JP2024018314A - Weather forecasting device and weather forecasting method - Google Patents

Abstract

[Problem] To obtain highly accurate weather-related indicators based on a plurality of weather-related forecasts.
A weather prediction device 20 includes a prediction target forecast data acquisition unit 21 that acquires prediction target forecast data including a plurality of forecast data that predicts at least one weather parameter including the prediction target at the time of the prediction target; A forecast observation data acquisition unit that acquires forecast observation data including observed values of meteorological parameters to be predicted at the time of the forecast target by inputting forecast data into a prediction model, the forecast observation data comprising a plurality of observed values or A predicted observation data acquisition unit 22 that includes a probability distribution of observed values, an observation value distribution calculation unit 23 that calculates a predicted observation data distribution that is a probability distribution of observed values, etc. based on the predicted observation data, and a predicted observation data distribution. and an index acquisition unit 24 that acquires the index by inputting the index into an index prediction model consisting of a classification model.
[Selection diagram] Figure 2

Description

The present invention relates to a weather prediction device and a weather prediction method for predicting predetermined weather-related indicators.

Conventionally, weather forecasters have judged and evaluated weather-related forecast values forecasted by meteorological organizations, and predetermined weather-related indicators have been forecasted. For example, indicators for the rainy season, which is a period of high precipitation, and the dry season, which is a period of low precipitation, have been predicted based on forecast values such as precipitation. Furthermore, Patent Document 1 discloses a technique for predicting spatiotemporally localized future weather condition fluctuations, including estimation errors, based on past time-series data related to weather. This technique provides an estimate of the prediction error based on the standard deviation.

Patent No. 5339317

Weather-related forecast values forecasted by meteorological organizations may deviate from actual observed values, so the accuracy of forecasting indicators based on forecast values was low. Furthermore, the forecast values forecasted by meteorological agencies are different for each of the plurality of meteorological agencies even if they target the same weather parameter, and it has been difficult to forecast indicators based on the plurality of forecast values.

Accordingly, one aspect of the present disclosure aims to obtain highly accurate weather-related indicators based on a plurality of weather-related forecasts.

In order to solve the above problems, a weather forecasting device according to one aspect of the present disclosure is a weather forecasting device that predicts a predetermined weather-related index, and targets meteorological parameters that are parameters representing specific weather. , a prediction target forecast data acquisition unit that acquires prediction target forecast data including a plurality of forecast data that predict one or more weather parameters including at least the prediction target weather parameter at the prediction target time, which is the target point of the index prediction; , a forecast observation data acquisition unit that acquires forecast observation data regarding a weather parameter to be predicted at the time of the forecast target by inputting the forecast data to be predicted into a prediction model, the forecast observation data being a plurality of observations of the weather parameter; a predicted observation data acquisition unit that includes a value or a predetermined analysis value based on the observed value, or a probability distribution of the observed value or analysis value; and a prediction that is a probability distribution of the observed value or analysis value based on the predicted observation data. an observation value distribution calculation unit that calculates the observation data distribution; and an indicator acquisition unit that acquires weather-related indicators by inputting the predicted observation data distribution into an indicator prediction model, each of which has a predetermined weather-related an index acquisition unit consisting of a classification model that classifies the predicted observed data represented by the predicted observed data distribution into two or more classes corresponding to the indicators, and the predicted model is based on a predetermined statistical analysis method. , is a regression model generated by machine learning using multiple forecast data as explanatory variables and observed data as an objective variable. The observation data is data that is a forecast of one or more weather parameters including weather parameters, and the observed data is an observed value of the weather parameter to be predicted at the time of the forecast target or a predetermined analysis value based on the observed value.

In order to solve the above problems, a weather forecasting method according to one aspect of the present disclosure is a weather forecasting method in a weather forecasting device that predicts a predetermined weather-related index, the weather forecasting method using a weather parameter that is a parameter representing a specific weather. Predicted target forecast data that is forecast data that predicts one or more weather parameters including at least the meteorological parameter of the predicted target at the forecasted time which is the target point of the prediction of the index. an acquisition step, and a forecast observation data acquisition step of obtaining forecast observation data regarding the meteorological parameter to be predicted at the time of the forecast target by inputting the forecast data to the prediction model, the forecast observation data being A predicted observation data acquisition step including a plurality of observation values or a predetermined analysis value based on the observation values, or a probability distribution of the observation values or analysis values, and a probability distribution of the observation values or analysis values based on the prediction observation data. an observation value distribution calculation step for calculating a predicted observation data distribution, and an index acquisition step for obtaining weather-related indicators by inputting the predicted observation data distribution into an index prediction model, each of which has a an index acquisition step consisting of a classification model that classifies the predicted observation data represented by the predicted observation data distribution into two or more classes corresponding to predetermined weather-related indicators, and the prediction model is based on a predetermined statistical analysis method is a regression model generated by machine learning based on forecast data as an explanatory variable and observed data as an objective variable. The observation data is data that is a forecast of one or more meteorological parameters including: , and the observed data is an observed value of the meteorological parameter to be predicted at the time of the forecast, or a predetermined analysis value based on the observed value.

