Water level prediction device, water level prediction method, and computer program

Abstract

To provide a water level prediction device that performs accurate water level prediction in a sewer pipe in which the relationship of the water level between a plurality of points is correctly maintained.SOLUTION: A water level prediction device 1 comprises: a water level relationship feature extraction unit 10 that receives water level information measured at a plurality of prediction points of a sewer pipe as input, and calculates a water level relationship feature amount indicating the relationship of the water level between the prediction points; and a plurality of water level learning prediction units 20A-20Z that train a plurality of water level prediction models for calculating a water level prediction value for each of the plurality of prediction points, by using weather information, measured values of the amount of precipitation, water level information at the plurality of prediction points, and time series data of the water level relationship feature amount, structure information of the sewer pipe, and topographic information, and calculate the water level prediction value for each of the plurality of prediction points by using the trained water level prediction model.SELECTED DRAWING: Figure 1

Description

[Technical field]

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

[0002] In recent years, there has been an increase in localized heavy rainfall over a short period of time, leading to frequent inland flooding within cities. In fact, when looking at the cost of flood damage (both external and internal flooding), the cost of damage from inland flooding accounts for roughly half of the total damage nationwide. In Tokyo, the cost of damage from inland flooding exceeds 90% of the total damage in the city, and in urban areas where levees are relatively well developed, inland flooding has become a new issue.

[0003] Research is being conducted into real-time prediction of water levels based on meteorological data as a measure against inland flooding. For example, a flood prediction system has been proposed that predicts the possibility of flooding and the depth of flooding within an area based on pre-stored surface data such as elevation within the predicted area, data on the amount of rainfall that has actually fallen in the past few hours from the current time, and data on the amount of rainfall predicted to fall in the next few hours. However, the phenomena leading up to flooding are complex, and further improvements in the accuracy of predictions are needed.

[0004] For example, a technology has been proposed that uses machine learning (multiple regression analysis, Kalman filter, neural network, random forest, support vector machine, etc.) to predict the rate at which water levels will rise from rainfall intensity and calculate the predicted time until inland flooding occurs. With this technology, the time until the water level reaches the level at which inland flooding will occur is predicted based on the predicted rate at which the water level will rise, but it is difficult for users to determine how much they can trust the predicted value.

[0005] For example, a technology has been proposed that selects one of several pre-calculated flood depth distribution maps as the future flood depth distribution map based on observed rainfall data and predicted rainfall data. If the model used to select the future flood depth distribution is not updated, it may not be able to respond to changes in precipitation phenomena due to recent climate change.

[0006] For example, when a simple water level prediction is made using deep learning and this predicted value is then used to further predict water levels between multiple points, it may not be possible to correctly interpolate the water levels between multiple points because the system does not learn about the relationship between water levels at each point (high and low water levels, which water level is higher, etc.). [Prior art documents] [Patent documents]

[0007] [Patent Document 1] JP 2020-201704 A [Patent Document 2] JP 2018-3473 A [Patent Document 3] JP 2019-96255 A Summary of the Invention [Problem to be solved by the invention]

[0008] In order to reduce damage caused by inland flooding, it is desirable to improve the accuracy of water level predictions in sewer pipes. In particular, when predicting water levels at multiple points and using the predicted values ​​to interpolate water levels at points between predicted water level points for purposes such as creating hazard maps, it is important to accurately predict the relationship between high and low water levels at the predicted points.

[0009] The embodiments of the present invention have been made in consideration of the above-mentioned circumstances, and aim to provide a water level prediction device, a water level prediction method, and a computer program that accurately predict water levels in sewer pipes while correctly maintaining the relationship between water levels at multiple points. [Means for solving the problem]

