closest to time T2 is excluded from the rainfall to be included in the total rainfall R12. From the above, in the example of Fig. 7, the total rainfall R12 associated with the water level difference ΔH12 and the time difference ΔT12 can be calculated by the following formula (1) as the sum of the rainfall in the time period from time T1-Td to time T2-Td. R12 = Ra + Rb + Rc (1)

[0072] Even in the case where the water level rises from H1 to H3 due to rainfall, the total rainfall R13 can be calculated in the same manner as in the example described with reference to FIG. 7. In this case, the water level difference is ΔH13=H3-H1, and the time difference is ΔT13=T3-T1. Considering the example of FIG. 7 as a base, the time T3 at which the water level H3 is reached is later than the time T2, so the time difference ΔT13 extends further to the right from ΔT12 shown in FIG. 7. Therefore, the total rainfall R13 associated with the water level difference ΔH13 and the time difference ΔT13 can be calculated by adding the rainfall Ra in the time period of the delay time Td before the time T1, as well as the rainfall in the time period from the time T2 to the time T3-Td, in the same manner as the total rainfall R12 shown in FIG. 7.

[0073] In addition, in the case where the water level rises from H1 to H3 due to rainfall, the total rainfall R23 from water level H2 to H3 can be calculated in the same manner as the example described with reference to Fig. 7. In this case, the water level difference is ΔH23=H3-H2, and the time difference is ΔT23=T3-T2. Based on the example of Fig. 7, the total rainfall R23 associated with the water level difference ΔH23 and the time difference ΔT23 can be calculated as the sum of the rainfall in the time period from time T2-Td to time T3-Td.

[0074] In this way, in this embodiment, all detection positions H1, H2, and H3 with height differences can be used to generate teacher data including the water level difference, time difference, and total rainfall. Note that if the water level does not reach the height of the specified water level H2 or H3, no data is acquired.

[0075] Moreover, data when the water level is decreasing can be obtained by, for example, performing the procedure described with reference to FIG. 7 in the order of H3, H2, and H1 from the highest to the lowest predetermined height.

[0076] In this way, the data accumulated by the rise and fall of the water level can be calculated as the time difference ΔTij when the water level detection position rises from a relatively low predetermined height Hi to a relatively high predetermined height Hj, and the rainfall Rij related to the rise of the water level. On the other hand, when the water level falls from the predetermined height Hj to the predetermined height Hi, the data can be calculated as the time difference ΔTji for the fall of the water level detection position height difference ΔHji, and the rainfall Rji related to the fall of the water level. If the accumulated height difference ΔH is the objective variable, and the time difference ΔT and rainfall R are explanatory variables, the appropriate input-output relationship of the prediction model can be obtained by supervised machine learning, and ΔH can be predicted from ΔT and R using the trained model.

[0077] In step S111, the teacher data creation unit 34 creates a teacher data set by associating the water level difference and time difference calculated in step S109 with the total rainfall calculated in step S110 (teacher data creation step). The teacher data creation unit 34 sets the water level difference and time difference calculated in step S109 as the above-mentioned learning water level difference information ΔHL and learning time difference information ΔTL, respectively, and sets the total rainfall calculated in step S110 as the above-mentioned learning rainfall information RL, and links these three pieces of information to create a set of teacher data sets. The teacher data creation unit 34 repeats this procedure for all the data acquired in steps S101 to S108 to create multiple teacher data sets.

[0078] In step S112, the model learning unit 35 uses the created teacher data set to learn the prediction model 4 through supervised machine learning (model learning step). The model learning unit 35 can apply various algorithms such as regression, decision tree, support vector machine, and neural network as supervised machine learning. The model learning unit 35 only needs to be able to learn the prediction model 4 using at least supervised learning, and there are no restrictions on the algorithm to be applied, but in this embodiment, an example in which multiple regression analysis is applied will be described.

[0079] The height difference ΔHL, which is the response variable, can be expressed by the multiple regression equation shown in the following equation (2) using the partial regression coefficient P for the rainfall RL and the partial regression coefficient Q for the time difference ΔTL. Here, P is the coefficient that represents the contribution of the rise in water level due to rainfall, and Q can be considered as the coefficient that represents the drop in water level due to water outflow. ΔHL = P × RL + Q × ΔTL (2)

[0080] In multiple regression analysis, a large number of explanatory variables and target variables are combined to find P and Q that minimize the sum of squared residuals (the sum of the squares of the differences between observed values ​​and predicted values).

[0081] The prediction is made when a change from absence to presence of water is detected at a given height Hn of the water detector 2 at a certain time Tn, and the water level Le at a time Te after ΔT thereafter is calculated using the partial regression coefficients P and Q calculated earlier in the following equation (3), where the amount of rainfall from time Tn-Td to Te-Td is ΔR. Le = Hn + P × ΔR + Q × ΔT (3)

[0082] In the case of multiple regression analysis, the time (lag time) Td required for the water level to rise after it flows into the location targeted for water level prediction from the time of rainfall observation can be calculated by setting multiple Td (for example, 10 minutes, 20 minutes, 30 minutes) and performing multiple regression analysis to automatically calculate the best coefficient of determination, i.e., the Td when the residual variation is the smallest relative to the total variation. Alternatively, the lag time Td can be manually set based on examples of observations of the time until runoff (the time difference between the peak rainfall and the peak water level). The lag time Td calculated in this way is a parameter that is uniquely determined by various conditions, such as the geographical conditions of each of the specific locations P1 to Pn targeted for water level prediction.

[0083] As a method for determining the delay time Td, other than the above method using the coefficient of determination as a criterion, for example, the following method can be mentioned. (1) Create multiple types of trained models 4A using multiple delay times Td. Apply actual rainfall information (rainfall information for forecasting ΔR) to each of the multiple trained models 4A created, and compare the obtained results (forecasted water level difference ΔL) with the detection status of the water detector 2 (water level difference information for learning ΔHL). As a result of the comparison, adopt the delay time Td used in the trained model 4A that produced the best prediction results. (2) When there is a large amount of training data, a portion of the training data can be used to train the prediction model 4, and the rest can be used for validation (cross-validation). Multiple types of trained models 4A with multiple delay times Td are created, and the delay time Td used in the trained model 4A that gave the best prediction result for the validation data through cross-validation is adopted.

[0084] When using other supervised machine learning algorithms (for example, neural networks), Td can be calculated from the coefficient of determination of multiple regression analysis or observed examples and used.

[0085] To improve prediction accuracy, observed rainfall can be used instead of predicted rainfall when creating prediction model 4 (determining P and Q in the case of multiple regression analysis). Other methods to improve prediction accuracy include calculating prediction model 4 using all data from rising and falling water levels to ensure a sufficient amount of data, or dividing the height intervals for calculation when there is sufficient data. It is also possible to use multiple machine learning algorithms in combination.

[0086] The model learning unit 35 provides the prediction model 4 that has completed learning in step S112 as a learned model 4A to the water level prediction device 5. When the process of step S112 is completed, this control flow ends.

[0087] 6 may not be performed as a series of processes. For example, the data acquisition process of steps S101 to S108 may be performed sequentially when a data acquisition condition is satisfied, such as when it is raining, and each process of teacher data creation and model learning of steps S109 to S112 may be performed periodically or when new data is acquired in the data acquisition process.

[0088] Next, the water level prediction method according to the first embodiment will be described with reference to Fig. 8. Fig. 8 is a flowchart of the water level prediction process according to the first embodiment. Each process in the flowchart shown in Fig. 8 is performed by the water level prediction device 5 according to the first embodiment.

