Water level prediction device and water level prediction program

Abstract

To determine, when predicting a water level of a river or the like through machine learning, using a mechanical method without using determinations of observatories or human settings, whether or not the water level has reached an issue reference for evacuation advisory and the like.SOLUTION: According to the present disclosure, In a verification stage, a probability distribution of prediction errors is calculated, and in a prediction stage, prediction errors are added to predicted water levels to calculate a probability distribution of the predicted water levels. Respective reaching probabilities of respective issue references are calculated based upon the probability distribution of the predicted water levels and respective water level ranges of the respective issue references. Further, it is determined that the respective reaching probabilities calculated as mentioned above have reached the respective issue references as compared with respective reaching probabilities obtained when the probability distribution of the predicted water levels is diagnosed to be a uniform distribution.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to a technology for predicting water levels of rivers and the like using machine learning. [Background technology]

[0002] Patent Document 1 and other publications disclose technology for predicting the water level of rivers and the like using machine learning. In the training stage, a water level prediction model such as a neural network is constructed using training data for water level learning. In the verification stage, verification data for verifying the water level prediction model is input into the water level prediction model, and the generalization performance of the water level prediction model is evaluated. In the prediction stage, test data for water level prediction is input into the water level prediction model, and the water level of rivers and the like is predicted. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-046710 Summary of the Invention [Problem to be solved by the invention]

[0004] Figure 1 shows the calculation process for the probability of reaching the warning water level using conventional technology. Here, it is conceivable that an evacuation advisory or other such advisory will be issued when the predicted water level exceeds the warning water level. However, because the predicted water level contains prediction errors, there is a high possibility that an evacuation advisory or other such advisory will not be issued. Therefore, it is conceivable to calculate the probability distribution of the predicted water level and then calculate the probability that the predicted water level will exceed the warning water level. This reduces the possibility that an evacuation advisory or other such advisory will not be issued, even if the predicted water level contains prediction errors.

[0005] Using the ensemble method, test data T is input into multiple water level prediction models M1, M2, M3, ... to calculate multiple predicted water levels P1, P2, P3, .... A histogram of the multiple predicted water levels P1, P2, P3, ... is then created to calculate the probability distribution of the predicted water levels. Furthermore, the probability that the predicted water level will exceed the warning water level is calculated, for example, to 20%. However, the judgment as to what probability the predicted water level must exceed the warning water level to reach the criteria for issuing an evacuation advisory or other warning varies from observation station to observation station, making it difficult to set this manually.

[0006] Therefore, in order to solve the above problem, the present disclosure aims to use machine learning to predict the water levels of rivers, etc., and to determine whether the criteria for issuing evacuation advisories, etc. have been reached using a mechanical method, rather than relying on the judgment of each observation station or artificial settings. [Means for solving the problem]

[0007] To solve the above problem, in the verification stage, the probability distribution of the prediction error is calculated, and in the prediction stage, the prediction error is added to the predicted water level to calculate the probability distribution of the predicted water level. Then, based on the probability distribution of the predicted water level and each water level range for each issuance criterion, the probability of each issuance criterion being reached is calculated. Furthermore, compared with each reach probability when the probability distribution of the predicted water level is considered to be a uniform distribution, each issuance criterion with a higher reach probability calculated above is determined to have been reached.

[0008] Specifically, the present disclosure relates to a water level prediction device that includes a model construction unit that constructs a water level prediction model using training data for water level learning, a prediction error probability distribution calculation unit that inputs verification data for verifying the water level prediction model into the water level prediction model and calculates the probability distribution of the prediction error, and a predicted water level probability distribution calculation unit that inputs test data for water level prediction into the water level prediction model, adds the prediction error to the predicted water level, and calculates the probability distribution of the predicted water level.

[0009] According to this configuration, the probability distribution of the prediction error is calculated in the verification stage, and then the probability distribution of the predicted water level is calculated in the prediction stage. Therefore, even if the predicted water level contains a prediction error, preprocessing can be performed to reduce the possibility of evacuation advisories, etc. not being issued.

[0010] The present disclosure also provides a water level prediction device that further includes an issuance standard reaching probability calculation unit that calculates the probability of reaching each issuance standard based on the probability distribution of the predicted water level and each water level range that satisfies each issuance standard for water level alerts.