[0010] A water level prediction device according to an embodiment includes a water level relationship feature extraction unit that receives water level information measured at multiple prediction points in a sewer pipe and calculates water level relationship features that indicate the water level relationship between the prediction points, and multiple water level learning and prediction units that use meteorological information, actual precipitation values, the water level information at the multiple prediction points, and time series data of the water level relationship features, structural information of the sewer pipe, and topographical information to learn multiple water level prediction models that calculate water level prediction values ​​for each of the multiple prediction points, and calculates water level prediction values ​​for each of the multiple prediction points using the learned water level prediction models. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 is a block diagram illustrating an example of the configuration of a water level prediction device according to the first embodiment. [Diagram 2] FIG. 2 is a diagram showing an example of the connection relationship between a plurality of points where water levels are measured. [Diagram 3] FIG. 3 is a diagram showing an example of actual measured water level values ​​and water level relationship feature amounts at the multiple points shown in FIG. [Figure 4] FIG. 4 is a diagram for explaining an example of a learned water level prediction model used in the water level learning and prediction unit of the water level prediction device shown in FIG. [Diagram 5] FIG. 5 is a diagram for explaining an example of a water level prediction model used in the water level learning and prediction unit of the water level prediction device shown in FIG. [Figure 6] FIG. 6 is a diagram illustrating an example of a manner in which the display unit displays a predicted water level value. [Figure 7] FIG. 7 is a flowchart for explaining an example of the operation of the water level prediction device of the first embodiment. [Figure 8] FIG. 8 is a block diagram illustrating an example of the configuration of the water level prediction device according to the second embodiment. [Figure 9] FIG. 9 is a diagram showing an example of the connection relationship between a plurality of points where water levels are measured. [Figure 10] FIG. 3 is a diagram showing an example of actual measured water levels at the multiple points shown in FIG. 2 and a water level relationship. [Figure 11] FIG. 11 is a diagram for explaining an example of the operation of the connection relationship reflecting unit of the water level prediction device of the second embodiment. [Figure 12] FIG. 12 is a flowchart for explaining an example of the operation of the water level prediction device of the second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Hereinafter, a water level prediction device, a water level prediction method, and a computer program according to embodiments will be described with reference to the drawings. The water level prediction device of this embodiment includes at least one processor, a memory in which a program executed by the processor is stored, and an auxiliary storage unit capable of storing various data. The water level prediction device can realize the various functions described below by software or a combination of software and hardware.

[0013] FIG. 1 is a block diagram illustrating an example of the configuration of a water level prediction device according to the first embodiment. The water level prediction device 1 of this embodiment includes a water level relation feature extraction unit 10 and a water level learning and prediction unit 20A-20Z for a plurality of positions AZ. The water level prediction device 1 of this embodiment is configured to be able to acquire various information from the outside, such as an external weather radar (external database DB1), analyzed rainfall data (external database DB2), water level data (external database DB3), sewer pipe structure information (external database DB4), and topographical information (external database DB5).

[0014] The water level relationship feature extraction unit 10 analyzes the water level relationship (e.g., the relationship between high and low water levels) from time series data (water level data) of water levels (actual measured values) input from outside, and outputs the water level relationship feature to the water level learning unit prediction unit 20.

[0015] The actual water level data input to the water level relation feature extraction unit 10 is a value measured by a water level gauge installed at multiple points in the sewer pipe. The water level value measured by the water level gauge may be the height from the bottom of the sewer pipe, or may be the height from any reference position such as 0 m above sea level. The water level relation feature extraction unit 10 may acquire the structure information of the sewer pipe from the sewer pipe structure information DB4. In this case, the water level relation feature extraction unit 10 may adjust the actual water level data taking into account the structure of the sewer pipe (e.g., the position of the pipe bottom or the reference position) at each point where the water level gauge is installed. The structure information of the sewer pipe may be stored in advance in an auxiliary storage unit (not shown). In this case, it is not necessary to acquire the structure information of the sewer pipe from the external sewer pipe structure information DB4.

[0016] FIG. 2 is a diagram showing an example of the connection relationship between a plurality of points where water levels are measured. FIG. 3 is a diagram showing an example of actual measured water levels at the multiple points shown in FIG. 2 and a water level relationship. Here, we will explain an example of the water level relationship (high-low water level relationship) between multiple points A, B, and C that are connected as shown in Figure 2. Point A is located upstream of points B and C. Points B and C are each located downstream of point A and are connected in parallel to point A.

[0017] When the measured water level at point A at a certain time is 3.1 m, the measured water level at point B at the same time is 2.8 m, and the measured water level at point C at the same time is 4.3 m, the relationship between the water levels at these points can be expressed by three values ​​(e.g., -1, 0, 1) as shown in FIG. 3. Here, the relationship between the water levels of multiple points A, B, and C is expressed as a 3-row, 3-column matrix in which each row corresponds to each of the multiple points A, B, and C. Each element of the matrix of the relationship is "-1" when the water level at the point corresponding to the column is higher than the water level at the point corresponding to the row, "1" when the water level at the point corresponding to the column is lower than the water level at the point corresponding to the row, and "0" when the water level at the point corresponding to the column is the same as the water level at the point corresponding to the row.

