JP7740642B2 - Photovoltaic power generation forecasting system, photovoltaic power generation forecasting method, photovoltaic power generation forecasting program, past forecast estimation device, past forecast estimation method, and past forecast estimation program - Google Patents

Description

The present invention relates to a solar power generation forecasting system, a solar power generation forecasting method, a solar power generation forecasting program, a past forecast estimation device, a past forecast estimation method, and a past forecast estimation program.

One power generation prediction system uses a neural network to predict future solar power generation based on past meteorological measurement data and future weather forecast data for a given point in time (see, for example, Patent Document 1).

Furthermore, one solar power generation prediction system takes sky images with a camera and predicts future solar power generation based on those sky images (see, for example, Patent Document 2).

Japanese Patent Application Laid-Open No. 2018-007312 Japanese Patent Application Laid-Open No. 2017-200360

In the power generation forecasting system described above, future weather forecast data at each point in time is used as input data for the neural network, so weather forecast data for many points in time over a long period of time is required as training data when machine learning the neural network.

However, while past measured weather data is relatively easy to obtain, it is difficult to obtain past weather forecast data. Weather forecast data providers are reluctant to provide past weather forecast data because it would reveal the accuracy of their weather forecasts. As a result, it is difficult to perform appropriate machine learning until weather forecast data has been collected over a long period of time. In other words, it is difficult to start accurately predicting solar power generation in a relatively short period of time. It is also possible to use measured weather data instead of weather forecast data, but in that case, there is a risk that machine learning will not be accurate because the values tend to differ between forecast data and measured data.

The present invention was made in consideration of the above-mentioned problems, and aims to provide a solar power generation forecasting system, a solar power generation forecasting method, and a solar power generation forecasting program that can begin to accurately predict solar power generation in a relatively short period of time when predicting solar power generation using a learning device that uses weather forecast data as input data, as well as a past forecast estimation device, a past forecast estimation method, and a past forecast estimation program that can be used for this machine learning.

The solar power generation prediction system according to the present invention includes a solar power generation prediction unit that predicts solar power generation using a first learning device that receives weather forecast data as input data, a past forecast estimation unit that estimates and generates past weather forecast history data from past actual weather history data, and a first machine learning processing unit that performs machine learning on the first learning device using the past weather forecast history data generated by the past forecast estimation unit.

The solar power generation prediction method according to the present invention comprises a solar power generation prediction step in which a first learning device uses weather forecast data as input data to predict solar power generation; a past forecast estimation step in which past weather forecast history data is estimated and generated from past actual weather history data; and a machine learning processing step in which machine learning is performed on the first learning device using the past weather forecast history data generated in the past forecast estimation step.

The solar power generation prediction program of the present invention causes a computer to execute a solar power generation prediction step of predicting solar power generation using a first learning device that uses weather forecast data as input data, a past forecast estimation step of estimating and generating past weather forecast history data from past actual weather history data, and a machine learning processing step of performing machine learning on the first learning device using the past weather forecast history data generated in the past forecast estimation step.

The past forecast estimation device according to the present invention includes a past forecast estimation unit that uses a learning device that receives past actual weather history data as input data to estimate and generate past weather forecast history data. Here, the past actual weather history data is actual weather history data from multiple points in time in the past, and the past weather forecast history data is weather forecast history data from multiple points in time in the past.

The past forecast estimation method according to the present invention includes a past forecast estimation step in which past historical weather data is used as input data to estimate and generate past historical weather forecast data using a learning device. Here, the past historical weather data is historical weather data from multiple points in time in the past, and the past historical weather forecast data is historical weather forecast data from multiple points in time in the past.

The past forecast estimation program of the present invention causes a computer to execute a past forecast estimation step in which a learning device uses past actual weather history data as input data to estimate and generate past weather forecast history data. Here, the past actual weather history data is actual weather history data from multiple points in time in the past, and the past weather forecast history data is weather forecast history data from multiple points in time in the past.

The present invention provides a solar power generation forecasting system, a solar power generation forecasting method, and a solar power generation forecasting program that can perform appropriate machine learning in a relatively short time when predicting solar power generation using a learning device that uses weather forecast data as input data, as well as a past forecast estimation device, a past forecast estimation method, and a past forecast estimation program that can be used for this machine learning.

The above and other objects, features, and advantages of the present invention will become more apparent from the following detailed description taken in conjunction with the accompanying drawings.