[0089] In step S201, information on a future predicted time Te at which the water level Le is to be predicted is acquired by the predicted time input unit 51. The information on the predicted time Te may be the same time information input collectively for all specific locations P1-Pn where the water detectors 2 are installed, or the specific locations P1-Pn for which the water level Le is to be predicted may be limited and selected, or a different predicted time Te may be input for each specific location P1-Pn. If the predicted time input unit 51 acquires information on at least one predicted time Te for at least one specific location P1-Pn, the process proceeds to step S202.

[0090] In step S202, a trained model 4A is set (model setting step). As the trained model 4A, a model generated in advance by the model generating device 3 can be used, or a model obtained by another method can be used.

[0091] Next, in step S203, water level prediction processing is started at each of the specific locations P1 to Pn for which water level prediction is instructed, and the processing of steps S204 to S209 is repeated until it is determined in step S210 that water level prediction processing for all locations has been completed.

[0092] In step S204, the prediction time difference calculation unit 54 calculates the time difference information ΔT for prediction. The prediction time difference calculation unit 54 calculates the time difference ΔT between the prediction time Te acquired in step S201 and the current time (implementation time) Tn at which the prediction is performed. Then, the information on the calculated time difference ΔT is output to the prediction unit 56 as the prediction time difference information ΔT, which is one of the explanatory variables.

[0093] In step S205, the forecast rainfall calculation unit 55 calculates the forecast rainfall information ΔR. The forecast rainfall calculation unit 55 refers to the forecast rainfall data DB stored in the forecast rainfall data storage unit 53, calculates the total rainfall ΔR for the time period of the forecast time difference ΔT calculated in step S204, and outputs the calculated total rainfall ΔR information to the prediction unit 56 as the forecast rainfall information ΔR. Note that, in the calculation of the total rainfall ΔR, the rainfall observed in the time period from the time Tn-Td, which is a predetermined delay time Td back from the implementation time Tn, to the time Te-Td, which is a predetermined delay time Td back from the forecast time Te, is also added, as in the method of creating the learning rainfall information RL described with reference to FIG. 7.

[0094] In step S206, the prediction unit 56 inputs the prediction time difference information ΔT calculated in step S204 and the prediction rainfall information ΔR calculated in step S205 as explanatory variables into the trained model 4A and outputs the objective variable. In this case, when the trained model 4A created from standardized training data is used, the prediction time difference information ΔT and the prediction rainfall information ΔR input as explanatory variables into the trained model 4A need to be standardized before being input into the trained model 4A. In the case where the trained model 4A is a multiple regression analysis model, the standardized prediction rainfall information ΔR and the prediction time difference information ΔT are multiplied by the corresponding partial regression coefficients P and Q, respectively, and added to calculate the objective variable (see the above formula (3)).

[0095] In step S207, the prediction unit 56 calculates the predicted water level difference ΔL based on the objective variable output from the trained model 4A (prediction step). Here, when the learning water level difference information ΔHL standardized during machine learning of the prediction model 4 as described above is used as the output information of the prediction model 4, the objective variable output from the trained model 4A also becomes a standardized value of the actual predicted water level difference ΔL. For this reason, the prediction unit 56 multiplies the objective variable by the standard deviation of the water level difference data used during learning and adds the average value to calculate the numerical value of the actual predicted water level difference ΔL.

[0096] In step S208, the predicted water level calculation unit 57 calculates the predicted water level Le at the prediction time Te based on the predicted water level difference ΔL calculated in step S207 (predicted water level calculation step). The predicted water level calculation unit 57 can obtain information on the detected water level Hn at the time Tn when the water level prediction is performed from the water detector 2, for example, and add the predicted water level difference ΔL to this detected water level Hn to calculate the predicted water level Le.

[0097] For example, in the example of graph B in Fig. 4, in the case of prediction graph C, the water detector 2 detects a water level of a predetermined height H1 at the implementation time Tn, so the predetermined height H1 is added to the predicted water level difference ΔL1 to calculate the predicted water level Le1 (=H1+ΔL1). Similarly, in the case of prediction graph D, the water detector 2 does not detect the water level at the implementation time Tn, so it can be determined that the water level is 0, so the height 0 is added to the predicted water level difference ΔL2 to calculate the predicted water level Le2 (=0+ΔL2). In other words, if the water detector 2 is unable to detect the water level Hn at the implementation time Tn, the predicted water level difference ΔL calculated by the prediction unit 56 becomes the predicted water level Le as is.

[0098] In step S209, the judgment unit 59 judges the dangerous water level LVj (j is an integer) at the prediction time Te. The judgment unit 59 holds judgment criteria information that is set in advance according to various conditions at the relevant point, for example, and can judge the dangerous water level LVj individually for each point. Specific examples of the dangerous water level LVj can include whether evacuation is necessary, whether flood prevention work is necessary, whether traffic restrictions are necessary, etc. For example, if the specific locations P1 to Pn that are the targets of water level prediction are roads, the dangerous water level LVj can be set to two types of levels, for example, "below the traffic restriction water level" and "above the traffic restriction water level."

[0099] The predicted water level calculation unit 57 and the determination unit 59 may notify residents around the location of the predicted water level Le and information on whether the water level is dangerous or not by using sound information or the like. The dangerous water level LVj may be set in multiple stages according to the predicted water level Le. In this case, the maximum number of subscripts j corresponds to the number of stages of the dangerous water level set in advance. In other words, the dangerous water level includes at least one stage depending on various conditions such as various environmental conditions for each of the specific locations P1 to Pn where the water level prediction is performed. The dangerous water level may also include multiple types of levels at the same location of the specific locations P1 to Pn.

[0100] Implementing a configuration for determining and reporting the dangerous water level LVj at the predicted time Te as in step S209 has the following advantages, for example. For example, consider a case where the specific locations P1 to Pn targeted for water level prediction are roads, and the content of the dangerous water level LVj is whether the road is flooded or not, and whether traffic restrictions due to flooding are necessary or not. In this case, if the start and end times of road restrictions due to road flooding can be predicted with high accuracy, road users can determine safe routes based on the prediction information, and road managers can operate roads efficiently. In this way, by being able to predict the dangerous water level LVj with high accuracy, it becomes possible to quickly implement highly reliable countermeasures based on the prediction information, and the safety of personnel in the vicinity of the specific locations P1 to Pn can be reliably improved.

[0101] In step S210, the prediction result table 58 is updated. In the prediction result table 58, for example, information on the predicted water level Le at the prediction time Te calculated in step S208, information on the dangerous water level LVj determined in step S209, etc. are recorded. The processes of steps S204 to S210 are repeated, and when water level predictions for all points in the specific location specified in step S201 are completed (S211), this control flow ends.

[0102] Fig. 9 is a diagram showing the results of an experiment using the water level prediction system 1 according to the first embodiment. The horizontal axis in Fig. 9 indicates time, which is the period from 14:10 on June 2nd to 04:50 on June 4th. One vertical axis in Fig. 9 indicates the amount of rainfall (mm) every 10 minutes, and these rainfall data are shown in Fig. 9 as a bar graph. The other vertical axis in Fig. 9 indicates the water level (m), which is shown in Fig. 9 as a solid line graph E and a dotted line graph F.

[0103] In the experiment in Figure 9, to determine the partial regression coefficients of the multiple regression analysis used in prediction model 4, water level gauge data from April 1, 2019 to December 31, 2022 was used, assuming that water detector 2 was located at heights of 0.03 m, 0.10 m, 0.20 m, and 0.25 m from the ground surface, and a total of 64 explanatory variables were used for 32 sets of water level rises and 32 sets of water level falls for six different delay times in 10-minute increments from 0 minutes to 50 minutes.