[0018] For example, for the element in row A, column A, the water level at the point corresponding to the row and the water level at the point corresponding to the column are equal, and the level relationship is set to "0." For the element in row A, column B, when comparing the water level at point A with the water level at point B, the water level at point A corresponding to the row is higher than the water level at point B corresponding to the column, and the level relationship is set to "1." For example, for the element in row A, column C, when comparing the water level at point A with the water level at point C, the water level at point A corresponding to the row is lower than the water level at point B corresponding to the column, and the level relationship is set to "-1." The water level relationship may be expressed as two values ​​(for example, -1 and 1, or 0 and 1), or the amount of height can be set as discrete values ​​and expressed as multiple values, such as four values.

[0019] The water level relation feature extraction unit 10 outputs the value of the water level relation (high-low relation) calculated from the time series data of the water level as described above as time series data of the water level relation feature. The water level relation feature extraction unit 10 may calculate the water level relation feature only when generating learning data (input data and teacher data) for the water level prediction model. When the time series data of the water level relation feature is used as input data for the water level prediction model when calculating the water level prediction value, the water level relation feature extraction unit 10 may calculate the water level relation feature when generating the learning data and when calculating the water level prediction value. The water level relation feature calculated by the water level relation feature extraction unit 10 is output as time series data associated with time information. The water level relation feature extraction unit 10 may record the calculated time series data of the water level relation feature in an auxiliary storage unit (not shown).

[0020] The water level learning and prediction units 20A-20Z are provided corresponding to each of the multiple water level prediction points AZ. Each of the multiple water level learning and prediction units 20A-20Z has multiple water level prediction models that calculate the water level prediction value of each water level prediction point AZ. The water level prediction model may be a model generated by machine learning using, for example, a genetic algorithm or a neural network, or may be a physical model.

[0021] 4 and 5 are diagrams for explaining an example of a learned water level prediction model used in the water level learning and prediction unit of the water level prediction device shown in FIG. An example of a water level prediction model using a neural network for machine learning is shown in Figures 4 and 5. The water level prediction model shown in Figure 4 has an input layer, multiple intermediate layers, and an output layer, and outputs a water level prediction value at the corresponding prediction point AZ in response to the input of various time-series data.

[0022] The water level prediction model shown in Fig. 4 is a machine-learned model that has been machine-learned using learning data including, for example, weather radar information, analyzed rainfall information, and water level information in sewer pipes (time series data of actual water level measurements), which are time-synchronized time series data with time stamps input from external databases DB1-DB5, sewer pipe structure information, topographical information, and time series data of water level-related features. Note that these learning data are merely examples, and other time series data may be used as learning data, and any of the above time series data may be omitted.

[0023] The past, present and future weather information input to the water level prediction model may include time series information on the distribution of rainfall and snowfall in an area observed by weather radar. The analyzed rainfall data is, for example, time series data on precipitation distribution obtained by combining observation data by weather radar and data from rain gauges across the country to analyze the amount of precipitation per unit time for each unit area. The water level information in the sewer pipe is time series data on the actual water level measured by water level gauges installed at multiple points (points A to Z) in the sewer pipe. The sewer pipe structure includes information such as pipe diameter and gradient at multiple points in the sewer pipe. The topography information includes information such as elevation and gradient at multiple points.

[0024] The water level prediction model is a machine learning model that is trained using the time series data of the actual water level (water level data) and the water level related feature amount of the time series data input from the water level related feature extraction unit 10 as teacher data, and outputs at least a water level predicted value after a predetermined time. The water level prediction model may be a machine learning model that uses the time series data of the actual water level and the water level related feature amount of the time series data input from the water level related feature extraction unit 10 as teacher data, and outputs time series data of the water level predicted value after a predetermined time and the predicted value of the water level related feature amount after the predetermined time. By using the water level data (actual measurement value) and the water level related feature amount after a predetermined time of the actual water level and the water level related feature amount used as input data for the water level prediction model as teacher data, learning with teacher data can be performed so that the water level prediction model outputs the water level predicted value and the predicted value of the water level related feature amount after a predetermined time.

[0025] When using a neural network, the water level and the water-level-related feature amount may be simultaneously learned from the learning data using, for example, a neural network structure as shown in Fig. 5. In addition, the learning model that outputs the predicted value of the water-level-related feature amount may be generated as a learning model other than the learning model that outputs the water level predicted value.

[0026] In addition, when training a machine learning model using a genetic algorithm or the like, for example, the water level and the water-level-related feature may be trained simultaneously by incorporating not only the water level error (e.g., squared error) but also the error of the water-level-related feature (e.g., cross-entropy loss) into the evaluation function.

