Realtime water level prediction device, realtime water level prediction method, and computer program

Abstract

To provide a realtime water level prediction device that provides highly accurate water level prediction information in real time, not only for water level measurement points, but also for unmeasured points.SOLUTION: The device according to an embodiment is configured to: divide a watershed into a plurality of sub-watersheds by a small drainage section including a water level measurement point on the basis of the connection relationship of a plurality of culverts in a sewer culvert network; calculate time-series data of water level prediction values at a water level measurement point by using a trained machine learning model; use, as a boundary condition of water levels in the plurality of sub-watersheds, time-series data of water level prediction values at the water level measurement point in a plurality of sub-watershed analysis models derived by dividing an outflow analysis model so as to calculate a water level prediction value at a water level unmeasured point included in each of the plurality of sub-watersheds; and input time-series data of rainfall measurement values and rainfall prediction values to a plurality of sub-outflow analysis models, so as to calculate the water level prediction value at the plurality of water level unmeasured points.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] An embodiment of the present invention relates to a real-time water level prediction device, a real-time water level prediction method, and a computer program. [Background technology]

[0002] In recent years, there have been many cases of heavy rainfall that falls locally and in a short period of time (localized torrential rain). Information released by the Japan Meteorological Agency and other organizations shows that the frequency and variability of extreme events such as torrential rainfall is increasing statistically. A typical type of damage caused by localized torrential rain is the frequent occurrence of inland flooding, where water overflows within cities.

[0003] To date, governments have taken measures to prevent flooding before it happens, such as building embankments, dredging river channels, improving revetments and constructing dams, mainly assuming flooding caused by the rise and collapse of large rivers. River flooding is called external flooding, and while measures to prevent external flooding have traditionally been focused on, it is thought that in the future measures that also take into account internal flooding will become important. In fact, when looking at the cost of flood damage (including external and internal flooding), internal flooding accounts for roughly half of the damage nationwide, but over 90% in Tokyo, meaning that internal flooding has become a new issue in urban areas where levees are relatively well developed.

[0004] There are many cases where there is a trade-off between inland flooding and external flooding. In other words, if storm water drainage pumps or pumps at storm water pumping stations attached to sewage treatment plants are used to discharge water into rivers in order to prevent inland flooding, the water level of the river will rise and the risk of external flooding will increase. On the other hand, if the start of the pumps is delayed in order to prevent external flooding, the risk of inland flooding will increase. As mentioned above, there is an essential trade-off between inland flooding and external flooding, so in order to fundamentally solve the problem, hard measures such as strengthening levees and setting up storage facilities according to the amount of rainfall are essential, and as a countermeasure, the construction of storm water storage facilities is being promoted.

[0005] On the other hand, in recent years, heavy rainfall of 70 to 100 mm / h or more, which greatly exceeds the planned rainfall (50 to 60 mm / h) based on the 5 to 10 year probability rainfall, has been observed relatively frequently, and it is often difficult in terms of time and financial resources to implement all of these hard measures. Therefore, in parallel with hard measures, it is necessary to quickly implement soft measures such as supporting self-help efforts to avoid flood damage by providing rainfall information and flood risk information, and efficient operation of storage facilities and drainage pumps. The main purpose of both hard and soft measures is to completely avoid flooding due to rainfall below the planned rainfall, and at the same time, to avoid and reduce flood damage due to flooding, for example by avoiding flooding above floor level, due to rainfall exceeding the planned rainfall. The Ministry of Land, Infrastructure, Transport and Tourism is promoting flood prevention and flood control measures that integrate such hard and soft measures. In light of the above, in the sewerage sector, which is responsible for responding to inland flooding, the concept of water level notification sewerage has been established in the Flood Control Act, and in urban areas with underground shopping malls, the installation and spread of main line water level gauges is progressing.

[0006] Furthermore, distributed runoff analysis methods (or distributed runoff analysis models) have been widely used in the past as a technique for comprehensively evaluating the risk of flooding throughout an urban area. Distributed runoff analysis is a method for tracking the flow of rainfall using a runoff analysis model that appropriately combines hydrological and hydraulic models (such as the Saint-Venant equations) using topographical information such as the land use pattern and elevation of an area, and civil engineering information such as the layout of sewer pipes. General-purpose analysis software for distributed runoff analysis is also widely used, and commercial packages of distributed runoff analysis software for representative models are also in use.

[0007] Traditionally, these distributed runoff analysis software programs have mainly been used offline to design facilities to prevent flooding and inundation, and were not primarily intended for the operational control of storm water pumps and storage facilities. However, in recent years, efforts have been made to use this software online for real-time hazard maps and flood prevention measures.

[0008] When considering the use of online distributed runoff analysis methods (or distributed runoff analysis models), it is necessary to input rainfall information in real time, and a method has been proposed in which a distributed runoff analysis model is used to input rainfall radar (precipitation radar) for each of several areas divided into a mesh pattern to make flood predictions for river channels.

[0009] On the other hand, distributed runoff analysis has a problem in terms of real-time processing. That is, it is difficult to process in real time due to the high calculation load, and it has been pointed out that real-time processing requires enormous computer resources.

[0010] In addition, methods derived from distributed runoff analysis models have been proposed that, separate from runoff analysis and flood prediction methods, involve data-driven estimation or prediction of water levels from actual measurement data of storm water inflow into sewage treatment plants and storm water drainage pumping stations, and water levels within sewer pipe networks.

[0011] For example, a method has been proposed to predict the inflow volume to a specific storm water pumping station or treatment plant, or the water level in a sewer pipe, using a data-driven statistical model or machine learning model (hereafter referred to as an AI model) based on past data. Methods using AI models have the advantage that they do not require physical or civil engineering information on the sewer pipe network, and models can be constructed through learning as long as data is available, and because they directly use past data, they generally tend to have high prediction accuracy. [Prior art documents] [Patent documents]

[0012] [Patent Document 1] JP 2019-159506 A [Patent Document 2] JP 2017-194344 A Summary of the Invention [Problem to be solved by the invention]

[0013] As mentioned above, distributed runoff analysis is difficult to calculate in real time, and the results of distributed runoff analysis tend to be less accurate than those of machine learning. In addition, prediction methods using AI models can only predict at measurement points where water levels and inflows are measured, and are not necessarily suitable for planar or spatial predictions.

[0014] The embodiments of the present invention have been made in consideration of the above-mentioned circumstances, and aim to provide a real-time water level prediction device, a real-time water level prediction method, and a computer program that provide accurate water level prediction information in real time not only for water level measured points but also for unmeasured points. [Means for solving the problem]

[0015] A real-time water level prediction device according to an embodiment is a real-time water level prediction device that calculates water level prediction values ​​for multiple water level unmeasured points for a watershed in which water level measurement values ​​at each of multiple water level measurement points in a sewer pipe network including multiple pipes and rainfall measurement values ​​at each of multiple rainfall measurement points are measured at a predetermined period, and includes a watershed division function unit that defines the watershed as multiple small drainage areas including each of the multiple pipes, and divides the watershed into multiple sub-watersheds by the small drainage areas including the water level measurement points based on the connection relationship of the multiple pipes in the sewer pipe network, and calculates water level prediction values ​​for the water level measurement points using time series data of water level measurement values ​​at the multiple water level measurement points and time series data of rainfall measurement values ​​at the multiple rainfall measurement points. The system comprises a water level prediction function unit that calculates time series data of the water level prediction value at the water level measurement point using a trained machine learning model that has been trained to calculate a water level prediction value, and an unmeasured water level water level prediction function unit that calculates in parallel time series data of the water level prediction value at the unmeasured water level points by using the time series data of the water level prediction value at the water level measurement point as a boundary condition for the water level in the multiple partial watersheds in multiple partial watershed analysis models obtained by dividing a runoff analysis model that performs runoff analysis of the watershed to calculate a water level prediction value at an unmeasured water level point included in each of the multiple partial watersheds, and inputting time series data of rainfall measurement values ​​and rainfall prediction values ​​at the multiple rainfall measurement points into a multiple partial runoff analysis model. [Brief description of the drawings]

[0016] [Figure 1] FIG. 1 is a block diagram illustrating an example of the configuration of a real-time water level prediction device according to an embodiment. [Diagram 2] FIG. 2 is a diagram illustrating an example of a procedure for constructing a distributed runoff analysis model used in the real-time water level prediction device of one embodiment. [Diagram 3] FIG. 3 is a diagram showing an example of a table based on information extracted from a flow rate calculation table. [Figure 4] FIG. 4 is a diagram showing an example of the connection relationship between multiple small drainage districts in a graph representation. [Diagram 5] FIG. 5 is a diagram showing an example of a table summarizing the correspondence between upstream nodes and downstream nodes in a plurality of small drainage districts. [Figure 6] FIG. 6 is a diagram showing an example of a plurality of partial drainage basins divided by water level measurement points. [Figure 7] FIG. 7 is a simplified diagram of an example of a watershed divided into sub-watersheds. [Figure 8] FIG. 8 is a diagram showing an example of a screen that visualizes the flood risk assessment result and presents it to the user. [Figure 9] FIG. 9 is a diagram for explaining an example of a process for adjusting (identifying) parameters and dividing regions in the real-time water level prediction device of the first modified example. [Figure 10] FIG. 10 is a diagram for explaining an example of a process for adjusting (identifying) parameters and dividing regions in a real-time water level prediction device of the second modified example. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0017] Hereinafter, a real-time water level prediction device, a real-time water level prediction method, and a computer program according to embodiments will be described with reference to the drawings.

[0018] The real-time water level prediction device of the present embodiment includes, for example, a computing device having at least one processor and a memory in which a program executed by the processor is stored. The real-time water level prediction device can be configured to realize various functions described below by software or a combination of software and hardware.