[0104] Furthermore, a water level of 0.03m or more is considered a road flooding event, and one road flooding event is defined as the time from when the water level reaches 0.03m or more until it falls below 0.03m again. However, if the water level falls below 0.03m and then rises to 0.03m or more again within two hours, it is considered to be one continuous road flooding event. In this example, the water level for road restrictions (i.e. an example of the danger water level LVj) can be set to 0.2m, which is used as a guideline by quite a few road administrators.

[0105] Using the above data, we performed machine learning on predictive model 4, and the best coefficients of determination and partial regression coefficients are as follows: Coefficient of determination: 0.77 Partial regression coefficient P: 0.452 Partial regression coefficient Q: -0.329

[0106] FIG. 9 shows the results of prediction of road flooding due to Typhoon No. 2 in June 2023 based on the detection of water and the detection of water disappearance from the detection unit of the water detector 2 installed at 0.03 m (low position) and 0.34 m (high position) using the values ​​of each of the above coefficients. Graph E shown by a solid line in FIG. 9 shows the predicted water level Le by the water level prediction device 5 of the water level prediction system 1 when a multiple regression analysis model using the above coefficients is applied as the trained model 4A. Graph F shown by a dotted line in FIG. 9 shows the actual water level observed by the water level gauge as a comparison target. In other words, graph E is an example, and graph F is a comparative example.

[0107] Referring to the amount of rainfall shown by the bar graph in FIG. 9, in this experiment, it rains in the first half of the time axis, and the amount of rainfall is measured. In particular, the rain is heavy in the early part of the time axis, and the rain gradually weakens. With this amount of rainfall, in graph E of the embodiment, in the section before time TA shown in FIG. 9, prediction is made based on information that water is detected by the water detector 2 in a low position, and in the section from time TA to time TB, prediction is made based on information that water is detected by the water detector 2 in a high position. In addition, in the section after time TB, the water level decreases, and prediction is made based on information that the water detector 2 in the high position detects the disappearance of water.

[0108] Comparing graph E of the embodiment with graph F of the comparative example, it can be seen that graph E of the embodiment is able to capture and predict the actual water level fluctuation trends shown in the comparative example based on rainfall information.

[0109] The effects of the water level prediction system 1 according to the first embodiment will be described below.

[0110] The model generation device 3 according to the first embodiment is a model generation device 3 for generating a trained model 4A for predicting water levels Le at specific locations P1 to Pn, and includes learning time difference information ΔTL relating to the time differences between times T1, T2, and T3 at which a change in the presence or absence of water at a plurality of specific water levels H1, H2, and H3 is detected by a water detector 2 that is installed at the specific locations P1 to Pn and detects a change in the presence or absence of water at the specific water levels H1, H2, and H3, and learning time difference information ΔTL relating to the time differences between times T1, T2, and T3 at which a change in the presence or absence of water at the specific water levels H1, H2, and H3 is detected, and learning time difference information ΔTL relating to the amount of rainfall at the specific locations P1 to Pn at which the change in the presence or absence of water at the specific water levels H1, H2, and H3 is detected. and a model learning unit 35 that uses the teacher dataset created by the teacher data creation unit 34 to perform machine learning using the learning time difference information ΔTL and the learning rainfall information RL as inputs to a prediction model 4 and the learning water level difference information ΔHL as an output of the prediction model 4 so that the input-output relationship of the prediction model 4 approaches the input-output relationship of the teacher dataset, thereby generating a trained model 4A.

[0111] In the model generating device 3 of the first embodiment, the detection result of the water detector 2 installed at one of the specific locations P1 to Pn designated for predicting the water level Le is used to perform machine learning of a prediction model for predicting the future water level difference ΔL at that location, and generate a trained model 4A. As described above, the water detector 2 is a device that does not transmit during a time period when there is no change in the presence or absence of water, and transmits information about the occurrence of a change to an external device via wireless or wired communication only when there is a change in the presence or absence of water at a certain height. Therefore, unlike conventional water level gauges, there is no need to transmit information related to the water level to an external device at regular intervals, and the value of the water level to be detected can be limited, so there is an advantage that the power consumption can be reduced compared to a water level gauge, and it can be operated for several years with, for example, a dry cell. In other words, the water detector 2 can grasp flooding regardless of the presence or absence of water, with low cost and space saving compared to a water level gauge, and can detect changes in the water level with a simple configuration.

[0112] Therefore, in the water level prediction system 1 according to the first embodiment, by using the water detector 2 to obtain a teacher data set used for machine learning of the prediction model 4 by the model generation device 3, it is possible to easily obtain a teacher data set sufficient for machine learning of the prediction model 4 in the specific locations P1 to Pn to be predicted. Then, by using the teacher data set obtained in this way, a trained model 4A capable of highly accurate water level prediction can be generated. As a result, the model generation device 3 according to the first embodiment makes it possible to perform highly accurate water level prediction using the water detector 2 having a simple configuration compared to a water level gauge. Furthermore, if the water level can be predicted using the water detector 2 that can grasp flooding at low cost and in a small space compared to a water level gauge, it will be easy to spread due to its low cost and space saving. Therefore, the spread of the water level prediction system 1 according to the first embodiment can be promoted.

[0113] Moreover, in the model generating device 3 according to the first embodiment, multiple regression analysis is used for the machine learning of the model learning unit 35. Among various methods of supervised machine learning, the advantages of using multiple regression analysis are as follows, for example.

[0114] [1] Excellent interpretability consistent with physical models The method of predicting water level based on accumulated rainfall and elapsed time in the water level prediction system 1 of the first embodiment reflects the mechanism (physical model) of actual water level changes, namely, the rise in water level due to rainfall flowing into the prediction point and the fall in water level due to water being discharged from the prediction point. Multiple regression analysis shows the degree of water level rise due to accumulated rainfall and the degree of water level fall over time as partial regression coefficients. This makes it easy to understand the effects of rainfall and time on water levels, which correspond to the physical model, and has the advantage of being excellent in explaining the prediction mechanism and making the prediction results easy to understand.

[0115] [2] Robustness Multiple regression analysis guarantees that water levels will increase at a constant rate due to accumulated rainfall and decrease at a constant rate over time, making the prediction results robust and less likely to be far off.

[0116] [3] Evaluation of statistical reliability Under the assumption that the training data follows a normal distribution, multiple regression analysis can show the prediction interval for the population, i.e., the range within which the predicted results are statistically possible.

[0117] The reason why logistic regression is used for classification in the water level prediction system according to the first embodiment is as follows. Supervised machine learning algorithms for classification include, for example, decision trees, support vector machines (SVMs), and ensemble learning using these. Decision trees are often used for qualitative variables and discard a small percentage of variables, while support vector machines are a method for obtaining a hyperplane with a method called the kernel trick to find the boundary that allows the best classification. These methods are primarily aimed at finding clear boundaries. In contrast, logistic regression is a method that aims to calculate the probability value of the classification target, and was adopted because the rainfall used in the prediction is a probabilistic event and it would be useful to be able to show probability values ​​in determining activities for flood prevention and traffic control.

[0118] In the model generating device 3 according to the first embodiment, it is preferable that the plurality of predetermined water levels detected by the water detector 2 at any one of the specific locations P1 to Pn include three or more types of predetermined water levels. Note that the example in Fig. 1 etc. illustrates a configuration for detecting three types of predetermined water levels H1, H2, and H3.

