Method for constructing a water level prediction model, program for constructing a water level prediction model, method for outputting water level prediction data, program for outputting water level prediction data, method for interpolating test data, and method for calculating the accuracy of water level predictions.

Abstract

The objective is to improve the accuracy of water level predictions by using a water level prediction model, taking into account the time required from rainfall in the river basin to water flow into the river. [Solution] As training data necessary for learning the water level at a certain time at a water level prediction point in a river area, at least one of the following is extracted from the rainfall data RD: (1) Total rainfall data showing the total rainfall from the start of rainfall to that time, in the past before that time, and in the future before that time at least one of the following: (however, it is reset to 0 when a predetermined time has elapsed from the start of no rainfall) and (2) Average rainfall data showing the average rainfall within a predetermined average period from that time, in the past before that time, and in the future before that time at least one of the following:

Description

[Technical Field]

[0001] This disclosure relates to a technology for learning and predicting water levels at water level prediction points in river areas. [Background technology]

[0002] A technique for learning and predicting water levels at water level prediction points in a river area using machine learning models such as neural networks is disclosed in Patent Document 1, etc.

[0003] First, a water level prediction model is constructed using training data. Specifically, as training data necessary to learn the water level at a given time at a water level prediction point, at least one of the following is extracted: (1) water level data showing the water level at the time and at least one of the past time at at least one of the water level prediction point and upstream of the water level prediction point; and (2) rainfall data showing the rainfall at the time, at least one of the past time and at least one of the future time at at least one of the water level prediction point and upstream of the water level prediction point. Then, a water level prediction model is constructed to predict the water level at the water level prediction point using at least one of the water level data and rainfall data.

[0004] Next, the water level at the prediction point is predicted using the water level prediction model constructed with training data. Specifically, as test data necessary to predict the water level at a given time at the prediction point, at least one of the following is input into the water level prediction model: (1) water level data showing the water level at the time and at least one of the past times at at least one of the prediction point and upstream of the prediction point, and (2) rainfall data showing the rainfall at the time, at least one of the past times and at least one of the future times at at least one of the prediction point and upstream of the prediction point. The water level prediction model then outputs the predicted water level data for that time at the prediction point. [Prior art documents] [Patent Documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2019-095240 [Overview of the Initiative] [Problems that the invention aims to solve]

[0006] Conventional techniques extract instantaneous rainfall data (analyzed rainfall, etc.) that shows the instantaneous rainfall at at least one of the following locations: the water level prediction point or a point upstream of the water level prediction point, at least one of the following instantaneous times: the time in question, the time in the past, or the time in the future.

[0007] However, instantaneous rainfall data does not take into account the time required from rainfall in the river area to its inflow into the river. Therefore, it is not possible to improve the accuracy of water level predictions by using a water level prediction model that takes into account the time required from rainfall in the river area to its inflow into the river.

[0008] Therefore, in order to solve the aforementioned problems, this disclosure aims to improve the accuracy of water level predictions by using a water level prediction model, taking into account the time required from rainfall in the river area to inflow into the river. [Means for solving the problem]

[0009] To solve the aforementioned problem, as training data necessary for learning the water level at a given time at a water level prediction point in a river area, at least one of the following is extracted from the rainfall data: (1) total rainfall data showing the total rainfall from the start of rainfall to that time, in the past before that time, and in the future before that time, at least one of the above

[0010] Specifically, this disclosure comprises, in order, a training data extraction step of extracting, as training data necessary for learning the water level at a water level prediction point in a river area, (1) water level data showing the water level at at least one of the water level prediction point and a point upstream of the water level prediction point at that time and in the past before that time, and (2) rainfall data showing the rainfall at at least one of the water level prediction point and a point upstream of the water level prediction point at that time and in the future before that time, and a water level prediction model construction step of constructing a water level prediction model for predicting the water level at the water level prediction point in the river area using the water level data and the rainfall data, wherein the training data extraction step includes, as the rainfall data, (1) the water level prediction point and the The method for constructing a water level prediction model is characterized by extracting at least one of the following: (1) instantaneous rainfall data showing the instantaneous rainfall at at least one of the times, past times, and future times, at least one of the water basins upstream of the water level prediction point; (2) total rainfall data showing the total rainfall from the start of rainfall to at least one of the times, past times, and future times, at least one of the times when rainfall has started (however, this is reset to 0 after a predetermined time has elapsed from the time when no rain falls); and (3) average rainfall data showing the average rainfall within a predetermined average period from at least one of the times, past times, and future times, at least one of the times when rainfall has started, at least one of the times when rainfall has started.

[0011] With this configuration, it is possible to construct a water level prediction model that improves the accuracy of water level prediction by using total rainfall data as the rainfall data and considering the time required from rainfall in the river area to inflow into the river. Furthermore, it is possible to construct a water level prediction model that improves the accuracy of water level prediction by using average rainfall data as the rainfall data, unlike with total rainfall data, without abruptly resetting the data to zero.

[0012] Furthermore, this disclosure provides a method for constructing a water level prediction model, characterized in that it further includes a water level prediction model selection step after the water level prediction model construction step, in which verification data for verifying the water level prediction model is input into multiple types of the water level prediction models constructed using multiple types of the training data, and a water level prediction model is selected that performs water level predictions that better meet the issuance criteria.

[0013] This configuration allows for the selection of the optimal water level prediction model from multiple water level prediction models constructed using multiple types of training data, such that the generalization performance is high near the water level peak (especially near the water level peak that serves as the basis for issuing evacuation advisories, etc.).