[0019] FIG. 1 is a block diagram illustrating an example of the configuration of a real-time water level prediction device according to an embodiment. The real-time water level prediction device of this embodiment includes a runoff analysis model input unit 1, a water level measurement point information input unit 2, a watershed division function unit 3, a partial runoff analysis model input / output unit 4, an AI water level prediction model input unit 5, a rainfall information input unit 6, a rainfall prediction function unit 7, a water level measurement information input unit 8, a water level measurement point water level prediction function unit 9, a water level unmeasured point water level prediction function unit 10, and an in-watershed pipeline predicted water level output unit 11. In this embodiment, the function blocks shown in dotted lines in Fig. 1 indicate processes performed offline, and the function blocks shown in solid lines indicate processes performed online (real time).

[0020] First, an example of processing performed offline in the real-time water level prediction device of this embodiment will be described. In the runoff analysis model input section 1, a group of mathematical expressions for runoff analysis, which is constructed offline for the basin (drainage area) to be analyzed, is defined in advance, for example, by a method to be described later. The runoff analysis model defined in the runoff analysis model input section 1 may be a distributed runoff analysis model, for example, a runoff analysis model that is already packaged and commercially available, such as INFOWORKS, MOUSE, SWMM, or a runoff analysis model such as NILIM that is generally disclosed by research institutes of the Ministry of Land, Infrastructure, Transport and Tourism, or a group of runoff analysis models created by the user using an analysis model generally known in the fields of hydraulics and hydrology that expresses the flow in a sewer pipe using partial differential equations (such as the Saint-Venant equation) and combines it with surface runoff.

[0021] FIG. 2 is a diagram illustrating an example of a procedure for constructing a distributed runoff analysis model used in the real-time water level prediction device of one embodiment. In this embodiment, a runoff analysis model is used that can be constructed more simply than the various runoff analysis models described above and is easy to understand for practitioners involved in sewerage projects. This is because a simplified model can calculate (process) faster than a general distributed runoff analysis model, and is therefore superior in real-time performance. Note that this embodiment aims to speed up calculations by dividing the watershed to be analyzed into multiple parts at water level measurement points and performing parallel processing, so it is not necessarily necessary to be limited to adopting the simplified runoff analysis model described below, and it is also possible to use the general distributed runoff analysis model described above.

[0022] The runoff analysis model input unit 1 of the real-time water level prediction device of this embodiment may be configured to construct a runoff analysis model, or may be configured to define an externally constructed runoff analysis model in the runoff analysis model input unit 1.

[0023] The configuration for constructing the runoff analysis model (runoff analysis model construction function unit) includes a flow calculation table / sewerage ledger DB 21, a runoff / inundation analysis model parameter table generation unit 22, and a runoff / inundation analysis model automatic construction unit 23.

[0024] The flow rate calculation table / sewerage ledger DB21 records planning information and maintenance information related to a sewerage pipe network including a plurality of sewer pipes in a target basin for constructing a runoff analysis model. In this embodiment, the flow rate calculation table / sewerage ledger DB21 stores information used for sewerage planning, called flow rate calculation table, which is usually held by a sewerage operator / manager such as a local government, and information on the sewerage ledger used for sewerage maintenance. It is preferable to store the information on the flow rate calculation table and the sewerage ledger in the flow rate calculation table / sewerage ledger DB21 as electronic data.

[0025] Items defined in the flow rate calculation table include, for example, items (1)-(12) below. (1) Number of the sub-drainage area (or the pipeline corresponding to the sub-drainage area) (line number) (2) The number of the sub-drainage area (or the connected pipeline) connected to the sub-drainage area from the upstream side (inflow line number) (3) Storm water drainage area and sewage drainage area of ​​each drainage district (4) Runoff coefficient of each drainage area (5) Length and total length of connecting sewer pipes in each sub-drainage district (6) Flow time of each drainage area (7) Maximum discharge of each drainage area (8) The planned volume of rainwater, planned volume of sewage, and special drainage flowing into each drainage district, and the sum of these (9) Cross-sectional diameter, slope, design flow velocity and design discharge of the connecting sewer of each sub-drainage area (10) Elevation of pipelines included in the sewerage register used for the maintenance and management of sewerage systems (11) Manhole model numbers connected to each sewer line included in the sewerage register (12) The ground surface elevation of the small drainage district in which each sewer line included in the sewerage register is buried. In addition, items (10)-(12) are often not included in the flow calculation tables used in sewerage planning, and if they are not included in the flow calculation tables, they may be omitted.

[0026] The sub-drainage districts refer to the areas into which the target area for water level prediction is divided based on the sewerage plan. The flow calculation table has, for example, flow calculation records corresponding to each of the multiple sub-drainage districts, and each flow calculation record includes the above items (1)-(12), etc.

[0027] The line number is an identification number of the sub-drainage area. The line number may be an identification number of a pipeline connecting the sub-drainage areas. The inflow line number is the line number of the sub-drainage area connected upstream of the sub-drainage area indicated by the line number. The area is the drainage area of ​​the sub-drainage area (or pipeline) indicated by the line number. The sewage drainage area is the area of ​​the range in which storm water and sewage are collected in the sub-drainage area.

[0028] The runoff coefficient indicates the runoff coefficient of the small drainage area indicated by the line number. The runoff coefficient is the ratio of rainwater that flows down the ground surface to the amount of rainfall. The runoff time indicates the runoff time of the small drainage area indicated by the line number. The runoff time is expressed as the sum of the inflow time of rainwater from the farthest point in the small drainage area to flow over the ground surface and into the drainage channel, and the flow time required for the inflowing rainwater to flow through the drainage channel to reach a certain point.

[0029] The extension indicates the length of the pipeline connecting the sub-drainage area indicated by the line number and the upstream sub-drainage area indicated by the inflow line number. The designed flow velocity indicates the designed value of the speed of water flowing through the pipeline connecting the sub-drainage area indicated by the line number and the upstream sub-drainage area indicated by the inflow line number. The cross-sectional pipe diameter indicates the cross-sectional diameter of the pipeline (sewer) connecting the sub-drainage area indicated by the line number and the upstream sub-drainage area indicated by the inflow line number.

[0030] In addition, when the flow velocity is undefined in the flow rate calculation table, the gradient and roughness of the pipeline may be used instead of the flow velocity. In this case, the flow velocity can be defined by the gradient and roughness of the pipeline by using the Manning equation described later.

[0031] The runoff / inundation analysis model parameter table generating unit 22 extracts information (parameters) for defining the sewer pipe network required to construct the runoff / inundation analysis model from the flow calculation table, and creates a table that compiles the extracted information for each sub-drainage area. The runoff / inundation analysis model parameter table generating unit 22 may extract necessary information (parameters) from the information stored in the flow calculation table / sewer register DB 21, and automatically generate a table using the extracted information, or may generate the table manually (for example, according to a command to generate a table including selected information).

[0032] FIG. 3 is a diagram showing an example of a table based on information extracted from a flow rate calculation table. Each row of the table shown in FIG. 3 corresponds to a sub-drainage district (or a sewer connected to the sub-drainage district), and each column of the table corresponds to a parameter extracted from the flow rate calculation table and sewerage register DB21.

[0033] In the example shown in Fig. 3, the first and second columns of the table show the identification number and connection information of each sub-drainage area (or the sewer pipe corresponding to each sub-drainage area). This information allows the connection relationship of the pipes in the sub-drainage area to be known.

[0034] The third and fourth columns of the table are parameters used to convert rainfall information into inflow information. The fifth column of the table is parameters used to calculate the inflow time. The sixth and seventh columns of the table are parameters used to calculate the capacity of the connecting sewer of the small drainage section.

[0035] The eighth and ninth columns of the table are the two parameters, roughness and slope, which are required for the calculation when assuming that the average flow velocity formula called Manning's formula or Kutter's formula is used to calculate the flow velocity. The tenth column of the table is the parameter for the pipeline elevation.

[0036] The gradient in the 9th column of the table can be calculated from the pipeline length in the 6th column and the pipeline elevation in the 10th column, but it is preferable to also input the pipeline elevation in the 10th column of the table so that correction can be made when there is a step junction at the connection point of the pipeline. The elevation value may be fixed to either the value at the upstream end of the pipeline or the value at the downstream end of the pipeline. Moreover, the 11th and 12th columns of the table are parameters (manhole diameter and manhole height) for calculating the capacity of the manhole corresponding to each sub-drainage area.

[0037] After the parameters necessary for runoff / inundation analysis are set by the runoff / inundation analysis model parameter table generation unit 22, the runoff / inundation analysis model automatic construction unit 23 generates a runoff / inundation analysis model for each sub-drainage area based on the connection information of the sub-drainage areas.

[0038] In this embodiment, the runoff / inundation analysis model automatic construction unit 23 reads the connection relationships of multiple sub-drainage areas defined in the runoff / inundation analysis model parameter table generation unit 22, and based on the connection relationships of the multiple sub-drainage areas, generates and outputs a group of equations (runoff / inundation analysis model for each sub-drainage area) that represent the continuity equation (water volume balance) and the movement equation (water movement) using parameters such as the flow arrival time and flow down time defined in the runoff / inundation analysis model parameter table generation unit 22.

[0039] The automatic runoff / inundation analysis model construction unit 23 first regards multiple sub-drainage areas as nodes in graph theory based on the information in the first and second columns of the table shown in Figure 3, and graphically represents the connection relationships between the multiple sub-drainage areas.

[0040] FIG. 4 is a diagram showing an example of the connection relationship between multiple small drainage districts in a graph representation. In the graph shown in Figure 4, "1" to "8" each represent the line number of a sub-drainage area. Here, the graph shows how multiple sub-drainage areas "1" to "8" are connected so that the sub-drainage area with line number "1" is the final node. QA1 to QA8 are the rainfall amounts converted into inflows for each of the small drainage areas numbered "1" to "8."

[0041] FIG. 5 is a diagram showing an example of a table summarizing the correspondence between upstream nodes and downstream nodes in a plurality of small drainage districts. As a method for expressing the connection relationships between a plurality of sub-drainage areas in an easier-to-understand manner, the correspondence relationships between upstream nodes (sub-drainage areas) and downstream nodes (sub-drainage areas) may be summarized in a table, as shown in FIG.

