JP7103237B2 - Flow rate prediction device and flow rate prediction method - Google Patents

Description

The present invention relates to a flow rate prediction device and a flow rate prediction method.

Conventionally, it has been widely practiced to predict the flow rate of a river based on meteorological information. Since river flow rate prediction requires skill and is highly difficult, it has been proposed to automate this and easily predict river flow rate.

In the automation of river flow rate prediction, for example, as disclosed in Patent Document 1, a prediction model using a neural network may be used.
The flow rate prediction device disclosed in Patent Document 1 includes a weighting ratio determining means for determining a weighting ratio that expresses the degree of influence of rainfall data at a certain location on the downstream flow rate for each of the rainfall data at a plurality of locations, and a weighting ratio. A model construction means for constructing a flow rate prediction model using the basin average rainfall data calculated from the corresponding rainfall data and the basin average rainfall data based on the determined weighting ratio are input to the constructed flow rate prediction model. It has a prediction means for generating predicted flow rate data related to the predicted flow rate.
The load ratio increases the weighted ratio of the rainfall data at the location where the increase / decrease in rainfall has a large effect on the increase / decrease in the downstream flow rate, and the load ratio of the rainfall data at the location where the increase / decrease in rainfall has a small effect on the increase / decrease in the downstream flow rate. It is determined by reducing the weight ratio. More specifically, the load ratio and the flow rate prediction model are evaluated based on the correlation coefficient, which is the degree of correlation between rainfall and flow rate, and the prediction error. The flow rate prediction model is learned by repeating learning and prediction.

In order to build a prediction model using a neural network as described above, learning data such as meteorological observation data and meteorological prediction data is required. However, data with a large flow rate and a large amount of rainfall that has a large effect on the flow rate, such as in the case of a typhoon or torrential rain, which requires particularly high prediction accuracy, is much smaller than data with a small flow rate and rainfall. .. Therefore, the prediction accuracy may not be improved when the flow rate and rainfall are large.

On the other hand, Patent Document 2 discloses that the flow rate of a river or the like at a predetermined time in the future is predicted by a flow rate prediction device using the following prediction model by a neural network. The prediction model is constructed based on the training data for constructing the prediction model by processing the forecast rainfall data, the actual rainfall data, the actual flow rate data, and the like. In the data for building the prediction model, the data in which the average rainfall, the flow rate or the change in the flow rate is below a certain value is deleted, and the data in which the average rainfall, the flow rate or the change in the flow rate is above a certain value is multiplied by a plurality of times. This addresses the bias in data regarding rainfall or flow.
In the flow rate prediction device of Patent Document 2, the average rainfall is a point average value obtained by adding up the rainfall at a plurality of points and dividing by the number of points.

JP-A-2007-205001 JP-A-2007-226450

In the flow rate prediction device of Patent Document 1, as described above, the prediction accuracy when the flow rate is large may not be desirable.
In addition, since learning and prediction are repeated, it takes a lot of time to learn the prediction model. Therefore, it is not easy to introduce the flow rate prediction device.

In the flow rate prediction device of Patent Document 2, data in which the average rainfall, the flow rate, or the change in the flow rate is a certain value or more is simply a plurality of times, that is, duplicated. That is, since the learning data with new information is not added, the amount of information of the learning data as a whole does not change.
Moreover, since the average rainfall is a point average value, geographical information is not reflected with high accuracy.
Furthermore, by comparing the flow rate or the amount of change in the flow rate in the data with a constant value, it is determined which data is to be deleted or multiplied by multiple values, but human arbitrariness intervenes in the determination of this constant value. As a result, an appropriate value may not be set.
From the above, there is still a possibility that the prediction accuracy when the flow rate is large is not desirable.

An object to be solved by the present invention is to provide a flow rate prediction device and a flow rate prediction method capable of highly accurate flow rate prediction even when the flow rate of a river is large and easy to introduce.

The present invention employs the following means in order to solve the above problems. That is, the present invention is a flow rate prediction device that predicts the flow rate of a specific river, and obtains meteorological observation data, meteorological prediction data, and flow rate data related to the river at a past time that is an arbitrary time in the past. As training data, the prediction model learning unit that generates a plurality of prediction models by ensemble learning a plurality of weak learners so as to predict the flow rate after the past time, and each of the plurality of prediction models are currently In, the meteorological observation data, the meteorological prediction data, and the flow rate data related to the river are input to output a provisional future predicted flow rate, and based on a plurality of the provisional future predicted flow rates. Provided is a flow rate prediction device including a flow rate prediction unit for predicting the future flow rate of the river.

Further, the present invention is a flow rate prediction method for predicting the flow rate of a specific river, and obtains meteorological observation data, meteorological prediction data, and flow rate data related to the river at a past time which is an arbitrary time in the past. As training data, a plurality of weak learners are ensemble-learned so as to predict the flow rate after the past time, a plurality of prediction models are generated, and each of the plurality of prediction models is used in the current river. The relevant meteorological observation data, the meteorological forecast data, and the flow rate data are input to output a provisional future predicted flow rate, and the future of the river is based on a plurality of the provisional future predicted flow rates. A flow rate prediction method for predicting a flow rate is provided.

According to the present invention, it is possible to provide a flow rate prediction device and a flow rate prediction method capable of highly accurate flow rate prediction even when the flow rate of a river is large and easy to introduce.

It is a block diagram of the flow rate prediction apparatus in embodiment of this invention. It is explanatory drawing about input / output of the prediction model learning part of the flow rate prediction apparatus in the said embodiment. It is explanatory drawing about input / output of the flow rate prediction part of the flow rate prediction apparatus in the said embodiment. Of the flow rate prediction method in the above embodiment, (a) is a flowchart at the time of learning the flow rate, and (b) is a flowchart at the time of predicting the flow rate. It is a block diagram of the flow rate prediction apparatus in the 1st modification of the said embodiment. It is explanatory drawing of the seasonal term data generation part of the flow rate prediction apparatus in the said 1st modification. It is a block diagram of the flow rate prediction apparatus in the 2nd modification of the said embodiment. It is explanatory drawing of the machine learning device and the feature extraction model in the said 2nd modification.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
The flow rate prediction device in the present embodiment is a flow rate prediction device that predicts the flow rate of a specific river, and is related to the river at a past time, which is an arbitrary time in the past, meteorological observation data, meteorological prediction data, and flow rate. Using the data as training data, the predictive model learning unit that generates multiple predictive models by ensemble learning a plurality of weak learners so as to predict the flow rate after the past time, and each of the plurality of predictive models are currently In, the meteorological observation data, meteorological forecast data, and flow rate data related to the river are input to output the provisional future predicted flow rate, and the future flow rate of the river based on multiple provisional future predicted flow rates. It is equipped with a flow rate prediction unit that predicts.