According to the above aspect, by inputting a plurality of forecast data to be predicted into a prediction model consisting of a regression model generated by machine learning with the plurality of forecast data as explanatory variables and observed data at the time of the prediction target as an objective variable. , forecast observation data can be obtained that takes into account fluctuations in the forecast data. In addition, because predicted observation data includes multiple observed values or analysis values (hereinafter referred to as "observed values, etc.") to be predicted, or probability distributions of observed values, etc., predicted observation data includes uncertainties in predicting observed values, etc. Contains the meaning. The predicted observed data distribution calculated based on the predicted observed data that includes uncertainty represents the variance of the observed values to be predicted, so the predicted observed data distribution can be input to the indicator prediction model. This makes it possible to obtain highly accurate indicators.

In the weather forecasting device according to another aspect, the predicted observation data acquisition unit may acquire the predicted observation data using a Monte Carlo dropout method.

According to the above aspect, it is possible to easily obtain predicted observation data that includes a plurality of observed values and appropriately represents the uncertainty in the observed values.

In the weather prediction device according to another aspect, the prediction model is a model configured based on Gaussian process regression, and the prediction observation data acquisition unit acquires prediction observation data including a probability distribution of observed values or analysis values. It may also be a thing.

According to the above aspect, it is possible to easily obtain predicted observation data that includes a probability distribution of observed values and appropriately represents uncertainty in observed values.

In the weather forecasting device according to another aspect, the predicted observation data distribution may be represented by the average value and standard deviation of observed values or analyzed values.

According to the above aspect, it is possible to suitably express the uncertainty of the observed value, etc. of the prediction target in the predicted observation data distribution. Then, by inputting the predicted observation data distribution configured in this way to the indicator prediction model, a highly accurate indicator can be obtained.

In the weather prediction device according to another aspect, the prediction model may be a regression model configured by any one of a Gaussian process, a random forest, a support vector machine, and a neural network.

According to the above aspect, since the prediction model is configured by the regression model exemplified above, it is possible to obtain predicted observation data that appropriately includes elements of error (uncertainty).

In the weather forecasting device according to another aspect, the indicator prediction model may be either a classification model that classifies the predicted observation data distribution based on a given threshold value or a classification model configured by a support vector machine. .

According to the above aspect, since the indicator prediction model is constituted by the classification model exemplified above, weather-related indicators can be easily obtained.

In the weather forecasting device according to another aspect, the weather parameter to be predicted is precipitation, the forecast data includes at least a forecast of precipitation, and the indicator prediction model is based on a given threshold regarding precipitation. It is also possible to output an index indicating whether the time to be predicted falls under either the rainy season or the dry season.

According to the above aspect, an indicator indicating either the rainy season or the dry season can be easily obtained with high accuracy by the prediction model and indicator prediction model constructed using precipitation as the meteorological parameter to be predicted.

According to one aspect of the present disclosure, it is possible to obtain highly accurate weather-related indicators based on a plurality of weather-related forecasts.

FIG. 2 is a block diagram showing the functional configuration of a predictive model generation device. FIG. 2 is a block diagram showing the functional configuration of a weather forecasting device. It is a diagram showing the hardware configuration of a weather forecasting device and a terminal. FIG. 2 is a diagram showing a schematic configuration of a prediction model. It is a figure showing the structure of forecast data as learning data used for learning and generation of a prediction model, and an example of data included. FIG. 2 is a diagram showing an example of the structure and included data of observation data as learning data used for learning and generating a predictive model. FIG. 2 is a diagram illustrating a schematic configuration of a prediction model, an index prediction model, and the flow of various data in the phase of predicting prediction observation data and acquiring weather-related indicators. FIG. 2 is a diagram illustrating an outline of a process for acquiring an index indicating a rainy season and a dry season using precipitation as a meteorological parameter to be predicted. It is a flowchart which shows the processing content of a weather prediction method.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In addition, in the description of the drawings, the same or equivalent elements are given the same reference numerals, and redundant description will be omitted.

FIG. 1 is a block diagram showing the functional configuration of a predictive model generation device 10 according to this embodiment. The prediction model generation device 10 is a device that generates a prediction model that predicts observed values based on weather parameters. The prediction model is a regression model generated by machine learning based on a predetermined statistical analysis method using a plurality of forecast data as explanatory variables and observation data as an objective variable.
Weather parameters are parameters representing specific weather conditions, such as precipitation, temperature, and atmospheric pressure.