[0119] In the model generating device 3, in order to calculate the learning time difference information ΔTL, the learning rainfall information RL, and the learning water level difference information ΔHL included in the teacher data set, the water detector 2 only needs to detect at least two predetermined water levels. If there are two detected predetermined water levels, the water level difference between these two predetermined water levels can be calculated, and the time required for this water level difference (time difference) can be obtained, and further, the total rainfall in the time period related to this time difference can be obtained. However, if three or more types of predetermined water levels can be detected, for example, if a flood occurs once in which the water level rises from the lowest predetermined water level to the highest predetermined water level, multiple water level differences between multiple predetermined water levels can be obtained using the detection result at that time. Then, multiple time difference information and total rainfall information corresponding to these multiple water level difference information can also be obtained, and as a result, multiple sets of teacher data sets can be created. In this way, if a configuration is capable of detecting three or more types of predetermined water levels, multiple teacher data sets can be created when a water level fluctuation phenomenon occurs once, so that teacher data sets can be created efficiently. Furthermore, as the number of detectable predetermined water levels increases, the number of sets of teacher data sets that can be created also increases, making it possible to create teacher data sets even more efficiently.

[0120] Furthermore, in the case of a configuration in which the water detector 2 detects three predetermined water levels H1, H2, and H3 at any one of the specific locations P1 to Pn, by placing one of the three water detectors at a certain threshold height (for example, the water level at which traffic is restricted) and placing the other two at heights below and above the threshold, the water detector placed below the threshold can be used to predict the rise in water level to the threshold, and the water detector placed above the threshold can be used to predict the fall of the water level that has risen above the threshold to the threshold. This makes it possible to predict, for example, the start and end of restrictions.

[0121] The water level prediction device 5 according to the first embodiment is a water level prediction device 5 for predicting the water level Le at specific locations P1 to Pn, and includes learning explanatory variables including learning time difference information ΔTL relating to the time difference between times T1, T2, T3 at which a change in the presence or absence of water at a plurality of specified water levels H1, H2, H3 was detected by a water detector 2 that is installed at the specific locations P1 to Pn and detects a change in the presence or absence of water at the specified water levels H1, H2, H3, and learning rainfall information RL relating to the rainfall at the specific locations P1 to Pn when the change in the presence or absence of water at the specified water levels H1, H2, H3 was detected; and learning water level difference information ΔHL relating to the water level difference between the plurality of specified water levels H1, H2, H3. The trained model 4A has acquired, through machine learning, a correspondence relationship with learning objective variables including the above, and a prediction unit 56 that inputs explanatory variables including prediction time difference information ΔT regarding the time difference from the implementation time Tn at which the prediction is performed to the prediction time Te at which the water level Te is to be predicted, and prediction rainfall information ΔR regarding the predicted rainfall for the period from the implementation time Tn to the prediction time Te, outputs information on the objective variable related to the information on the explanatory variables, and predicts the water level difference ΔL for the period from the implementation time Tn to the prediction time Te based on the objective variable, and a predicted water level calculation unit 57 that calculates the water level Le at the prediction time Te based on the water level difference ΔL predicted by the prediction unit 56.

[0122] As already explained as an effect of the model generating device 3, by using the water detector 2 that can grasp flooding at low cost and in a small space compared to a water level gauge and can detect changes in water level with a simple configuration, it is possible to easily obtain a teacher data set sufficient for machine learning of the prediction model 4 at the specific locations P1 to Pn to be predicted, and it can be obtained at low cost. Using the teacher data set acquired in this way, a trained model 4A capable of predicting water levels with high accuracy can be generated. Then, by using the trained model 4A generated in this way in the water level prediction device 5, it becomes possible to perform water level prediction with high accuracy. As a result, the water level prediction device 5 according to the first embodiment makes it possible to perform water level prediction with high accuracy using the water detector 2 that has a simple configuration compared to a water level gauge. Furthermore, if the water level can be predicted using the water detector 2 that can grasp flooding at low cost and in a small space compared to a water level gauge, it will be easy to spread because of its low cost and space saving. Therefore, the spreadability of the water level prediction system 1 according to the first embodiment can be promoted.

[0123] The water level prediction device 5 according to the first embodiment also includes a determination unit 59 that determines whether the water level Le at the prediction time Te calculated by the predicted water level calculation unit 57 is at a dangerous water level LVj or not.

[0124] With this configuration, it is possible to determine not only the future predicted water level Le at the specific locations P1 to Pn being predicted, but also whether the predicted water level Le has reached the dangerous water level LVj, making it possible to take measures based on the water level prediction more quickly.

[0125] Furthermore, the water level prediction system 1 according to the first embodiment makes it possible to perform highly accurate water level predictions using the water detector 2, which has a simpler configuration than a water level gauge, as described above. This makes it easier to predict flooding and inundation in urban areas and road areas, which was previously difficult to predict using conventional water level prediction methods that use water level gauges.

[0126] [Second embodiment] A water level prediction system 1A according to the second embodiment will be described with reference to FIGS.

[0127] The overall configuration of the water level prediction system 1A according to the second embodiment is basically the same as that of the water level prediction system 1 according to the first embodiment described with reference to Fig. 1. However, there are differences in the model generating device 3A and the water level prediction device 5A. Also, it differs from the first embodiment in that it includes a plurality of prediction models 41-4i and trained models 4A1-4Ai (i is an integer).

[0128] 10 is a functional block diagram of a model generating device 3A according to the second embodiment. As shown in FIG.

[0129] The rainfall pattern determination unit 36 ​​determines the rainfall pattern RPi (i is any integer) associated with the time period of the calculated learning time difference information ΔTL based on the rainfall data DA stored in the rainfall data storage unit 33. The rainfall pattern determination unit 36 ​​receives rainfall pattern information manually input from weather information, for example. The rainfall pattern can be set to normal, which is the default, or to weather warnings such as typhoon and short-term heavy rain.

[0130] Alternatively, the rainfall pattern determination unit 36 ​​may be configured to classify and hold information on rainfall trends, such as time-series data on rainfall amounts in a plurality of rainfall patterns, such as typhoons, torrential rain, prolonged rain, and sudden rain, in advance. In this case, the rainfall pattern determination unit 36 ​​can select one rainfall pattern RPi that corresponds to the rainfall data DA to be determined from the plurality of rainfall patterns, for example, by comparing the rainfall data DA with the rainfall trend information. The rainfall pattern determination unit 36 ​​outputs information on the selected one rainfall pattern RPi to the teacher data creation unit 34A.

[0131] The teacher data creation unit 34A creates a plurality of teacher data sets corresponding to the plurality of rainfall patterns RPi, respectively. The teacher data creation unit 34A creates a plurality of teacher data sets including one each of learning time difference information ΔTL, learning rainfall information RL, and learning water level difference information ΔHL. Then, based on the rainfall pattern RPi (i is any integer) determined by the rainfall pattern determination unit 36 ​​using information corresponding to each teacher data set, each teacher data set can be classified into a plurality of rainfall patterns RPi.

[0132] The model learning unit 35A uses a plurality of teacher data sets classified for each of the plurality of rainfall patterns RPi created by the teacher data creation unit 34A to learn a plurality of prediction models 41, 42, ... 4i to generate a plurality of trained models 41A, 4A2, ... 4Ai (see FIG. 11) corresponding to the plurality of rainfall patterns RPi. The plurality of trained models 4A1, 4A2, ... 4Ai are water level prediction models that are adapted to the plurality of rainfall patterns RPi that are preset in the rainfall pattern determination unit 36. In other words, the number of types i of the trained models 4A1 to 4Ai is the same as the number of classifications i of the rainfall patterns RPi determined in the rainfall pattern determination unit 36.

[0133] In addition, the model generation device 3A according to the second embodiment can also perform the same process as the model generation process of the first embodiment described with reference to Figs. 6 and 7 to generate multiple trained models 41A, 4A2, ..., 4Ai. However, at the timing after step S109 in Fig. 6 is performed, the rainfall pattern determination unit 36 ​​performs a process of determining the rainfall pattern RPi related to the time period of the learning time difference information ΔTL calculated in step S109. In step S111, the teacher data creation unit 34A performs a process of creating multiple teacher data sets corresponding to the determined multiple rainfall patterns RPi. In step S112, the model learning unit 35A uses the multiple teacher data sets classified for the multiple rainfall patterns RPi in step S111 to learn the multiple prediction models 41, 42, ..., 4i, and generates the multiple trained models 41A, 4A2, ..., 4Ai corresponding to the multiple rainfall patterns RPi.