FIG. 1 is a block diagram showing the configuration of a photovoltaic power generation prediction system according to an embodiment of the present invention. FIG. 2 is a block diagram showing the configuration of the past forecast estimation unit 22 in the photovoltaic power generation prediction system shown in FIG. FIG. 3 is a block diagram showing the configuration of the photovoltaic power generation amount prediction unit 25 in the photovoltaic power generation amount prediction system shown in FIG. FIG. 4 is a block diagram showing the configuration of the solar radiation amount predicting unit 41 in the solar power generation amount predicting unit 25 shown in FIG.

The following describes an embodiment of the present invention with reference to the accompanying drawings.

Figure 1 is a block diagram showing the configuration of a solar power generation prediction system according to an embodiment of the present invention. The solar power generation prediction system shown in Figure 1 is configured with a single information processing device (such as a personal computer), but the processing units and units described below may be distributed across multiple information processing devices that are capable of data communication with each other. Furthermore, such multiple information processing devices may include a GPU (Graphics Processing Unit) that processes specific calculations in parallel.

The solar power generation prediction system shown in Figure 1 is a system that predicts the amount of power generated by, for example, a specific solar power generation facility, and includes a storage device 1, a communication device 2, a user interface 3, and an arithmetic processing device 4.

Storage device 1 is a non-volatile storage device such as flash memory or a hard disk, and stores various data and programs.

Here, the storage device 1 stores a past forecast estimation program 11 and a solar power generation prediction program 12, and also stores a parameter dataset 13 (parameters such as the neural network coupling coefficients (weights) and biases described below) as needed.

Communication device 2 is a device capable of data communication, such as a network interface, peripheral device interface, or modem, and performs data communication with other devices as needed.

The user interface 3 includes a display device 3a, such as a display that displays operation screens, and an input device 3b, such as a keyboard, that accepts user operations.

In FIG. 1, the user interface 3 is connected to the arithmetic processing unit 4 via an internal bus or peripheral device interface. However, if the arithmetic processing unit 4 is instead built into a server on a network (such as the Internet or a local area network (LAN)), the user interface 3 serves as a user interface device for a terminal device that can communicate with the server, and functions as a user interface for the processing units and units implemented in the arithmetic processing unit 4 through data communication between the server and the terminal device.

The arithmetic processing unit 4 is a computer equipped with a CPU (Central Processing Unit), ROM (Read Only Memory), RAM (Random Access Memory), etc., and operates as various processing units by loading programs from the ROM, storage device 1, etc. into the RAM and executing them on the CPU. The RAM is used as work memory for these processing units.

Here, by executing program 11, the processing device 4 operates as a past actual data acquisition unit 21, a past forecast estimation unit 22, and a machine learning processing unit 23.

The past actual data acquisition unit 21 acquires actual measurement data for specific items (temperature, humidity, solar radiation, cloud area ratio, etc.) in past weather forecasts. For example, the past actual data acquisition unit 21 downloads actual measurement data from a specific server using the communication device 2, or reads actual measurement data that has been stored in advance in the storage device 1.

The past forecast estimation unit 22 estimates and generates past weather forecast history data using a learning device that uses past actual weather history data as input data. Here, this past actual weather history data is actual weather history data from multiple points in time in the past, and this past weather forecast history data is weather forecast history data from those multiple points in time in the past.

In this embodiment, the past forecast estimation unit 22 uses past actual weather history data, cloud area ratio history data, and solar position history data as input data and estimates and generates past weather forecast history data using a learning device.

In addition, this learning device is trained by machine learning using past actual weather history data and past weather forecast history data of specific weather forecast providers as learning data, and the past forecast estimation unit 22 estimates and generates past weather forecast history data for specific weather forecast providers. In other words, for each of multiple weather forecast providers, past weather forecast history data specific to that weather forecast provider (values for each item of weather forecast for multiple future points in time viewed from past points in time) is individually generated.

Figure 2 is a block diagram showing the configuration of the past forecast estimation unit 22 in the solar power generation prediction system shown in Figure 1.

For example, as shown in FIG. 2, the past forecast estimation unit 22 includes an encoder 31, encoders 32-1 to 32-3, a feature vector connection unit 33, a latent space encoder 34, and a decoder 35.

The encoder 31 has one or more layers of neural networks (here, a fully connected neural network), and uses the neural network to convert the above-mentioned historical weather data (input data) into feature vectors.