[0027] When the water level learning and prediction unit 20A-20Z calculates a water level prediction value at a prediction point AZ using a learned model, it inputs the time-synchronized time series data with time stamps, namely weather radar information, analyzed rainfall information, and water level data information in the sewer pipe, as well as sewer pipe structure information and topography information, into the learned model, and outputs the time series data of the water level prediction value to the display unit 30 (external).

[0028] The water level learning and prediction units 20A-20Z may output the time series data of the predicted value of the water level related feature amount to the display unit 30 together with the time series data of the water level predicted value calculated by the learned water level prediction model. When the time series data of the predicted value of the water level related feature amount is not necessary, the time series data of the predicted value of the water level related feature amount does not have to be output from the water level learning and prediction units 20A-20Z to the display unit 30.

[0029] The water level learning prediction unit 20A-20Z may further positively utilize the time series data of the predicted value of the water level relation characteristic amount to correct the time series data of the water level predicted value. In other words, when the high-low relationship of the water level predicted value does not match the high-low relationship of the predicted value of the water level relation characteristic amount, the water level predicted value may be corrected based on the predicted value of the water level relation characteristic amount.

[0030] For example, when the difference in elevation between water level predicted values ​​that does not match the elevation relationship of the predicted values ​​of the water-level relationship feature is below a threshold, the water level learning prediction unit 20A-20Z may correct the water level prediction value by assuming that the elevation relationship indicated by the predicted value of the water-level relationship feature is correct, by multiplying the water level at a point where the water level is predicted to be high based on the predicted value of the water-level relationship feature by a coefficient greater than 1, and by multiplying the water level at a point where the water level is predicted to be low based on the predicted value of the water-level relationship feature by a coefficient less than 1.

[0031] The display unit 30 can display time series data of the water level prediction value and the prediction value of the water level relation characteristic amount output from the water level learning and prediction unit 20A-20Z. The display unit 30 may display the time series data of the water level prediction value and the prediction value of the water level relation characteristic amount as numerical values, or may display values ​​such as a graph in a visually recognizable manner.

[0032] FIG. 6 is a diagram illustrating an example of a manner in which the display unit displays a predicted water level value. The display unit 30 may display the water level up to the present and the future water level prediction value for each water level prediction point in a graph as shown in Fig. 6, for example, using the time series data of the water level prediction value output by the water level learning and prediction units 20A-20Z. The graph in Fig. 6 has time on the horizontal axis and water level on the vertical axis, and shows the time change of the water level up to the present at any water level prediction point and the time change of the future water level prediction value. Note that the display of the water level prediction value shown in Fig. 6 is one example, and the display mode of the water level prediction value on the display unit 30 is not limited to this. The display unit 30 may display the time series data of the prediction value of the water level-related feature amount in addition to the time series data of the water level prediction value.

[0033] As described above, the water level prediction device of this embodiment can correctly predict the high and low water level relationships at multiple water level prediction points by using a trained model that uses the water level relationship feature values ​​as learning data. As a result, the high and low water level relationships at multiple water level prediction points can be correctly maintained, so that it is possible to calculate highly accurate sewer water level prediction values ​​for points other than the prediction points by correction processing.

[0034] Next, an example of the operation of the water level prediction device of this embodiment will be described. FIG. 7 is a flowchart for explaining an example of the operation of the water level prediction device of the first embodiment. First, the water-level relation feature extraction unit 10 calculates the water-level relation feature amount (step SA1). That is, the water-level relation feature extraction unit 10 calculates and outputs the water-level relation feature amount of the time-series data from the water-level measurement value, which is the time-series data input from the outside. At this time, the water-level relation feature extraction unit 10 outputs the time-series data in which the value of the water-level relation feature amount is associated with the time information so as to be synchronized with the time information of the input time-series data.

[0035] Next, the water level learning and prediction units 20A-20Z each use the time-synchronized time series data with time stamps input from outside, which is a combination of weather radar, analyzed rainfall, water level data in the sewer pipe, sewer pipe structure, and topographical information, as well as the time series data of water-level-related features input from the water-level-related feature extraction unit 10, as learning data, and learn a water level prediction model that outputs water level prediction values ​​and prediction values ​​of water-level-related features for each water level prediction point AZ (step SA2).