FIG. 1 is a block diagram of the flow rate prediction device according to the present embodiment. FIG. 2 is an explanatory diagram regarding input / output of the prediction model learning unit of the flow rate prediction device according to the present embodiment. FIG. 3 is an explanatory diagram regarding input / output of the flow rate prediction unit of the flow rate prediction device according to the present embodiment.
The flow rate prediction device 1 in the present embodiment predicts the future flow rate of a specific river.

The flow rate prediction device 1 is an information processing device such as a personal computer. The flow rate prediction device 1 includes a meteorological observation data storage unit 11, a weather prediction data storage unit 12, a flow rate storage unit 13, a data processing unit 14, a learning unit 15, a prediction model parameter storage unit 16, and a prediction unit 17. The learning unit 15 includes a predictive model learning unit 20. The prediction unit 17 includes a flow rate prediction unit 30.
Among the components of the flow rate prediction device 1, the data processing unit 14, the prediction model learning unit 20, and the flow rate prediction unit 30 may be, for example, software or a program executed by the CPU in the information processing device. Further, the meteorological observation data storage unit 11, the meteorological prediction data storage unit 12, the flow rate storage unit 13, and the prediction model parameter storage unit 16 are realized by storage devices such as semiconductor memories and magnetic disks provided inside and outside the information processing device. It may have been done.

The input data of the flow rate prediction device 1 includes meteorological observation data 2, meteorological prediction data 3, and flow rate data 4.
The meteorological observation data 2 is meteorological data actually observed at an observatory or the like. Meteorological observation data 2 includes at least rainfall. The meteorological observation data 2 may include the temperature, the amount of solar radiation, the amount of snowfall, the amount of snowfall, and the like in addition to the amount of rainfall. In this embodiment, the amount of rainfall is particularly determined by the observatory data 2o observed at the observatories provided in each area and the area included as an observation target observed by radar observation or the like based on the latitude and longitude on the map. It has mesh data of 2 m divided in a mesh pattern. The meteorological observation data 2 related to the river to be predicted by the flow rate prediction device 1 from the past to the present is stored in the meteorological observation data storage unit 11.
The weather forecast data 3 is future predicted weather data at the time of forecasting the weather, which is provided by, for example, a weather forecast service. The meteorological prediction data 3 also includes at least the amount of rainfall, like the meteorological observation data 2. The weather prediction data 3 may include the temperature, the amount of solar radiation, the amount of snowfall, the amount of snowfall, and the like in addition to the amount of rainfall. In the present embodiment, particularly, the amount of rainfall includes mesh data 3 m as in the meteorological observation data 2. The weather prediction data 3 related to the river to be predicted by the flow rate prediction device 1 from the past to the present is stored in the weather prediction data storage unit 12.
The flow rate data 4 is the flow rate data actually observed at an observation station or the like. The flow rate data 4 of the river to be predicted by the flow rate prediction device 1 from the past to the present is stored in the flow rate storage unit 13.

The flow rate prediction unit 30 predicts the future flow rate of the river based on the data stored in the meteorological observation data storage unit 11, the weather prediction data storage unit 12, and the flow rate storage unit 13. In order to make this prediction effectively, in particular, the prediction model learning unit 20 is equipped with a machine learning device as will be described later, and machine learning data including meteorological observation data 2, meteorological prediction data 3, and flow rate data 4. Machine learning is performed by inputting to the learner to generate prediction model parameters for predicting the future flow rate of the river.
That is, the flow rate prediction device 1 is roughly classified into two operations of learning the flow rate of the river and predicting the flow rate. In order to simplify the explanation, first, each component of the flow rate prediction device 1 at the time of learning the flow rate will be described, and then the behavior of each component at the time of predicting the flow rate will be described.

At the time of learning the flow rate, the machine learner assumes the first past time (past time) which is an arbitrary time in the past, and each data based on the first past time, that is, at the time of the first past time. Actual weather observation status (meteorological observation data at past time), weather prediction status at the time of the first past time (weather prediction data at the past time), and flow rate at the time of the first past time (past) The event that actually occurred or was predicted in the first past time, such as the flow rate data at the time), is input. For these inputs, the machine learner has a second past, which is a time relative to the first past time, for example, one hour, two hours, or several hours after the first past time. Predict the flow rate in the time zone up to the time.
Since the second past time is the past time at the time when the machine learner is trained, the flow rate in the time zone from the first past time to the second past time is actually observed and used as data. exist. This actually observed flow rate data after the first past time is compared with the flow rate after the first past time predicted by the machine learning device, and the predicted flow rate is close to the actually observed flow rate data. The machine learner is learned so that

That is, the meteorological observation data 2 used as an input to the machine learner at the time of learning the flow rate is the meteorological observation data before the first past time (meteorological observation at the past time) stored in the meteorological observation data storage unit 11. Data) 2A. Further, the weather prediction data 3 used as an input to the machine learner at the time of learning the flow rate is the weather prediction data (meteorology at the past time) stored in the weather prediction data storage unit 12 after the first past time. Prediction data) 3A. Further, the flow rate data 4 used as an input to the machine learning device at the time of learning the flow rate is the flow rate data (flow rate data at the past time) 4A before the first past time.
As the weather forecast data 3A after the first past time, it is desirable to use the latest weather forecast data 3 that is before the first past time and as close to the first past time as possible.
The result output by the machine learner to which the above value is input, that is, the predicted flow rate after the first past time (flow rate after the past time) 5A is the correct answer value of the river after the first past time. Is compared with the actual flow rate data (flow rate data at the past time) 4C.
These meteorological observation data 2A before the first past time, weather prediction data 3A after the first past time, flow rate data 4A before the first past time, and correct value, that is, flow rate data 4C after the first past time. Is the learning data of the machine learner.