[0134] 11 is a functional block diagram of a water level prediction device 5A according to the second embodiment. As shown in FIG.

[0135] The rainfall pattern determination unit 60 determines the rainfall pattern RPi (i is any integer) related to the time period of the calculated prediction time difference information ΔT based on the predicted rainfall data DB stored in the predicted rainfall data storage unit 53. The types of multiple rainfall patterns stored in the rainfall pattern determination unit 60 are similar to those stored in the rainfall pattern determination unit 36 ​​of the model generating device 3A in FIG. 10. The rainfall pattern determination method can also be the same as that of the rainfall pattern determination unit 36. The rainfall pattern determination unit 60 outputs information on the selected one rainfall pattern RPi to the prediction unit 56A.

[0136] The prediction unit 56A selects one of the multiple trained models 4A1, 4A2, ... 4Ai according to the desired rainfall pattern RPi for prediction, and uses it to predict the water level difference ΔL. As a result, the prediction unit 56A selects a model that is optimal for the rainfall pattern in the time period up to the prediction time Te from the multiple water level prediction models that are compatible with the multiple rainfall patterns RPi that are preset in the rainfall pattern determination unit 36, and uses it for water level prediction.

[0137] In addition, the water level prediction device 5A according to the second embodiment can also perform the same process as the water level prediction process of the first embodiment described with reference to FIG. 8 to predict the water level Le at the prediction time Te using a trained model according to the rainfall pattern RPi at the prediction time Te. However, in step S202 in FIG. 8, a process is performed in which a plurality of trained models 4A1, 4A2, . . . 4Ai generated in advance by the model generation device 3A are set, or a process is performed in which those acquired by other methods are set as the plurality of trained models 4A1, 4A2, . . . 4Ai. In addition, at a timing after the execution of step S204, the rainfall pattern determination unit 60 performs a process of determining one rainfall pattern RPi (i is any one of integers) related to the time period of the prediction time difference information ΔT calculated in step S204. In step S206, the prediction unit 56A selects one of the trained models 4A1, 4A2, . . . 4Ai according to the determined rainfall pattern RPi, and uses it to output the target variable.

[0138] The effects of the water level prediction system 1A according to the second embodiment will be described. In the model generation device 3A according to the second embodiment, the teacher data creation unit 34A creates multiple teacher data sets corresponding to multiple rainfall patterns RPi, and the model learning unit 35A generates multiple trained models 4A1-4Ai corresponding to the multiple rainfall patterns RPi using the multiple teacher data sets created by the teacher data creation unit 34A.

[0139] With this configuration, multiple trained models 4A1 to 4Ai corresponding to multiple rainfall patterns RPi can be generated, making it possible to generate a variety of water level prediction models that are suitable for various rainfall patterns, thereby generating models with higher prediction accuracy.

[0140] In addition, the water level prediction device 5A of the second embodiment has a plurality of trained models 4A1 to 4Ai corresponding to a plurality of rainfall patterns RPi, and the prediction unit 56A selects one of the plurality of trained models 4A1 to 4Ai according to the rainfall pattern RPi predicted for the time period up to the prediction time Te and uses it to predict the water level difference ΔL.

[0141] With this configuration, an appropriate model can be selected from the trained models 4A1 to 4Ai according to the predicted rainfall pattern RPi, making it possible to use different water level prediction models that are suitable for various rainfall patterns RPi, thereby further improving the accuracy of water level prediction.

[0142] Furthermore, the model generation device 3A of the second embodiment has configurations in common with the model generation device 3 of the first embodiment in that it utilizes a water detector 2, uses multiple regression analysis for the machine learning of the prediction models 41 to 4i, and detects three or more specified water levels at the same location among the specific locations P1 to Pn, and therefore can achieve the same effects as the first embodiment with respect to these configurations.

[0143] Furthermore, the water level prediction device 5A of the second embodiment has a common configuration with the water level prediction device 5 of the first embodiment in that it predicts water levels using trained models 4A1 to 4Ai generated using a water detector 2, and is equipped with a judgment unit 59 that judges a dangerous water level LVj based on a predicted water level Le, and therefore can achieve the same effects as the first embodiment with respect to these configurations.

[0144] <Modification of the second embodiment> The water level prediction system 1A according to the second embodiment includes a plurality of prediction models 41-4i and trained models 4A1-4Ai (i is an integer). In the water level prediction system 1A, the prediction models 41-4i for machine learning are selected based on the rainfall pattern RPi determined by the rainfall pattern determination unit 36 ​​of the model generating device 3A. Similarly, the trained models 4A1-4Ai used for water level prediction are selected based on the rainfall pattern RPi determined by the rainfall pattern determination unit 60 of the water level prediction device 5A. In other words, the water level prediction system 1A according to the second embodiment is characterized by a configuration in which a plurality of models are used selectively, and the rainfall pattern RPi is applied as information for selecting one model from the plurality of models.

[0145] Therefore, like the second embodiment, a configuration is adopted in which a plurality of models are used, while unlike the second embodiment, a configuration is possible in which information other than the rainfall pattern RPi is applied as information for selecting one model from the plurality of models. For example, a configuration may be adopted in which a model is used depending on the range of the water level to be predicted. For example, the teacher data used for machine learning of the prediction model is divided into a group below a predetermined water level and a group above the predetermined water level, and each is used for machine learning of a different prediction model. This makes it possible to obtain a trained model used when predicting a water level below a predetermined water level and a trained model used when predicting a water level above the predetermined water level. In this case, when predicting the water level, for example, information on the detected water level Hn at the time Tn at which the water level prediction is performed is obtained from the water detector 2, and one trained model that has been machine-learned using teacher data in a range including the obtained detected water level Hn can be selected and used for the water level prediction.

[0146] In the water level prediction system 1 of the first embodiment, the model generating device 3 and the water level prediction device 5 may also be provided with a function for determining the rainfall pattern RPi, similar to the rainfall pattern determining units 36 and 60 of the second embodiment. In this case, the rainfall pattern RPi can be added to the explanatory variables of the prediction model 4. This makes it possible to reflect the rainfall pattern RPi in the water level prediction even in the configuration of the first embodiment using a single prediction model 4, making it possible to obtain the same accuracy of water level prediction as in the second embodiment.

[0147] [Third embodiment] The third embodiment will be described with reference to FIGS.

[0148] Fig. 12 is a diagram showing the overall configuration of a water level prediction system 1B according to the third embodiment. As shown in Fig. 12, the overall configuration of the water level prediction system 1B according to the third embodiment is basically the same as that of the water level prediction system 1 according to the first embodiment described with reference to Fig. 1. However, there are differences in the model generating device 3B and the water level prediction device 5B. Another difference is that the output of the water level prediction device 5B is the dangerous water level LVj at the prediction time Te.

[0149] 13 is a functional block diagram of a model generating device 3B according to the third embodiment. As shown in FIG. 13, the model generating device 3B has a dangerous water level determining unit 37.

[0150] The dangerous water level determination unit 37 determines the dangerous water level LVj (j is an integer) in the time period in which the water level was detected, based on the information of the detected predetermined water levels H1, H2, H3 among the detection information stored in the detection information storage unit 32. The determination method of the dangerous water level determination unit 37 is similar to that of the determination unit 59 of the water level prediction devices 5, 5A shown in Figures 3 and 11. The dangerous water level determination unit 37 outputs information on the determined dangerous water level LVj to the teacher data creation unit 34B.