[0036] When the learning of the water level prediction model is completed, the water level learning and prediction unit 20A-20Z uses various time series data input from outside as input data and calculates the water level prediction value and the prediction value of the water level related feature amount using the learned water level prediction model (step SA3). That is, the water level learning and prediction unit 20A-20Z inputs the time-synchronized time series data with time stamps, such as weather radar, analyzed rainfall, water level data in the sewer pipe (water level information), and data combining the water level related feature amount, information on the sewer pipe structure, and topographical information, into the learned water level prediction model, and calculates the data in time series of the water level prediction value and the prediction value of the water level related feature amount. The water level learning and prediction unit 20A-20Z may correct the water level prediction value using the calculated prediction value of the water level related feature amount.

[0037] The water level learning and prediction units 20A-20Z output the calculation results based on the learned water level prediction model to the outside (for example, the display unit 30). The water level learning and prediction units 20A-20Z may be configured to record the time series data calculated based on the learned water level prediction model in an auxiliary storage unit (not shown) and accumulate the time series data of the calculated prediction value in the auxiliary storage unit.

[0038] According to the water level prediction device and water level prediction method of the present embodiment, a machine learning model that corresponds to the complex phenomena leading up to flooding and changes in precipitation patterns due to climate change can be generated by using a trained water level prediction model that has undergone machine learning using various time series data such as water level data and water level relationship features. As a result, according to the water level prediction device and water level prediction method of the present embodiment, the high and low water levels of multiple prediction points can be correctly predicted using an individual water level prediction model for each of multiple prediction points that has learned the high and low water level relationships.

[0039] Furthermore, according to the water level prediction device and the water level prediction method of this embodiment, the errors of the respective water level prediction models can be corrected based on the high and low relationship of the predicted water level values.

[0040] Furthermore, according to the water level prediction device and water level prediction method of this embodiment, it is possible to correctly maintain the high-low relationship between the water level prediction values ​​of multiple water level prediction points as described above, and it is possible to calculate a highly accurate water level prediction value in a sewer pipe by interpolating the water levels at points other than the prediction point using the water level prediction value calculated by the water level prediction device of this embodiment.

[0041] That is, according to this embodiment, it is possible to provide a water level prediction device, a water level prediction method, and a computer program that accurately predict the water level in a sewer pipe while correctly maintaining the relationship between water levels at multiple points.

[0042] Next, a water level prediction device, a water level prediction method, and a computer program according to a second embodiment will be described in detail with reference to the drawings. In the following description, the same components as those in the first embodiment will be denoted by the same reference numerals and will not be described.

[0043] FIG. 8 is a block diagram illustrating an example of the configuration of the water level prediction device according to the second embodiment. The water level prediction device 1 of this embodiment differs from the first embodiment described above in that a connection relationship reflecting unit 11 is provided.

[0044] FIG. 9 is a diagram showing an example of the connection relationship between a plurality of points where water levels are measured. This diagram shows the physical connections between multiple points AE in a sewer pipe. Point A is located upstream of points B and C. Points B and C are both located downstream of point A and are connected to point A in parallel. Point D is located upstream of point E. The group of points AC and the group of point DE are not connected.

[0045] FIG. 10 is a diagram showing an example of actual measured water level values ​​and water level relationship feature amounts at the multiple points shown in FIG. When the actual water level measurement value at point A at a certain time is 3.1 m, the actual water level measurement value at point B at the same time is 2.8 m, the actual water level measurement value at point C at the same time is 4.3 m, the actual water level measurement value at point D at the same time is 1.5 m, and the actual water level measurement value at point E at the same time is 2.4 m, the high and low water level relationship at these points can be expressed by three values ​​(e.g., -1, 0, 1) as shown in Figure 10. Note that the connection state of multiple points AE is not taken into consideration here, and only the high and low water level relationship is shown.

[0046] 10, the elevation relationship of the multiple points AE is represented, for example, as a 5-by-5 ​​matrix in which each of the multiple points AE corresponds to each row and column. Each element of the elevation relationship matrix is ​​"-1" when the water level at the point corresponding to the column is higher than the water level at the point corresponding to the row, "1" when the water level at the point corresponding to the column is lower than the water level at the point corresponding to the row, and "0" when the water level at the point corresponding to the column is the same as the water level at the point corresponding to the row, by comparing the water level at the point corresponding to the column with the water level at the point corresponding to the row.