As shown in FIG. 1, the predictive model generation device 10 is composed of a computer including a processor 101, and functionally includes a forecast data acquisition section 11, an observed data acquisition section 12, a predictive model generation section 13, and a model output section. 14. Each of these functional units 11 to 14 is realized by loading a program P1 (prediction model generation program) into the processor 101, which will be explained with reference to FIG. 3, and executing the program. A description of each functional section will be given later. Note that in this embodiment, each of the functional units 11 to 14 is configured in the weather forecasting device 20, but they may be configured in a distributed manner in a plurality of computers.

The prediction model generation device 10 also includes storage means such as a forecast data storage section 15, an observation data storage section 16, and a prediction model storage section 17. In the example shown in FIG. 1, these storage units 15, 16, and 17 are configured in the predictive model generation device 10, but they may be configured in other devices configured to be accessible from the predictive model generation device 10. may be configured. A description of each storage section will be given later.

FIG. 2 is a block diagram showing the functional configuration of the weather forecasting device 20 according to this embodiment. The weather prediction device 20 is a device that predicts predetermined weather-related indicators.

The weather prediction device 20 is composed of a computer including a processor 101, and functionally includes a prediction target forecast data acquisition unit 21, a prediction observation data acquisition unit 22, an observed value distribution calculation unit 23, an index acquisition unit 24, and an index output unit. 25. Each of these functional units 21 to 25 is realized by loading a program P1 (weather prediction program) into the processor 101, which will be explained with reference to FIG. 3, and executing the program. A description of each functional section will be given later. Note that in this embodiment, each of the functional units 21 to 25 is configured in the weather forecasting device 20, but they may be configured in a distributed manner in a plurality of computers.

The weather prediction device 20 also includes storage means such as a prediction target forecast data storage section 26 and a prediction model storage section 17. Although these storage units 26 and 27 are configured in the weather prediction device 20 in the example shown in FIG. 2, they may be configured to be accessible from the weather prediction device 20. Further, the predictive model generation device 10 and the weather prediction device 20 may be configured as one device. A description of each storage section will be given later.

FIG. 3 is a diagram showing an example of the hardware configuration of the predictive model generation device 10 and the weather prediction device 20, and shows a computer 100 that functions as the predictive model generation device 10 or the weather prediction device 20.

As an example, the computer 100 includes a processor 101, a main storage section 102, an auxiliary storage section 103, and a communication section 104 as hardware components.

Processor 101 is a computing device that executes an operating system and application programs. Examples of processors include a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit), but the type of processor 101 is not limited to these. For example, processor 101 may be a combination of dedicated circuits. The dedicated circuit may be a programmable circuit such as an FPGA (Field-Programmable Gate Array), or may be another type of circuit.

The main storage unit 102 is a device that stores programs for implementing the predictive model generation device 10 or the weather forecasting device 20, calculation results output from the processor 101, and the like. The main storage unit 102 includes, for example, at least one of ROM (Read Only Memory) and RAM (Random Access Memory).

The auxiliary storage unit 103 is generally a device that can store a larger amount of data than the main storage unit 102. The auxiliary storage unit 103 is configured by a nonvolatile storage medium such as a hard disk or flash memory. The auxiliary storage unit 103 stores a program P1 and various data for causing the computer 100 to function as the predictive model generation device 10 or the weather forecasting device 20. In this embodiment, a prediction model generation program for realizing the prediction model generation device 10 and a weather prediction program for realizing the weather prediction device 20 are implemented as a program P1. Furthermore, when each of the storage units 15 to 17 and 26 to 27 is included in either the predictive model generation device 10 or the weather forecasting device 20, each of the storage units 15 to 17 and 26 to 27 is included in the main storage unit 102, It may be configured in either the auxiliary storage unit 103 or other storage elements.

The communication unit 104 is a device that performs data communication with other computers via a communication network. The communication unit 104 is configured by, for example, a network card or a wireless communication module.

Each functional element of the predictive model generation device 10 and the weather forecasting device 20 is realized by loading a corresponding program P1 onto the processor 101 or the main storage unit 102 and causing the processor 101 to execute the program. The program P1 includes codes for realizing each functional element of the predictive model generation device 10 or the weather forecasting device 20. The processor 101 operates the communication unit 104 according to the program P1 to read and write data in the main storage unit 102 or the auxiliary storage unit 103. Through such processing, each functional element of the predictive model generation device 10 and the weather forecasting device 20 is realized.

The program P1 may be provided after being permanently recorded on a tangible recording medium such as a CD-ROM, a DVD-ROM, or a semiconductor memory. Alternatively, the program P1 may be provided via a communication network as a data signal superimposed on a carrier wave.