[0151] The teacher data creation unit 34B creates multiple teacher data sets including one each of learning time difference information ΔTL, learning rainfall information RL, and learning dangerous water level information ΔLVj. The learning dangerous water level information ΔLVj is input from the dangerous water level determination unit 37 and is information on the learning time difference information ΔTL and the dangerous water level LVj corresponding to the learning rainfall information RL.

[0152] The model learning unit 35B uses the teacher data set created by the teacher data creation unit 34B to learn the prediction model 4 and generate a trained model 4A. The model learning unit 35B uses the learning time difference information ΔTL and learning rainfall information RL of the teacher data as inputs to the prediction model 4 and the learning dangerous water level information ΔLVj as output of the prediction model 4, and performs machine learning so that the input / output relationship of the prediction model 4 approaches the input / output relationship of the teacher data set, thereby generating a trained model 4A.

[0153] In addition, the model generating device 3B according to the third embodiment can also generate a trained model 4A by performing processing similar to the model generating processing of the first embodiment described with reference to Fig. 6 and Fig. 7. However, in step S109 in Fig. 6, instead of the learning time difference information ΔTL, the dangerous water level determination unit 37 determines the dangerous water level LVj in the time period in which the water level was detected based on the information of the predetermined water levels H1, H2, and H3 in each data set of the detection information acquired in the processing of steps S101 to S108, and the teacher data creation unit 34B calculates the learning dangerous water level information ΔLVj. In step S111, the teacher data creation unit 34B performs processing to create a teacher data set using the learning dangerous water level information ΔLVj calculated in step S109 instead of the learning time difference information ΔTL. Then, in step S112, the model learning unit 35B uses the learning time difference information ΔTL and learning rainfall information RL from the teacher data created in step S111 as input to the prediction model 4, and uses the learning dangerous water level information ΔLVj as output of the prediction model 4, to learn the prediction model 4 and generate a learned model 4A.

[0154] Fig. 14 is a functional block diagram of a water level prediction device 5B according to the third embodiment. As shown in Fig. 14, the prediction unit 56B inputs explanatory variables including prediction time difference information ΔT and prediction rainfall information ΔR to the trained model 4A, and causes the trained model 4A to output a response variable related to the input explanatory variables. The prediction unit 56B predicts a dangerous water level LVj at a prediction time Te based on the response variable output by the trained model 4A. Each time the prediction unit 56B predicts a new dangerous water level LVj at a prediction time Te, the prediction unit 56B adds and stores this information in the prediction result table 58 to update it.

[0155] The water level prediction device 5B according to the third embodiment can also predict the dangerous water level LVj at the prediction time Te by performing a process similar to the water level prediction process of the first embodiment described with reference to Fig. 8. In step S202 in Fig. 8, a trained model 4A previously generated by the model generating device 3B is set, or a trained model acquired by another method is set as the trained model 4A. In addition, instead of steps S207, S208, and S209, a prediction unit 56B performs a process of calculating the dangerous water level LVj at the prediction time Te based on the objective variable output from the trained model 4A in step S206.

[0156] The effect of the water level prediction system 1B according to the third embodiment will be described. In the model generating device 3 according to the third embodiment, the dangerous water level LVj (learning dangerous water level information ΔLVj) is used as a learning objective variable used by the model learning unit 35B in the machine learning of the prediction model 4. Also, in the water level prediction device 5B according to the third embodiment, information on the dangerous water level LVj at the prediction time Te is calculated as an objective variable output from the trained model 4A by the prediction unit 56B.

[0157] If the future water level Le in the specific locations P1 to Pn to be predicted can be predicted, it is possible to determine whether the water level has reached the dangerous water level LVj based on the predicted water level Le. If such a determination can be made as quickly as possible, it is advantageous in taking measures based on the water level prediction. For example, if the specific location is a road or a residential area and flooding due to a rise in water level occurs, it is considered necessary to take measures such as evacuation, flood prevention work, and traffic restrictions as quickly as possible. Therefore, if the configuration is such that the dangerous water level LVj is directly output from the trained model 4A as in the third embodiment, it is possible to more quickly determine whether the predicted water level at the desired prediction time Te is dangerous, which is effective in taking measures based on the water level prediction.

[0158] Furthermore, the model generation device 3B according to the third embodiment has a common configuration with the model generation device 3 according to the first embodiment in that it performs machine learning of the prediction model 4 using detection information from the water detector 2 to generate a trained model 4A. Furthermore, the objective variable of the teacher data used in the machine learning of the prediction model 4 is information related to water level, called training dangerous water level information ΔLVj. Therefore, similar to the first embodiment, it is possible to achieve the effect of enabling highly accurate water level prediction using a water detector 2 having a simpler configuration than a water level meter.

[0159] In addition, the model generation device 3B of the third embodiment has a configuration in common with the model generation device 3 of the first embodiment in that it uses multiple regression analysis for the machine learning of the prediction model 4 and detects three or more specified water levels at the same location among the specific locations P1 to Pn, and therefore can achieve the same effects as the first embodiment with respect to these configurations as well.

[0160] Moreover, the water level prediction device 5B according to the third embodiment has a common configuration with the water level prediction device 5 of the first embodiment in that it predicts information related to the water level, called the dangerous water level LVj, by using a trained model 4A generated by using a water detector 2. Therefore, like the first embodiment, it is possible to achieve the effect of enabling highly accurate water level prediction by using a water detector 2 having a simpler configuration compared to a water level meter.

[0161] In the water level prediction system 1B according to the third embodiment, the model generating device 3B and the water level prediction device 5B may be provided with a function for determining the rainfall pattern RPi, similar to the rainfall pattern determining units 36 and 60 in the second embodiment. In this case, the model generating device 3B can generate a plurality of trained models 4A1-4Ai according to a plurality of rainfall patterns RPi, and the water level prediction device 5B can select one model from the plurality of trained models 4A1-4Ai according to the rainfall pattern at the prediction time Te, and predict the dangerous water level LVj using the selected model. With this configuration, as in the second embodiment, more accurate predictions that are compatible with various rainfall patterns RPi are possible.

[0162] The present embodiment has been described above with reference to specific examples. However, the present disclosure is not limited to these specific examples. Any design modifications made by a person skilled in the art to these specific examples are also included within the scope of the present disclosure as long as they have the features of the present disclosure. The elements of each of the above-mentioned specific examples and their arrangements, conditions, shapes, etc. are not limited to those exemplified and can be changed as appropriate. The combination of each of the elements of each of the above-mentioned specific examples can be changed as appropriate as long as no technical contradiction occurs.

[0163] In the above embodiment, the time difference information and rainfall information (which may also include the rainfall pattern RPi in the modified example of the first embodiment) are used as input information (explanatory variables) for the prediction model 4 and the trained model 4A, but other information that may improve the estimation accuracy of the water level prediction may be added. Examples of such information include information related to the season, information related to the latest temperature and humidity, and weather warnings. [Explanation of symbols]

[0164] 1, 1A, 1B Water level prediction system 2. Water detector 3, 3A, 3B Model Generator 4, 41-4i Prediction model 4A, 4A1~4Ai trained models 5, 5A, 5B Water level prediction device 34, 34A, 34B Teacher Data Creation Department 35, 35A, 35B Model Learning Section 56, 56A, 56B Prediction Section 57 Predicted Water Level Calculation Section 59 Judgment section ΔHL Water level difference information for learning ΔTL Learning time difference information RL Rainfall information for learning RPi Rainfall Pattern ΔT Time difference information for prediction ΔR Prediction Rainfall Information Le predicted water level ΔL Predicted water level difference LVj Danger water level ΔLVj Dangerous water level information for learning P1~Pn Specific locations

Claims

1. A model generating device for generating a model for predicting a water level at a specific location, comprising: a teacher data creation unit that creates a teacher data set including learning time difference information relating to the time difference between the times when a change in the presence or absence of water at a plurality of predetermined water levels is detected by a water detector that is installed at the specific location and detects a change in the presence or absence of water at the predetermined water level, learning rainfall information relating to the amount of rainfall at the specific location when the change in the presence or absence of water at the predetermined water level is detected, and learning water level difference information relating to the water level difference between the plurality of predetermined water levels; a model learning unit that performs machine learning using the teacher data set created by the teacher data creation unit, with the learning time difference information and the learning rainfall information as inputs to the model and the learning water level difference information as an output of the model, so that the input / output relationship of the model approaches the input / output relationship of the teacher data set, thereby generating a trained model; Equipped with Model generation device.