From the meteorological observation data storage unit 11, the meteorological prediction data storage unit 12, and the flow rate storage unit 13, the data processing unit 14 includes meteorological observation data 2A before the first past time, and meteorological prediction data 3A after the first past time. The flow rate data 4A before the first past time and the flow rate data 4C after the first past time are acquired and converted into the input / output format of the machine learner.
The data processing unit 14 transmits each of the converted data to the learning unit 15.

The learning unit 15 receives each data from the data processing unit 14.
The learning unit 15 trains the machine learner by the prediction model learning unit 20 to generate a learned prediction model.
In the present embodiment, the machine learner is learned by ensemble learning. Ensemble learning is a learning method using a plurality of weak learners as machine learners, that is, weak learners with a certain degree of correlation with truly correct regression. Rather than predicting using one prediction model that may not be sufficiently accurate in prediction, ensemble learning trains multiple weak learners under different conditions to generate multiple prediction models, each of which is used. It is based on the idea of improving the accuracy of prediction by combining multiple predicted results in a prediction model.
That is, the predictive model learning unit 20 includes a plurality of weak learners 21, and the predictive model learning unit 20 ensemble learns the plurality of weak learners 21 to generate a plurality of predictive models 31.

In this embodiment, boosting is applied as a learning method for ensemble learning. In boosting, the prediction accuracy of the weak learner 21 as a whole is improved by learning and combining the plurality of weak learners 21 one by one in order. More specifically, when the weak learner 21 is trained, the predicted value (in the present embodiment, the predicted flow rate 5A after the first past time) is determined by the weak learner 21 that has already been learned before that. For data having a large error from the correct answer value (flow data 4C after the first past time in the present embodiment), learning is performed so that the error is preferentially reduced. For example, in the weak learner 21 that has already been learned, if the error between the predicted value 5A and the correct answer value 4C in the first data is small and the error in the second data is large, the weak learner 21 to be learned next is the second. Prioritize and emphasize the data of 2. The data may be prioritized by, for example, weighting each data, and for the data considered to have the error eliminated in the already learned weak learner 21, the next weak learner 21 It may be ignored in learning.

Based on the above idea, the prediction model learning unit 20 first trains the first weak learner 21A using the learning data received from the data processing unit 14. The predicted value 5A and the correct answer value 4C, which are the outputs of the first weak learner 21A, are compared by, for example, a comparator 23, and the result is reflected in the parameters provided inside the first weak learner 21A. The first weak learner 21A is learned by a method such as making the learner 21A. When it is determined that the learning of the first weak learner 21A is completed, if there is data with a large error between the predicted value 5A and the correct answer value 4C during the learning of the first weak learner 21A, this is prioritized. Then, the second weak learner 21B is trained.
The second weak learner 21B is also learned by reflecting the comparison result of the predicted value 5A and the correct answer value 4C inside the second weak learner 21B, similarly to the first weak learner 21A. When it is determined that the learning of the second weak learner 21B is completed, if there is data with a large error between the predicted value 5A and the correct answer value 4C during the learning of the second weak learner 21B, this is prioritized. Then, the third weak learner 21C is further trained. The above process is repeated until the learning of all the weak learners 21 is completed.

As a result of learning each of the plurality of weak learners 21 under different conditions by the above-mentioned learning, when the same input is given to each of the plurality of weak learners 21, the plurality of weak learners 21 Make different predictions. In the present embodiment, the flow rate prediction device 1 inputs the weather observation data 2, the weather prediction data 3, and the flow rate data 4 to predict the flow rate 6, and therefore, the difference in this condition is, for example, the difference in the flow rate. That is, the case where the flow rate is large and the case where the flow rate is small are included. For example, one weak learner 21 performs learning focusing on the case where the flow rate is small, and another weak learner 21 performs learning focusing on the case where the flow rate is large.
Here, in the case where a particularly high prediction accuracy is required, for example, when it is assumed that the flow rate is large at the time of an abnormality such as a typhoon or a torrential rain, the data corresponding to this is that the flow rate corresponding to normal times is small. It is much less than the data. Therefore, it is basically not easy to perform learning such that the prediction accuracy becomes high when the flow rate is large. On the other hand, especially in the present embodiment, since boosting is applied as a learning method for ensemble learning, for example, when the flow rate is large in the first weak learner 21A, the error between the predicted value 5A and the correct answer value 4C When becomes large, in another weak learner 21 which is subsequently learned, learning can be performed with an emphasis on the case where the flow rate is large.

As described above, the prediction model learning unit 20 trains a plurality of weak learners 21. When the learning is completed, the prediction model learning unit 20 stores the parameters adjusted by the learning of each weak learner 21 in the prediction model parameter storage unit 16 as the prediction model parameters. The prediction model parameters stored in the prediction model parameter storage unit 16 are acquired by the prediction unit 17, which will be described later, and a plurality of prediction models 31 that predict the flow rate corresponding to each weak learner 21 are realized.
That is, the prediction model learning unit 20 generates a learned prediction model 31 in which appropriate learning parameters have been learned, which is used as a program module that is a part of artificial intelligence software.

Next, the behavior of each component at the time of predicting the flow rate will be described.

At the time of predicting the flow rate, the meteorological observation data 2, the meteorological prediction data 3, and the flow rate data 4 based on the present are input to the prediction model 31 for which the learning has been completed in order to predict the future flow rate of the river. .. That is, the input of the prediction model 31 is before the present, that is, current and past meteorological observation data (current meteorological observation data) 2B, and after the present, that is, future meteorological prediction data (current meteorological prediction data) 3B. , And the current flow data (current flow data) 4B.
The data processing unit 14 obtains the meteorological observation data 2B before the present, the future meteorological prediction data 3B, and the flow rate data 4B before the present from the meteorological observation data storage unit 11, the meteorological prediction data storage unit 12, and the flow rate storage unit 13. Acquire and convert each into the input format of the flow rate prediction unit 30.
The data processing unit 14 transmits each of the converted data to the prediction unit 17.