[0014] Furthermore, this disclosure is a water level prediction model construction program that causes a computer to sequentially execute each processing step of the water level prediction model construction method described above.

[0015] This configuration makes it possible to provide a program that has the effects described above.

[0016] Furthermore, this disclosure provides a water level prediction data output method, characterized in that, when predicting the water level at the water level prediction point in the river area using the water level prediction model constructed by the water level prediction model construction method described above, the method includes a water level prediction data output step in which the water level prediction model outputs water level prediction data indicating the predicted water level at the water level prediction point at that time, inputting the water level data and rainfall data into the water level prediction model as test data necessary for predicting the water level at a certain time, and outputting water level prediction data from the water level prediction model.

[0017] This configuration allows for improved water level prediction accuracy by using a water level prediction model constructed with total rainfall data, taking into account the time required from rainfall in the river basin to its inflow into the river. Furthermore, unlike with total rainfall data, it is possible to improve water level prediction accuracy by using a water level prediction model constructed with average rainfall data without abruptly resetting the data to zero.

[0018] In addition, the present disclosure is a water level prediction data output program for causing a computer to execute the water level prediction data output step included in the water level prediction data output method described above.

[0019] According to this configuration, a program having the effects described above can be provided.

[0020] In order to solve further problems, as test data, when the water level data at the water level prediction location is missing, in interpolating the water level data at the missing time at the water level prediction location, the rainfall and actual water level situation at a prediction time arbitrarily before the missing time is reflected and then output, and interpolation is performed using the water level prediction data at the missing time arbitrarily after the prediction time at the water level prediction location.

[0021] Specifically, the present disclosure provides a test data interpolation method characterized by including a test data interpolation step of interpolating the water level data at the missing time at the water level prediction location in the test data extracted by the water level prediction data output method described above, using the water level prediction data at the missing time arbitrarily after the prediction time at the water level prediction location, which is output by the water level prediction data output step at a prediction time arbitrarily before the missing time.

[0022] According to this configuration, even when the water level test data at the water level prediction location is missing, by reflecting the rainfall and actual water level situation and interpolating the water level test data at the water level prediction location, the water level prediction accuracy at the water level prediction location can be improved.

[0023] In order to solve further problems, as verification data, water level data, past observed rainfall data, and future forecast rainfall data are extracted, and the ratio of those that fall within the error range around the water level observation data among the water level prediction data is calculated as the water level prediction hit rate.

[0024] Specifically, this disclosure is a method for calculating the accuracy of a water level prediction, comprising, in order: a verification data extraction step of extracting water level data and past observed rainfall data and future forecast rainfall data as rainfall data, as verification data for verifying the water level prediction model constructed by the water level prediction model construction method described above; a water level prediction data output step of inputting the water level data, past observed rainfall data and future forecast rainfall data for verifying the water level at multiple times at the water level prediction point as verification data into the water level prediction model and outputting water level prediction data from the water level prediction model that shows the predicted water level at the water level prediction point at those multiple times; and a water level prediction accuracy calculation step of calculating the proportion of the water level prediction data for those multiple times at the water level prediction point that falls within the error range around the water level observation data at the water level prediction point at those multiple times as the accuracy of the prediction of the water level at the water level prediction point.

[0025] With this configuration, since future observed rainfall data cannot be extracted as test data during the prediction phase, even when future forecast rainfall data is extracted, the reliability of the water level prediction data can be visualized by calculating the water level prediction accuracy rate during the verification phase.

[0026] Furthermore, the inventions disclosed above can be combined as much as possible. [Effects of the Invention]

[0027] Thus, this disclosure shows that by taking into account the time required from rainfall in the river area to inflow into the river, it is possible to improve the accuracy of water level prediction using a water level prediction model. [Brief explanation of the drawing]

[0028] [Figure 1] This figure shows the configuration of the water level prediction device disclosed herein. [Figure 2] This figure shows the procedure for water level prediction processing in this disclosure. [Figure 3] This figure shows the total rainfall data and average rainfall data disclosed herein. [Figure 4] This figure shows the configuration of the water level prediction device disclosed herein. [Figure 5] This diagram shows the procedure for selecting a water level prediction model in this disclosure. [Figure 6] This figure shows a specific example of the synthesis process of training data as described in this disclosure. [Figure 7] This figure shows a specific example of the selection process for the water level prediction model disclosed herein. [Figure 8] This figure shows examples of prior art and the water level prediction process described herein. [Figure 9] This figure shows examples of prior art and the water level prediction process described herein. [Figure 10] This figure shows examples of prior art and the water level prediction process described herein. [Figure 11] This figure shows examples of prior art and the water level prediction process described herein. [Figure 12] This figure shows the configuration of the water level prediction device disclosed herein. [Figure 13] This figure shows the procedure for the test data interpolation process described herein. [Figure 14] This figure shows a specific example of the water level test data interpolation process described herein. [Figure 15] This figure shows a specific example of the rainfall test data interpolation process described herein. [Figure 16] This figure shows the configuration of the water level prediction device disclosed herein. [Figure 17] This diagram shows the procedure for calculating the accuracy rate of water level predictions as disclosed herein. [Figure 18] This figure shows a specific example of the water level prediction error range setting process disclosed herein. [Figure 19] This figure shows a specific example of the water level prediction accuracy calculation process disclosed herein. [Modes for carrying out the invention]

[0029] Embodiments of the present disclosure will be described with reference to the attached drawings. The embodiments described below are examples of the implementation of the present disclosure, and the present disclosure is not limited to these embodiments.