[0042] In the example shown in Fig. 5, the nodes corresponding to the rows of the table are downstream, and the nodes corresponding to the columns are upstream, with the column for nodes in a connected relationship being marked with "1." Here, the following connections are shown as examples: a connection in which nodes are connected in the order of "1", "2", "3", "6", and "8" from the upstream side; a connection in which nodes are connected in the order of "4" and "3", a connection in which nodes are connected in the order of "5" and "6", and a connection in which nodes are connected in the order of "7" and "8".

[0043] By expressing in a matrix how each drainage basin is connected, the connection relationships between nodes can be summarized in a quantitative and easy-to-interpret format. Note that, because the essential connection information is defined in columns 1 and 2 of the table shown in Figure 3, when automatically constructing a runoff / inundation analysis model, it is not necessarily necessary to go through the representation format shown in Figure 5.

[0044] The runoff / inundation analysis model automatic construction unit 23 reads the connection information for each sub-drainage area from the information in the first and second columns of the table shown in Fig. 3, for example, and then applies the basic equation of the runoff / inundation analysis model to each sub-drainage area. Here, the basic equation used in the runoff / inundation analysis model automatic construction unit 23 is based on an equation that expresses the balance of water volume, called the equation of continuity, and a runoff analysis model is constructed by combining this with an equation for rainfall inflow and an equation for flow movement (equation of movement). A typical runoff analysis model is, for example, the following equation (1). QA * (t+Tin * )=FA * ×A * ×RA * (t) (1)

[0045] Here, R.A. * is the rainfall input (rainfall intensity) for each sub-drainage area. As described later, it may be converted from MP radar rainfall intensity, or it may be ground rainfall data installed in each sub-drainage area. * is the area of ​​each drainage basin, as defined in the third column of the table in Figure 3. FA *is the average runoff coefficient of each drainage area, as defined in the fourth column of the table in Figure 3. * means the average inflow time for each drainage area, and is the value obtained by subtracting the flow time calculated by the Manning formula or the like from the flow arrival time in the fifth column of the table shown in Figure 3, or the flow time calculated by using the pipe length in the sixth column from the assumed average flow velocity. In other words, the average inflow time Tin * is expressed by the following equation (2). Tin * =Inflow time = Delivery time - Downflow time (pipe length / average flow velocity) (2) QA, which is the output of equation (1) * represents the inflow to each drainage area node.

[0046]

number

[0047] Equation (4) is a simplification of the so-called "equation of motion," where K is the inverse of the flow time, and the flow time can be calculated using the flow velocity (or a value equivalent to the flow velocity) and the pipeline length in the table shown in Figure 3. If the flow velocity is defined in a flow rate calculation table, equation (4) can be used as is, but generally, the flow velocity is calculated using an equation such as the Manning equation, and then K in equation (4) is found, or the flow rate Q is calculated directly using the Manning equation.* (t) is calculated. In the latter case, Q is calculated as follows: * (t) can be calculated.

[0048]

number

[0049] By using equations (5) to (8) instead of equation (4), it is possible to approximately simulate the phenomenon in which the flow velocity changes with the change in water level, and the flow time also changes accordingly. * and storage volume S * It should be noted that since there is a one-to-one correspondence between the two, calculating the storage volume and the water level are essentially the same thing. When applying the Manning's equation, in what is called Manning's mean velocity equation, the gradient I in equation (5) means the pipeline gradient, but in reality this gradient means the hydraulic gradient, so if we approximate the hydraulic gradient in some way, it becomes possible to simulate a phenomenon in which a rise in the downstream water level lengthens the flow time from upstream to downstream, ultimately resulting in backwater. One method for doing so is to replace the pipeline gradient I in equation (6) with the hydraulic gradient Id, and calculate Id using the following equation.

[0050]

number

[0051] By making at least a correction like equation (9), it becomes possible to simulate the phenomenon in which the hydraulic gradient becomes smaller and the flow velocity slows down when the water level downstream becomes higher than the water level upstream, which makes it possible to simulate the so-called backwater phenomenon in which the water level propagates from the downstream side to the upstream side.Note that equation (9) gives an approximation of the hydraulic gradient from the pipe gradient and the pipe water level, but when connecting sewer pipes with different pipe diameters, it is better to also take into account the step at the connection point, so equation (10) may be used instead of equation (9).

[0052]

number

[0053] In addition, a model for predicting flood risk can be created by using an equation equivalent to equation (9) expressed using a "full pipe rate" indicating the ratio of the storage volume to the capacity of the sewer pipe, instead of a "water level". This is because the storage volume and the water level can be converted into each other, and the essential concept is the same. For example, a flood risk can be warned by issuing an alarm indicating that the risk of flooding increases when the sewer pipe exceeds the full pipe. In reality, flooding only occurs when the storage volume exceeds the full pipe and then exceeds the capacity that can be stored in the manhole (when water overflows from the sewer pipe to the ground surface). Therefore, in order to make a more precise flood prediction, it is more preferable to also take into account the manhole capacity. For this reason, in this embodiment, the water level is used in equation (9).

[0054] However, directly considering a manhole as a structure would make the equations for runoff analysis complicated, so here we adopt a method that mimics an idea called the Priceman slot, which is used in calculations when the pipe is overfull in general distributed runoff analysis. The Priceman slot used in the numerical calculations for distributed runoff analysis is a pipe that becomes a pressure pipe rather than an open pipe when it is full, so to avoid the need to mix open pipe hydraulic calculations and pressure pipe hydraulic calculations in order to deal with this directly, we consider an open pipe with a thin slot at the top (a pipe with a slot at the top that is not a circular pipe when viewed in cross section but a flask-shaped pipe) and devised a way to perform calculations when the pipe is overfull using only the hydraulic calculations for the open pipe, and this is a method used in general runoff analysis.

[0055] The simplified runoff analysis model of the above equations (1) to (10) does not require solving partial differential equations (such as the Saint-Venant equation) as in distributed runoff analysis, and can perform simplified runoff analysis by simply solving the ordinary differential equation in equation (3). Therefore, the above-mentioned approach (to avoid mixing hydraulic calculations of open conduits and hydraulic calculations of pressure pipes) is not directly necessary, but when using equation (9) or (10), a virtual pipe similar to the Priceman slot is considered. In other words, instead of directly considering a structure called a manhole, a sewer pipe is assumed with a plate-shaped virtual slot on top of the sewer pipe that matches the capacity of the manhole, and equation (9) is used to calculate the water level of equation (10) in the assumed sewer pipe. This makes it possible to calculate the hydraulic gradient from when the sewer pipe exceeds the full capacity to when it exceeds the manhole capacity without changing the shape of equation (9) or (10).

[0056] In addition, since the storage volume is determined by giving the elevation of the ground surface, the elevation of the pipe top, and the length of the pipe, and then determining the width of the virtual slot, the width of the virtual slot that matches the manhole capacity can be uniquely determined. However, since the water level measured in the actual manhole does not necessarily match the water level in the virtual slot, if the water level at the water level measurement point described below exceeds the full pipe, it is necessary to calculate the storage volume in the manhole from the measured water level and then convert it to the water level in the virtual slot.

[0057] Furthermore, flooding occurs when the manhole capacity is exceeded (when water overflows from the sewer pipe to the ground surface), and in reality, the elevation of the ground surface at that time is exceeded. However, if flooding is determined based only on the manhole capacity, it is not necessary to use the virtual slot width as an adjustment parameter and make the elevation when the manhole capacity is exceeded match the elevation of the ground surface. In particular, since increasing the virtual slot width generally speeds up the calculation process, if there are issues with real-time performance, the slot width may be used as an adjustable parameter. However, unless there is a special reason, it is sufficient to determine the virtual slot width as described above so as to match the overall capacity.

[0058] Also, here, a method that mimics the idea of ​​the Priceman slot makes it possible to perform runoff analysis from when full pipe capacity is exceeded until manhole capacity is exceeded. However, in order to implement this more simply, it is also possible to directly use the idea of ​​full pipe rate and change the definition of full pipe capacity from "sewer pipe capacity" to "sewer pipe capacity + manhole capacity" and perform runoff analysis using a simple model that predicts flood risk.

[0059] The automatic runoff / inundation analysis model construction unit 23 has a function of generating a group of equations corresponding to each of the small drainage areas by connecting the above equations (1) to (10) according to the graph structure of FIG. 4 or the matrix of FIG. 5. The group of equations generated by the automatic runoff / inundation analysis model construction unit 23 is an example of a runoff analysis model input by the runoff analysis model input unit 1. If the above-mentioned simple runoff analysis model is used, a runoff analysis model can be automatically constructed by simply preparing a table such as that shown in FIG. 3, which is a table in which the information recorded in the flow rate calculation table and the sewerage register is slightly organized in advance, and there are advantages in that the engineering man-hours for model construction are reduced and faster processing is possible. The runoff analysis model used in the real-time water level prediction device of this embodiment is not limited to the above-mentioned simple runoff analysis model, and a so-called distributed runoff analysis model used in general runoff analysis may be constructed.

[0060] Information on water level measurement points is input to the water level measurement point information input unit 2. In this embodiment, the water level measurement point information input unit 2 inputs information on L water level measurement points in a target drainage basin.

[0061] FIG. 6 is a diagram showing an example of a plurality of partial drainage basins divided by water level measurement points. In FIG. 6, for example, the position indicated by the black circle is the water level measurement point. A water level gauge is installed at the water level measurement point in the pipeline, and the water level in the pipeline is measured. When actually inputting the data, in order to mechanically (automatically) process the subsequent steps, it is preferable to input the information on the water level measurement point by setting a flag such as "1" in the information on the pipe (or the corresponding small drainage area) in which the water level gauge is set in the table in FIG. 3. The input of the water level measurement point information is not limited to the above, and any form of input may be used as long as it is known in which pipe the water level gauge is installed.

[0062] The watershed division function unit 3 divides the watershed into a plurality of sub-watersheds based on the information of the water level measurement points. In this embodiment, the watershed division function unit 3 divides the runoff analysis model inputted by the runoff analysis model input unit 1 into K sub-watershed runoff analysis models based on the information obtained from the water level measurement point information input unit 2, by a method to be described later.