The prediction unit 17 receives each data from the data processing unit 14.
The prediction unit 17 predicts the future flow rate 6 of the river by the flow rate prediction unit 30 in which a plurality of prediction models 31 are built therein.
Each of the plurality of prediction models 31 is constructed corresponding to each of the plurality of weak learners 21 learned in the prediction model learning unit 20. For example, each of the first prediction model 31A, the first prediction model 31B, and the first prediction model 31C corresponds to the first weak learner 21A, the second weak learner 21B, and the third weak learner 21C, respectively. ing. The flow rate prediction unit 30 acquires the prediction model parameters corresponding to each weak learner 21 stored in the prediction model parameter storage unit 16, and builds a prediction model 31 corresponding to each weak learner 21 based on each of these. do.
The flow rate prediction unit 30 inputs each data received from the data processing unit 14 into each of the plurality of prediction models 31.
As described above, the plurality of weak learners 21 are learned under different conditions. That is, the prediction model 31 corresponding to each of the plurality of weak learners 21 outputs different values even if the same data is input to each of them.

The flow rate prediction unit 30 includes a coupler 32. The flow rate prediction unit 30 inputs the provisional future predicted flow rate 5B, which is the output of the plurality of prediction models 31, into the coupler 32. The coupler 32 calculates and predicts the future flow rate 6 based on these plurality of provisional future predicted flow rates 5B. The coupler 32 predicts the future flow rate 6 by, for example, calculating the sum, average, and weighting sum of a plurality of provisional future predicted flow rates 5B. When the coupler 32 calculates the weight sum, the weight depends on the performance of each weak learner 21, for example, the weight of the weak learner 21 having a low error rate, that is, a good regression accuracy is determined to be large. Will be done.

Next, the flow rate prediction method by the flow rate prediction device 1 will be described with reference to FIGS. 1 to 3 and FIG. FIG. 4A is a flowchart at the time of learning the flow rate, and FIG. 4B is a flowchart at the time of predicting the flow rate.
This flow rate prediction method is a flow rate prediction method for predicting the flow rate of a specific river, and learns meteorological observation data, meteorological prediction data, and flow rate data related to the river at an arbitrary time in the past. As data, multiple weak learners are ensemble-learned to predict the flow rate after the past time, multiple prediction models are generated, and each of the multiple prediction models is the current weather related to the river. Observation data, meteorological forecast data, and flow rate data are input to output a provisional future predicted flow rate, and the future flow rate of the river is predicted based on the plurality of the provisional future predicted flow rates.

First, the operation of each component of the flow rate prediction device 1 at the time of learning the flow rate will be described.
From the meteorological observation data storage unit 11, the meteorological prediction data storage unit 12, and the flow rate storage unit 13, the data processing unit 14 includes meteorological observation data 2A before the first past time, and meteorological prediction data 3A after the first past time. The flow rate data 4A before the first past time and the flow rate data 4C after the first past time are acquired, and each is converted into the input / output format of the weak learner 21.
The data processing unit 14 transmits each of the converted data to the learning unit 15.
The learning unit 15 receives each data from the data processing unit 14.
The learning unit 15 trains the weak learner 21 by the prediction model learning unit 20 to generate the learned prediction model 31 (step S1). In the present embodiment, the prediction model learning unit 20 uses the weather observation data 2A before the first past time, the weather prediction data 3A after the first past time, the flow rate data 4A before the first past time, and the correct answer value. A plurality of weak learners 21 are ensemble-learned so as to predict the predicted flow rate 5A after the first past time, using the flow rate data 4C after the first past time as training data.
As described above, the prediction model learning unit 20 trains a plurality of weak learners 21. In the present embodiment, since boosting is applied as a learning method for ensemble learning, a plurality of weak learning devices are used until it is determined that the learning of all the weak learning devices 21 is completed by the determination process shown in step S3. The vessels 21 are learned one by one in order. When the learning is completed, the prediction model learning unit 20 stores the parameters adjusted by the learning of each weak learner 21 in the prediction model parameter storage unit 16 as the prediction model parameters (step S5).

Next, the behavior of each component at the time of predicting the flow rate will be described.
The data processing unit 14 obtains the meteorological observation data 2B before the present, the future meteorological prediction data 3B, and the flow rate data 4B before the present from the meteorological observation data storage unit 11, the meteorological prediction data storage unit 12, and the flow rate storage unit 13. Acquire and convert each into the input format of the flow rate prediction unit 30.
The data processing unit 14 transmits each of the converted data to the prediction unit 17.
The prediction unit 17 receives each data from the data processing unit 14.
The flow rate prediction unit 30 acquires the prediction model parameters corresponding to each weak learner 21 stored in the prediction model parameter storage unit 16, and builds a prediction model 31 corresponding to each weak learner 21 based on each of these. (Step S11).
The flow rate prediction unit 30 inputs each data received from the data processing unit 14 into each of the plurality of prediction models 31. On the other hand, each of the plurality of prediction models 31 outputs a provisional future predicted flow rate of 5B (step S13).
The flow rate prediction unit 30 acquires a plurality of provisional future predicted flow rates 5B output by the plurality of prediction models 31 and inputs them to the coupler 32. The coupler 32 calculates and predicts the future flow rate 6 based on the plurality of provisional future predicted flow rates 5B (step S15).

Next, the effects of the flow rate prediction device 1 and the flow rate prediction method will be described.