[0030] (Procedure for water level prediction processing in this disclosure) Figure 1 shows the configuration of the water level prediction device disclosed in this disclosure. Figure 2 shows the procedure for the water level prediction process disclosed in this disclosure. A technology for learning and predicting water levels at water level prediction points in a river area using a machine learning model such as a neural network is disclosed in the following description of this disclosure.

[0031] The water level prediction device W is equipped with a rainfall data conversion unit 1, a training data extraction unit 2, and a water level prediction model construction unit 3 to execute the water level prediction model construction method (steps S1 to S3). It constructs a water level prediction model M and can be realized by installing the water level prediction model construction program (steps S1 to S3) shown in Figure 2 onto a computer.

[0032] The water level prediction device W is equipped with a rainfall data conversion unit 1, a test data extraction unit 4, and a water level prediction data output unit 5 to execute the water level prediction data output method (steps S4 to S6). It can be realized by applying the water level prediction model M and installing the water level prediction data output program (steps S4 to S6) shown in Figure 2 onto a computer.

[0033] In the following, we will first describe the training phase processing (steps S1-S3) of the water level prediction model M disclosed herein, then describe the prediction phase processing (steps S4-S6) of the water level prediction model M disclosed herein, and finally describe examples of the prior art and the water level prediction processing disclosed herein.

[0034] The training data extraction unit 2 extracts the following as training data necessary for learning the water level at a given time at a water level prediction point in the river area: (1) water level data WL showing the water level at at least one of the water level prediction point and at least one of the points upstream of the water level prediction point, at the time in question and at least one of the points in the past before that time; and (2) rainfall data RD showing the rainfall at at least one of the water level prediction point and at least one of the water basin upstream of the water level prediction point, at the time in question, at least one of the points in the past before that time and at least one of the points in the future (step S2). Here, for the rainfall data RD, "upstream water basin" may also include "upstream point only".

[0035] Here, it is desirable for the training data extraction unit 2 to extract rainfall and water level data from periods with a high correlation to flood water levels at water level prediction points in the river area, and to discard rainfall and water level data from periods with a low correlation.

[0036] For example, the training data extraction unit 2 may, as training data, (1) set a water level peak discrimination threshold that rejects periods of no water level fluctuation and then select a period for extracting water level data; (2) set a water level peak extraction width that rejects periods of no water level fluctuation and then select a period for extracting water level data; (3) set a predetermined water level period for extracting periods of rising and falling water levels and then select a period for extracting water level data; (4) set a predetermined rainfall period that takes into account the time it takes for rainfall to reach the water level prediction point and then select a period for extracting rainfall data; or (5) set a predetermined cumulative period that takes into account the time it takes for rainfall to flow from the soil or reservoir into the river and then select a period for extracting rainfall data (see Patent Document 1).

[0037] The water level prediction model construction unit 3 uses water level data WL and rainfall data RD to construct a water level prediction model M for predicting water levels at water level prediction points in the river area (step S3).

[0038] The test data extraction unit 4 extracts the following test data necessary for predicting the water level at a water level prediction point in the river area at a given time: (1) water level data WL showing the water level at at least one of the water level prediction point and at least one of the points upstream of the water level prediction point, at the time in question and at least one of the points in the past before that time; and (2) rainfall data RD showing the rainfall at at least one of the water level prediction point and at least one of the water basin upstream of the water level prediction point, at the time in question, at least one of the points in the past before that time and at least one of the points in the future (step S5). Here, for the rainfall data RD, "upstream water basin" may also include "upstream points only".

[0039] The water level prediction data output unit 5 inputs water level data WL and rainfall data RD as test data to the water level prediction model M (step S5), and outputs water level prediction data P from the water level prediction model M that shows the predicted water level at the water level prediction point in the river area at that time (step S6).

[0040] The total rainfall data and average rainfall data of this disclosure are shown in Figure 3. The rainfall data conversion unit 1 generates at least one of the following as rainfall data RD: (1) instantaneous rainfall data showing the instantaneous rainfall (analyzed rainfall, etc.) at at least one of the time, past time, and future time at at least one of the water level prediction point and the water basin upstream of the water level prediction point; (2) total rainfall data showing the total rainfall from the start of rainfall to at least one of the time, past time, and future time at at least one of the time, past time, and future time at at least one of the time, past time, and future time at at least one of the time, past time, and future time at at least one of the time, past time, and future time at at least one of the time, past time, and future time. (However, it is reset to 0 when a predetermined time has elapsed from the time when no rain falls); and (3) average rainfall data showing the average rainfall within a predetermined averaging period at at least one of the time, past time, and future time at at least one of the time, past time, and future time at at least one of the time, past time, and future time at at least one of the time, past time, and future time. The training data extraction unit 2 and the test data extraction unit 4 extract at least one of the following as rainfall data RD: instantaneous rainfall data, total rainfall data, and average rainfall data (steps S2, S5). Here, the instantaneous rainfall data, total rainfall data, and average rainfall data may be the "spatial average value for the entire upstream basin" or the "spatial sum for the entire upstream basin."