[0047] For example, for the element in row A, column A, the water level at the point corresponding to the row and the water level at the point corresponding to the column are equal for both points A, and the level relationship is set to "0." For the element in row A, column D, when comparing the water level at point A with the water level at point D, the water level at point A corresponding to the row is higher than the water level at point D corresponding to the column, and the level relationship is set to "1." For example, for the element in row D, column B, when comparing the water level at point D with the water level at point B, the water level at point D corresponding to the row is lower than the water level at point B corresponding to the column, and the level relationship is set to "-1." As in the first embodiment described above, the water level relationship may be expressed as two values ​​(e.g., -1, 1), or the amount of height can be set as discrete values ​​and expressed as multiple values ​​such as four values.

[0048] The water level relation feature extraction unit 10 outputs the value of the high-low relation calculated from the water level data of the time series data as described above as a time series water level relation feature. The time series data of the water level relation feature output from the water level relation feature extraction unit 10 is supplied to the connection relation reflection unit 11.

[0049] The water-level relation feature extraction unit 10 only needs to calculate the water-level relation feature when generating learning data for the water-level prediction model. The time series data calculated by the water-level relation feature extraction unit 10 may be recorded in an auxiliary storage device as time series data with a time stamp associated with time information.

[0050] The connection relationship reflecting unit 11 receives sewer pipe connection information included in the sewer pipe structure information DB4 and time series data of water level relation feature. The sewer pipe connection information is, for example, information indicating the physical connection relationship between multiple prediction points of a sewer pipe. The connection relationship reflecting unit 11 acquires, for example, the connection relationship of point AE shown in FIG. 9 from the sewer pipe structure information DB4 as connection information based on a tree structure. The connection information based on the tree structure may include information on the upstream and downstream sides in addition to connection information between points (information between the tree structure to which point AC is connected and the tree structure to which point DE is connected). The connection relationship reflecting unit 11 refers to the connection information of the sewer pipes and calculates a connection water-level relationship feature amount based on the water-level relationship feature amount input from the water-level relationship feature extracting unit 10.

[0051] Fig. 11 is a diagram for explaining an example of the operation of the connection relationship reflecting unit of the water level prediction device of the second embodiment. Fig. 11 shows an example of the connection water level relationship feature value when the sewer pipe connection information shown in Fig. 9 and the water level relationship feature value shown in Fig. 10 are input to the connection relationship reflecting unit 11.

[0052] The connection relationship reflecting unit 11 performs processing such as grouping the water level relationship feature values ​​indicating the high and low water level relationship between the prediction points that are physically connected, and invalidating (NA: not available) the water level relationship feature values ​​indicating the high and low water level relationship between the prediction points that are not physically connected, and calculates the connection water level relationship feature values. For example, in the example shown in FIG. 11, a plurality of water level relationship feature values ​​(3 rows and 3 columns) indicating the high and low water level relationship between the points AC are grouped, a plurality of water level relationship feature values ​​(2 rows and 2 columns) indicating the high and low water level relationship between the points DE are grouped, and a water level relationship feature value indicating the high and low water level relationship between any of the points AC and any of the points DE is invalid (NA). That is, the time series data of the connection water level relationship feature values ​​may include the value of the connection water level relationship feature values ​​between the water level prediction points and the information of the group set based on the connection information of the points.

[0053] The connection relationship reflecting unit 11 supplies the time series data of the calculated connection water level relationship feature amount to the water level learning and prediction units 20A-20Z. The water level learning and prediction units 20A-20Z are provided corresponding to each of the plurality of water level prediction points AZ, similarly to the first embodiment described above. The water level learning and prediction units 20A-20Z have a water level prediction model for the water level prediction point AZ.

[0054] In this embodiment, the time series data of the connection water level relationship feature is used as the learning data of the water level prediction model instead of the time series data of the water level relationship feature. That is, the water level learning and prediction unit 20A-20Z is a machine-learned model that has been machine-learned using, for example, time-synchronized time series data with time stamps input from the external databases DB1-DB5, such as weather radar information, analyzed rainfall information, water level data information in the sewer pipe, and learning data of the connection water level relationship feature, sewer pipe structure information, and topographical information. Each water level prediction model has an input layer, multiple intermediate layers, and an output layer, and outputs a water level prediction value at the corresponding prediction point AZ in response to input of time series data such as weather radar information, analyzed rainfall information, and water level data information in the sewer pipe. The learned water level prediction model may be configured to output time series data of the prediction value of the connection water level relationship feature together with the time series data of the water level prediction value.