The flow rate prediction device in the present embodiment is a flow rate prediction device 1 that predicts the flow rate 6 of a specific river, and is related to the river at a past time, which is an arbitrary time in the past, meteorological observation data 2A, and meteorological prediction data. Prediction model learning unit 20 that generates a plurality of prediction models 31 by ensemble learning a plurality of weak learners 21 so as to predict a flow rate 5A after the past time using 3A and flow data 4A and 4C as training data. And, each of the plurality of prediction models 31 is input with the current meteorological observation data 2B, the meteorological prediction data 3B, and the flow rate data 4B related to the river to output the provisional future predicted flow rate 5B. It is provided with a flow rate prediction unit 30 that predicts the future flow rate 6 of the river based on the provisional future prediction flow rate 5B of the above.
Further, this flow rate prediction method is a flow rate prediction method for predicting the flow rate 6 of a specific river, and is related to the river at a past time, which is an arbitrary time in the past, meteorological observation data 2A, meteorological prediction data 3A, Using the flow data 4A and 4C as training data, a plurality of weak learners 21 are ensemble-learned so as to predict the flow rate 5A after the past time, a plurality of prediction models 31 are generated, and the plurality of prediction models 31 are generated. Each of them inputs the current meteorological observation data 2B, the meteorological prediction data 3B, and the flow rate data 4B related to the river to output the provisional future prediction flow rate 5B, and a plurality of the provisional future predictions. The future flow rate 6 of the river is predicted based on the flow rate 5B.
According to the above configuration, a plurality of weak learners 21 are ensemble-learned to generate a plurality of prediction models 31, thereby predicting the future flow rate 6 of the river. That is, a plurality of weak learners 21 are trained to generate a plurality of prediction models 31 under different conditions including a difference in the flow rate, and a future flow rate 6 is predicted based on the plurality of results predicted by each of them. For this reason, even for data that is expected to have a large flow rate but may not be available in large numbers, such as in the case of a typhoon or torrential rain, a weak learner that emphasizes this and makes predictions. 21 exists. Therefore, it is possible to predict the flow rate with high accuracy even when the flow rate of the river is large.
Further, since the preparation before the introduction is completed by learning each weak learner 21, the introduction is easy.

In particular, in the present embodiment, when learning each weak learner 21, the data processing unit 14 only converts the data input format and output format, and changes the data content itself such as data deletion and duplication. Not. Therefore, it is difficult for human arbitrariness to intervene in the learning of the weak learner 21, and the loss of information from the data is suppressed.
As a result, the effect of high accuracy of flow rate prediction can be achieved more efficiently.

Further, the prediction model learning unit 20 ensemble learns each of the plurality of weak learners 21 by boosting.
For example, in the case of a typhoon or torrential rain, the predicted value is 5A when the flow rate is large in a weak learner 21 for data that is expected to have a large flow rate but may not be obtained in large numbers. The error between the correct answer value and the correct answer value of 4C becomes large, and there is a possibility that sufficient learning is not performed.
Even in such a case, according to the above configuration, in any other weak learner 21 that is learned after the weak learner 21, sufficient learning is performed by the weak learner 21. Learning is done with an emphasis on data that increases the flow rate that did not exist.
Therefore, it is possible to predict the flow rate with high accuracy even when the flow rate of the river is large.

In addition, the learning data includes meteorological observation data 2A before the past time related to the river, meteorological prediction data 3A after the past time related to the river, flow rate data 4A before the past time of the river, and the river as the correct answer value. This is the flow rate data 4C after the past time of And input the current and previous flow data 4B of the river.
According to the above configuration, the flow rate prediction device 1 and the flow rate prediction method can be appropriately realized.

[First Modified Example of Embodiment]
Next, with reference to FIGS. 5 and 6, a first modification of the flow rate prediction device 1 and the flow rate prediction method shown as the above-described embodiment will be described. FIG. 5 is a block diagram of the flow rate prediction device in the first modification. FIG. 6 is an explanatory diagram of a seasonal data generation unit of the flow rate prediction device in the first modification. The flow rate prediction device 40 and the flow rate prediction method in the first modification are different in that the learning data further includes seasonal data having different values depending on the season.

The flow rate prediction device 40 includes a seasonal data generation unit 43. When the date and time are input, the seasonal data generation unit 43 generates information on the season corresponding to the date and time as seasonal data. More specifically, the seasonal term data generation unit 43 includes a continuous function that outputs different values throughout the year according to the date and time, and when the date and time are input, the output of this continuous function is used as seasonal term data. Output.
Since December 31st, which is the end of the year, and January 1st, which is the beginning of the year, are dates before and after each other, it is considered that they are almost the same seasonally. Therefore, the above continuous function is also a periodic function in which the value is repeated so that the value is continuous between the output corresponding to December 31st and the output corresponding to January 1st, that is, the value is repeated with a cycle of one year. ..
As such a function, the sin function 65 and the cos function 66 can be applied. These functions 65 and 66 are periodic functions that are continuous and whose values are repeated with a period of one year. Furthermore, since at least one of the values of the sin function 65 and the cos function 66 has different values for different dates and times within a year, when these functions 65 and 66 are used together, the date and time are input. The corresponding seasonal data is uniquely determined.
The seasonal data generation unit 43 transmits the output to which the above functions 65 and 66 are applied to the input date and time to the data processing unit 44.

At the time of learning the flow rate, when the date and time data 7 corresponding to the first past time is input, the seasonal data generation unit 43 outputs the seasonal data corresponding to the first past time and processes the data. It is transmitted to the unit 14.
The data processing unit 44 includes the seasons corresponding to the first past time received from the seasonal data generation unit 43 in addition to the data acquired from the meteorological observation data storage unit 11, the weather prediction data storage unit 12, and the flow rate storage unit 13. The term data is converted as necessary and transmitted to the learning unit 45.
The learning unit 45 receives each data from the data processing unit 44.
The learning unit 45 trains the weak learner 51 by the prediction model learning unit 50 to generate the learned prediction model 61. In this modification, the prediction model learning unit 50 corresponds to the meteorological observation data before the first past time, the meteorological prediction data after the first past time, the flow rate data before the first past time, and the first past time. The seasonal term data and the correct answer value, that is, the flow rate data after the first past time are used as training data, and a plurality of weak learners 51 are ensemble-learned so as to predict the predicted flow rate after the first past time.
When the learning is completed, the prediction model learning unit 50 stores the parameters adjusted by the learning of each weak learner 51 in the prediction model parameter storage unit 46 as the prediction model parameters.