[0041] In the upper part of Figure 3, total rainfall data is extracted, showing the total rainfall from the start of rainfall to the time of concern (instantaneous rainfall data is indicated by a left arrow, and total rainfall data is indicated by a right arrow). However, the total rainfall data is reset to 0 after a predetermined time has elapsed from the start of no rainfall. Here, the "predetermined elapsed time" should be set to construct a water level prediction model M with improved water level prediction accuracy. In the lower part of Figure 3, average rainfall data is extracted, showing the average rainfall within a predetermined averaging period up to the time of concern (instantaneous rainfall data is indicated by a left arrow, and average rainfall data is indicated by a right arrow). Here, the "predetermined averaging period" should be set to construct a water level prediction model M with improved water level prediction accuracy. Note that since the total rainfall data is greater than the average rainfall data, it is desirable to standardize either the total rainfall data or the average rainfall data so that the water level prediction data P based on the total rainfall data is approximately equal to the water level prediction data P based on the average rainfall data.

[0042] Thus, by using total rainfall data as the rainfall data RD, a water level prediction model M can be constructed that improves the accuracy of water level prediction by considering the time required from rainfall in the river area to inflow into the river. Furthermore, by using average rainfall data as the rainfall data RD, a water level prediction model M can be constructed that improves the accuracy of water level prediction without abruptly resetting the data to zero, unlike with total rainfall data.

[0043] Furthermore, by considering the time required from rainfall in the river basin to inflow into the river, the water level prediction accuracy can be improved using a water level prediction model M constructed with total rainfall data as rainfall data RD. In contrast to total rainfall data, the water level prediction accuracy can also be improved using a water level prediction model M constructed with average rainfall data as rainfall data RD without abruptly resetting the data to zero.

[0044] (Procedure for selecting a water level prediction model in this disclosure) The configuration of the water level prediction device disclosed herein (in particular, the water level prediction model selection process) is also shown in Figure 4. The procedure for the water level prediction model selection process disclosed herein is shown in Figure 5. The water level prediction device W also includes a verification data extraction unit 6 and a water level prediction model selection unit 7, and the water level prediction model selection program shown in Figure 5 can also be realized by installing it on a computer.

[0045] The training data extraction unit 2 extracts multiple types of training data (step S11, see Figure 6). The water level prediction model construction unit 3 uses the multiple types of training data to construct multiple types of water level prediction models M1, ..., Mn (step S11).

[0046] The verification data extraction unit 6 extracts the following verification data necessary for verifying the water level at a water level prediction point in the river area at a given time: (1) water level data WL showing the water level at at least one of the water level prediction point and at least one of the points upstream of the water level prediction point, at that time and at least one of the points in the past and before that time; and (2) rainfall data RD showing the rainfall at at least one of the water level prediction point and at least one of the water basin upstream of the water level prediction point, at that time, at least one of the points in the past and before that time (step S12). Here, for the rainfall data RD, "upstream water basin" may also include "upstream point only". The extraction period for the verification data is the same as that for the training data.

[0047] The water level prediction model selection unit 7 inputs verification data for verifying the water level prediction model M into multiple types of water level prediction models M1, ..., Mn, which are constructed using multiple types of training data (step S12), and selects a water level prediction model (for example, M1) that better satisfies the issuance criteria (i.e., has a smaller prediction error near the water level peak) (step S13).

[0048] A specific example of the training data synthesis process described in this disclosure is shown in Figure 6. The training data extraction unit 2 (and the verification data extraction unit 6) must extract water level data WL and arbitrarily extract the converted rainfall data (at least one of instantaneous rainfall data, total rainfall data, and average rainfall data) as rainfall data RD.

[0049] In the upper left column of Figure 6, only the water level data WL from time t1 to time t2 is extracted as the first training data, and the rainfall data RD is not extracted. In the lower left column of Figure 6, the water level data WL and instantaneous rainfall data from time t1 to time t2 are extracted as the second training data. In the upper middle column of Figure 6, the water level data WL and total rainfall data from time t1 to time t2 are extracted as the third training data. In the lower middle column of Figure 6, the water level data WL, instantaneous rainfall data and total rainfall data from time t1 to time t2 are extracted as the fourth training data. In the upper right column of Figure 6, the water level data WL and average rainfall data from time t1 to time t2 are extracted as the fifth training data. In the lower right column of Figure 6, the water level data WL, instantaneous rainfall data and average rainfall data from time t1 to time t2 are extracted as the sixth training data.

[0050] A specific example of the water level prediction model selection process in this disclosure is shown in the left column of Figure 7. The water level prediction model selection unit 7 should select a water level prediction model (e.g., M1) when the predicted time of the water level rise period near the water level peak is the same as the actual time of the water level rise period. Even if the predicted time of the water level rise period near the water level peak is not the same as the actual time of the water level rise period, a water level prediction model (e.g., M1) should be selected according to the following policy.

[0051] In other words, as in Model B, the water level prediction model is considered superior when the predicted time of the water level rise period is earlier than the actual time of the water level rise period. On the other hand, as in Model A, the water level prediction model is considered inferior when the predicted time of the water level rise period is later than the actual time of the water level rise period. In this way, the possibility of delays in issuing evacuation advisories and other warnings is reduced.

[0052] A specific example of the water level prediction model selection process in this disclosure is also shown in the right column of Figure 7. The water level prediction model selection unit 7 should select a water level prediction model (e.g., M1) when the predicted water level near the water level peak (peak water level, etc.) is the same as the actual water level near the water level peak. Even if the predicted water level near the water level peak (peak water level, etc.) is not the same as the actual water level near the water level peak, a water level prediction model (e.g., M1) should be selected according to the following policy.

[0053] In other words, as in Model B, the water level prediction model is considered superior when the predicted water level near the peak is higher than the actual water level near the peak. On the other hand, as in Model A, the water level prediction model is considered inferior when the predicted water level near the peak is lower than the actual water level near the peak. In this way, the possibility of the criteria for issuing evacuation advisories not being met is reduced.