[0063] FIG. 7 is a simplified diagram of an example of a watershed divided into sub-watersheds. As mentioned above, a sewer pipe network can be represented as a graph with each sewer pipe as an edge (branch) and each connection point as a node (vertex) as shown in Figure 4, so a sewer pipe network corresponds to one graph in graph theory. On the other hand, since the locations of water level gauges correspond to pipes (edges), there is a possibility that the locations of water level gauges will be bridges in the graph (bridges: points at which, when cut, the graph becomes two unconnected graphs (graphs that have no path between them and cannot be traced)). Even if a single water level gauge installation pipe is not a bridge, the group of pipes (a collection of edges) at multiple water level gauge installation locations may become cut edges (edges in which the two vertices at both ends of the edge are elements of different graphs in two divided unconnected graphs).

[0064] For example, the two water level installation locations in the left diagram of Figure 7 are both bridges. Therefore, if several water level gauges are installed, it is expected that the entire sewer network graph will be separated into several disconnected graphs. Note that, depending on the number and locations of water level gauges, even if water level gauges are installed, it may not be possible to generate a disconnected graph; however, most sewer networks usually have a graph structure in which water does not form loops, merging from upstream to downstream and flowing into pumping stations and treatment plants, so even a single water level gauge installation location can usually separate the upstream and downstream parts, so it often forms a bridge. If an unconnected graph cannot be generated, it is preferable to deal with this by adjusting the installation locations or number of water level gauges so that an unconnected graph can be generated. However, if it is difficult to increase the number of water level gauges, it is acceptable to consider a coarse sewer pipe network graph that ignores small pipes less than a certain pipe diameter (for example, pipe diameter less than 80 cm) in the sewer pipe network and achieve unconnection (i.e., ignoring points where thin pipes are connected).

[0065] In the example shown in Figure 7, the two water level installation locations are both bridges, so the sewer pipe network graph is divided into three disconnected graphs. In addition, in the flow calculation table and sewer division plan, the small drainage areas are divided so that they correspond 1:1 to each sewer pipe, so the entire analysis basin can be divided into sub-basins by dividing the basin by the small drainage areas corresponding to each sewer pipe. In addition, in Figure 7, the image is shown of separating the small drainage areas within the entire basin, so the small drainage areas appear to correspond to the nodes (vertices) of the sewer pipe connection points, but in the actual flow calculation table and division plan, they correspond to the edges (branches) corresponding to the sewer pipes. Therefore, the pipe at the water level measurement point is defined overlapping in two sub-basins, but since water flows from upstream to downstream, the small drainage area corresponding to the pipe is assigned to the sub-basin where the water level measurement point pipe is on the upstream side.

[0066] In the example shown in Figure 7, a partial network graph of three disconnected sewer pipes is generated by the installation locations of two water level gauges, and three sub-basins can be defined by corresponding small drainage districts to this.

[0067] Here, the case where the above-mentioned simple runoff analysis model is used has been described, but when a distributed runoff analysis model is used, the sewer pipe network is expressed as a graph, so the basic idea of ​​division is the same, and a subgraph of a plurality of unconnected sewer pipe networks can be obtained. However, when a distributed runoff analysis model is applied, if the correspondence between the drainage basin and the small drainage area cannot be obtained, the drainage basin may be divided using a division method such as the Thiessen method, for example, to obtain correspondence between the partial drainage basin and the small drainage area. In addition, when radar rainfall, which will be described later, is input, the radar rainfall data provides rainfall data for each mesh of a predetermined size (e.g., 250m x 250m mesh, etc.), so the drainage basin division function unit 3 may divide the drainage basin for each mesh, regarding the pipe closest to each mesh as the area that the pipe is responsible for.

[0068] The watershed division function unit 3 matches a group of equations corresponding to each of the small drainage areas supplied from the runoff analysis model input unit 1 to each of multiple sub-watersheds based on information from water level measurement points, and outputs a runoff analysis model (partial runoff analysis model) for each sub-watershed.

[0069] The partial runoff analysis model input / output unit 4 holds the runoff analysis model for each sub-basin (partial runoff analysis model) output from the basin division function unit 3 so that it can be used in real-time calculation, and sets it as a model to be used in real-time calculation described later. The partial runoff analysis model input / output unit 4 holds, for example, runoff analysis models corresponding to three sub-basins shown in Fig. 7 as independent models. In this embodiment, the partial runoff analysis model input / output unit 4 has K input / output units 41-4K that are in a state in which each of the partial runoff analysis models for the K sub-basins divided by the basin division function unit 3 is input and the corresponding partial runoff analysis model can be output.

[0070] The AI ​​water level prediction model input unit 5 learns data-driven models such as machine learning models and time series models using water level data from water level measurement points and past rainfall data, and has the learned models as a set of formulas. The AI ​​water level prediction model input unit 5 can perform real-time calculations using the learned models. In this embodiment, the AI ​​water level prediction model input unit 5 builds an AI prediction model (a set of formulas for prediction) that can predict the predicted water level of L water level measurement points from rainfall data from M locations and water level data from L water level measurement points in advance using a method described below, and makes it available.

[0071] The model learned by the AI ​​water level prediction model input unit 5 may be any model as long as it is a so-called data-driven model. For example, the model learned by the AI ​​water level prediction model input unit 5 may be a model used in so-called classical time series analysis, such as an ARX model or an ARMAX model, or a transfer function model called a BJ (Box Jenkins) model that includes them. In addition, a method such as deep learning may be used, and when the amount of rainfall in an area divided into meshes is input as rainfall data from a weather radar or the like, a CNN-type deep learning model in which a CNN (convolutional neural network) often used in the image processing field is applied to two-dimensional data of mesh-shaped rainfall may be learned. Alternatively, it may be a recurrent neural network such as LSTM, which is said to be effective in time series information processing such as voice time series data analysis.

[0072] In addition, when using a transfer function model such as the BJ model, if the rainfall amount of a mesh-shaped area such as a weather radar is input as rainfall data, there is a high possibility that a problem of so-called multicollinearity will occur, rather than using an algorithm such as the usual least squares method, since there is a strong correlation between the input rainfall and the BJ model. For this reason, an algorithm that can perform regularization and reduction in dimension to deal with the problem of multicollinearity may be applied to the transfer function model. For example, an algorithm such as ridge regression, lasso regression, or elastic net regression may be used as a regularization method, and a method such as principal component regression (PCR) or partial least squares (PLS) may be used as a reduction method. In addition, an algorithm such as the auxiliary variable method or the prediction error method used in system identification in the field of control theory may be used. In any case, the AI ​​water level prediction model input unit 5 creates a trained AI model with parameters determined by some model and learning algorithm, and is made ready to input input data. Here, for example, a simple ARX model such as the one shown below is shown as an example of a model used in the AI ​​water level prediction model input unit 5.

[0073] H(t+1)=a0×H(t)+a1×H(t-1)+…+an×H(tn)+ b10×R1(t)+b11×R1(t-1)+…+b1m×R1(tl)+ …… bM0×RM(t)+bM1×RM(t-1)+…+bMl×RM(tl) (11) In equation (11), H(·) represents the water level at one of L water level measurement points, and R1(·), R2(·) ..., RM(·) represent rainfall data at M points. Additionally, n is the lag order of the autoregressive model, meaning up to n steps before a specified interval (e.g., 5 minutes), and l is the lag order of the external input (rainfall), meaning l steps before the specified interval. Additionally, a0, ... an, b10, ... bMl are regression coefficients of the ARX model, and the values ​​of these coefficients are determined using past rainfall and water level data with one of the algorithms mentioned above.

[0074] The AI ​​water level prediction model input unit 5 constructs (learns) a model such as that shown in equation (11) for each of the L water level measurement locations, and makes it available in real time. Note that the learning performed by the AI ​​water level prediction model input unit 5 can also be online learning. In this embodiment, it is assumed that the above processing (functional blocks indicated by dotted lines in FIG. 1) is constructed offline in advance.

[0075] An example of a process performed online (in real time) in the real-time water level prediction device of this embodiment will be described below.

[0076] The rainfall information input unit 6 inputs time series data (rainfall data) of rainfall amount or rainfall intensity measured in real time at multiple locations (e.g., M locations) at a predetermined period (e.g., 1-minute period or 5-minute period). The rainfall data may be data from a ground rain gauge installed in the basin, but in recent years, the Ministry of Land, Infrastructure, Transport and Tourism has installed X-band MP radars nationwide and distributes composite rainfall data such as C-band radar as XRAIN, so the rainfall information input unit 6 may continuously input mesh-shaped rainfall time series data such as XRAIN. When inputting mesh-shaped rainfall time series data, the rainfall information input unit 6 inputs time series data of rainfall intensity (or rainfall amount) of multiple mesh-shaped areas including the entire target basin in real time. This allows the radar rainfall data to be seamlessly connected to the real-time runoff analysis in the real-time water level prediction device of this embodiment.

[0077] The rainfall prediction function unit 7 generates prediction information of rainfall (or rainfall intensity) corresponding to a period for which a predicted water level is to be calculated (for example, a period from t0 to t0+Lp from the current time t0 to a predetermined prediction length Lp). In this embodiment, the rainfall prediction function unit 7 uses the rainfall time series data input by the rainfall information input unit 6 to calculate time series data of rainfall prediction values ​​of a predetermined period for a predetermined future period in rainfall measurement values ​​at M locations. For example, when it is desired to predict the water level one hour ahead, rainfall prediction is performed for about one hour or so, which is necessary for water level prediction up to one hour ahead. Note that, since there is a delay in inflow time and flow down time between rainfall and water level, the time series data of rainfall prediction values ​​may be data for a period slightly shorter than the prediction period. For example, when predicting the water level about 10 to 20 minutes ahead, it may not be necessary to perform rainfall prediction by the rainfall prediction function unit 7.