[0011] According to this configuration, the probability of reaching each issuance criterion is calculated based on the probability distribution of the predicted water level and the water level range of each issuance criterion, so that even if the predicted water level contains prediction errors, preprocessing can be performed to reduce the possibility of evacuation advisories, etc. not being issued.

[0012] The present disclosure also provides a water level prediction device that further includes an issuance standard reaching / not reaching determination unit that determines that each issuance standard has been reached for each issuance standard for which the reaching probability calculated by the issuance standard reaching probability calculation unit is higher than each reaching probability when the probability distribution of the predicted water level is considered to be a uniform distribution.

[0013] According to this configuration, the probability thresholds for each issuance criterion are mechanically calculated for each probability of reaching each issuance criterion, so it is possible to determine whether the criteria for issuing an evacuation advisory, etc. have been reached using a mechanical method, without relying on the judgment of each observation station or artificial settings.

[0014] The present disclosure also provides a water level prediction program for causing a computer to execute the processing steps performed by each processing unit included in the water level prediction device described above.

[0015] According to this configuration, it is possible to provide a program having the above-described effects. [Effects of the Invention]

[0016] In this way, the present disclosure uses machine learning to predict water levels of rivers, etc., and can determine whether the criteria for issuing evacuation advisories, etc. have been reached using a mechanical method, without relying on the judgment of each observation station or artificial settings. [Brief explanation of the drawings]

[0017] [Figure 1] FIG. 10 is a diagram showing a calculation process for the probability of reaching a warning water level according to the prior art. [Figure 2] FIG. 1 is a diagram illustrating a configuration of a water level prediction device according to the present disclosure. [Figure 3] FIG. 10 is a diagram showing the procedure of the water level prediction process of the present disclosure. [Figure 4] FIG. 10 is a diagram illustrating a calculation process of the probability distribution of prediction errors according to the present disclosure. [Figure 5] FIG. 10 is a diagram illustrating a calculation process of the probability distribution of prediction errors according to the present disclosure. [Figure 6] FIG. 10 is a diagram showing the calculation process of the probability distribution of predicted water levels according to the present disclosure. [Figure 7] FIG. 10 is a diagram illustrating a calculation process for each arrival probability of each issuance standard in the present disclosure. [Figure 8] FIG. 10 is a diagram illustrating a calculation process for each arrival probability of each issuance standard in the present disclosure. [Figure 9] FIG. 10 is a diagram illustrating a process for determining whether each of the issuance criteria of the present disclosure has been reached. [Figure 10] FIG. 10 is a diagram illustrating a process for determining whether each of the issuance criteria of the present disclosure has been reached. [Figure 11] FIG. 10 is a diagram illustrating an example of determining whether each of the issuance criteria of the present disclosure has been reached. DETAILED DESCRIPTION OF THE INVENTION

[0018]

[0023] The following embodiments of the present disclosure will be described with reference to the accompanying drawings. The embodiments described below are examples of implementation of the present disclosure, and the present disclosure is not limited to the following embodiments.

[0019] (Configuration of the water level prediction device of the present disclosure) The configuration of the water level prediction device of the present disclosure is shown in Figure 2. The procedure for the water level prediction process of the present disclosure is shown in Figure 3. The water level prediction device W comprises a water level prediction model M, a model construction unit 1, a prediction error probability distribution calculation unit 2, a predicted water level probability distribution calculation unit 3, an issuance standard reaching probability calculation unit 4, and an issuance standard reaching / not reaching determination unit 5, and can be realized by installing the water level prediction program of Figure 3 on a computer.

[0020] The model construction unit 1 constructs a water level prediction model M using training data L for water level learning (step S1). The training data L is data on water levels (especially high water levels) of rivers, etc., and data on rainfall (especially heavy rainfall) in river areas, etc., excluding verification data V from the learning data. The water level prediction model M is a neural network, etc.

[0021] The forecast error probability distribution calculation unit 2, the forecast water level probability distribution calculation unit 3, the issuance standard reaching probability calculation unit 4, and the issuance standard reaching / not reaching determination unit 5 execute the following processes (steps S2 to S5).