At the time of predicting the flow rate, when the date and time data 7 corresponding to the present is input, the seasonal data generation unit 43 outputs the seasonal data corresponding to the present and transmits it to the data processing unit 14.
The data processing unit 44 adds the data acquired and converted from the meteorological observation data storage unit 11, the weather prediction data storage unit 12, and the flow rate storage unit 13, and the seasonal term corresponding to the present received from the seasonal term data generation section 43. The data is transmitted to the prediction unit 47.
The prediction unit 47 receives each data from the data processing unit 44.
The flow rate prediction unit 60 acquires the prediction model parameters corresponding to each weak learner 51 stored in the prediction model parameter storage unit 46, and builds a prediction model 61 corresponding to each weak learner 51 based on each of these. do.
The flow rate prediction unit 60 inputs each data received from the data processing unit 44 into each of the plurality of prediction models 61. On the other hand, each of the plurality of prediction models 61 outputs a provisional future prediction flow rate.
The flow rate prediction unit 60 acquires a plurality of provisional future predicted flow rates output by the plurality of prediction models 61 and inputs them to the coupler. The coupler calculates and predicts the future flow rate 6 based on these plurality of provisional future predicted flow rates.

In the flow rate prediction device 40 of the present modification, the training data further includes seasonal term data having different values depending on the season, and the flow rate prediction unit 60 indicates the seasonal term corresponding to the present in each of the plurality of prediction models 61. The data is further input as learning data to output a provisional future predicted flow rate, and the future flow rate 6 of the river is predicted based on a plurality of provisional future predicted flow rates.
Snowfall during the winter months does not melt immediately, but melts when temperatures rise in April or May. Therefore, river flow is affected by snowmelt and tends to increase, especially in April and May.
According to the above configuration, the weak learner 51 can be trained in consideration of such seasonal factors such as snowmelt, and therefore, the future flow rate 6 is predicted in consideration of the seasonal factors. be able to.

Further, the seasonal term data is an output value of the continuous functions 65 and 66, and the periodic functions 65 and 66 whose values are repeated with a cycle of one year.
According to the above configuration, the flow rate prediction device 40 can be appropriately realized.

Needless to say, the first modification has other effects similar to those of the above-described embodiment.

[Second variant of the embodiment]
Next, with reference to FIGS. 7 and 8, a second modification of the flow rate prediction device 1 and the flow rate prediction method shown as the above-described embodiment will be described. FIG. 7 is a block diagram of the flow rate prediction device in the second modification. FIG. 8 is an explanatory diagram of the machine learning device and the feature extraction model in the second modification. This second modification is a further modification of the first modification. The flow rate prediction device 70 and the flow rate prediction method in the second modification are provided with a mesh feature amount learning unit that inputs mesh data and causes a machine learner to perform deep learning using the features of the mesh data as feature amounts. The learning device is different in that the mesh data features are input as training data instead of the mesh data.

At the time of learning the flow rate, the data processing unit 44 acquired and received from the meteorological observation data storage unit 11, the weather prediction data storage unit 12, the flow rate storage unit 13, and the seasonal data generation unit 43, as in the second modification. Each data is converted and transmitted to the learning unit 75.
The learning unit 75 includes a mesh feature amount learning unit 80, a mesh feature amount parameter storage unit 82, and a mesh feature amount extraction unit 83. The mesh feature amount learning unit 80 and the mesh feature amount extraction unit 83 may be, for example, software or a program executed by the CPU in the information processing apparatus. Further, the mesh feature amount parameter storage unit 82 may be realized by a storage device such as a semiconductor memory or a magnetic disk provided inside or outside the information processing device.

The mesh feature amount learning unit 80 receives the meteorological observation data before the first past time and the meteorological prediction data before the first past time from the data processing unit 44.
The mesh feature amount learning unit 80 includes a machine learning device 81. The machine learning device 81 performs convolutional deep learning, and is particularly realized by a convolutional autoencoder. The machine learning device 81 generates trained models 84 and 91, which will be described later, in which appropriate learning parameters have been learned, which are used as a program module that is a part of artificial intelligence software.
In this modification, the output of the feature extraction model 84 is used as an input for learning the weak learner 86 in the prediction model learning unit 85. Therefore, the learning of the machine learning device 81 is performed prior to the learning of the weak learning device 86.
As described above, the meteorological observation data 2 and the meteorological prediction data 3 include mesh data. The mesh feature amount learning unit 80 inputs the mesh data included in each of the meteorological observation data before the first past time and the meteorological prediction data before the first past time into the machine learning device 81 as input mesh data 81a.

The machine learning device 81 includes a convolution processing unit 81b, a fully connected unit 81c, and a deconvolution processing unit 81d.
The convolution processing unit 81b executes convolution processing, pooling processing, and the like on the input mesh data 81a to generate a feature map to be input to the fully connected unit 81c. The value of the feature map is stored in each input node of the fully connected portion 81c.
In the fully connected portion 81c, the weighted sum is calculated from the input node to the output node via the intermediate node based on the coupling load set between the nodes.
The value of each output node of the fully connected unit 81c is input to the deconvolution processing unit 81d. In the deconvolution processing unit 81d, processing symmetrical to that of the convolution processing unit 81b is executed. That is, the input mesh data 81a is compressed to a low dimension by the convolution processing unit 81b up to the fully connected portion 81c, but the deconvolution processing unit 81d operates so as to be restored from the compressed state to the low dimension. do. As a result, the output mesh data 81e having the same resolution as the input mesh data 81a is generated.
In the machine learning device 81, machine learning is performed so that the output mesh data 81e becomes mesh data that reproduces the input mesh data 81a.

As described above, the machine learning device 81 is a convolution autoencoder, and the input mesh data 81a is compressed to a low dimension in the convolution processing unit 81b, input to the fully connected unit 81c, compressed to the maximum, and then deconvolved. The processing unit 81d restores the reproduction mesh data 81e. That is, the fully connected portion 81c is a portion in which the features of the input mesh data 81a are in the most compressed state, and the value of the output node of the fully connected portion 81c abstractly condenses the features of the input mesh data 81a. It can be treated as a feature amount, that is, a mesh data feature amount 81f.