2. The teacher data creation unit creates a plurality of teacher data sets corresponding to a plurality of rainfall patterns, The model learning unit generates a plurality of trained models corresponding to the plurality of rainfall patterns by using the plurality of teacher data sets. The model generating device according to claim 1 .

3. The teacher data creation unit creates a plurality of teacher data sets corresponding to a plurality of water level ranges, The model learning unit generates a plurality of trained models corresponding to the plurality of water level ranges using the plurality of teacher data sets. The model generating device according to claim 1 .

4. A multiple regression analysis is used for the machine learning of the model learning unit. The model generating device according to any one of claims 1 to 3.

5. The plurality of predetermined water levels include three or more types of the predetermined water levels. The model generating device according to any one of claims 1 to 3.

6. A water level prediction device for predicting a water level at a specific location, a trained model that acquires, by machine learning, a correspondence between learning explanatory variables including learning time difference information relating to the time difference between the times when a change in the presence or absence of water at a plurality of predetermined water levels is detected by a water detector that is installed at the specific location and detects a change in the presence or absence of water at the predetermined water level, and learning rainfall information relating to the rainfall at the specific location when the change in the presence or absence of water at the predetermined water level is detected, and learning objective variables including learning water level difference information relating to the water level difference at the plurality of predetermined water levels; a prediction unit that inputs explanatory variables including prediction time difference information relating to the time difference from an implementation time to a prediction time at which the water level is to be predicted, and prediction rainfall information relating to a predicted rainfall for the period from the implementation time to the prediction time, into the trained model, outputs information on a response variable related to the information on the explanatory variables, and predicts the water level difference for the period from the implementation time to the prediction time based on the response variable; a predicted water level calculation unit that calculates a water level at the predicted time based on the water level difference predicted by the prediction unit; Equipped with Water level prediction device.

7. The trained model has a plurality of trained models corresponding to a plurality of rainfall patterns, The prediction unit selects one of the plurality of trained models according to a rainfall pattern predicted for a time period up to the prediction time, and uses the selected one for predicting the water level difference. The water level prediction device according to claim 6.

8. The trained model has a plurality of trained models corresponding to a plurality of water level ranges, The prediction unit selects one of the plurality of trained models according to the water level at the implementation time and uses the selected model to predict the water level difference. The water level prediction device according to claim 6.

9. a determination unit that determines whether the water level is at a dangerous water level based on information about the water level at the predicted time calculated by the predicted water level calculation unit, The water level prediction device according to any one of claims 6 to 8.

10. 1. A method for generating a model for predicting a water level at a specific location, comprising the steps of: a training data creation step for creating a training data set including learning time difference information relating to the time difference between the times when a change in the presence or absence of water at a plurality of predetermined water levels is detected by a water detector installed at the specific location and configured to detect a change in the presence or absence of water at the predetermined water level, learning rainfall information relating to the amount of rainfall at the specific location when the change in the presence or absence of water at the predetermined water level is detected, and learning water level difference information relating to the water level difference at the plurality of predetermined water levels; a model learning step of performing machine learning using the teacher data set created in the teacher data creation step, with the learning time difference information and the learning rainfall information as inputs of the model and the learning water level difference information as an output of the model, so that the input / output relationship of the model approaches the input / output relationship of the teacher data set, thereby generating a trained model; Including, Model generation method.

11. A model generation program for generating a model for predicting a water level at a specific location, comprising: a training data creation function for creating a training data set including learning time difference information relating to the time difference between the times when a change in the presence or absence of water at a plurality of predetermined water levels is detected by a water detector installed at the specific location and detecting a change in the presence or absence of water at the predetermined water level, learning rainfall information relating to the amount of rainfall at the specific location when a change in the presence or absence of water at the predetermined water level is detected, and learning water level difference information relating to the water level difference between the plurality of predetermined water levels; a model learning function that uses the teacher data set created by the teacher data creation function to perform machine learning using the learning time difference information and the learning rainfall information as inputs to the model and the learning water level difference information as an output of the model so that the input / output relationship of the model approaches the input / output relationship of the teacher data set, thereby generating a trained model; to cause a computer to execute Model generator.

12. A water level prediction method for predicting a water level at a specific location, comprising: a model setting step of setting a trained model that has been acquired by machine learning a correspondence between learning explanatory variables including learning time difference information relating to the time difference between the times when a change in the presence or absence of water at a plurality of predetermined water levels is detected by a water detector that is installed at the specific location and detects a change in the presence or absence of water at the predetermined water level, and learning rainfall information relating to the rainfall at the specific location when the change in the presence or absence of water at the predetermined water level is detected, and learning objective variables including learning water level difference information relating to the water level difference at the plurality of predetermined water levels; a prediction step of inputting explanatory variables including prediction time difference information relating to the time difference from the implementation time to the prediction time at which the water level is to be predicted and prediction rainfall information relating to the predicted rainfall for the period from the implementation time to the prediction time into the trained model set in the model setting step, outputting information on a response variable related to the information on the explanatory variables, and predicting the water level difference for the period from the implementation time to the prediction time based on the response variable; a predicted water level calculation step of calculating a water level at the predicted time based on the water level difference predicted in the prediction step; Including, Water level prediction methods.

13. A water level prediction program for predicting a water level at a specific location, a model setting function for setting a trained model that has been acquired by machine learning a correspondence between learning explanatory variables including learning time difference information relating to the time difference between the times when a change in the presence or absence of water at a plurality of predetermined water levels is detected by a water detector that is installed at the specific location and detects a change in the presence or absence of water at the predetermined water level, and learning rainfall information relating to the rainfall at the specific location when the change in the presence or absence of water at the predetermined water level is detected, and learning objective variables including learning water level difference information relating to the water level difference at the plurality of predetermined water levels; a prediction function for inputting explanatory variables including prediction time difference information relating to the time difference from the implementation time to the prediction time at which the water level is to be predicted and prediction rainfall information relating to the predicted rainfall for the period from the implementation time to the prediction time into the trained model set by the model setting function, outputting information on a response variable related to the information on the explanatory variables, and predicting the water level difference for the period from the implementation time to the prediction time based on the response variable; a predicted water level calculation function for calculating a water level at the predicted time based on the water level difference predicted by the prediction function; to cause a computer to execute Water level prediction program.