[0022] (Calculation process of probability distribution of predicted water level according to the present disclosure) The calculation process of the probability distribution of prediction errors of the present disclosure is shown in Figures 4 and 5. The prediction error probability distribution calculation unit 2 inputs verification data V for verifying the water level prediction model M into the water level prediction model M and calculates the probability distribution of prediction errors (step S2). The verification data V includes data on water levels (especially high water levels) of rivers, etc. and data on rainfall (especially heavy rainfall) in river areas, etc., excluding the training data L from the learning data. The calculation methods in Figures 4 and 5 are different.

[0023] In Fig. 4, the prediction error probability distribution calculation unit 2 excludes training data L from the learning data and inputs verification data V into the water level prediction model M, calculates the predicted water level P, and calculates a time series of prediction errors based on the predicted water level P and the teacher data S. Then, based on the time series of prediction errors, a histogram of prediction errors is created and the probability distribution of the prediction errors is calculated. Here, the prediction error range is between -1.0 m and +1.0 m, with 0.0 m at the center.

[0024] In Figure 5, since it may be difficult to prepare multiple types of learning data when constructing a water level prediction model M, only one type of learning data is prepared, and the method of dividing the training data L and the verification data V is devised for each type. That is, the forecast error probability distribution calculation unit 2 excludes training data L1, L2, and L3 from the learning data and inputs the verification data V1, V2, and V3 into the water level prediction models M1, M2, and M3, calculates predicted water levels P1, P2, and P3, and calculates a time series of each forecast error based on the predicted water levels P1, P2, and P3 and each teacher data. Then, based on the time series of each forecast error, a histogram of each forecast error is created and forecast error probability distributions E1, E2, and E3 are calculated. Furthermore, the forecast error probability distribution is calculated by adding up the forecast error probability distributions E1, E2, and E3. Here, the forecast error band is between -1.0 m and +1.0 m, with 0.0 m as the center.

[0025] The calculation process for the probability distribution of predicted water levels according to the present disclosure is shown in Figure 6. The predicted water level probability distribution calculation unit 3 inputs test data T for water level prediction into the water level prediction model M, adds the prediction error probability distribution E to the predicted water level P, and calculates the probability distribution of the predicted water level (step S3). The test data T includes data on the water level of rivers, etc., and data on rainfall in river areas, etc., for the last few hours before the prediction time.

[0026] In Figure 6, the predicted water level probability distribution calculation unit 3 inputs test data T into the water level prediction model M and calculates a time series of predicted water level P. Then, a certain prediction error probability distribution E is added to the predicted water level P at each time to calculate the probability distribution of the predicted water level P at each time. Here, the predicted water level P at time 3:00 is 3.0m, and the prediction error range at time 3:00 is centered around 3.0m and is greater than or equal to 3.0m-1.0m=2.0m and less than or equal to 3.0m+1.0m=4.0m.

[0027] In this way, the probability distribution of the prediction error is calculated in the verification stage, and then the probability distribution of the predicted water level is calculated in the prediction stage.Therefore, even if the predicted water level contains a prediction error, preprocessing can be performed to reduce the possibility of evacuation advisories, etc. not being issued.

[0028] (Calculation process of each arrival probability of each issuance standard of the present disclosure) The calculation process for the probability of reaching each issuance criterion of the present disclosure is shown in Figures 7 and 8. The issuance criterion arrival probability calculation unit 4 calculates the probability of reaching each issuance criterion based on the probability distribution of the predicted water level and each water level range that satisfies each issuance criterion for water level alerts (step S4). That is, the calculation calculates the probability of reaching each issuance criterion as the total probability of the probability distribution of the predicted water level or the total frequency of the histogram of the predicted water level in each water level range that satisfies each issuance criterion. The issuance criterion water level range A includes data such as the riverbed water level L0, the flood defense unit standby water level L1, the flood warning water level L2, the evacuation decision water level L3, the flood danger water level L4, and the crest water level L5. The issuance criteria are different in Figures 7 and 8.

[0029] In Figure 7, the probability distribution of predicted water levels is the same as that in Figure 6, with the flood defense unit standby water level L1 being lower than 2.0 m, the flood danger water level L4 being higher than 4.0 m, and the flood warning water level L2 and the evacuation decision water level L3 being relatively far apart, with the peak of the probability distribution of predicted water levels in between. The issuance standard reaching probability calculation unit 4 calculates the following: probability of not reaching the issuance standard = 0%, probability of reaching the flood defense unit standby = 30%, probability of reaching the flood warning = 60%, probability of reaching the evacuation decision = 10%, and probability of reaching the flood danger = 0%. However, even if the flood warning reaching probability = 60% is relatively high, whether the flood warning issuance standard has been reached remains to be determined by the issuance standard reaching determination unit 5. Based on the probability distribution of predicted water levels, the determination that the flood warning issuance standard has been reached is considered correct.