[0078] The rainfall prediction function unit 7 may predict rainfall in any way, and a prediction model such as data-driven machine learning may be created in advance, as in the case of the water level prediction described above, but since some short-term rainfall prediction information is distributed as nowcast, such rainfall prediction service data may be used. When short-term rainfall prediction service information is not available, the rainfall prediction function unit 7 may directly extrapolate (zeroth-order extrapolation) the rainfall inputted in the rainfall information input unit 6 and use it, or, when there is a tendency for rainfall to become stronger, may linearly extrapolate (first-order extrapolation) in consideration of risk and use it. Alternatively, the rainfall prediction function unit 7 may create a simple autoregressive model to generate time series data of simplified rainfall prediction values.

[0079] Furthermore, when using the final water level prediction results as hazard map information for flood risk, it may be necessary to predict a scenario of how the flood risk will change for each assumed rainfall, rather than focusing on accurately predicting the actual water level. In such cases, the rainfall prediction function unit 7 may create a scenario of rainfall intensity assumed as a future rainfall prediction, and use this as rainfall prediction information.

[0080] The rainfall prediction information generated by the rainfall prediction function unit 7, together with the rainfall measurement information input by the rainfall information input unit 6, can be used as input information for the water level measurement point water level prediction function unit 9 described later.

[0081] The water level measurement information input unit 8 inputs time series data of water level measured in real time by water level measuring instruments (water level sensors) installed at L locations at a predetermined cycle (for example, 1-minute cycle or 5-minute cycle). The information input by the water level measurement information input unit 8 is used as input information for the water level measurement point water level prediction function unit 9, which will be described later. In this embodiment, the water level measurement information input unit 8 inputs water level time series data of L locations obtained online (at a predetermined cycle) in real time.

[0082] The water level measurement point water level prediction function unit 9 outputs, in real time (at a predetermined cycle), the predicted water level at L water level measurement points as time series data over a predetermined prediction period, using the time series data of the predicted rainfall value output by the rainfall prediction function unit 7 (including the measured time series data (rainfall measurement information) of the rainfall data input by the rainfall information input unit 6) and the time series data of the water level measurement value input by the water level measurement information input unit 8. In this embodiment, the water level measurement point water level prediction function unit 9 inputs the time series data of the predicted rainfall value output by the rainfall prediction function unit 7 (including the rainfall measurement information) and the measured water level time series data at L water level measurement points, and outputs the time series data of the predicted water level over a predetermined future period (prediction period) of the measurement points at L points, using the AI ​​prediction model supplied from the AI ​​water level prediction model input unit 5.

[0083] The water level measurement point water level prediction function unit 9 extracts one or more explanatory variables from M pieces of rainfall data and L pieces of water level measurement data for the AI ​​water level prediction model, and inputs past rainfall data (t0-Ls to t0) from a predetermined time Ls to the present (e.g., the time point at which the water level prediction value is calculated) t0, time series data of rainfall from the past to the future (t0-Ls to t0+Lp) including future rainfall prediction information (e.g., time series data of rainfall prediction values) obtained from the rainfall prediction function unit 7, and time series data of water level measurement values ​​in the past (t0-Ls to t0), into the extracted explanatory variables, and calculates time series data of water level prediction values ​​at L locations up to a predetermined prediction length (t0+Lp) at a predetermined period (Tc).

[0084] At this time, the water level measurement point water level prediction function unit 9 uses the water level prediction model input from the AI ​​water level prediction model input unit 5. When an ARX model (regression model) such as the above-mentioned equation (11) is used as the water level prediction model, the water level measurement point water level prediction function unit 9 can calculate the predicted water level at the measurement point for a specified period by repeatedly using the equation of the water level prediction model.

[0085] For example, if the water level measurement point water level prediction function unit 9 predicts the water level up to one hour ahead at a predetermined cycle of five minutes, using equation (11) once will calculate the predicted water level five minutes ahead, and using it twice will calculate the predicted water level ten minutes ahead. Therefore, the water level measurement point water level prediction function unit 9 can predict the water level up to one hour ahead by repeating the use of equation (11) 12 times.

[0086] In addition, in equation (11), for example, when the degree of autoregression n=6 and the degree of external input (rainfall) l=12, when the water level measurement point water level prediction function unit 9 calculates the predicted water level 5 minutes ahead in the first step, it is necessary to input measured water level time series data of 6 points from the present to 5×6=30 minutes ago as the measured water level, and it is necessary to input measured rainfall time series data from the present to 5×12=60 minutes ago as the rainfall information. When the water level measurement point water level prediction function unit 9 calculates the predicted water level 10 minutes ahead in the second step, the water level predicted in the first step is used as the current water level data, and the time series data obtained by adding the predicted water level 5 minutes ahead to the measured water level from the present to 25 minutes ago is input to the prediction model (regression equation) of equation (11). For rainfall information, the water level measurement point water level prediction function unit 9 generates and inputs time series data of rainfall prediction values ​​by combining the rainfall prediction value data for the next 5 minutes output from the rainfall prediction function unit 7 with the measured rainfall data from the present to 55 minutes ago. This enables the water level measurement point water level prediction function unit 9 to calculate the predicted water level for the next 10 minutes.

[0087] The water level measurement point water level prediction function unit 9 repeats the same operations as the first and second steps 12 times (up to the 12th step) to generate time series data of water level prediction in 5-minute increments up to 1 hour (60 minutes ahead). In this case, generally, the predicted rainfall data output from the rainfall prediction function unit 7 is also required up to 60 minutes ahead, but since there is usually a time delay between rainfall and water level corresponding to the inflow time and flow down time, the coefficients of the rainfall data at the current time b10, ..., bM0 and the coefficients of the most recent rainfall data such as 5 minutes ago b11, ..., bM1 are often almost 0. Therefore, for example, in the case where there is a 20-minute delay between rainfall and water level (i.e., the coefficients up to b14, ..., bM4 are 0), if there is a rainfall prediction value up to 40 minutes ahead, the water level can be predicted up to 1 hour ahead.

[0088] As described above, the water level measurement point water level prediction function unit 9 can predict the water level at L water level measurement points at a predetermined cycle using a water level prediction model such as equation (11). It goes without saying that the water level prediction model used in the water level measurement point water level prediction function unit 9 is not limited to an ARX model (regression model) such as equation (11), and as mentioned above, a neural network model trained by deep learning or the like may also be used. Regardless of the water level prediction model used, the water level measurement point water level prediction function unit 9 can obtain water level prediction time series data up to a predetermined prediction period.

[0089] The water level prediction function unit 10 at the unmeasured water level performs runoff analysis by parallel calculation for each sub-basin using the predicted water level (predicted water level value) at the water level measurement point as a boundary condition, and calculates the predicted water level (predicted water level value) at the unmeasured point. Specifically, the water level prediction function unit 10 at the unmeasured water level has water level prediction units 101-10K corresponding to each of the multiple sub-basins, and inputs time series data of rainfall data (measured values) input from the rainfall information input unit 6 at a predetermined cycle, time series data of rainfall amount prediction values ​​output from the rainfall prediction function unit 7, time series data of water level measurement values ​​input from the water level measurement information input unit 8, and time series data of water level prediction values ​​at the measurement points output from the water level measurement point water level prediction function unit 9, and predicts the water level at the unmeasured point by performing runoff analysis for each of the K sub-basins using a runoff analysis model for each sub-basin input from the partial runoff analysis model input / output unit 4.

[0090] The unmeasured point water level prediction function unit 10, for example, inputs M pieces of time series data on rainfall over a predetermined period (t0-Ls to t0+Lp) from the past to the future at a predetermined cycle (Tp=k×Tc (k is an integer greater than or equal to 1)) into the runoff analysis model, and at the same time sets the time series data of the water level measurement values ​​of L water level measurement points and the L water level prediction values ​​for the prediction period (t0+Lp) obtained by the water level measurement point water level prediction function unit 9 as boundary conditions of the partial watershed (partial watershed sewer pipe network graph), and performs runoff analysis in real time divided into K partial watersheds using the runoff analysis model, and calculates the water level prediction values ​​for the prediction period (t0+Lp) of the NL unmeasured water level points.

[0091] In this embodiment, the water level prediction function unit 10 of the water level unmeasured point uses the partial runoff analysis model for each of K partial basins supplied from the partial runoff analysis model input / output unit 4, uses the time series data of the rainfall prediction value corresponding to each partial basin in the predicted rainfall time series data (including rainfall measurement information) output from the rainfall prediction function 7, and at the same time sets the time series data of the water level prediction value at L measurement points output from the water level measurement point water level prediction function unit 9 as a boundary condition, and calculates the predicted water level in all sewer pipes of all target basins by executing the K partial runoff analysis models in parallel by the water level prediction unit 101-10K. The water level prediction function unit 10 of the water level unmeasured point may further use the water level measurement value of the water level measurement point supplied from the water level measurement information input unit 8 as a boundary condition.

[0092] In this embodiment, the water level prediction function unit 10 for unmeasured points sets the time series data of water level measurement values ​​input by the water level measurement information input unit 8 and the water level prediction data for measurement points which is the output of the water level measurement point water level prediction function unit 9 as boundary conditions for runoff analysis for each sub-catchment area.

[0093] In runoff analysis, rainfall information is often used as input to calculate the water level of each sewer pipe in the target basin and the flow rate through the sewer pipe. In contrast, the real-time water level prediction device of this embodiment inputs not only rainfall information but also water level information measured at water level measurement points to calculate the water level and flow rate of each sewer pipe in the target basin. Furthermore, by replacing the calculated water level calculated in runoff analysis with the measured water level at the water level measurement point, the accuracy of the water level at the calculation point is clearly improved. In other words, the real-time water level prediction device of this embodiment can be considered to perform feedback compensation for the calculated water level (calculated water level) so that the error between the measured water level and the calculated water level becomes zero.

[0094] As described above, in the real-time water level prediction device of this embodiment, water level information at measurement points becomes more reliable, so the accuracy of the calculated water level can be improved by correcting the water level at an unmeasured point calculated using the value at the measurement point (measured water level).

[0095] That is, in the real-time water level prediction device of this embodiment, not only water level information from the measurement point is used, but also the value is predicted as a boundary condition by the AI ​​model to predict the predicted water level at the measurement point in the future. This improves the accuracy of future water level predictions at unmeasured points, as described below.