Referring again to FIG. 1, each functional unit and each storage unit of the predictive model generation device 10 will be explained. The forecast data acquisition unit 11 acquires learning forecast data including a plurality of forecast data. The forecast data is data that predicts one or more weather parameters including at least the weather parameters to be predicted in the prediction model. Forecast data is data obtained by predicting weather parameters at the time of the forecast target, which is the time of the target of the forecast, before the forecast target time.

FIG. 4 is a diagram showing a schematic configuration of a prediction model. As shown in FIG. 4, the prediction model M1 is generated by machine learning based on a predetermined statistical analysis method, using learning forecast data d11 including a plurality of forecast data as an explanatory variable and observation data d12 as an objective variable. It is constructed by a regression model.

In this embodiment, as an example, the prediction model M1 uses precipitation as a weather parameter to be predicted, so the observation data d12 includes the observed value of precipitation. Further, the prediction model M1 uses one or more meteorological parameters including at least the amount of precipitation as forecast data in order to predict the amount of precipitation. Specifically, the forecast data includes precipitation, temperature, and atmospheric pressure.

The forecast data acquisition unit 11 may acquire a plurality of forecast data based on forecasts of each of a plurality of meteorological agencies as the learning forecast data d11. Specifically, the forecast data acquisition unit 11 may acquire the learning forecast data d11 from the forecast data storage unit 15. The forecast data storage section 15 is a storage means that stores in advance learning forecast data d11 consisting of forecast data predicted by a weather organization.

FIG. 5 is a diagram showing an example of the structure and included data of the learning forecast data d11 stored in the forecast data storage unit 15. As shown in FIG. 5, the forecast data storage unit 15 stores learning forecast data d11, which is forecast data that predicts weather parameters at the time of the forecast target, which is the time point of the forecast target. Further, the learning forecast data d11 stores a plurality of forecast data (forecast data 1, forecast data 2, forecast data 3) that predict weather parameters at the time of the same forecast target. These forecast data include forecast values that are predicted, for example, 14 days before the forecast target time when the forecast is output. Further, the first forecast data includes precipitation, temperature, and atmospheric pressure as meteorological parameters to be forecasted.

Since meteorological agencies independently forecast weather parameters, forecast data based on the forecasts of each of the plurality of meteorological agencies includes forecast values of different weather parameters. That is, the forecast data acquisition unit 11 acquires a plurality of forecast data as exemplified as forecast data 1, forecast data 2, and forecast data 3 in FIG. Learning forecast data d11 including fluctuations caused by this will be obtained.

Further, the forecast data acquisition unit 11 may acquire a plurality of forecast data based on each of a plurality of forecasts by one meteorological agency. A meteorological agency may output multiple or distributed forecast data regarding one forecast target, for example, using a method such as ensemble forecasting. The forecast data storage unit 15 may store, as learning forecast data d11, a plurality of forecast data output by one meteorological agency regarding the same forecast target. Also in this case, the forecast data acquisition unit 11 acquires learning forecast data d11 including fluctuations by acquiring a plurality of forecast data regarding the same forecast target.

Referring again to FIG. 1, the observation data acquisition unit 12 acquires observation data d12, which is the observed value of the meteorological parameter to be predicted at the time of the forecast target or a predetermined analysis value based on the observed value (note that the observed value (or analysis values may be abbreviated as "observed values, etc." hereinafter).
The observation data d12 may be an observed value of a weather parameter to be predicted in the prediction model M1, or may be an analytical value obtained by a predetermined calculation based on the observed value.

Specifically, the observation data acquisition unit 12 may acquire observation data d12 from the observation data storage unit 16. The observation data storage unit 16 is a storage unit that previously stores observation data d12 observed by, for example, a meteorological agency.

FIG. 6 is a diagram showing an example of the structure of observation data stored in the observation data storage unit 16 and the data included therein. As shown in FIG. 6, the observation data storage unit 16 stores the observed value of precipitation, which is a meteorological parameter to be predicted by the prediction model M1, at a time when the forecast is to be made. A pair of learning forecast data d11 associated with the same forecast target time and observation data d12 of the prediction target constitutes one learning data for machine learning of the prediction model M1.

Referring again to FIGS. 1 and 4, the predictive model generation unit 13 performs machine learning on the predictive model M1 using the learning forecast data d11 as an explanatory variable and the observed data d12 as an objective variable, based on a predetermined statistical analysis method. Generated by The prediction model M1 may be, for example, a regression model configured by any one of a Gaussian process, a random forest, a support vector machine, and a neural network. All of these exemplary regression models are well known in the art to those skilled in the art. The predictive model generation unit 13 updates the weights and the like constituting the predictive model M1 based on a machine learning method, for example, an error backpropagation learning method. That is, as shown by the arrow ud, the prediction model generation unit 13 generates error information er between the model output data mo output from the prediction model M1 and the observation data d12 in response to the input of the learning forecast data d11. The prediction model M1 is trained by updating the weights and the like included in the prediction model M1.