This modification conceptually utilizes the above-mentioned characteristics of the convolutional autoencoder. That is, the input mesh data 81a is obtained by inputting each of the input mesh data 81a to the feature extraction models 84 and 91, which are the machine learning devices 81 for which learning has been completed, and extracting the value of the output node of the fully connected portion 81c. It is converted into the mesh data feature amount 81f. The prediction model learning unit 85 makes each weak learner 86 ensemble learn using the mesh data feature amount 81f converted from each mesh data instead of each mesh data.
Therefore, in this modification, the portion from the convolution processing portion 81b to the fully connected portion 81c surrounded by the alternate long and short dash line in FIG. 8 corresponds to the feature extraction models 84 and 91 which are trained models. Therefore, the value of each parameter corresponding to this portion, that is, the value of the weight of each filter used in the convolution processing unit 81b, the value of each coupling load of the fully connected portion 81c, and the like are used as the mesh feature amount parameters. It is transmitted to and stored in the parameter storage unit 82.

As described above, the mesh feature amount parameter storage unit 82 stores each parameter of the machine learning device 81 whose learning has been completed by the mesh feature amount learning unit 80 as a mesh feature amount parameter.

The mesh feature extraction unit 83 acquires the mesh feature parameter of the machine learning device 81 stored in the mesh feature parameter storage unit 82, and constructs the feature extraction model 84 corresponding to the machine learning device 81.
The mesh feature amount extraction unit 83 inputs each mesh data included in the meteorological observation data before the first past time and the meteorological prediction data before the first past time into the feature extraction model 84 as input mesh data 81a, and inputs each. The mesh data feature amount 81f corresponding to the mesh data 81a is generated.
The mesh feature amount extraction unit 83 transmits each mesh data feature amount 81f to the prediction model learning unit 85.

The prediction model learning unit 85 trains the weak learner 86 to generate a trained prediction model 93. In this modification, the prediction model learning unit 85 is replaced with the meteorological observation data before the first past time in which the mesh data is replaced with the corresponding mesh data feature amount 81f, and the mesh data feature amount 81f in which the mesh data is replaced. Meteorological prediction data after the first past time, flow rate data before the first past time, seasonal term data corresponding to the first past time, and correct answer value, that is, flow rate data after the first past time are used as training data. Ensemble learning of a plurality of weak learners 86 so as to predict the predicted flow rate after the first past time.
When the learning is completed, the prediction model learning unit 85 stores the parameters adjusted by the learning of each weak learner 86 in the prediction model parameter storage unit 46 as the prediction model parameters.

At the time of predicting the flow rate, the data processing unit 44 acquired and received from the meteorological observation data storage unit 11, the weather prediction data storage unit 12, the flow rate storage unit 13, and the seasonal data generation unit 43, as in the first modification. Each data is transmitted to the prediction unit 77.
The prediction unit 77 receives each data from the data processing unit 44.
The prediction unit 77 includes a mesh feature extraction unit 90 having a feature extraction model 91. The mesh feature extraction unit 90 and the feature extraction model 91 have the same configuration as the mesh feature extraction unit 83 and the feature extraction model 84 provided in the learning unit 75.
The mesh feature extraction unit 90 inputs each mesh data included in the current and previous weather observation data and future weather prediction data received by the prediction unit 77 into the feature extraction model 91 as input mesh data 81a, and inputs each. The mesh data feature amount 81f corresponding to the mesh data 81a is generated.
The mesh feature amount extraction unit 90 transmits each mesh data feature amount 81f to the flow rate prediction unit 92.

The flow rate prediction unit 92 acquires the prediction model parameters corresponding to each weak learner 86 stored in the prediction model parameter storage unit 46, and builds a prediction model 93 corresponding to each weak learner 86 based on each of these. do.
The flow rate prediction unit 60 inputs each data received from the data processing unit 44 into each of the plurality of prediction models 93. Here, in the meteorological observation data before the present and the meteorological prediction data in the future, the mesh data included therein is replaced with the corresponding mesh data feature amount 81f and then input to each prediction model 93. In contrast, each of the plurality of prediction models 93 outputs a tentative future prediction flow rate.
The flow rate prediction unit 92 acquires a plurality of provisional future predicted flow rates output by the plurality of prediction models 93 and inputs them to the coupler. The coupler calculates and predicts the future flow rate 6 based on these plurality of provisional future predicted flow rates.

In the flow rate prediction device 70 of this modification, each of the meteorological observation data 2 and the meteorological prediction data 3 includes mesh data 81a in which the areas included as observation and prediction targets are divided into a mesh pattern by latitude and longitude on the map. The mesh feature amount learning unit 80, which inputs the mesh data 81a and causes the machine learner 81 to perform deep learning using the features of the mesh data 81a as the feature amount 81f, and the feature extraction models 84, 91, which are the machine learners 81 for which the learning has been completed. On the other hand, the prediction model learning unit 85 includes mesh feature extraction units 83 and 90 that input mesh data 81a and extract the feature amount 81f of the mesh data 81a, and the prediction model learning unit 85 replaces each of the mesh data 81a. Using the feature amount 81f obtained by inputting each of the mesh data 81a into the mesh feature amount extraction unit 83, a plurality of weak learners 86 are ensemble-learned, and the flow rate prediction unit 92 substitutes for each of the mesh data 81a. The future flow rate 6 of the river is predicted by using the feature amount 81f obtained by inputting each of the mesh data 81a into the mesh feature amount extraction unit 92.
When the river basin is large, the meteorological mesh may be large. Further, when it is desired to input values in many time series as meteorological observation data 2 and meteorological prediction data 3, the number of input dimensions becomes enormous. In such a case, if the mesh data is directly input to the weak learner 86 and the weak learner 86 is subjected to ensemble learning, the learning efficiency may not be good and highly accurate prediction may not be possible.
According to the above configuration, instead of the mesh data included in the meteorological observation data 2 and the meteorological prediction data 3, the features of these mesh data are reflected and the mesh data features are compressed to reduce the amount of data. The weak learner 86 is trained using the quantity 81f. As a result, the weak learner 86 can be trained more efficiently, and highly accurate prediction with higher generalization performance becomes possible.