[0055] For example, the water level prediction model is a machine learning model that is trained using time series data of the actual water level measurement value and time series data of the connection water level relation feature as teacher data, and outputs at least a water level prediction value after a predetermined time. The water level prediction model may also be a machine learning model that uses time series data of the actual water level measurement value and time series data of the connection water level relation feature as teacher data, and outputs time series data of the water level prediction value after a predetermined time and the prediction value of the connection water level relation feature after a predetermined time. By using the water level data (actual measurement value) and the water level relation feature after a predetermined time as teacher data for the actual water level measurement value and the connection water level relation feature used as input data for the water level prediction model, learning with teacher data can be performed so that the water level prediction model outputs the water level prediction value and the prediction value of the connection water level relation feature after a predetermined time.

[0056] When the water level learning and prediction unit 20A-20Z calculates a water level prediction value at a prediction point AZ using a learned water level prediction model, it inputs, for example, time-synchronized time series data with input time stamps, such as weather radar information, analyzed rainfall information, and water level data information in the sewer pipe, sewer pipe structure information, and topography information, into the learned model, and outputs at least the time series data of the water level prediction value to the display unit 30 (external).

[0057] The water level learning and prediction unit 20A-20Z may output time series data of the water level prediction value and the connection water level relation feature amount calculated by the learned water level prediction model to the display unit 30 (external). When the time series data of the prediction value of the connection water level relation feature amount is not necessary, the water level learning and prediction unit 20A-20Z does not need to output the time series data of the prediction value of the connection water level relation feature amount to the display unit 30.

[0058] The water level learning prediction unit 20A-20Z may further positively utilize the time series data of the predicted value of the connection water level relation characteristic amount to correct the time series data of the water level predicted value. In other words, when the high-low relationship of the water level predicted value does not match the high-low relationship of the predicted value of the connection water level relation characteristic amount, the water level predicted value may be corrected based on the predicted value of the connection water level relation characteristic amount.

[0059] For example, when the elevation difference between water level predicted values ​​that do not match the elevation relationship of the predicted values ​​of the connection water-level relationship feature is below a threshold, the water level learning prediction unit 20A-20Z may assume that the elevation relationship indicated by the predicted value of the connection water-level relationship feature is correct, and may correct the water level prediction value by multiplying the water level at a point where the water level is predicted to be high based on the predicted value of the connection water-level relationship feature by a coefficient greater than 1, and multiplying the water level at a point where the water level is predicted to be low based on the predicted value of the connection water-level relationship feature by a coefficient less than 1.

[0060] Next, an example of the operation of the water level prediction device of this embodiment will be described. FIG. 12 is a flowchart for explaining an example of the operation of the water level prediction device of the second embodiment. First, the water-level relation feature extraction unit 10 calculates the water-level relation feature amount (step SB1). That is, the water-level relation feature extraction unit 10 calculates and outputs the water-level relation feature amount of the time-series data from the water-level data, which is time-series data input from the outside. At this time, the water-level relation feature extraction unit 10 outputs time-series data in which the value of the water-level relation feature amount is associated with time information so as to be synchronized with the time information of the input time-series data.

[0061] Next, the connection relationship reflection unit 11 refers to the sewer pipe connection information input from outside, calculates time series data of the connection water level relationship feature based on the time series data of the water level relationship feature input from the water level relationship feature extraction unit 10, and outputs it to the water level learning unit prediction unit 20A-10Z (step SB2).

[0062] Next, the water level learning and prediction units 20A-20Z each use the time-synchronized time series data with time stamps input from outside, which is a combination of weather radar, analyzed rainfall, water level data in the sewer pipe, sewer pipe structure, and topographical information, as well as the time series data of the connection water-level relationship features input from the connection relationship reflection unit 11, as learning data, and learn a water level prediction model that outputs water level prediction values ​​and prediction values ​​of the connection water-level relationship features for each water level prediction point AZ (step SB3).

[0063] When the learning of the water level prediction model is completed, the water level learning and prediction unit 20A-20Z uses various time series data input from outside as input data and calculates the water level prediction value and the prediction value of the water level related feature amount using the learned water level prediction model (step SB4). That is, the water level learning and prediction unit 20A-20Z inputs the time-synchronized time series data with time stamps, such as weather radar, analyzed rainfall, water level data in the sewer pipe, and the connected water level related feature amount, information on the sewer pipe structure, and topographical information, into the learned water level prediction model, and calculates the data in time series of the water level prediction value and the connected water level related feature amount after a predetermined time. The water level learning and prediction unit 20A-20Z may correct the water level prediction value using the calculated connected water level related feature amount.