The model output unit 14 outputs the prediction model M1 generated by the prediction model generation unit 13. Specifically, the model output unit 14 causes the predictive model storage unit 17 to store the learned predictive model M1. The prediction model storage unit 17 is a storage means for storing the prediction model M1 that has been learned or is in the learning process.

Next, each functional unit and each storage unit of the weather forecasting device 20 will be explained with reference to FIGS. 2 and 7. FIG. 7 is a diagram illustrating a schematic configuration of a prediction model, an index prediction model, and the flow of various data in the phase of prediction of prediction observation data and acquisition of weather-related indicators.

The prediction target forecast data acquisition unit 21 acquires the prediction target forecast data d21. The prediction target forecast data d21 includes a plurality of forecast data in which one or more weather parameters including at least the weather parameter to be predicted are predicted at the prediction target time, which is the target point of time for predicting the weather-related index. In this embodiment, the forecast data included in the prediction target forecast data d21 includes forecast values of precipitation, temperature, and atmospheric pressure.

Specifically, the prediction target forecast data acquisition unit 21 may acquire the prediction target forecast data d21 from the prediction target forecast data storage unit 26. The prediction target forecast data storage unit 26 is a storage means that stores the prediction target forecast data d21 in advance.

The prediction target forecast data d21 has the same format as the learning forecast data d11. That is, as illustrated in FIG. 5, when the learning forecast data d11 consists of three forecast data, as will be explained later with reference to FIG. 8, the prediction target forecast data d21 is the learning forecast data d11. Similarly, it consists of three forecast data. The prediction target forecast data d21 may include a plurality of forecast data forecasted by each of a plurality of meteorological agencies. Further, the forecast target forecast data d21 may include a plurality of forecast data based on a plurality of forecast values from one meteorological agency.

The forecast observation data acquisition unit 22 inputs the forecast target forecast data d21 into the prediction model M1, thereby obtaining forecast observation data including the observed value of the weather parameter to be predicted at the time of the forecast target or a predetermined analysis value based on the observed value. Get d22. The predictive observation data acquisition unit 22 can refer to and acquire the learned predictive model M1 stored in the predictive model storage unit 17.

The predicted observation data d22 includes a plurality of observation values md. Further, the predicted observation data d22 may include a probability distribution of md such as observed values. Although there are no limitations on the method for acquiring a plurality of observed values, etc. md or by using a probability distribution, for example, the predicted observation data acquisition unit 22 uses a Monte Carlo dropout method to obtain the predicted observed data d22 consisting of a plurality of observed values, etc. md. You may obtain it. Monte Carlo dropout is a method to obtain multiple outputs that make up a distribution by randomly switching neurons in a neural network or randomly eliminating parts of the hidden layer. By acquiring the predicted observation data d22 in this manner, it is possible to easily obtain predicted observation data that includes a plurality of observed values and appropriately represents the uncertainty in the observed values.

Further, when the prediction model M1 is a model configured based on Gaussian process regression, the prediction observation data acquisition unit 22 may acquire prediction observation data d22 consisting of a probability distribution of observed values, etc. md. . By acquiring the predicted observation data d22 in this way, it is possible to easily obtain predicted observation data that includes the probability distribution of the observed values and appropriately represents the uncertainty in the observed values.

The observed value distribution calculation unit 23 calculates a predicted observed data distribution d23, which is a probability distribution of observed values md, based on the predicted observed data d22. The predicted observed data distribution d23 is expressed, for example, by the mean value E and standard deviation σ of the observed values, etc. md. Specifically, the observation value distribution calculation unit 23 calculates the predicted observation data distribution d23 based on the plurality of observation values md. Further, the observed value distribution calculation unit 23 may obtain the predicted observed data distribution d23 based on the distribution of observed values etc. md.

In this way, the uncertainty of the observed values to be predicted can be suitably expressed in the predicted observed data distribution d23. Then, by inputting the predicted observation data distribution d23 configured in this way to the indicator prediction model M2, a highly accurate indicator can be obtained.

The index acquisition unit 24 acquires a weather-related index d24 by inputting the predicted observation data distribution d23 into the index prediction model M2. The indicator prediction model M2 consists of a classification model that classifies the predicted observation data represented by the predicted observation data distribution d23 into two or more classes, each of which corresponds to a predetermined indicator d24 related to weather.

Although the type of classification model is not limited, the indicator prediction model M2 may be either a classification model that classifies the predicted observed data distribution d23 based on a given threshold value and a classification model configured by a support vector machine. .

When the index prediction model M2 is configured by a classification model that classifies the predicted observed data distribution d23 based on a given threshold, the threshold may be an arbitrarily set value, and the predicted observed data distribution It may be set based on the past performance of the combination of d23 and the index.