Encoder 32-1 has a single or multiple layer neural network (here, a fully connected neural network) that converts cloud area ratio history data (input data) into a feature vector. Encoder 32-2 has a single or multiple layer neural network (here, a fully connected neural network) that converts time history data (input data) into a feature vector. Encoder 32-3 has a single or multiple layer neural network (here, a fully connected neural network) that converts sun position history data (input data) into a feature vector.

Note that the input data to the neural networks in encoders 31, 32-1 to 32-3, etc. is normalized appropriately (for example, to a range of 0 to 1). Furthermore, periodic input data (sun position and time) is normalized using sine and cosine functions.

The actual meteorological measurements (temperature, humidity, wind direction, wind speed, etc.), cloud area ratio values, and solar position (elevation angle and azimuth angle) values at each point in time are identified, and the actual meteorological measurements for multiple points in time are treated as historical meteorological data, the cloud area ratio values for those multiple points in time are treated as cloud area ratio data, and the solar position values for those multiple points in time are treated as historical solar position data.

The feature vector concatenation unit 33 concatenates the feature vectors generated by the encoders 31, 32-1 to 32-3 into a single feature vector.

The latent space encoder 34 includes a neural network that maps the concatenated feature vectors obtained by the feature vector concatenation unit 33 in latent space.

The decoder 35 is equipped with a neural network that converts the feature vectors mapped by the latent space encoder 34 into weather forecast prediction history data (output data).

The machine learning processing unit 23 performs machine learning on the learning device (here, the above-mentioned neural network) in the past forecast estimation unit 22 using an existing learning method (such as backpropagation). Specifically, the machine learning processing unit 23 uses learning data containing multiple sets of pairs of input data and output data for the learning device to update the parameters of the learning device (here, the weights and biases of the neural network).

Once a certain period of input data has been obtained and machine learning has been completed by the learning device (here, the neural network described above) in the past forecast estimation unit 22, the past forecast estimation unit 22 can estimate past forecast values for weather forecast items (temperature, humidity, wind direction, wind speed, etc.) from past measured values, making it possible to obtain a large number of past forecast values in a relatively short period of time, and these forecast values are used for machine learning in the machine learning processing unit 26 described below.

Furthermore, by executing the program 12, the calculation processing device 4 operates as an input data acquisition unit 24, a solar power generation amount prediction unit 25, and a machine learning processing unit 26.

The input data acquisition unit 24 acquires input data for the solar power generation facility for which solar power generation output is to be predicted. The input data includes, for example, sky image history data, actual weather history data, power generation condition history data, satellite forecast data, weather forecast data, and cloud area ratio history data. For example, the input data acquisition unit 24 downloads the input data from a specific server using the communication device 2, or reads the input data that has been previously stored in the storage device 1.

The solar power generation prediction unit 25 predicts solar power generation using a learning device that uses weather forecast data as input data.

In this embodiment, the solar power generation prediction unit 25 predicts solar power generation using a learning device that uses weather forecast data (forecast values for each item in the weather forecast for multiple future points in time from the present) and sky image history data as multimodal input data, and this learning device (a) generates solar radiation prediction data from the sky image history data, and (b) generates a predicted value for solar power generation from the solar radiation prediction data and weather forecast data. Note that the sky image history data is image data of a series of sky images obtained by photographing the sky at multiple points in time with a photographing device from the power generation facility that is the target for predicting solar power generation.

Figure 3 is a block diagram showing the configuration of the solar power generation prediction unit 25 in the solar power generation prediction system shown in Figure 1. Figure 4 is a block diagram showing the configuration of the solar radiation prediction unit 41 in the solar power generation prediction unit 25 shown in Figure 3.

For example, as shown in FIG. 3, the solar power generation amount prediction unit 25 includes a solar radiation amount prediction unit 41, encoders 42-1 to 42-4, a data linking unit 43, and a power generation amount prediction unit 44.

The solar radiation prediction unit 41 is equipped with a learning device that derives predicted solar radiation values for multiple future points in time from input data such as sky image history data, weather forecast data, cloud area ratio forecast data, and time data. The solar radiation prediction unit 41 is common to all solar power generation facilities (i.e., it is machine-trained to be common to multiple solar power generation facilities). However, the encoders 51 and 52 in the solar radiation prediction unit 41, which will be described later, are specific to the solar power generation facility and differ for each solar power generation facility (i.e., it is machine-trained to be different for each solar power generation facility).