[0030] In Figure 8, the probability distribution of predicted water levels is the same as that shown in Figure 6, with the flood defense unit standby water level L1 being lower than 2.0 m, the flood danger water level L4 being higher than 4.0 m, and the flood warning water level L2 and the evacuation decision water level L3 being relatively close to each other, with the peak of the probability distribution of predicted water levels sandwiched between them. The issuance standard reaching probability calculation unit 4 calculates the following: probability of not reaching the issuance standard = 0%, probability of reaching the flood defense unit standby water level = 40%, probability of reaching the flood warning water level = 20%, probability of reaching the evacuation decision water level = 40%, and probability of reaching the flood danger water level = 0%. However, even if the probability of reaching the flood defense unit standby water level and the evacuation decision water level = 40% is relatively high, the determination of whether the issuance standard for the flood defense unit standby water level and the evacuation decision water level has been reached awaits the determination by the issuance standard reaching determination unit 5. Based on the probability distribution of predicted water levels, it is considered that the determination that the flood defense unit standby water level and the evacuation decision water level have been reached is incorrect, and the determination that the flood warning water level has been reached is correct.

[0031] In this way, the probability of reaching each issuance criterion is calculated based on the probability distribution of the predicted water level and each water level range for each issuance criterion, so even if the predicted water level contains prediction errors, preprocessing can be performed to reduce the possibility of evacuation advisories, etc. not being issued.

[0032] (Processing for determining whether each of the standards for issuing orders disclosed herein has been met) The process of determining whether each of the issuance criteria of the present disclosure has been reached is shown in Figures 9 and 10. The issuance criteria reach determination unit 5 determines that each issuance criterion has been reached for each issuance criterion whose reach probability calculated by the issuance criteria reach probability calculation unit 4 is higher than the reach probability when the probability distribution of the predicted water level is considered to be a uniform distribution (step S5). The issuance criteria are different in Figures 9 and 10.

[0033] In Fig. 9, the probability distribution of the predicted water level is the probability distribution of Fig. 6, and each water level L0 to L5 is the same as Fig. 7. The issuance standard reaching determination unit 5 regards the probability distribution of the predicted water level as a uniform distribution, and determines each water level range L m ~L m+1 Each probability threshold [%] for each arrival probability = (L m+1 -L m) / (L5-L0)*100 is calculated. Then, the probability threshold for not reaching the issuance criteria is calculated as 30%, the probability threshold for flood defense corps standby is calculated as 20%, the probability threshold for flood warning is calculated as 25%, the probability threshold for evacuation decision is calculated as 15%, and the probability threshold for flood risk is calculated as 10%. As a result, the probability of reaching the flood warning and flood defense corps standby is calculated as 60% and 30%, which are higher than the probability thresholds for flood warning and flood defense corps standby is calculated as 25% and 20%, and it is determined that the criteria for issuing a flood warning and flood defense corps standby have been reached. In other words, the probability distribution of flood warning and flood defense corps standby is biased compared to the uniform probability distribution of all issuance criteria.

[0034] In Fig. 10, the probability distribution of the predicted water level is the probability distribution of Fig. 6, and each water level L0 to L5 is the same as Fig. 8. The issuance standard reaching determination unit 5 regards the probability distribution of the predicted water level as a uniform distribution, and determines each water level range L m ~L m+1 Each probability threshold [%] for each arrival probability = (L m+1 -L m ) / (L5-L0)*100 is calculated. Then, the probability threshold for not issuing a flood warning = 2.5%, the probability threshold for flood defense unit standby = 45%, the probability threshold for flood warning = 5%, the probability threshold for evacuation decision = 45%, and the probability threshold for flood risk = 2.5% are calculated. As a result, it is determined that the probability of a flood warning being issued = 20% is higher than the probability threshold for a flood warning = 5%, and the criteria for issuing a flood warning have been met. In other words, the probability distribution of a flood warning is biased compared to the uniform probability distribution of all issuance criteria.