In addition, when the indicator prediction model M2 is configured by a classification model configured by a support vector machine, the indicator prediction model M2 is It may also be generated.

The index output unit 25 outputs the index d24 acquired by the index acquisition unit 24. Although the output mode is not limited, the index output unit 25 outputs the acquired index d24 by storing it in a predetermined storage means, displaying it on a predetermined display device, or transmitting it to a predetermined device. You may.

Next, with reference to FIG. 8, a specific example of weather prediction processing for predicting weather-related indicators will be described. FIG. 8 is a diagram illustrating an outline of a process for acquiring an index indicating a rainy season and a dry season using precipitation as a meteorological parameter to be predicted. Further, the forecast data includes at least the amount of precipitation, and includes the amount of precipitation, temperature, and atmospheric pressure.

The prediction target forecast data acquisition unit 21 acquires the prediction target forecast data d21. The prediction target forecast data d21 is based on forecast data output 14 days before the prediction target time, with “2022/6/1 0:00” as the prediction target time. The prediction target forecast data d21 includes forecast data 1, 2, and 3. These plurality of forecast data may be a plurality of forecast data forecasted by one meteorological agency, or may be forecast data forecasted by each of a plurality of meteorological agencies.

The forecast observation data acquisition unit 22 inputs the forecast target forecast data d21 into the prediction model M1, thereby acquiring forecast observation data d22 including the observed values of the weather parameters of the forecast target at the time of the forecast target. The predicted observation data d22 shown in FIG. 8 includes, for example, a plurality of observed values md outputted from the prediction model M1 by a method such as Monte Carlo dropout. The observed value md is a predicted value of precipitation at the time of prediction.

The observed value distribution calculation unit 23 calculates a predicted observed data distribution d23, which is a probability distribution of observed values md included in the predicted observed data d22. The predicted observed data distribution d23 is represented by the average value E and standard deviation σ of observed values. For example, using an arbitrarily set weight n, the predicted observed data distribution d23(f(t)) is expressed by the following equation (1), as shown in FIG.
f(t)=E(t)+nσ(t)...(1)
t in equation (1) represents time.

The index acquisition unit 24 acquires a weather-related index d24 by inputting the predicted observation data distribution d23 into the index prediction model M2. In the example shown in FIG. 8, the index prediction model M2 is configured by a classification model that classifies the predicted observed data distribution d23 based on a threshold value. Specifically, the index prediction model M2 sets the threshold regarding precipitation to 1 mm, and the upper limit of precipitation that takes distribution into account, which is calculated based on the predicted observation data distribution d23 (formula (1)), exceeds the threshold. If it exceeds the threshold, index 1 (rainy season) is output, and if it does not exceed the threshold, index 0 (dry season) is output.

Next, with reference to FIG. 9, an example of the operation of the weather forecasting device 20 of this embodiment will be described. FIG. 9 is a flowchart showing the processing details of the weather prediction method performed by the weather prediction device 20.

In step S1, the prediction target forecast data acquisition unit 21 acquires the prediction target forecast data d21.

In step S2, the prediction observation data acquisition unit 22 inputs the prediction target forecast data d21 into the prediction model M1. Then, in step S3, the predicted observation data acquisition unit 22 acquires predicted observation data d22 including the observed values of the meteorological parameters of the prediction target at the time of the prediction target. The predicted observation data d22 includes a plurality of observed values, etc., or a probability distribution of observed values, etc.

In step S4, the observed value distribution calculation unit 23 calculates a predicted observed data distribution d23, which is a probability distribution of the observed values md represented by the predicted observed data d22.

In step S5, the index acquisition unit 24 inputs the predicted observation data distribution d23 to the index prediction model M2. Then, in step S6, the index acquisition unit 24 acquires the weather-related index d24 as the output of the index prediction model M2.

In step S7, the index output unit 25 outputs the index d24 acquired by the index acquisition unit 24.

According to the weather forecasting device 20, the weather forecasting method, and the weather forecasting program of the present embodiment described above, a regression model generated by machine learning using a plurality of forecast data as explanatory variables and observed data at the time of prediction as an objective variable. By inputting a plurality of target forecast data into a prediction model, prediction observation data that takes into account fluctuations in the forecast data can be obtained. Moreover, since the predicted observation data includes a plurality of observed values to be predicted or a probability distribution of the observed values, etc., the predicted observation data includes uncertainty in prediction of the observed values. The predicted observed data distribution calculated based on the predicted observed data that includes uncertainty represents the variance of the observed values to be predicted, so the predicted observed data distribution can be input to the indicator prediction model. This makes it possible to obtain highly accurate indicators.

The present invention has been described above in detail based on the embodiments thereof. However, the present invention is not limited to the above embodiments. The present invention can be modified in various ways without departing from the gist thereof.