The sky image history data, weather forecast data, cloud area ratio forecast data, and time data are data that respectively indicate sky images, weather forecasts, cloud area ratio forecasts, and times at multiple points in time up to the present. These sky images are generated by scaling and cropping the captured image so that it has a specified number of pixels. These weather forecasts, cloud area ratio forecasts, and times are normalized as appropriate.

For example, as shown in FIG. 4, the solar radiation prediction unit 41 includes an encoder 51, an encoder 52, a feature vector connection unit 53, and a solar radiation prediction value calculation unit 54.

Encoder 51 includes a convolutional neural network (CNN) that converts the blank image history data into a feature vector. The CNN in encoder 51 includes a long short-term memory (LSTM) layer and a global average pooling (GAP) layer.

Encoder 52 has a neural network including an LSTM layer, and this neural network converts weather forecast data (wind speed, wind direction, etc. at specific times), cloud area ratio forecast data (forecast values of cloud area ratio at those specific times), and solar position data (elevation angle and azimuth angle at those specific times) into feature vectors as input data.

The feature vector concatenation unit 53 concatenates the feature vectors generated by the encoders 51 and 52 to create a single feature vector.

The solar radiation prediction value calculation unit 54 has a single or multiple layer neural network (here, a fully connected neural network), and uses this neural network to derive a solar radiation prediction value from the feature vector obtained by the feature vector connection unit 53.

Here, this predicted value of solar radiation is, for example, global solar radiation (GHI:
Global Horizontal Irradiance (GHIR) forecasts.

Also, encoder 42-1 in Figure 3 has a neural network including an LSTM layer, and this neural network converts power generation condition history data indicating power generation conditions at multiple past points in time into a feature vector. These power generation conditions include solar radiation history (actual solar radiation data at multiple points in time up to the present), output history such as output voltage, output current, and output power (output voltage, output current, output power, etc. at the above-mentioned multiple points in time up to the present), temperature history (temperature at the solar power generation facility at the above-mentioned multiple points in time up to the present), and solar panel specifications (type of mounting and solar panel at the power generation facility, etc.). Encoder 42-1 is specific to the solar power generation facility and differs for each solar power generation facility (in other words, it is machine-learned to be different for each solar power generation facility).

Encoder 42-2 has a neural network including an LSTM layer, and uses this neural network to convert satellite forecast data indicating weather forecast values for multiple future points in time based on images taken by weather satellites at multiple past and/or present points in time into feature vectors. These weather forecast values include solar radiation (GHI, DNI: Direct Normal Irradiance, EBH, DHI: Diffuse Horizontal Irradiance, etc.), cloud area ratio, cloud opacity, solar position (elevation angle and azimuth angle), temperature, etc. Encoder 42-2 is common to all solar power generation facilities (i.e., it has been machine-learned to be common to multiple solar power generation facilities).

Encoder 42-3 has a neural network including an LSTM layer, and uses this neural network to convert weather forecast data indicating weather forecast values for multiple future points in time into feature vectors. The weather forecast values include temperature, air pressure, humidity, cloud cover ratio, precipitation, wind direction, wind speed, etc. Encoder 42-3 is common to all solar power generation facilities (i.e., it has been machine-trained to be common to multiple solar power generation facilities).

Encoder 42-4 has a neural network including an LSTM layer, and uses this neural network to convert cloud area ratio forecast data indicating cloud area ratio forecast values at multiple future points in time and time data into feature vectors. Note that the time data is the time value (such as the time of day) at the multiple points in time mentioned above. Encoder 42-4 is common to all solar power generation facilities (i.e., it has been machine-trained to be common to multiple solar power generation facilities).

The data concatenation unit 43 concatenates the solar radiation prediction value obtained by the solar radiation prediction unit 41 and the feature vectors generated by the encoders 42-1 to 42-4 to create a single feature vector.

The power generation prediction unit 44 is equipped with a neural network, which derives predicted power generation values at multiple future points in time from the feature vectors obtained by the data linking unit 43. The predicted power generation values are predicted values for the power generation amount of the solar power generation facility, and may be values for power or energy. Here, the multiple future points in time are points that arrive at predetermined time intervals, such as 5-minute intervals, 15-minute intervals, or 30-minute intervals, within a period such as 3 hours, 6 hours, or 24 hours from the present time.