[0054] In this way, from multiple types of water level prediction models M1, ..., Mn constructed using multiple types of training data, the optimal water level prediction model M can be selected so as to enhance generalization performance near the water level peak (especially near the water level peak that serves as the basis for issuing evacuation advisories, etc.).

[0055] (Examples of prior art and water level prediction processing in this disclosure) Figure 8 shows examples of the conventional technology and the water level prediction processing described herein. In the conventional technology, a water level prediction model M is constructed using only water level data WL and analyzed rainfall data. In this disclosure, a water level prediction model M is constructed using water level data WL and total rainfall data.

[0056] The document then shows the temporal changes in actual water level data, water level prediction data P (water level prediction result 6 hours later), analyzed rainfall data, and total rainfall data during rainfall. In the conventional technology, the water level prediction data P exceeds the evacuation decision water level, unlike the actual water level data. On the other hand, in this disclosure, the water level prediction data P is approaching the evacuation decision water level, similar to the actual water level data.

[0057] Figure 9 also shows examples of the conventional technology and the water level prediction processing of this disclosure. In the conventional technology, a water level prediction model M is constructed using only water level data WL and analyzed rainfall data. In this disclosure, a water level prediction model M is constructed using water level data WL and analyzed total rainfall data.

[0058] The document then shows the temporal changes in actual water level data, water level prediction data P (water level prediction result 6 hours later), analyzed rainfall data, and total rainfall data during rainfall. In the conventional technology, the water level prediction data P exceeds the evacuation decision water level, unlike the actual water level data. On the other hand, in this disclosure, the water level prediction data P is approaching the evacuation decision water level, similar to the actual water level data.

[0059] Figure 10 also shows examples of the conventional technology and the water level prediction processing of this disclosure. In the conventional technology, a water level prediction model M is constructed using only water level data WL and analyzed rainfall data. In this disclosure, a water level prediction model M is constructed using water level data WL and mean rainfall data.

[0060] The document then shows the temporal changes in actual water level data, water level prediction data P (water level prediction result 6 hours later), analyzed rainfall data, and average rainfall data during rainfall. In the conventional technology, the water level prediction data P exceeds the evacuation decision water level, unlike the actual water level data. On the other hand, in this disclosure, the water level prediction data P slightly exceeds the evacuation decision water level, unlike the actual water level data.

[0061] Figure 11 also shows examples of the conventional technology and the water level prediction processing of this disclosure. In the conventional technology, a water level prediction model M is constructed using only water level data WL and analyzed rainfall data. In this disclosure, a water level prediction model M is constructed using water level data WL and analyzed mean rainfall data.

[0062] The data then shows the temporal changes in actual water level data, water level prediction data P (water level prediction result 6 hours later), analyzed rainfall data, and average rainfall data during rainfall. In the conventional technology, the water level prediction data P exceeds the evacuation decision water level, unlike the actual water level data. On the other hand, in this disclosure as well, the water level prediction data P exceeds the evacuation decision water level, unlike the actual water level data.

[0063] (Procedure for test data interpolation processing in this disclosure) The configuration of the water level prediction device of this disclosure (particularly the test data interpolation process) is also shown in Figure 12. The procedure for the test data interpolation process of this disclosure is shown in Figure 13. A specific example of the water level test data interpolation process of this disclosure is shown in Figure 14. A specific example of the rainfall test data interpolation process of this disclosure is shown in Figure 15. The water level prediction device W also includes a test data interpolation unit 8, and the test data interpolation program shown in Figure 13 can also be realized by installing it on a computer.

[0064] The test data interpolation unit 8 interpolates the water level data WL at the time of the missing data at the water level prediction point when the water level data WL at the water level prediction point is missing (step S21) (step S22). Here, the test data interpolation unit 8 interpolates using the water level prediction data P at the time of the missing data, which was output by the water level prediction data output unit 5 at a prediction time arbitrarily before the time of the missing data, and which occurred at a time arbitrarily after the prediction time at the water level prediction point (step S22). The test data extraction unit 4 inputs the test data (including the water level data WL at the time of the missing data interpolated by the test data interpolation unit 8, and the rainfall data RD at the time of the missing data will be described later) into the water level prediction model M (step S22).

[0065] Furthermore, if water level data WL is missing at the upstream location, even if water level data WL is available at the water level prediction location, the water level prediction data P at the water level prediction location cannot be output. Therefore, the water level data WL at the time of the missing data at the upstream location is interpolated using the water level prediction data P at the upstream location.

[0066] Furthermore, the arbitrary time set by the test data interpolation unit 8 is equal to the time interval at which the water level prediction data output unit 5 outputs the water level prediction data P. In other words, the rainfall forecast for a short period of time is highly accurate, and the arbitrary time set by the test data interpolation unit 8 is short.

[0067] In the first section of Figure 14, in order to predict the water level at 12:30 (30 minutes from the current time), the following test data are extracted: water level data of 3.0m at 11:00 (1 hour prior to the current time), water level data of 4.0m at 11:30 (30 minutes prior to the current time), and water level data of 5.0m at 12:00 (the current time). In sections 3-5 of Figure 14, some water level data is missing as test data.

[0068] In the second section of Figure 14, the predicted water level 30 minutes from the current time is output as follows: predicted water level data of 2.9m for 11:00 (one hour prior to the current time), predicted water level data of 3.9m for 11:30 (30 minutes prior to the current time), and predicted water level data of 4.8m for 12:00 (the current time). In the second section of Figure 14, the arbitrary time set by the test data interpolation unit 8 is 30 minutes.