[0035] In this way, the probability thresholds for each issuance criterion are mechanically calculated for each probability of reaching each issuance criterion, so it is possible to determine whether the criteria for issuing an evacuation advisory, etc. have been reached using a mechanical method, without relying on the judgment of each observation station or artificial settings.

[0036] An example of determining whether each of the issuance criteria disclosed in this disclosure has been reached is shown in Figure 11. At time mm / d3 2:30 to 7:30 (mm indicates the month, and d1 to d5 indicate the day), the peak water level was observed, and the predicted water level six hours later follows the actual water level, but there is still a prediction error.

[0037] Here, the riverbed water level L0 = -1.00 m, the flood defense unit standby water level L1 = 0.90 m, the flood warning water level L2 = 1.60 m, the evacuation decision water level L3 = 2.20 m, the flood danger water level L4 = 3.35 m, and the crest water level L5 = 5.20 m. The probability thresholds for failure to issue a warning are 31%, for flood defense unit standby water level L4 = 11%, for flood warning water level L10%, for evacuation decision water level L19%, and for flood danger water level L5 = 30%. Furthermore, the determination of whether the warning criteria have been met is performed as follows:

[0038] At time mm / d3 3:00, the flood warning arrival probability of 78.9% exceeds the flood warning probability threshold of 10% for the first time. Meanwhile, at the same time mm / d3 3:00, the actual water level of 1.88 m exceeds the flood warning water level L2 of 1.60 m for the first time. In other words, it is possible to predict six hours in advance that the flood warning issuance criteria will be reached.

[0039] At time mm / d3 7:30, the flood risk arrival probability of 99.2% exceeds the flood risk probability threshold of 30% for the first time. Meanwhile, 30 minutes earlier at time mm / d3 7:00, the actual water level of 4.28 m exceeds the flood risk level L4 of 3.35 m for the first time. In other words, it is possible to predict 5 hours and 30 minutes in advance that the flood risk warning criteria will be reached.

[0040] In this way, the present disclosure, which sets a probability threshold for the issuance criteria, can limit the predicted error in the arrival time to about ±30 minutes, whereas the prior art, which compares the predicted water level with the warning water level, widens the predicted error in the arrival time to an error greater than ±30 minutes. [Industrial Applicability]

[0041] The water level prediction device and water level prediction program disclosed herein use machine learning to predict the water level of rivers, etc., and can determine whether the criteria for issuing evacuation advisories, etc. have been reached using a mechanical method, rather than relying on the judgment of each observation station or artificial settings. [Explanation of symbols]

[0042] W: Water level prediction device M, M1, M2, M3: Water level prediction models 1: Model construction section 2: Forecast error probability distribution calculation section 3: Calculation of predicted water level probability distribution 4: Calculation unit for probability of arrival of announcement standard 5: Determining whether the standard for issuing an order has been reached L, L1, L2, L3: training data V, V1, V2, V3: validation data T: Test data A: Standard water level range for issuing a warning P, P1, P2, P3: Forecast water levels S: Teacher data E, E1, E2, E3: Forecast error probability distribution

Claims

1. a model construction unit that constructs a water level prediction model using training data for water level learning; a prediction error probability distribution calculation unit that inputs verification data for verifying the water level prediction model into the water level prediction model and calculates a probability distribution of prediction errors; a predicted water level probability distribution calculation unit that inputs test data for water level prediction into the water level prediction model, adds the prediction error to the predicted water level, and calculates a probability distribution of the predicted water level; A water level prediction device comprising:

2. an issuance standard reaching probability calculation unit that calculates the probability of reaching each issuance standard based on the probability distribution of the predicted water level and each water level range that satisfies each issuance standard for water level warning; The water level prediction device according to claim 1 , further comprising:

3. an issuance standard attainment determination unit that determines that each of the issuance standards has been reached for which the attainment probability calculated by the issuance standard attainment probability calculation unit is higher than the attainment probability when the probability distribution of the predicted water level is considered to be a uniform distribution; The water level prediction device according to claim 2, further comprising:

4. A water level prediction program for causing a computer to execute each processing step performed by each processing unit included in the water level prediction device according to any one of claims 1 to 3.

Images