Further, the machine learning device 81 and the feature extraction models 84 and 91 are realized by a convolutional autoencoder.
According to the above configuration, the flow rate prediction device 70 can be appropriately realized.

Needless to say, this second modification has other effects similar to those of the embodiment and the first modification already described.

The flow rate prediction device and the flow rate prediction method of the present invention are not limited to the above-described embodiment and each modification described with reference to the drawings, and various other modifications can be considered within the technical scope thereof. Be done.

For example, in the above embodiment and each modification, the predictive model learning unit ensemble learns each of the plurality of weak learners by boosting, but the present invention is not limited to this. That is, the predictive model learning unit may perform ensemble learning of each of the plurality of weak learners by bagging.
In bagging, each weak learner is made to learn by extracting a small amount of data from all the training data and inputting it. That is, in bagging, unlike boosting, each weak learner is learned in parallel. As already explained, the data in which the flow rate is expected to be large, such as in the case of a typhoon or torrential rain, is much smaller than the data in which the flow rate is small. However, for example, by extracting these small and large flow rate data and training one or more weak learners by specializing in the large flow rate data, the prediction accuracy is high when the flow rate is large. Predictive models can be generated.
As described above, even when bagging is adopted instead of boosting, the prediction accuracy when the flow rate is large can be improved by extracting a small amount of data.

In addition to this, as long as the gist of the present invention is not deviated, the configurations given in the above-described embodiment and each modification can be selected or changed to other configurations as appropriate.

1, 40, 70 Flow forecaster 2 Meteorological observation data 2A Meteorological observation data before the first past time (meteorological observation data at the past time)
2B Meteorological observation data before the present (current meteorological observation data)
2m, 3m mesh data 3 Meteorological prediction data 3A Meteorological prediction data after the first past time (weather prediction data at the past time)
3B Future weather forecast data (current weather forecast data)
4 Flow rate data 4A Flow rate data before the first past time (flow rate data at the past time)
4B Flow data before the present (current flow data)
4C Flow rate data after the first past time, correct answer value (flow rate data at the past time)
5A Predicted flow rate after the first past time (flow rate after the past time)
5B Temporary future predicted flow rate 6 Future flow rate 7 Date and time data 16,46 Predictive model parameter storage unit 20, 50, 85 Predictive model learning unit 21, 51, 86 Weak learner 30, 60, 92 Flow rate prediction unit 31, 61, 93 Predictive model 43 Seasonal data generation unit 65, 66 Function 80 Mesh feature amount learning unit 81 Machine learner 81a Input mesh data (mesh data)
81e Output mesh data 81f Mesh data Feature (feature)
82 Mesh feature parameter storage unit 83, 90 Mesh feature extraction unit 84, 91 Feature extraction model

Claims (9)

A flow rate predictor that predicts the flow rate of a specific river.
A plurality of weak learners so as to predict the flow rate after the past time by using the meteorological observation data, the weather prediction data, and the flow rate data related to the river at an arbitrary time in the past as learning data. Predictive model learning unit that generates multiple predictive models by learning the ensemble,
Each of the plurality of prediction models is input with the current meteorological observation data, the meteorological prediction data, and the flow rate data related to the river to output a provisional future predicted flow rate, and a plurality of the said ones. A flow rate prediction unit that predicts the future flow rate of the river based on the provisional future forecast flow rate,
A flow rate predictor. The flow rate prediction device according to claim 1, wherein the prediction model learning unit ensemble learns each of the plurality of weak learners by boosting. The flow rate prediction device according to claim 1, wherein the prediction model learning unit ensemble learns each of the plurality of weak learners by bagging. The learning data includes meteorological observation data before the past time related to the river, meteorological prediction data after the past time related to the river, flow rate data before the past time of the river, and correct answer values. It is the flow data after the past time of the river of
The flow rate prediction unit inputs into each of the plurality of prediction models the current and previous meteorological observation data related to the river, the future weather prediction data related to the river, and the current and previous flow rate data of the river. , The flow rate prediction device according to any one of claims 1 to 3. The training data further includes seasonal data having different values depending on the season.
The flow rate prediction unit further inputs the seasonal data corresponding to the present to each of the plurality of prediction models as training data to output the provisional future predicted flow rate, and causes the plurality of the provisional future to be output. The flow rate prediction device according to any one of claims 1 to 4, which predicts the future flow rate of the river based on the predicted flow rate of the above. The flow rate predictor according to claim 5, wherein the seasonal data is a continuous function and is an output value of a periodic function whose value is repeated with a cycle of one year. Each of the meteorological observation data and the meteorological prediction data includes mesh data in which the areas included as observation and prediction targets are divided into a mesh pattern by latitude and longitude on a map.
A mesh feature learning unit that inputs the mesh data and causes the machine learner to perform deep learning using the features of the mesh data as features.
A mesh feature extraction unit that inputs the mesh data to the feature extraction model of the machine learning device for which learning has been completed and extracts the feature amount of the mesh data.
With
The prediction model learning unit ensembles the plurality of weak learners by using the feature amount obtained by inputting each of the mesh data into the mesh feature amount extraction unit instead of each of the mesh data. Learn and
The flow rate prediction unit predicts the future flow rate of the river by using the feature amount obtained by inputting each of the mesh data into the mesh feature amount extraction unit instead of each of the mesh data. , The flow rate prediction device according to any one of claims 1 to 6. The flow rate prediction device according to claim 7, wherein the machine learning device and the feature extraction model are realized by a convolutional autoencoder. It is a flow rate prediction method that predicts the flow rate of a specific river.
A plurality of weak learners so as to predict the flow rate after the past time by using the meteorological observation data, the weather prediction data, and the flow rate data related to the river at the past time which is an arbitrary time in the past as learning data. Ensemble learning to generate multiple predictive models,
Each of the plurality of prediction models is input with the current meteorological observation data, the meteorological prediction data, and the flow rate data related to the river to output a provisional future predicted flow rate, and a plurality of the said ones. A flow rate prediction method for predicting the future flow rate of the river based on a provisional future predicted flow rate.

Images