14. A water level prediction system for predicting a water level at a specific location, comprising: A water detector that is installed at the specific location and detects a change in the presence or absence of water at a predetermined water level; a model generating device that generates a model for predicting the water level at the specific location based on a detection result regarding a change in the presence or absence of water at a plurality of predetermined water levels detected by the water detector; A water level prediction device that predicts a water level at the specific location using the trained model generated by the model generation device; Equipped with The model generating device comprises: a teacher data creation unit that creates a teacher data set including learning time difference information relating to the time difference between the times when the water detector, which is installed at the specific location and detects a change in the presence or absence of water at the specific water level, detects each of the changes in the presence or absence of water at the multiple specific water levels, learning rainfall information relating to the rainfall at the specific location when the change in the presence or absence of water at the specific water level is detected, and learning water level difference information relating to the water level difference at the multiple specific water levels; a model learning unit that performs machine learning using the teacher data set created by the teacher data creation unit, with the learning time difference information and the learning rainfall information as inputs to the model and the learning water level difference information as an output of the model, so that the input / output relationship of the model approaches the input / output relationship of the teacher data set, thereby generating the trained model; Equipped with The water level prediction device includes: the trained model is a model in which a correspondence between explanatory variables for learning including the time difference information for learning relating to the time difference between the times when the change in the presence or absence of water at the multiple specified water levels is detected by the water detector that is installed at the specific location and detects the change in the presence or absence of water at the specified water level, and the rainfall information for learning relating to the rainfall at the specific location when the change in the presence or absence of water at the specified water level is detected, and a target variable for learning including the water level difference information for learning relating to the water level difference at the multiple specified water levels is obtained by machine learning; a prediction unit that inputs explanatory variables including prediction time difference information relating to the time difference from an implementation time to a prediction time at which the water level is to be predicted, and prediction rainfall information relating to a predicted rainfall for the period from the implementation time to the prediction time, into the trained model, outputs information on a response variable related to the information on the explanatory variables, and predicts the water level difference for the period from the implementation time to the prediction time based on the response variable; a predicted water level calculation unit that calculates a water level at the predicted time based on the water level difference predicted by the prediction unit; Equipped with Water level prediction system.

15. A trained model in which the correspondence between learning explanatory variables including learning time difference information regarding the time difference between the times when a change in the presence or absence of water at multiple specified water levels is detected by a water detector installed at a specific location for predicting water levels and detecting a change in the presence or absence of water at the specified water level, and learning rainfall information regarding the rainfall at the specific location when a change in the presence or absence of water at the specified water level is detected, and learning objective variables including learning water level difference information regarding the water level difference between the multiple specified water levels.

16. A model generating device for generating a model for predicting a water level at a specific location, comprising: a teacher data creation unit that creates a teacher data set including learning time difference information relating to the time difference between the times when a change in the presence or absence of water at a plurality of specified water levels is detected by a water detector that is installed at the specific location and detects a change in the presence or absence of water at the specified water level, learning rainfall information relating to the amount of rainfall at the specific location when the change in the presence or absence of water at the specified water level is detected, and learning dangerous water level information relating to the dangerous water level during the time period related to the learning time difference information; a model learning unit that performs machine learning using the teacher data set created by the teacher data creation unit, with the learning time difference information and the learning rainfall information as inputs to the model and the learning dangerous water level information as an output of the model, so that the input / output relationship of the model approaches the input / output relationship of the teacher data set, thereby generating a trained model; Equipped with Model generation device.

17. A water level prediction device for predicting a water level at a specific location, a trained model that acquires, by machine learning, a correspondence between learning explanatory variables including learning time difference information relating to the time difference between the times when a change in the presence or absence of water at a plurality of specified water levels is detected by a water detector that is installed at the specific location and detects a change in the presence or absence of water at the specified water level, and learning rainfall information relating to the rainfall at the specific location when the change in the presence or absence of water at the specified water level is detected, and learning objective variables including learning dangerous water level information relating to the dangerous water level in the time period related to the learning time difference information; a prediction unit that inputs explanatory variables including prediction time difference information relating to the time difference from the implementation time of the prediction to the prediction time at which the water level is to be predicted and prediction rainfall information relating to the predicted rainfall for the period from the implementation time to the prediction time into the trained model, outputs information on a response variable related to the information on the explanatory variables, and predicts the dangerous water level at the prediction time based on the response variable; Equipped with Water level prediction device.

18. 1. A method for generating a model for predicting a water level at a specific location, comprising the steps of: a training data creation step for creating a training data set including learning time difference information relating to the time difference between the times when a change in the presence or absence of water at a plurality of specified water levels is detected by a water detector installed at the specific location and configured to detect a change in the presence or absence of water at the specified water level, learning rainfall information relating to the amount of rainfall at the specific location when the change in the presence or absence of water at the specified water level is detected, and learning dangerous water level information relating to the dangerous water level during the time period related to the learning time difference information; a model learning step of performing machine learning using the teacher data set created in the teacher data creation step, with the learning time difference information and the learning rainfall information as inputs to the model and the learning dangerous water level information as an output of the model, so that the input / output relationship of the model approaches the input / output relationship of the teacher data set, thereby generating a trained model; Including, Model generation method.

19. A model generation program for generating a model for predicting a water level at a specific location, comprising: a training data creation function for creating a training data set including learning time difference information relating to the time difference between the times when a change in the presence or absence of water at a plurality of specified water levels is detected by a water detector installed at the specific location and detecting a change in the presence or absence of water at the specified water level, learning rainfall information relating to the amount of rainfall at the specific location when the change in the presence or absence of water at the specified water level is detected, and learning dangerous water level information relating to the dangerous water level during the time period related to the learning time difference information; a model learning function that uses the teacher data set created by the teacher data creation function to perform machine learning using the learning time difference information and the learning rainfall information as inputs to the model and the learning dangerous water level information as an output of the model so that the input / output relationship of the model approaches the input / output relationship of the teacher data set, thereby generating a trained model; to cause a computer to execute Model generator.

20. A water level prediction method for predicting a water level at a specific location, comprising: a model setting step of setting a trained model that has been acquired by machine learning a correspondence between learning explanatory variables including learning time difference information relating to the time difference between the times when a change in the presence or absence of water at a plurality of specified water levels is detected by a water detector that is installed at the specific location and detects a change in the presence or absence of water at the specified water level, and learning rainfall information relating to the rainfall at the specific location when the change in the presence or absence of water at the specified water level is detected, and learning objective variables including learning danger water level information relating to the danger water level in the time period related to the learning time difference information; a prediction step of inputting explanatory variables including prediction time difference information relating to the time difference from the implementation time of the prediction to the prediction time at which the water level is to be predicted and prediction rainfall information relating to the predicted rainfall for the period from the implementation time to the prediction time into the trained model set in the model setting step, outputting information on a response variable related to the information on the explanatory variables, and predicting the dangerous water level at the prediction time based on the response variable; Including, Water level prediction methods.

21. A water level prediction program for predicting a water level at a specific location, a model setting function for setting a trained model that has been acquired by machine learning a correspondence between learning explanatory variables including learning time difference information relating to the time difference between the times when a change in the presence or absence of water at a plurality of specified water levels is detected by a water detector that is installed at the specific location and detects a change in the presence or absence of water at the specified water level, and learning rainfall information relating to the rainfall at the specific location when the change in the presence or absence of water at the specified water level is detected, and learning objective variables including learning danger water level information relating to the danger water level in the time period related to the learning time difference information; a prediction function for inputting explanatory variables including prediction time difference information relating to the time difference from the implementation time of the prediction to the prediction time at which the water level is to be predicted and prediction rainfall information relating to the predicted rainfall for the period from the implementation time to the prediction time into the trained model set by the model setting function, outputting information on a response variable related to the information on the explanatory variables, and predicting the dangerous water level at the prediction time based on the response variable; to cause a computer to execute Water level prediction program.

22. A trained model in which the correspondence between learning explanatory variables including learning time difference information regarding the time difference between the times when changes in the presence or absence of water at multiple specified water levels are detected by a water detector installed at a specific location for predicting water levels and detecting changes in the presence or absence of water at the specified water level, and learning rainfall information regarding the rainfall at the specific location when a change in the presence or absence of water at the specified water level is detected, and learning objective variables including learning danger water level information regarding the danger water level in the time period related to the learning time difference information.

Images