[0096] Runoff analysis is usually used as an analysis, and is not used with particular consideration of calculating future information. Since the water level can be calculated by inputting rainfall into the runoff analysis model, it is possible to calculate the future water level accordingly by predicting future rainfall and inputting it into the runoff analysis model. However, the accuracy depends on the accuracy of the rainfall prediction and the accuracy of the runoff analysis model (the error between the actual phenomenon and the phenomenon calculated by the model, also called modeling error), and it is generally difficult to predict the water level with good accuracy. Even if the accuracy of runoff analysis is improved by making corrections using measured water levels, future information on the water level at the measurement point cannot be obtained unless runoff analysis is performed, so the accuracy of the prediction of the future water level cannot be improved as it is. In other words, when inputting not only rainfall but also the water level at the measurement point into the runoff analysis model, it is difficult to improve the accuracy of the future predicted water level unless the future water level at the measurement point is provided separately from outside.

[0097] In the real-time water level prediction device of this embodiment, future water level information is predicted by an AI model, and the predicted water level information is provided from outside the runoff analysis model. Doing so has the following two main advantages.

[0098] The first advantage is that, generally, when data is measured, data-driven models such as AI models tend to calculate predictions more accurately than physical models such as runoff analysis. For this reason, it is highly likely that the predicted water level at the measurement point calculated by the AI ​​model will be more accurate than the predicted water level (at the measurement point) calculated by the runoff analysis model using predicted rainfall as input. Therefore, at least at the water level measurement point, it is considered that calculating the predicted water level using the AI ​​model is more accurate than the predicted water level obtained as a result of runoff analysis using predicted rainfall as input. Therefore, by inputting more reliable water level information not only for the present time (= the time when the prediction is being made) but also for the future into the prediction model, it is expected that the accuracy of the predicted water level at unmeasured points will also improve. In other words, it is expected that the prediction accuracy of the water level of all pipelines in the target basin, including measured and unmeasured points, will improve.

[0099] Another advantage is that runoff analysis calculations can be performed for each sub-basin by inputting the predicted water level as a boundary condition. Therefore, by performing runoff analysis calculations for multiple sub-basins in parallel, the time required for runoff analysis can be significantly improved. As a result, the real-time water level prediction device of this embodiment makes it possible to perform real-time calculations of water level predictions, and the runoff analysis results can be used to provide information such as real-time hazard maps.

[0100] The intra-basin pipeline predicted water level output unit 11 outputs and presents, for example, information on predicted water levels of all sewer pipes over a predetermined future period provided by the unmeasured point water level prediction function unit 10. That is, the intra-basin pipeline predicted water level output unit 11 outputs, to a monitoring screen or the like, water level prediction information of measured points predicted by an AI model and water level prediction information of unmeasured points calculated by runoff analysis for each partial basin, making it possible to present future water level information to a user in real time.

[0101] The intra-basin pipe predicted water level output unit 11 may simply present predicted water levels for all sewer pipes, but if it is intended to be used as a flood prediction function, it may determine and present the flooding state from the predicted water levels. In other words, the intra-basin pipe predicted water level output unit 11 may obtain time series data of predicted water level values ​​at points where the water level is not measured and time series data of predicted water level values ​​at points where the water level is measured, and output a result of determining the flooding risk in multiple pipes based on the obtained water level prediction values. The intra-basin pipeline predicted water level output unit 11 can make a judgment of the risk of flooding, for example, for each pipeline, such as: if the water level prediction value (or the capacity prediction value converted from the water level prediction value) is less than the capacity of each pipeline (less than full) (water level smaller than each pipe diameter), it is normal; if the water level prediction value (or the capacity prediction value converted from the water level prediction value) exceeds the capacity of each pipeline (is equal to or greater than the capacity) but is less than the capacity plus the manhole capacity (the sum of the capacity of each pipeline and the capacity of the manhole), it is a state where there is no flooding but the risk of flooding is increasing; if the water level prediction value (or the capacity prediction value converted from the water level prediction value) exceeds the sum of the capacity of each pipeline and the capacity of the manhole (is equal to or greater than the sum of the capacity of each pipeline and the capacity of the manhole), it is a state where the risk of flooding is extremely high.

[0102] FIG. 8 is a diagram showing an example of a screen that visualizes the flood risk assessment result and presents it to the user. Furthermore, the basin pipe predicted water level output unit 11 may have a flood risk hazard map display function that displays, for example, a sewer pipe network that can individually identify sewer pipes corresponding to N water level gauge locations on a map of the target basin, and displays a flood risk hazard map of the basin based on the water level of each sewer pipe. For example, the basin pipe predicted water level output unit 11 may determine a display mode such as a display color corresponding to each judgment result, such as blue for less than full pipes, yellow for full pipes or more and less than the capacity including manhole capacity, and red for full pipes or more and the amount including manhole capacity is exceeded, and provide information in real time by coloring each pipe along with a diagram of the basin as shown in FIG. 8, or by coloring the small drainage districts corresponding to each pipe.

[0103] In addition, the intra-basin pipeline predicted water level output unit 11 may provide a hazard map that visualizes the flood risk to a cloud server, and the cloud server may provide the hazard map to a user terminal, thereby providing real-time flood hazard map information. In this case, the function of the real-time water level prediction device of this embodiment may be installed in the cloud server.

[0104] As described above, in the real-time water level prediction device of this embodiment, future water level values ​​at water level measurement points are calculated using an AI model, and the future water level values ​​calculated by the AI ​​model are set as external information and boundary conditions of the runoff analysis model, making it possible to determine future water levels at points where the water level has not been measured by parallel runoff analysis calculations for each sub-basin into which the basin is divided.

[0105] As a result, according to the real-time water level prediction device of this embodiment, the above series of processes makes it possible to predict water levels in sewer pipes by effectively utilizing time series data of water level measurement values ​​from water level gauges in sewer pipes, which are being increasingly introduced, and to predict inundation based on water level predictions, and it becomes possible to provide water level prediction results and inundation prediction results in real time that are more accurate than water levels predicted by runoff analysis that uses only rainfall and predicted rainfall as input.

[0106] In other words, according to this embodiment, it is possible to provide a real-time water level prediction device, a real-time water level prediction method, and a computer program that provide accurate water level prediction information in real time not only for water level measured points but also for unmeasured points.

[0107] Furthermore, according to the real-time water level prediction device, real-time water level prediction method, and computer program of this embodiment, it is possible to provide an accurate real-time hazard map by predicting flooding.

[0108] In addition, according to the real-time water level prediction device, real-time water level prediction method, and computer program of this embodiment, by using a simplified runoff analysis model, not only is real-time performance improved, but runoff and inundation analysis can be performed using only information such as flow rate calculation tables and plot division plans that are normally held by sewerage operators.

[0109] In addition, according to the real-time water level prediction device, real-time water level prediction method, and computer program of this embodiment, radar rainfall data such as XRAIN (X-band MP radar rainfall) can be seamlessly incorporated into real-time runoff and inundation analysis.

[0110] Next, a number of modified examples of the real-time water level prediction device according to the above-described embodiment will be described. In the following description, the same components as those in the above-described embodiment will be denoted by the same reference numerals and description thereof will be omitted. (First Modification) For example, a function of identifying and adjusting parameters of a runoff analysis model may be added as an additional function to the real-time water level prediction device of the above-described embodiment.

[0111] The real-time water level prediction device of the above embodiment combines a function to predict the water level at a water level measurement point by AI and a function to predict the water level at a point where the water level is not measured by using the runoff analysis model with the predicted water level value as a boundary condition. As mentioned above, the AI ​​model is constructed by learning from actual data using accumulated data from the past, whereas the runoff analysis model is constructed using physical information. In other words, the former is an inductive modeling method that constructs a model to match an actual phenomenon, whereas the latter is a deductive modeling method based on physical phenomena.

[0112] In general, deductive modeling methods based on physical laws are superior to inductive modeling methods based on real data in that they can build models even without actual data such as measured data. However, unlike inductive modeling, deductive modeling methods based on physical laws do not reflect actual measured and observed results in the model, so local accuracy tends to be lower than inductive modeling. One method known to resolve this tendency is to adjust physical models using observed data, and this is called data assimilation.

[0113] In data assimilation, the objects to be adjusted are often state variables contained in a physical model (e.g., water level, flow rate, temperature, pressure, etc.), but parameters of the physical model (e.g., pipe resistance, pipe roughness, resistance, structural information of the physical model, etc.) can also be subject to data assimilation by regarding these as unchanging state variables (state variables whose differential value = 0).

[0114] From the perspective of data assimilation, in the real-time water level prediction device of the above-mentioned embodiment, the process of correcting (replacing) the water level at the water level measurement point calculated by runoff analysis with the actually observed calculated water level is equivalent to performing data assimilation.

[0115] On the other hand, as mentioned above, the runoff analysis model includes various parameters, and these parameters are set by extracting information from flow calculation tables, sewerage registers, etc., as mentioned above. However, some of these parameters, such as roughness, include uncertainties compared to physical structure parameters such as pipe diameter and pipeline length, and although typical values ​​are given, they may not necessarily match the actual situation. In addition, although the accuracy of the values ​​of the physical structure parameters is extremely high, they may not necessarily match the actual situation due to reasons such as incorrect input, differences between the planning stage and the construction stage, and information not being updated when repairs are performed.

[0116] Therefore, in order to perform runoff analysis with higher accuracy, it is preferable to adjust these parameters as well. For example, a function for performing data assimilation of parameters offline in advance before performing runoff analysis in real time may be added to the real-time water level prediction device.

[0117] That is, instead of providing an initially constructed runoff analysis model to the runoff analysis model input unit 1 in Fig. 1, parameters may be adjusted in advance. In this case, the data used for data assimilation are rainfall data at M rainfall measurement locations input from the rainfall information input unit 6 and water level data at N water level measurement locations input from the water level measurement information input unit 8. The rainfall data at M locations and the water level data at N locations are past time-series data stored in a data server or the like.