[0064] The water level learning and prediction units 20A-20Z output the calculation results based on the learned water level prediction model to the outside (for example, the display unit 30). The water level learning and prediction units 20A-20Z may be configured to record the time series data calculated based on the learned water level prediction model in an auxiliary storage unit (not shown) and accumulate the time series data of the calculated prediction value in the auxiliary storage unit.

[0065] According to the water level prediction device and water level prediction method of the present embodiment, as in the first embodiment, the high and low water level relationships of a plurality of water level prediction points can be correctly predicted by using a machine learning model that has learned the high and low water level relationships. As a result, it is possible to correctly maintain the high and low water level prediction values ​​of a plurality of water level prediction points, and the water level prediction value calculated by the water level prediction device of this embodiment can be used to calculate a highly accurate water level prediction value in a sewer pipe by correction processing even at points other than the prediction point.

[0066] Moreover, according to the water level prediction device of the present embodiment, the amount of data of the water level relation feature used for learning the water level prediction model is reduced, and only the water level relation feature between points that are physically connected to each other is used as learning data, so that the water level prediction model can effectively learn the water level relation feature. As a result, in the water level prediction device of the present embodiment, the water level learning model becomes compact, the amount of memory required for learning and calculation can be reduced, and the prediction accuracy of the water level can be improved.

[0067] As described above, the water level prediction device of this embodiment can correctly predict the relationship between the water levels of multiple water level prediction points by using a trained model that uses water level feature values ​​as learning data. As a result, the relationship between the water levels of multiple water level prediction points can be correctly maintained, so that it is possible to calculate highly accurate water level prediction values ​​in sewer pipes by correction processing for points other than the prediction points.

[0068] That is, according to this embodiment, it is possible to provide a water level prediction device, a water level prediction method, and a computer program that perform highly accurate prediction of water levels in sewer pipes while correctly maintaining the relationship between water levels at multiple points.

[0069] 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.

[0070] 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]

[0071] 1...water level prediction device, 10...water level related feature extraction unit, 11...connection relationship reflection unit, 20A-10Z...water level learning and prediction unit, 30...display unit, AZ...water level prediction point, DB1-DB5...database (external database)

Claims

1. a water level relationship feature extraction unit that receives water level information measured at a plurality of prediction points in a sewer pipe and calculates a water level relationship feature that indicates a water level relationship between the prediction points; a plurality of water level learning and prediction units that learn a plurality of water level prediction models that calculate a predicted water level value for each of the plurality of prediction points using meteorological information, actual precipitation values, the water level information at the plurality of prediction points, and time series data of the water level-related features, structural information of the sewer pipe, and topographical information, and that calculate a predicted water level value for each of the plurality of prediction points using the learned water level prediction models.

2. 2. The water level prediction device according to claim 1, wherein the water level relationship feature extraction unit further receives structural information of the sewer pipe and calculates the water level relationship between the prediction points taking into account the structural information.

3. The water level prediction device according to claim 1 , wherein the water level learning and prediction unit learns the water level prediction model using time series data of the water level information and the water level related feature amount at a plurality of the prediction points as teacher data.

4. The water level prediction device according to claim 1 , wherein the trained water level prediction model calculates and outputs the water level prediction value and the prediction value of the water level relationship feature amount.

5. The water level prediction device according to claim 4 , wherein the water level learning and prediction unit is capable of correcting the water level prediction value based on the predicted value of the water level related feature amount.

6. A connection relationship reflecting unit receives the water level relationship feature and connection information indicating a physical connection relationship between the plurality of prediction points of the sewer pipe, and outputs a connection water level relationship feature extracted from the water level relationship feature between the physically connected prediction points based on the connection information, The water level prediction device according to claim 1 , wherein the water level learning and prediction unit learns the water level prediction model by using the connected water level relation feature amount instead of the water level relation feature amount.

7. The water level prediction device according to claim 6, wherein the connection relationship reflection unit calculates the connection water level relationship feature value having a reduced amount of data compared to the water level relationship feature value by invalidating the water level relationship feature value between the prediction points that are not physically connected.

8. Water level information measured at a plurality of prediction points in a sewer pipe is input, and a water level relationship feature quantity indicating a water level relationship between the prediction points is calculated; learning a plurality of water level prediction models that calculates a water level prediction value for each of the plurality of prediction points using meteorological information, actual precipitation measurements, the water level information at the plurality of prediction points, time series data of the water level related feature amount, structural information of the sewer pipe, and topographical information; A water level prediction method, comprising: calculating a water level prediction value for each of a plurality of prediction points using the trained water level prediction model.

9. A computer program for causing a computer to execute the water level prediction method according to claim 8.

Images