[0069] In the third panel of Figure 14, the water level data for 11:00 (3.0m) and 11:30 (4.0m) are missing. Therefore, these are interpolated using the predicted water level data for 11:00 (2.9m) and 11:30 (3.9m). In the fourth panel of Figure 14, the water level data for 11:30 (4.0m) and 12:00 (5.0m) are missing. Therefore, these are interpolated using the predicted water level data for 11:30 (3.9m) and 12:00 (4.8m). In the fifth panel of Figure 14, the water level data for 11:30 (4.0m) is missing. Therefore, this is interpolated using the predicted water level data for 11:30 (3.9m).

[0070] In this way, even when water level test data is missing in the river area, the accuracy of water level prediction in the river area can be improved by reflecting rainfall and actual water level conditions and then interpolating the water level test data from the river area.

[0071] Then, the missing data can be interpolated using water level prediction data P, which is output at a prediction time shortly before the missing data time (the time interval for water level prediction), reflecting rainfall and actual water level conditions, and then output at a missing data time shortly after the prediction time (the time interval for water level prediction) in the river area.

[0072] When rainfall data RD is missing at a water level prediction point (step S23), the test data interpolation unit 8 interpolates the rainfall data RD for the missing time at the water level prediction point (step S24). Here, the test data interpolation unit 8 interpolates using the forecast rainfall data RF for the missing time at the water level prediction point (step S24). The test data extraction unit 4 inputs the test data (including the rainfall data RD for the missing time interpolated by the test data interpolation unit 8, and the water level data WL for the missing time as described above) into the water level prediction model M (step S24).

[0073] Furthermore, if rainfall data RD is missing in the upstream basin, even if rainfall data RD is available at the water level prediction point, the water level prediction data P at the water level prediction point cannot be output. Therefore, the rainfall data RD for the missing time in the upstream basin is interpolated using the forecast rainfall data RF from the upstream basin.

[0074] In the first section of Figure 15, in order to predict the water level at 12:30 (30 minutes from the current time), rainfall data of 30 mm at 11:00 (1 hour prior to the current time), rainfall data of 40 mm at 11:30 (30 minutes prior to the current time), and rainfall data of 50 mm at 12:00 (the current time) are extracted as test data. In sections 3 to 5 of Figure 15, some rainfall data is missing as test data.

[0075] In the second row of Figure 15, the forecast rainfall data RF (short-term precipitation forecast) is entered as follows: forecast rainfall data of 29 mm for 11:00 (one hour before the current time), forecast rainfall data of 39 mm for 11:30 (30 minutes before the current time), and forecast rainfall data of 48 mm for 12:00 (the current time).

[0076] In the third panel of Figure 15, the rainfall data for 11:00 (30 mm) and 11:30 (40 mm) are missing. Therefore, these are interpolated using the forecast rainfall data for 11:00 (29 mm) and 11:30 (39 mm). In the fourth panel of Figure 15, the rainfall data for 11:30 (40 mm) and 12:00 (50 mm) are missing. Therefore, these are interpolated using the forecast rainfall data for 11:30 (39 mm) and 12:00 (48 mm). In the fifth panel of Figure 15, the rainfall data for 11:30 (40 mm) is missing. Therefore, this is interpolated using the forecast rainfall data for 11:30 (39 mm).

[0077] In this way, even when rainfall test data for river areas is missing, the accuracy of water level prediction in river areas can be improved by incorporating forecast rainfall data (RF) for river areas and then interpolating the rainfall test data for river areas.

[0078] (Procedure for calculating the accuracy rate of water level predictions in this disclosure) The configuration of the water level prediction device disclosed herein (particularly the water level prediction accuracy calculation process) is also shown in Figure 16. The procedure for the water level prediction accuracy calculation process disclosed herein is shown in Figure 17. The water level prediction device W also includes a water level prediction error range setting unit 9 and a water level prediction accuracy calculation unit 10, and the water level prediction accuracy calculation program shown in Figure 17 can also be realized by installing it on a computer.

[0079] The training data extraction unit 2 extracts, as training data necessary for learning the water level at a given time at a water level prediction point, past observed rainfall data RO for at least one of the water level prediction point and at least one of the upstream watersheds from the water level prediction point, and rainfall data RO for "future observations" from that time onward.

[0080] The test data extraction unit 4 extracts, as test data necessary for predicting the water level at a given time at a water level prediction point, the following: past observed rainfall data RO at the water level prediction point and at least one of the upstream areas of the water basin from the water level prediction point, at the time in question and at least one of the past areas prior to that time; and "future forecast" rainfall data RF from that time onward.

[0081] However, because the "future forecast" rainfall data RF is less accurate than the "future observation" rainfall data RO, the water level "prediction" data P can deviate significantly from the water level "observation" data. Therefore, this disclosure visualizes the reliability of the water level "prediction" data P.

[0082] Here, the training data is known data for water level learning and cannot be applied to the calculation of the water level prediction accuracy rate H, as it is set to approximately 100%. The test data is unknown data for water level prediction, but it is data specific to a particular rainfall situation and therefore cannot be applied to the calculation of the water level prediction accuracy rate H. On the other hand, the validation data is unknown data for validating the water level prediction model M and is not data specific to a particular rainfall situation, so it can be applied to the calculation of the water level prediction accuracy rate H.