[0118] The runoff analysis model input unit 1 may have an offline runoff analysis model learning function that performs runoff analysis of the entire basin using these past time-series data, for example, past rainfall data, and performs data assimilation to reduce the analysis error (measurement error) between the calculated water level, which is the runoff analysis result at L water level gauge installation locations obtained as a result of the analysis, and the actually measured water level. In other words, the runoff analysis model input unit 1 may adjust some parameters stored in the table generated by the runoff / inundation analysis model parameter table generation unit 22, using the sum of the squared error or absolute value error of the analysis error at the water level gauge installation locations over a predetermined period as a minimization evaluation index.

[0119] The runoff analysis model input unit 1 may, for example, use time series data of M rainfall measurement values ​​accumulated over a specified past period (Lstore) as input to the runoff analysis model, perform normal runoff analysis for the past period (Lstore), and adjust and assimilate parameters included in the runoff analysis model so that the error between the time series data of calculated water level data for L locations in the past (Lstore) obtained as a result of the runoff analysis and the time series data of L water level measurement values ​​in the past period (Lstore) is reduced.

[0120] Among the parameters shown in FIG. 3, the parameters that are considered to have a relatively high degree of uncertainty are considered to be roughness, runoff coefficient, and flow arrival time, and therefore some of these parameters are selected as parameters for adjustment. Then, the runoff analysis model input unit 1 adjusts the selected parameters so as to reduce the analysis error. In this case, a so-called data assimilation method may be used. Data assimilation methods include batch-type algorithms and sequential-type algorithms. As a batch-type processing method, for example, a method such as the four-dimensional variation method is known. As a sequential method, methods such as a particle filter and an ensemble Kalman filter are widely known.

[0121] Since the real-time water level prediction device of this modified example adjusts parameters in advance, a batch-type algorithm such as the four-dimensional calculus of variations may be used, or a sequential particle filter or ensemble Kalman filter may also be used, so any data assimilation method may be adopted.

[0122] Furthermore, if the purpose is to adjust (identify) parameters, it is not necessary to adopt a data assimilation method. For example, it is possible to treat it as an optimization problem as an optimization evaluation function for minimizing analysis errors, and adjust (identify) parameter values ​​using an appropriate optimization method. The runoff analysis model input unit 1 may identify and learn parameters such as roughness, runoff coefficient, and flow arrival time using heuristic optimization methods called metaheuristics, such as genetic algorithms (GA) and particle swarm optimization (PSO).

[0123] FIG. 9 is a diagram for explaining an example of a process for adjusting (identifying) parameters and dividing regions in the real-time water level prediction device of the first modified example. The real-time water level prediction device of this modified example differs from the above embodiment in that it classifies the time series data of the measured water level at the water level measurement point into data measured in the past (data already stored in a data server, etc.) and data measured in real time, and uses the former for data assimilation and the latter for real-time runoff analysis. The data measured in real time (the former) may include data measured in a predetermined past period going back from the start of the calculation of the real-time runoff analysis, and the data measured in the past (the latter) may be data obtained by excluding the data measured in real time from all the measurement data.

[0124] In principle, real-time performance is not required in the data assimilation that the runoff analysis model input unit 1 performs in advance. Therefore, even if the calculation load is large, the entire basin is analyzed over time and the parameter values ​​of the runoff analysis model are adjusted in advance. In addition, in runoff analysis for each subbasin that requires real-time performance, the calculation load can be reduced and the processing speed can be improved by using only data measured in real time. As described above, the real-time water level prediction device of this modified example can obtain the same effects as the above-mentioned embodiment, and by adding a process of adjusting the parameters of the runoff analysis model in advance while maintaining the same real-time performance as the above-mentioned embodiment, more accurate water level prediction and inundation prediction can be achieved.

[0125] In other words, according to this modified example, it is possible to provide a real-time water level prediction device, a real-time water level prediction method, and a computer program that provide accurate water level prediction information in real time not only for water level measured points but also for unmeasured points.

[0126] (Second Modification) The second modified real-time water level prediction device is a modified version of the first modified real-time water level prediction device, in which the runoff analysis model adjusted based on measurement data is replaced with an AI model.

[0127] When the runoff analysis model is adjusted to fit the actual measurement data, the water level calculated by the runoff analysis model is considered to be close to the actual water level. Therefore, in the real-time water level prediction device of this modified example, the runoff analysis model input unit 1 virtually regards the runoff analysis model whose parameters have been adjusted by data assimilation as the actual system, inputs various rainfall time series data to generate water level time series data, and uses pairs of the input rainfall time series data and the generated water level time series data to construct an AI model by a method such as machine learning.

[0128] For example, the runoff analysis model input unit 1 generates time series data of water levels as output data through runoff analysis using time series data of rainfall as input data for the runoff analysis model, and constructs a trained AI model through machine learning using the input data and output data.

[0129] In this case, since the data used to construct (learn) the AI ​​model in the runoff analysis model input unit 1 is the data used for runoff analysis, the rainfall data may be not only actual rainfall measurements but also virtual rainfall, and the water level data may be not only water level data from L measurement points but also water level data from N points in the entire pipeline. In this way, the runoff analysis model input unit 1 can construct an AI model that approximates the parameter-adjusted runoff analysis model.

[0130] By constructing an AI model as described above, virtual rainfall data can be input to the AI ​​model, so that, for example, in addition to normal rainfall data, data for various rainfall scenarios such as heavy rainfall and continuous rainfall can be used for learning. This makes it possible to confirm in advance what kind of flooding phenomena will be expressed in response to rainfall in the AI ​​model.

[0131] Furthermore, since the processing speed of the AI ​​model during real-time calculations is overwhelmingly faster than that of the runoff analysis model, even in cases where the number of water level measurement points L is small and the target basin cannot be divided into small parts, the runoff analysis model input unit 1 can ensure real-time performance by constructing the AI ​​model in advance.

[0132] In addition, when performing real-time runoff analysis for each sub-basin as described in the above embodiment, when the runoff analysis model input unit 1 constructs an AI model, it is necessary to learn the AI ​​model so that it can be separated at L water level measurement points. This can be realized by imposing constraints on the structure of the AI ​​model. However, since the online calculation (water level inundation prediction calculation) of the AI ​​model can be performed at an extremely high speed, if it is laborious to separate and learn the AI ​​model, this AI model may be constructed collectively for the entire basin. However, in that case, it is necessary to have an AI model structure that allows the measured water levels at L points to be input in real-time water level / inundation prediction. Such an AI model is, for example, an autoregressive transfer function model having an autoregressive term such as the formula (11), or an AI model of a type that uses past data of measurement points such as a recurrent neural network or LSTM.

[0133] As described above, the real-time water level prediction device of the second modified example has a function of replacing the runoff analysis model with an AI model after prior parameter adjustment in the runoff analysis model input unit 1 described in the first modified example. Therefore, the runoff analysis model input unit 1 inputs the AI ​​model constructed in place of the runoff analysis model to the watershed division function unit 3 as a runoff analysis surrogate model.

[0134] According to this modified example, the same effects can be obtained as with the real-time water level prediction device of the first modified example described above, and it becomes possible to perform real-time water level prediction and inundation prediction even in cases where there are few locations where water level gauges are installed and it is difficult to ensure real-time performance.

[0135] In other words, according to this modified example, it is possible to provide a real-time water level prediction device, a real-time water level prediction method, and a computer program that provide accurate water level prediction information in real time not only for water level measured points but also for unmeasured points.

[0136] (Third Modification) The real-time water level prediction device of the third modified example adjusts the parameters of the runoff analysis model simultaneously while performing real-time inundation prediction. In the real-time water level prediction devices of the first and second modified examples, water level data from L water level measurement points is divided into past data (data already stored in a data server, etc.) and current data (data measured in real time), parameters are adjusted using the past data, and real-time runoff analysis is performed using the current data. In contrast, the real-time water level prediction device of the third modified example performs real-time parameter adjustment and real-time runoff analysis simultaneously. To achieve this, the real-time water level prediction device of this modified example divides L water level measurement points into Lb water level measurement points for basin division and La (=L-Lb) water level measurement points for data assimilation, and uses the respective measurement data.

[0137] The drainage basin division function unit 3 divides the drainage basin into a plurality of sub-basins using a part (La pieces) of time-series data of a plurality (L pieces) of water level measurement values ​​as boundary condition setting data.

[0138] The water level prediction function unit 10 for unmeasured points has a function of adjusting the parameters included in the partial runoff analysis model so as to minimize the error between the learning data (Lai (i=1, 2, ..., K La=La1+La2+...+Lak)) present in the divided multiple partial watersheds and the Lai pieces of water level prediction value data calculated by the partial watershed analysis model corresponding to the partial watershed, using Lb (=L-La) pieces of time series data of multiple (L) water level measurement values ​​excluding the data for setting boundary conditions as learning data.

[0139] FIG. 10 is a diagram for explaining an example of a process for adjusting (identifying) parameters and dividing regions in a real-time water level prediction device of the third modified example. In FIG. 10, there are L (= 3) water level measurement points, but the watershed division function unit 3 divides the watershed into two using Lb (= 1) water level measurement points for watershed division, and the unmeasured point water level prediction function unit 10 uses La (= 3-1 = 2) water level measurement points as water level measurement points for data assimilation. That is, in this modified example, the multiple water level measurement points are defined in advance as water level measurement points for watershed division and water level measurement points for data assimilation. Then, the unmeasured point water level prediction function unit 10 performs runoff analysis by parallel calculation for the two partial watersheds shown in FIG. 10 while performing the real-time inundation prediction calculation described in the above embodiment.

[0140] At this time, since the water level at the water level measurement point for data assimilation has been measured, the water level prediction function unit 10 for points where the water level is not measured uses the measurement data of this water level measurement point to perform data assimilation of parameters using a data assimilation method, while also predicting the water level at points where the water level is not measured. Since data assimilation in this case is performed in real time, for example, by using sequential data assimilation such as the ensemble Kalman filter or particle filter described above, data assimilation and water level prediction can be performed in parallel.