The machine learning processing unit 26 performs machine learning on the learning device (here, a neural network) in the solar power generation prediction unit 25 using an existing learning method (such as backpropagation). Specifically, the machine learning processing unit 26 uses learning data containing multiple sets of pairs of input data and output data for the learning device to update the parameters of the learning device (here, the weights and biases of the neural network).

The machine learning processing unit 26 performs machine learning in advance on the solar radiation prediction unit 41 individually for each power generation facility. Here, a solar radiation meter is installed at each power generation facility, and this solar radiation meter obtains actual measured data on the amount of solar radiation at each power generation facility, which is used as learning data for this machine learning.

The machine learning processing unit 26 performs machine learning on the learner in the solar power generation prediction unit 25 using the past weather forecast history data generated by the past forecast estimation unit 22 as input data (the above-mentioned weather forecast data).

Next, we will explain the operation of the above-mentioned solar power generation prediction system.

(a) Solar power generation forecast

As described above, the input data acquisition unit 24 acquires sky image history data, power generation condition history data, satellite forecast data, weather forecast data, and cloud area ratio forecast data as input data, and the solar power generation amount prediction unit 25 derives predicted solar power generation amounts at multiple future points in time from the input data.

(b) Machine learning by the solar power generation forecasting unit 25

(b1) Generating training data

First, the past actual data acquisition unit 21 acquires historical weather data, cloud area ratio data, time history data, and solar position data for multiple past points in time. For example, the past actual data acquisition unit 21 acquires historical weather data and cloud area ratio data for multiple past points in time from an external server, and generates time history data and solar position history data for those multiple points in time.

Next, the past forecast estimation unit 22 derives weather forecast history data corresponding to the actual weather history data for multiple past points in time from the acquired actual weather history data, cloud area ratio history data, time history data, and solar position history data.

In addition, cloud area ratio forecast data and satellite forecast data may also be generated by the past forecast estimation unit 22 in the same manner as weather forecast data (i.e., the above-mentioned weather forecast history data).

(b2) Machine Learning

The solar radiation prediction unit 41 is shared by multiple solar power generation facilities. The machine learning processing unit 26 first performs machine learning for the solar radiation prediction unit 41. At this time, pairs of sky image history data and actual solar radiation measurements for each solar power generation facility are used as learning data, and the learning data for multiple solar power generation facilities is compiled and used for the machine learning of the solar radiation prediction unit 41.

Next, pairs of sky image history data, weather forecast data (i.e., the above-mentioned weather forecast history data), cloud area ratio forecast data, satellite forecast data, and power generation condition history data for multiple past points in time, and actual power generation amounts at multiple future points in time as viewed from those multiple past points in time, are used as learning data, and the machine learning processing unit 26 uses this learning data to perform machine learning on the solar power generation amount prediction unit 25. At this time, the encoders 51 and 52 of the solar radiation prediction unit 41 are not updated, and the results of the machine learning by the solar radiation prediction unit 41 are left as they are.

As described above, learning data is generated based on actual measurement data obtained from each of multiple solar power generation facilities, and the past forecast estimation unit 22 and solar power generation prediction unit 25, which are common to the multiple solar power generation facilities, are machine-learned based on the learning data.

As described above, according to the above embodiment, the past forecast estimation unit 22 estimates and generates past weather forecast history data using a learning device that uses past actual weather history data as input data. Here, the past actual weather history data is actual weather history data from multiple points in time in the past, and the past weather forecast history data is weather forecast history data from multiple points in time in the past. Then, the machine learning processing unit 26 performs machine learning on the learning device in the solar power generation prediction unit 25 using the past weather forecast history data generated by the past forecast estimation unit 22.

This allows for accurate prediction of solar power generation using a learning machine that uses weather forecast data as input data, making it possible to quickly estimate appropriate past weather forecast data (i.e., weather forecast history data). This allows for appropriate machine learning to be performed in a relatively short time, enabling accurate prediction of solar power generation.

Furthermore, forecast values can be estimated at any point in the past, and the past forecast estimation unit 22 can also estimate weather forecasts for locations (power generation facilities) where weather forecasts are not localized in the first place, so a sufficient amount of learning data can be obtained for any location (power generation facility), allowing appropriate machine learning to be performed.