[0083] A specific example of the water level prediction error range setting process in this disclosure is shown in Figure 18. The verification data extraction unit 6 extracts water level data WL, past observed rainfall data RO, and future observed rainfall data RO as verification data for verifying the water level prediction model M (step S31), but does not extract future forecast rainfall data RF.

[0084] The verification data extraction unit 6 inputs water level data WL, past observed rainfall data RO, and future observed rainfall data RO into the water level prediction model M as verification data to verify the water level at multiple times at the water level prediction point (step S32). The water level prediction error range setting unit 9 obtains water level prediction data P, which shows the predicted water level at the water level prediction point for those multiple times, from the water level prediction model M (step S32).

[0085] The water level prediction error range setting unit 9 calculates the predicted error probability distribution E of the water level at the water level prediction location, and sets the error range of the water level prediction data P with respect to the water level observation data in the water level prediction hit rate calculation step S35 (step S33). Here, the water level prediction error range setting unit 9 calculates the predicted error probability distribution E for each command criterion for water level warning at the water level prediction location and / or for each period length from the predicted output time in the water level prediction data output step S32 to each prediction target time, and sets the error range of the water level prediction data P (step S33).

[0086] In the column of "prediction error after each N hours" in FIG. 18, the predicted error probability distribution E of the water level at the water level prediction location is calculated for the observed water level less than the command reference water level A of 0 m or more, the observed water level greater than or equal to the command reference water level A and less than the command reference water level B, and the observed water level greater than or equal to the command reference water level B. In the column of "error range after each N hours" in FIG. 18, for the observed water level less than the command reference water level A of 0 m or more, the observed water level greater than or equal to the command reference water level A and less than the command reference water level B, and the observed water level greater than or equal to the command reference water level B, among the predicted error probability distributions E, e lоw-min above e lоw-max below, and e mid-min above e mid-max below, and e high-min above e high-max below, are set as the error range for the water level prediction data P with respect to the water level observation data. Here, taking the minimum value and the maximum value of the predicted error probability distribution E as 0% and 100%, 0.1% ≦ e high-min ≦ e mid-min ≦ e lоw-min ≦ 20%, and 80% ≦ e lоw-max ≦ e mid-max ≦ e high-max ≦ 99.9% are satisfied. That is, the error range of the water level prediction data P is set wider for higher water levels, narrower for lower water levels, and in the middle for medium water levels. The upward and downward swing ranges of the error range of the water level prediction data P may be equal or different according to the distribution shape of the predicted error probability distribution E.

[0087] Thus, when using machine learning to predict water levels in river areas, during the validation phase, it is possible to calculate the probability distribution E of the prediction error for water levels, set the error range P for the water level prediction data relative to the water level observation data, and prepare to calculate the water level prediction accuracy H.

[0088] Furthermore, since future observed rainfall data RO cannot be extracted as test data during the prediction phase, even when extracting future forecast rainfall data RF, the water level prediction accuracy rate H can be prepared during the verification phase to support the determination of whether each issuance criterion has been reached.

[0089] A specific example of the water level prediction accuracy calculation process in this disclosure is shown in Figure 19. The verification data extraction unit 6 extracts water level data WL, past observed rainfall data RO, and future forecast rainfall data RF as verification data for verifying the water level prediction model M (step S31), but does not extract future observed rainfall data RO.

[0090] The verification data extraction unit 6 inputs water level data WL, past observed rainfall data RO, and future forecast rainfall data RF into the water level prediction model M as verification data to verify the water level at multiple times at the water level prediction point (step S34). The water level prediction data output unit 5 outputs water level prediction data P, which shows the predicted water level at the water level prediction point for those multiple times, from the water level prediction model M (step S34).

[0091] The water level prediction accuracy calculation unit 10 calculates the water level prediction accuracy rate H at the water level prediction point as the proportion of the water level prediction data P at the water level prediction point at the water level prediction point at the water level prediction point at the water level prediction point at the water level prediction point at the water level prediction point at the water level prediction point at the water level prediction point at the water level prediction point at the water level prediction point at the water level prediction point at the water level prediction point at the water level prediction point at the water level prediction point at the water level prediction point (step S35).

[0092] In the "Predicted Data for Each N Hours Later" column of Figure 19, water level prediction data P for multiple time points at the water level prediction point is output for observed water levels of 0m or more but less than the warning threshold water level A, observed water levels of warning threshold water level A or more but less than the warning threshold water level B, and observed water levels of warning threshold water level B or higher. For observed water levels of 0m or more but less than the warning threshold water level A, there are 80 data points that fall within the error range, 10 data points that are below the error range, and 10 data points that are above the error range. For observed water levels of warning threshold water level A or more but less than the warning threshold water level B, there are 30 data points that fall within the error range, 15 data points that are below the error range, and 15 data points that are above the error range. For observed water levels of warning threshold water level B or higher, there are 10 data points that fall within the error range, 10 data points that are below the error range, and 10 data points that are above the error range.

[0093] In the "Accuracy of Prediction After Each N Hours" column of Figure 19, the accuracy of water level prediction H at the water level prediction point is calculated for observed water levels of 0m or more but less than the warning threshold water level A, observed water levels of warning threshold water level A or more but less than the warning threshold water level B, and observed water levels of warning threshold water level B or higher. For observed water levels of 0m or more but less than the warning threshold water level A, the accuracy of water level prediction is 80 / (80+10+10)×100=80%. For observed water levels of warning threshold water level A or more but less than the warning threshold water level B, the accuracy of water level prediction is 30 / (30+15+15)×100=50%. For observed water levels of warning threshold water level B or higher, the accuracy of water level prediction is 10 / (10+10+10)×100=33%.