The gist of the present invention will be shown below.
[1]
A weather prediction device that predicts predetermined weather-related indicators,
Forecast data in which a weather parameter that is a parameter representing a specific weather is the target of forecasting, and one or more of the weather parameters including at least the meteorological parameter to be predicted are predicted at the time of the prediction target which is the target time point of the prediction of the index. a prediction target forecast data acquisition unit that acquires prediction target forecast data including a plurality of prediction target forecast data;
A forecast observation data acquisition unit that acquires forecast observation data regarding the weather parameter of the prediction target at the time of the forecast target by inputting the forecast target forecast data into a prediction model, the forecast observation data comprising: a predicted observation data acquisition unit that includes a plurality of observed values or a predetermined analytical value based on the observed values, or a probability distribution of the observed values or the analytical values;
an observed value distribution calculation unit that calculates a predicted observed data distribution that is a probability distribution of the observed values or the analyzed values based on the predicted observed data;
An indicator acquisition unit that obtains weather-related indicators by inputting the predicted observation data distribution into an indicator prediction model, wherein the indicator prediction model is divided into two or more classes, each of which corresponds to a predetermined weather-related indicator, an index acquisition unit comprising a classification model for classifying the predicted observed data represented by the predicted observed data distribution,
The prediction model is
A regression model generated by machine learning based on a predetermined statistical analysis method, using multiple forecast data as explanatory variables and observed data as an objective variable,
The forecast data is data that predicts one or more of the weather parameters including at least the weather parameter to be predicted at the time of the forecast target, which is the time of the target of the forecast,
The observation data is an observed value of the meteorological parameter of the prediction target at the time of the forecast target or a predetermined analysis value based on the observed value,
Weather forecasting device.
[2]
The predicted observation data acquisition unit acquires the predicted observation data using a Monte Carlo dropout method.
The weather prediction device according to [1].
[3]
The prediction model is a model configured based on Gaussian process regression,
The predicted observation data acquisition unit acquires the predicted observation data including a probability distribution of the observed value or the analyzed value.
The weather prediction device according to [1].
[4]
The predicted observed data distribution is represented by the average value and standard deviation of the observed value or the analyzed value,
The weather forecasting device according to any one of [1] to [3].
[5] Limitation and enumeration of specific examples of predictive models.
The prediction model is a regression model composed of one of a Gaussian process, a random forest, a support vector machine, and a neural network.
The weather forecasting device according to any one of [1] to [4].
[6]
The indicator prediction model is either a classification model that classifies the predicted observed data distribution based on a given threshold, or a classification model configured by a support vector machine.
The weather forecasting device according to any one of [1] to [5].
[7]
The meteorological parameter to be predicted is precipitation,
The forecast data includes at least a forecast of precipitation,
The indicator prediction model outputs an indicator indicating whether the prediction target time corresponds to either a rainy season or a dry season, based on a given threshold regarding precipitation.
The weather forecasting device according to any one of [1] to [6].
[8]
A weather prediction method in a weather prediction device that predicts a predetermined weather-related index,
Forecast data in which a weather parameter that is a parameter representing a specific weather is the target of forecasting, and one or more of the weather parameters including at least the meteorological parameter to be predicted are predicted at the time of the prediction target which is the target time point of the prediction of the index. a prediction target forecast data acquisition step of acquiring prediction target forecast data that is
a forecast observation data acquisition step of acquiring forecast observation data regarding the weather parameter of the prediction target at the time of the forecast target by inputting the forecast target forecast data into a prediction model, the forecast observation data comprising: inputting the forecast target forecast data into a prediction model; a predicted observation data acquisition step, which includes a plurality of observed values or a predetermined analytical value based on the observed values, or a probability distribution of the observed values or the analytical values;
an observed value distribution calculation step of calculating a predicted observed data distribution that is a probability distribution of the observed value or the analyzed value based on the predicted observed data;
an index acquisition step of acquiring a weather-related index by inputting the predicted observation data distribution into an index prediction model, the index prediction model dividing the weather-related index into two or more classes, each of which corresponds to a predetermined weather-related index; an index acquisition step consisting of a classification model for classifying the predicted observed data represented by the predicted observed data distribution,
The prediction model is
A regression model generated by machine learning based on a predetermined statistical analysis method with forecast data as an explanatory variable and observed data as an objective variable,
The forecast data is data that predicts one or more of the weather parameters including at least the weather parameter to be predicted at the time of the forecast target, which is the time point of the forecast target;
The observation data is an observed value of the meteorological parameter of the prediction target at the time of the forecast target or a predetermined analysis value based on the observed value,
Weather forecasting methods.