[0141] According to this modified real-time water level prediction device, when there are many water level gauge setting locations or when the computational load of runoff analysis is not too high due to simple runoff analysis, by adjusting model parameters online instead of coarsening the number of watershed divisions, it is possible to make real-time water level predictions and inundation predictions with higher accuracy.

[0142] In other words, according to this modified example, it is possible to provide a real-time water level prediction device, a real-time water level prediction method, and a computer program that provide accurate water level prediction information in real time not only for water level measured points but also for unmeasured points.

[0143] (Fourth Modification) The real-time water level prediction device of the fourth modified example combines the functions of the real-time water level prediction devices of the first to third modified examples. That is, the real-time water level prediction device of this modified example performs parameter adjustment and identification offline in advance in the runoff analysis model input unit 1, and performs parameter identification, water level prediction, and inundation prediction online in real time in the unmeasured water level prediction function unit 10.

[0144] As explained in the second modified example, the AI ​​model constructed by the runoff analysis model input unit 1 has parameters as the AI ​​model. For example, in the case of a regression model, the parameters are regression coefficients, and in the case of a neural network model, the parameters are connection weights, etc. These parameters have already been learned, but when measurement information (parameters) of the water level measurement points for data assimilation shown in FIG. 10 is obtained online, for example, these parameters can be sequentially identified online using an online learning algorithm. For example, in the case of a regression model, sequential processing such as the sequential least squares method and the Kalman filter are known, so the parameter values ​​can be learned online using these algorithms. In this way, the method of the second modified example and the method of the third modified example can be combined.

[0145] The real-time water level prediction device of this variant can achieve the same effects as the real-time water level prediction devices of the first to third variants described above, and although prior adjustments are required, it is possible to achieve real-time water level predictions and flood predictions with even higher accuracy than the real-time water level prediction device of the third variant.

[0146] In other words, according to this modified example, it is possible to provide a real-time water level prediction device, a real-time water level prediction method, and a computer program that provide accurate water level prediction information in real time not only for water level measured points but also for unmeasured points.

[0147] The program according to the present embodiment may be transferred in a state where it is stored in an electronic device, or in a state where it is not stored in an electronic device. In the latter case, the program may be transferred via a network, or in a state where it is stored in a storage medium. The storage medium is a non-transitory tangible medium. The storage medium is a computer-readable medium. The storage medium may be in any form, such as a CD-ROM or a memory card, as long as it is capable of storing a program and is computer-readable.

[0148] Although some embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope and spirit of the invention, and are included in the scope of the invention and its equivalents described in the claims. [Explanation of symbols]

[0149] 1...Runoff analysis model input section, 2...Water level measurement point information input section, 3...Bassin division function section, 4...Partial runoff analysis model input / output section, 41-4K...Input / output section, 5...AI water level prediction model input section, 6...Rainfall information input section, 7...Rainfall prediction function section, 8...Water level measurement information input section, 9...Water level measurement point water level prediction function section, 10...Water level unmeasured point water level prediction function section, 101-10K...Water level prediction section, 11...Intra-basin pipeline predicted water level output section, 21...Flow rate calculation table / sewerage register DB, 22...Runoff / inundation analysis model parameter table generation section, 23...Runoff / inundation analysis model automatic construction section

Claims

1. A real-time water level prediction device for a watershed where water level measurements at a plurality of water level measurement points in a sewer pipe network including a plurality of pipes and rainfall measurements at a plurality of rainfall measurement points are measured at a predetermined interval, the real-time water level prediction device calculates water level prediction values ​​at a plurality of unmeasured water level points, a watershed division function unit that divides the watershed into a plurality of small drainage areas each including a plurality of the pipes, and divides the watershed into a plurality of sub-watersheds by the small drainage areas each including the water level measurement point based on a connection relationship between the plurality of pipes in the sewer pipe network; a water level prediction function unit that calculates time series data of water level prediction values ​​at the water level measurement points using time series data of water level measurement values ​​at the multiple water level measurement points and time series data of rainfall measurement values ​​at the multiple rainfall measurement points, using a trained machine learning model that has been trained to calculate water level prediction values ​​at the water level measurement points; a water level prediction function unit for calculating in parallel time series data of the water level prediction value at the water level measurement point as a boundary condition of the water level in the plurality of partial basins, and inputting time series data of rainfall measurement values ​​and rainfall prediction values ​​at the plurality of rainfall measurement points into a plurality of partial runoff analysis models obtained by dividing a runoff analysis model for analyzing the runoff of the basin so as to calculate a water level prediction value at a water level unmeasured point included in each of the plurality of partial basins; A real-time water level prediction device equipped with

2. 2. A real-time water level prediction device according to claim 1, wherein the partial runoff analysis model is a model that calculates a water level prediction value by taking into account the capacity of the sewer included in the partial drainage basin and the capacity of a manhole above the sewer.

3. A model construction function unit generates a group of equations based on a continuity equation representing the water balance and a motion equation representing the water motion corresponding to each of the plurality of small drainage areas, using the table that defines connection information of the plurality of small drainage sections using the identification numbers of each of the plurality of small drainage areas included in the watershed, the table including the area and runoff coefficient of the small drainage area, the flow arrival time to the small drainage area, the flow down time or information capable of calculating the flow down time of the plurality of small drainage sections, and the capacity or information capable of calculating the capacity of the sewer connecting the plurality of small drainage sections, The real-time water level prediction device according to claim 1 , wherein the partial runoff analysis model includes the group of equations for the small drainage districts included in the corresponding partial watersheds.

4. A model construction function unit generates a group of equations based on a continuity equation representing the water balance and a motion equation representing the water motion corresponding to each of the plurality of small drainage areas, using the table that defines connection information of the plurality of small drainage sections using the identification numbers of each of the plurality of small drainage areas included in the watershed, the table including the area and runoff coefficient of the small drainage area, the flow arrival time to the small drainage area, the flow down time or information capable of calculating the flow down time of the plurality of small drainage sections, the capacity or information capable of calculating the capacity of the sewer connecting the plurality of small drainage sections, and the capacity or information capable of calculating the capacity of the manhole in the small drainage area, The real-time water level prediction device according to claim 2 , wherein the partial runoff analysis model includes the group of equations for the small drainage areas included in the corresponding partial watershed.

5. The model construction function unit includes: a water level calculation function that calculates the storage volume of the culvert included in the plurality of small drainage areas, sets a water level prediction value as the water level of the culvert when the storage volume is less than the capacity of the culvert, and calculates the water level of the culvert by assuming a state in which a virtual slot that matches the capacity of the manhole is connected to the upper part of the culvert when the storage volume is equal to or greater than the capacity of the culvert and less than the sum of the capacity of the culvert and the capacity of the manhole; A real-time water level prediction device as described in claim 4, further comprising a flow rate calculation function that calculates a flow rate taking into account the hydraulic gradient by calculating a hydraulic gradient using the difference in water levels of the multiple pipes calculated by the water level calculation function.

6. A real-time water level prediction device as described in claim 1, further comprising a model construction function unit that inputs time series data of rainfall measurement values ​​accumulated over a specified past period into the runoff analysis model, and constructs the runoff analysis model by adjusting parameters so as to reduce the error between the time series data of water levels obtained by performing runoff analysis of the specified past period and the time series data of water level measurement values ​​for the specified past period.

7. The real-time water level prediction device of claim 5, wherein the model construction function unit generates time series data of water levels as output data by runoff analysis using time series data of rainfall as input data for the runoff analysis model, constructs a trained AI model through machine learning using the input data and the output data, and outputs the AI ​​model in place of the runoff analysis model.

8. the watershed division function unit divides the watershed into a plurality of sub-watersheds using a part of the data of the plurality of water level measurement values ​​as boundary condition setting data, and The real-time water level prediction device of claim 1, wherein the unmeasured water level prediction function unit has a function of adjusting parameters contained in the partial runoff analysis model using learning data other than the boundary condition setting data of multiple water level measurement values ​​so as to minimize the error between the learning data present in the multiple divided partial watersheds and the water level prediction value data calculated by the partial watershed analysis model corresponding to the partial watershed.

9. 2. The real-time water level prediction device according to claim 1, wherein the time series data of rainfall measurements is time series data of rainfall measurements in a plurality of mesh-shaped areas including the watershed.

10. A real-time water level prediction device as described in claim 1, further comprising a predicted water level output unit that acquires time series data of water level prediction values ​​at the points where the water level is not measured and time series data of water level prediction values ​​at the water level measured points, and outputs a judgment result of the risk of flooding in multiple sewer systems based on the acquired water level prediction values.

11. A real-time water level prediction method for a watershed where water level measurements at a plurality of water level measurement points in a sewer pipe network including a plurality of pipes and rainfall measurements at a plurality of rainfall measurement points are measured at a predetermined interval, the method comprising the steps of: calculating water level prediction values ​​at a plurality of unmeasured water level points; The watershed is divided into a plurality of sub-watersheds by the sub-watersheds including the water level measurement points based on a connection relationship between the plurality of pipes in the sewer pipe network, and the watershed is divided into a plurality of sub-watersheds by the sub-watersheds including the water level measurement points. Calculating time series data of water level prediction values ​​at the water level measurement points using time series data of water level measurement values ​​at the plurality of water level measurement points and time series data of rainfall measurement values ​​at the plurality of rainfall measurement locations, using a trained machine learning model trained to calculate water level prediction values ​​at the water level measurement points; A real-time water level prediction method in which a runoff analysis model that performs runoff analysis of a basin is divided to calculate predicted water level values ​​at unmeasured water level points included in each of the multiple sub-basins, and in multiple sub-basin analysis models, time series data of the predicted water level values ​​at the water level measurement points is used as boundary conditions for the water levels in the multiple sub-basins, and time series data of rainfall measurement values ​​and predicted rainfall values ​​at the multiple rainfall measurement points are input into the multiple partial runoff analysis models, thereby calculating time series data of the predicted water level values ​​at the multiple unmeasured water level points in parallel.

12. A computer program for causing a computer to execute the real-time water level prediction method according to claim 11.

Images