[0094] Thus, since future observed rainfall data RO cannot be extracted as test data during the prediction stage, even when extracting future forecast rainfall data RF, the reliability of the water level prediction data P can be visualized by calculating the water level prediction accuracy rate H during the verification stage.

[0095] Furthermore, since future observed rainfall data RO cannot be extracted as test data during the prediction phase, even when extracting future forecasted rainfall data RF, the water level prediction accuracy rate H can be calculated during the verification phase to support the determination of whether each issuance criterion has been reached. [Industrial applicability]

[0096] The water level prediction model construction method, water level prediction model construction program, water level prediction data output method, and water level prediction data output program disclosed herein can improve the accuracy of water level predictions by using a water level prediction model, while taking into account the time required from rainfall in the river area to inflow into the river.

[0097] The test data interpolation method disclosed herein can improve the accuracy of water level predictions at water level prediction points even when water level test data is missing at those points. The water level prediction accuracy calculation method disclosed herein can visualize the reliability of water level prediction data by calculating the water level prediction accuracy rate. [Explanation of symbols]

[0098] W: Water level prediction device M, M1, Mn: Water level prediction model 1: Rainfall data conversion unit 2: Training data extraction unit 3: Water Level Prediction Model Construction Department 4: Test data extraction unit 5: Water level prediction data output unit 6: Data extraction unit for verification 7: Water Level Prediction Model Selection Section 8: Test data interpolation section 9: Water level prediction error range setting section 10: Water level prediction accuracy calculation unit WL: Water level data RD: Rainfall data P: Water level prediction data RO: Observed rainfall data RF: Forecast rainfall data E: Prediction error probability distribution H: Water level prediction accuracy rate V: Verification data

Claims

1. A training data extraction step involves extracting the following as training data necessary for learning the water level at a given time at a water level prediction point in a river area: (1) water level data showing at least one of the water levels at the water level prediction point and at least one of the water levels upstream of the water level prediction point, at the time in question and at least one of the water levels in the past prior to that time; and (2) rainfall data showing at least one of the rainfall amounts at the water level prediction point and at least one of the rainfall amounts in the past prior to that time and at least one of the rainfall amounts in the future prior to that time, at the time in question and at least one of the rainfall amounts upstream of the water level prediction point. A water level prediction model construction step involves constructing a water level prediction model for predicting the water level at the water level prediction point in the river area using the water level data and the rainfall data, They are provided in order, The training data extraction step extracts at least one of the following as rainfall data: (1) instantaneous rainfall data showing the instantaneous rainfall at at least one of the time, past time, and future time at at least one of the water level prediction point and the water basin upstream of the water level prediction point; (2) total rainfall data showing the total rainfall from the start of rainfall to at least one of the time, past time, and future time at at least one of the time, past time, and future time at at least one of the time, past time, and future time at at least one of the time, past time, and future time at the water level prediction point and the water basin upstream of the water level prediction point (however, it is reset to 0 when a predetermined time has elapsed from the time when no rain falls); and (3) average rainfall data showing the average rainfall within a predetermined average period at at least one of the time, past time, and future time at at least one of the time, past time, and future time at at least one of the time, past time, and future time. A method for constructing a water level prediction model, characterized by the following features.

2. A water level prediction model selection step involves inputting verification data for verifying the water level prediction model into multiple types of water level prediction models constructed using multiple types of training data, and selecting the water level prediction model that best fulfills the issuance criteria. Further provided after the water level prediction model construction step A method for constructing a water level prediction model according to claim 1, characterized in that...

3. A water level prediction model construction program for causing a computer to sequentially execute each processing step comprising the water level prediction model construction method described in claim 1 or 2.

4. In order to predict the water level at the water level prediction point in the river area using the water level prediction model constructed by the water level prediction model construction method according to claim 1 or 2, the water level prediction data output step includes inputting the water level data and the rainfall data into the water level prediction model as test data necessary for predicting the water level at a certain time at the water level prediction point, and outputting water level prediction data from the water level prediction model that shows the predicted water level at that time at the water level prediction point. A method for outputting water level prediction data, characterized by comprising the following features.

5. A water level prediction data output program for causing a computer to perform the water level prediction data output step included in the water level prediction data output method according to claim 4.

6. In the test data extracted by the water level prediction data output method according to claim 4, when the water level data at the water level prediction point is missing, the water level data for the time of the missing data at the water level prediction point is interpolated using the water level prediction data output at the prediction time, which was output at an arbitrary time before the time of the missing data, and which is interpolated using the water level prediction data for the time of the missing data after the arbitrary time after the prediction time at the water level prediction point. A test data interpolation method characterized by comprising the following:

7. The water level prediction model construction method according to claim 1 or 2 includes a verification data extraction step of extracting the water level data and past observed rainfall data and future forecast rainfall data as rainfall data, as verification data for verifying the constructed water level prediction model, The water level prediction data output step involves inputting the water level data, past observed rainfall data, and future forecast rainfall data for verifying the water level at multiple time points at the water level prediction point as verification data, into the water level prediction model, and outputting water level prediction data from the water level prediction model that shows the predicted water level at those multiple time points. A water level prediction accuracy calculation step, in which the proportion of the water level prediction data for multiple time points at the water level prediction point that falls within the error range around the water level observation data for multiple time points at the water level prediction point is calculated as the water level prediction accuracy at the water level prediction point, A method for calculating the accuracy rate of water level prediction, characterized by comprising the following in order.

Images