10 Prediction model generation device 11 Forecast data acquisition unit 12 Observation data acquisition unit 13 Prediction model generation unit 14 Model output unit 15 Forecast data storage unit 16 Observation data storage unit 17 Prediction model storage unit 20 Weather prediction device 21 Forecast target forecast data acquisition unit 22 Forecast observation data acquisition section 23 Observation value distribution calculation section 24 Index acquisition section 25 Index output section 26 Forecast target forecast data storage section d11 Learning forecast data d12 Observation data d21 Forecast target forecast data d22 Forecast observation data d23 Forecast observation data distribution d24 Indicator M1 Prediction model M2 Indicator prediction model P1 Program

Claims (8)

A weather prediction device that predicts predetermined weather-related indicators,
Forecast data in which a weather parameter that is a parameter representing a specific weather is the target of forecasting, and one or more of the weather parameters including at least the meteorological parameter to be predicted are predicted at the time of the prediction target which is the target time point of the prediction of the index. a prediction target forecast data acquisition unit that acquires prediction target forecast data including a plurality of prediction target forecast data;
A forecast observation data acquisition unit that acquires forecast observation data regarding the weather parameter of the prediction target at the time of the forecast target by inputting the forecast target forecast data into a prediction model, the forecast observation data comprising: a predicted observation data acquisition unit that includes a plurality of observed values or a predetermined analytical value based on the observed values, or a probability distribution of the observed values or the analytical values;
an observed value distribution calculation unit that calculates a predicted observed data distribution that is a probability distribution of the observed values or the analyzed values based on the predicted observed data;
An indicator acquisition unit that obtains weather-related indicators by inputting the predicted observation data distribution into an indicator prediction model, wherein the indicator prediction model is divided into two or more classes, each of which corresponds to a predetermined weather-related indicator, an index acquisition unit comprising a classification model for classifying the predicted observed data represented by the predicted observed data distribution,
The prediction model is
A regression model generated by machine learning based on a predetermined statistical analysis method, using multiple forecast data as explanatory variables and observed data as an objective variable,
The forecast data is data that predicts one or more of the weather parameters including at least the weather parameter to be predicted at the time of the forecast target, which is the time of the target of the forecast,
The observation data is an observed value of the meteorological parameter of the prediction target at the time of the forecast target or a predetermined analysis value based on the observed value,
Weather forecasting device. The predicted observation data acquisition unit acquires the predicted observation data using a Monte Carlo dropout method.
The weather prediction device according to claim 1. The prediction model is a model configured based on Gaussian process regression,
The predicted observation data acquisition unit acquires the predicted observation data including a probability distribution of the observed value or the analyzed value.
The weather prediction device according to claim 1. The predicted observed data distribution is represented by the average value and standard deviation of the observed value or the analyzed value,
The weather prediction device according to claim 1. The prediction model is a regression model composed of one of a Gaussian process, a random forest, a support vector machine, and a neural network.
The weather prediction device according to claim 1. The indicator prediction model is either a classification model that classifies the predicted observed data distribution based on a given threshold, or a classification model configured by a support vector machine.
The weather prediction device according to claim 1. The meteorological parameter to be predicted is precipitation,
The forecast data includes at least a forecast of precipitation,
The indicator prediction model outputs an indicator indicating whether the prediction target time corresponds to either a rainy season or a dry season, based on a given threshold regarding precipitation.
The weather prediction device according to claim 1. A weather prediction method in a weather prediction device that predicts a predetermined weather-related index,
Forecast data in which a weather parameter that is a parameter representing a specific weather is the target of forecasting, and one or more of the weather parameters including at least the meteorological parameter to be predicted are predicted at the time of the prediction target which is the target time point of the prediction of the index. a prediction target forecast data acquisition step of acquiring prediction target forecast data that is
a forecast observation data acquisition step of acquiring forecast observation data regarding the weather parameter of the prediction target at the time of the forecast target by inputting the forecast target forecast data into a prediction model, the forecast observation data comprising: inputting the forecast target forecast data into a prediction model; a predicted observation data acquisition step, which includes a plurality of observed values or a predetermined analytical value based on the observed values, or a probability distribution of the observed values or the analytical values;
an observed value distribution calculation step of calculating a predicted observed data distribution that is a probability distribution of the observed value or the analyzed value based on the predicted observed data;
an index acquisition step of acquiring a weather-related index by inputting the predicted observation data distribution into an index prediction model, wherein the index prediction model divides the weather into two or more classes, each of which corresponds to a predetermined weather-related index; an index acquisition step consisting of a classification model for classifying the predicted observed data represented by the predicted observed data distribution,
The prediction model is
A regression model generated by machine learning based on a predetermined statistical analysis method with forecast data as an explanatory variable and observed data as an objective variable,
The forecast data is data that predicts one or more of the weather parameters including at least the weather parameter to be predicted at the time of the forecast target, which is the time of the target of the forecast,
The observation data is an observed value of the meteorological parameter of the prediction target at the time of the forecast target or a predetermined analysis value based on the observed value,
Weather forecasting methods.

Images