Various changes and modifications to the above-described embodiments will be apparent to those skilled in the art. Such changes and modifications may be made without departing from the spirit and scope of the subject matter and without diminishing its intended advantages. Thus, it is intended that such changes and modifications be included within the scope of the claims.

For example, a device equipped with the past actual data acquisition unit 21, past forecast estimation unit 22, and machine learning processing unit 23 in the above embodiment may be used as a past forecast estimation device. In that case, the past forecast estimation device does not need to be equipped with the input data acquisition unit 24, solar power generation amount prediction unit 25, and machine learning processing unit 26. Furthermore, the past weather forecast data obtained by the past forecast estimation device (and the past forecast estimation method executed in the past forecast estimation device) can be used for purposes other than predicting solar power generation.

In addition, in the above embodiment, programs 11 and 12 may each be recorded on a portable, computer-readable recording medium and installed on the device from that recording medium.

The present invention can be applied, for example, to predicting solar power generation.

11 Past forecast estimation program 12 Photovoltaic power generation amount prediction program 22 Past forecast estimation unit 23 Machine learning processing unit (an example of a second machine learning processing unit)
25 Photovoltaic power generation amount prediction unit 26 Machine learning processing unit (an example of a first machine learning processing unit)

Claims (10)

a photovoltaic power generation amount prediction unit that predicts the photovoltaic power generation amount using a first learning device that uses weather forecast data as input data;
a past forecast estimation unit that estimates and generates past weather forecast history data from past actual weather history data;
a first machine learning processing unit that performs machine learning of the first learning device using the past weather forecast history data generated by the past forecast estimation unit;
A solar power generation prediction system comprising: the solar power generation amount prediction unit predicts the solar power generation amount using the first learning device, which receives as input data the weather forecast data and sky image history data obtained by photographing the sky with an imaging device from a power generation facility that is a target for predicting the solar power generation amount;
The first learning device (a) generates solar radiation prediction data from the sky image history data, and (b) generates a predicted value of the solar power generation amount from the solar radiation prediction data and the weather forecast data;
The solar power generation amount prediction system according to claim 1, Further comprising a second machine learning processing unit,
the past forecast estimation unit estimates and generates the past weather forecast history data from the past actual weather history data using a second learning device;
the second machine learning processing unit performs machine learning of the second learner;
The solar power generation amount prediction system according to claim 1, a photovoltaic power generation amount prediction step of predicting the photovoltaic power generation amount using a first learning device that uses weather forecast data as input data;
a past forecast estimation step of estimating and generating past weather forecast history data from past actual weather history data;
a machine learning processing step of performing machine learning of the first learning device using the past weather forecast history data generated in the past forecast estimation step;
A method for predicting solar power generation amount, comprising: On the computer,
a photovoltaic power generation amount prediction step of predicting the photovoltaic power generation amount using a first learning device that uses weather forecast data as input data;
a past forecast estimation step of estimating and generating past weather forecast history data from past actual weather history data;
a machine learning processing step of performing machine learning of the first learning device using the past weather forecast history data generated in the past forecast estimation step;
A solar power generation prediction program that executes the above. a past forecast estimation unit that estimates and generates past weather forecast history data using a learning device that uses past actual weather history data as input data;
The past actual weather history data is actual weather history data at a plurality of points in time in the past,
The past weather forecast history data is weather forecast history data at a plurality of past points in time;
A past forecast estimation device characterized by the above. The past forecast estimation device described in claim 6, characterized in that the past forecast estimation unit estimates and generates the past weather forecast history data using a learning device, using the past actual weather history data, cloud area ratio history data, and solar position history data as input data. the learning device is machine-learned using the past actual weather history data and past weather forecast history data of a specific weather forecast provider as learning data;
the past forecast estimation unit estimates and generates the past weather forecast history data for the specific weather forecast provider;
7. The past forecast estimation device according to claim 6, wherein: a past forecast estimation step of estimating and generating past weather forecast history data using a learning device that uses past actual weather history data as input data;
The past actual weather history data is actual weather history data at a plurality of points in time in the past,
The past weather forecast history data is weather forecast history data at a plurality of past points in time;
A past forecast estimation method characterized by: On the computer,
a learning device that uses past actual weather history data as input data to estimate and generate past weather forecast history data;
The past actual weather history data is actual weather history data at a plurality of points in time in the past,
The past weather forecast history data is weather forecast history data at a plurality of past points in time;
A past forecast estimation program that features: