JP6984780B1 - Information processing equipment, information processing system and information processing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information processing apparatus capable of easily adjusting a parameter for determining a configuration of a learning model and a parameter for learning a learning model to an appropriate value. SOLUTION: A first input receiving unit that accepts input of a selected learning model from a plurality of different types of learning models, a second input receiving unit that accepts input of parameters related to the selected learning model, and the parameters. However, the first determination unit for determining whether or not the first condition for determining whether or not the selected learning model can be learned and the above-mentioned when the parameter satisfies the first condition. An information processing apparatus including a learning unit that learns the selected learning model using parameters. [Selection diagram] FIG. 19

Description

The present invention relates to an information processing apparatus, an information processing system and an information processing method.

Devices are known that use machine learning to predict future demand for energy and commodities.

For example, Patent Document 1 describes a device for predicting electric power demand using a neural network and the like. Further, Patent Document 2 describes a device and the like for extracting a day similar to the power demand of the predicted target day among the past days by using a decision tree.

In a conventional device such as these, the designer of the device usually generates a learning model after setting parameters for determining the configuration of the learning model and parameters for learning the learning model.

Patent No. 3360520 Patent No. 6187003

By the way, the operator of the device may want to update the learning model to one that can better reproduce the actual value, for example, when the error between the actual value and the predicted value by the device exceeds the allowable range. In such a case, it is conceivable to adjust the parameters set by the designer of the device.

However, setting and adjusting such parameters to appropriate values requires considerable experience and expertise. Therefore, it is generally difficult for the operator of the device to adjust the parameters by himself / herself. On the other hand, requesting the designer of the device to adjust the parameters requires a great deal of time and cost.

An object of the present invention is to provide an information processing apparatus capable of easily adjusting parameters for determining a configuration of a learning model and parameters for learning a learning model to appropriate values.

One invention for achieving the above object is a first input receiving unit that accepts input of a selected learning model from a plurality of different types of learning models, and a first input receiving unit that accepts input of parameters related to the selected learning model. 2 The input receiving unit, the first determination unit for determining whether the parameter satisfies the first condition for determining whether the selected learning model can be learned, and the parameter being the first. An information processing apparatus including a learning unit that learns the selected learning model using the parameters when the condition of 1 is satisfied. Other features of the invention will be clarified by the description herein.

INDUSTRIAL APPLICABILITY According to the present invention, it is possible to provide an information processing apparatus capable of easily adjusting a parameter for determining a configuration of a learning model and a parameter for learning a learning model to an appropriate value.

It is a figure explaining the information processing system 1. It is a figure explaining the hardware configuration of the learning processing apparatus 3. It is a figure explaining the learning data 60. It is a figure explaining the functional block of the learning processing apparatus 3. It is a figure explaining the flow of the process of generating a learning model. It is a figure explaining the hardware composition of the demand forecasting apparatus 4. It is a figure explaining the functional block of the demand forecasting apparatus 4. It is a figure explaining the flow of the process for predicting the electric power demand. It is a figure explaining the transition of an error. It is a figure explaining the hardware configuration of the information processing apparatus 5. It is a figure explaining the information of a learning model 61. It is a figure explaining the structure of the model of a neural network. It is a figure explaining the functional block of an information processing apparatus 5. It is a figure which shows the example of the influence degree of a predictor. It is a figure explaining the flow until the learning model of the information processing apparatus 5 is generated. It is a figure explaining the learning model selection screen 8. It is a figure explaining the parameter setting screen 9. It is a figure explaining the parameter setting screen 9. It is a figure explaining the pop-up screen 90. It is a figure explaining the pop-up screen 91. It is a figure explaining the transition of an error. It is a figure explaining the flow of the process until the parameter P0 is stored.

== Embodiment ==
<< Information processing system >>
FIG. 1 is a diagram showing a schematic configuration of the information processing system 1 of the present embodiment. The information processing system 1 includes a sensor 2, a learning processing device 3, a demand forecasting device 4, and an information processing device 5. The sensor 2, the learning processing device 3, the demand forecasting device 4, and the information processing device 5 are communicably connected via the network NW. The network NW is, for example, a public line such as the Internet, a dedicated line such as a WAN (Wide Area Network) or a LAN (Local Area Network), or the like.

<Sensor>
The sensor 2 is a device for observing temperature, humidity, atmospheric pressure, wind direction, wind speed, etc., and is provided outdoors. The sensor 2 outputs the observation result to the demand forecasting device 4 and the information processing device 5.

<Learning processing device 3>
The learning processing device 3 is a device that generates a learning model used when the demand forecasting device 4 described later predicts demand such as electric power demand.

[Hardware configuration of learning processing device 3]
FIG. 2 is a diagram showing an example of the hardware of the learning processing device 3. The exemplary learning processing device 3 includes a processor 300, a main storage device 301, an auxiliary storage device 302, an input device 303, an output device 304, and a communication device 305. It should be noted that the learning processing device 3 illustrated, in whole or in part, is virtual information provided by using virtualization technology, process space separation technology, or the like, for example, like a virtual server provided by a cloud system. It may be realized by using processing resources. Further, all or a part of the functions provided by the learning processing device 3 may be realized by, for example, a service provided by the cloud system via API (Application Programming Interface) or the like.

The processor 300 includes, for example, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), an FPGA (Field Programmable Gate Array), an ASIC (Application Specific Integrated Circuit), an AI (Artificial Intelligence) chip, and the like. It is configured using.

The main storage device 301 is a device for storing programs and data, and is, for example, a ROM (Read Only Memory), a RAM (Random Access Memory), a non-volatile memory (NVRAM (Non Volatile RAM)), and the like.

The auxiliary storage device 302 includes, for example, an SSD (Solid State Drive), a hard disk drive, an optical storage device (CD (Compact Disc), DVD (Digital Versatile Disc), etc.), a storage system, an IC card, an SD card, or an optical recording device. A reading / writing device for a recording medium such as a medium, a storage area for a cloud server, and the like. Programs and data can be read into the auxiliary storage device 302 via the reading device of the recording medium or the communication device 305. Programs and data stored (stored) in the auxiliary storage device 302 are read into the main storage device 301 at any time.

The auxiliary storage device 302 stores the learning data 60, which will be described in detail later, and the parameter P0.

The input device 303 is an interface for receiving input from the outside, and is, for example, a keyboard, a mouse, a touch panel, a card reader, a pen input type tablet, a voice input device, and the like.

The output device 304 is an interface that outputs various information such as processing progress and processing results. The output device 304 is, for example, a display device (liquid crystal monitor, LCD (Liquid Crystal Display), graphic card, etc.) that visualizes the above-mentioned various information, and a device (voice output device (speaker, etc.)) that visualizes the above-mentioned various information. , A device (printing device, etc.) that converts the above various information into characters.

The input device 303 and the output device 304 configure a user interface for receiving and presenting information with the user.

The communication device 305 is a device that realizes communication with other devices. The communication device 305 is a wired or wireless communication interface that realizes communication with other devices via a communication network such as the Internet, and is, for example, a NIC (Network Interface Card), a wireless communication module, or a USB. Modules, etc.

For example, an operating system, a file system, a DBMS (DataBase Management System) (relational database, NoSQL, etc.), a KVS (Key-Value Store), or the like may be introduced in the learning processing device 3.

Each function provided in the information processing system 1 is performed by the processor 300 of the learning processing device 3 reading and executing a program stored in the main storage device 301, or by hardware (FPGA) constituting the information processing system 1. , ASIC, AI chip, etc.). The information processing system 1 stores various types of information (data), for example, as a database table or a file managed by a file system.

[Learning data]
FIG. 3 is a diagram illustrating the learning data 60 of the present embodiment stored in the auxiliary storage device 302 of the learning processing device 3. The training data 60 includes a plurality of data 60a. One data 60a is arranged in one column of FIG.

The learning data 60 of the present embodiment is data in which the actual values of the maximum temperature, the minimum temperature, the day of the week, and the electric power demand are associated with each other. Here, the maximum temperature, the minimum temperature, and the day of the week are predictors, and the power demand is the correct answer data.

In the present embodiment, the training data 60 includes actual values for each of the 366 days from January 1, 2020 to December 31, 2020.

[Parameter P0]
The parameter P0 is a parameter related to the learning model. Details will be described later, but when the learning model is a neural network, the parameters P0 include the number of elements in the input layer, the number of hidden layers, the number of elements in the hidden layer, the predictors used in the training data, the learning period, and the kernel function. Includes learning coefficient, number of learnings and moment term.

[Functional block of learning processing device 3]
FIG. 4 is a diagram showing a functional block of the learning processing device 3 of the present embodiment. In the learning processing device 3 of the present embodiment, the acquisition unit 310, the learning unit 311, the storage unit 312, the acquisition unit 313, and the storage unit 314 are realized by the processor 300 executing a predetermined program. ..

The acquisition unit 310 acquires the learning data 60 and the parameter P0 stored in the auxiliary storage device 302.

The learning unit 311 generates a learning model by performing learning based on the learning data 60. Here, the type of learning model is not particularly limited. In this embodiment, a neural network is used as a type of learning model. The type of learning model may be selectable from a plurality of candidates by the operator of the information processing system 1 (hereinafter, simply referred to as "operator").

When the learning model is a neural network, the learning unit 311 inputs the predictor to the input layer and outputs the power demand from the output layer. The power demand output from the output layer is compared with the correct answer data, and the parameters (corresponding to the parameter P3) used for the arithmetic processing are optimized so that the output power demand becomes the correct answer data.

The parameter P3 is, for example, a weight between elements (coupling coefficient), a coefficient of an activation function used in each element, and the like. The method of optimizing the parameter P3 is not particularly limited, but the learning unit 311 optimizes each parameter P3 by using, for example, an error back propagation method.

The storage unit 312 stores the parameter P3 optimized by the learning unit 311 in the auxiliary storage device 402 of the demand forecasting device 4, which will be described later. Specifically, the storage unit 312 stores the parameter P3 used for the calculation processing of the prediction.

The acquisition unit 313 acquires the parameter P0 related to the learning model. Although the details will be described later, the parameter P0 acquired by the acquisition unit 313 is the parameter P0 adjusted by the information processing apparatus 5.

The storage unit 314 stores the parameter P0 acquired by the acquisition unit 313. As a result, the parameter P0 adjusted by the information processing apparatus 5 is stored in the storage unit 314.

<Process to learn the learning model>
FIG. 5 is a flowchart illustrating a flow of processing in which the learning processing device 3 learns the learning model. The process of learning the learning model includes steps S101 to S103.

First, in step S101, the acquisition unit 310 acquires the learning data 60 and the parameter P0. The learning data 60 shown in FIG. 3 is an example of the learning data acquired here by the acquisition unit 310.

Next, in step S102, the learning unit 311 generates a learning model based on the learning data 60.

Next, in step S103, the storage unit 312 stores the parameter P3 used for the arithmetic processing of the learning model optimized by the learning unit 311 in the auxiliary storage device 402 of the demand forecasting device 4.
<Demand forecaster 4>
The demand forecasting device 4 applies the parameter P3 optimized by the learning processing device 3 to the learning model, and based on the output result of the sensor 2, the power demand on the forecast target date (corresponding to the second forecast target date). It is a device that predicts.

[Hardware configuration of demand forecaster 4]
FIG. 6 is a diagram showing an example of the hardware of the demand forecasting device 4. The illustrated demand forecasting device 4 includes a processor 400, a main storage device 401, an auxiliary storage device 402, an input device 403, an output device 404, and a communication device 405.

The configurations of the processor 400, the main storage device 401, the auxiliary storage device 402, the input device 403, the output device 404, and the communication device 405 of the demand forecasting device 4 are the processor 300, the main storage device 301, and the auxiliary storage of the learning processing device 3, respectively. Since the configuration is the same as that of the device 302, the input device 303, the output device 304, and the communication device 305, the description of the common parts will be omitted.

[Auxiliary storage device]
Parameter P0 and parameter P3 are stored in the auxiliary storage device 402. The parameter P0 is the same as the parameter P0 stored in the auxiliary storage device 302 of the learning processing device 3. The parameter P3 is a parameter optimized by the learning unit 311 of the learning processing device 3.

[Functional block of demand forecasting device 4]
FIG. 7 is a diagram showing a functional block of the demand forecasting device 4 of the present embodiment. In the demand forecasting device 4 of the present embodiment, when the processor 400 executes a predetermined program, the acquisition unit 410, the output unit 411, the acquisition unit 412, the alarm unit 413, the acquisition unit 414, and the storage unit 415 and is realized.

The acquisition unit 410 (corresponding to the second acquisition unit) acquires the parameter P0 stored in the auxiliary storage device 402 and the parameter P3. The acquisition unit 410 further acquires the forecast value and the actual value of the predictor (corresponding to the second predictor) on the forecast target date.

In the present embodiment, the acquisition unit 410 acquires the maximum temperature, the minimum temperature, and the day of the week of the forecast target day from the weather forecast via the network NW as forecast values of the predictor before forecasting. Further, before the prediction, the acquisition unit 410 acquires the maximum temperature, the minimum temperature, and the day of the week of the prediction target day from the sensor 2 via the network NW as actual values of the prediction factor.

The output unit 411 (corresponding to the second output unit) uses the parameter P0 acquired by the acquisition unit 410, the parameter P3, and the forecast value of the predictor on the forecast target date to predict the power demand on the forecast target date. Is output. Further, after the prediction target date, the output unit 411 outputs the predicted value of the power demand on the prediction target date by using the learning model 61 and the actual value of the prediction factor on the prediction target date.

The acquisition unit 412 (corresponding to the third acquisition unit) acquires the actual value of the electric power demand on the forecast target date. In the present embodiment, the acquisition unit 412 acquires the actual value of the electric power demand announced by the electric power company via the network NW after the forecast target date. The actual value of the electric power demand is output to the alarm unit 413 described later.

The alarm unit 413 (corresponding to the second alarm unit) issues an alarm when the error between the predicted value of the power demand and the actual value of the power demand satisfies a predetermined condition. Here, the error of the predicted value of the power demand with respect to the actual value of the power demand is the difference between the predicted value of the power demand and the actual value of the power demand. Further, as the predicted value of the electric power demand here, either a predicted value output using the forecast value of the predictor or a predicted value output using the actual value of the predictor may be used.

In the present embodiment, as the predicted value of the electric power demand here, the predicted value output by using the actual value of the predictor is used. The details will be described later, but the error of the predicted value of the power demand in this case with respect to the actual value of the power demand (hereinafter, may be simply referred to as “error”) is not caused by the predictor but is caused only by the learning model. It is something to do.

Here, as the predetermined condition for the error, for example, any of the examples of the conditions shown below can be used.

One example is when the absolute value of the error exceeds a predetermined value. Another example is the case where the number of times the error is positive or negative exceeds the predetermined number of times in the predetermined number of predictions. Another example is the case where the error variation exceeds a predetermined value in a predetermined number of predictions. Another example is the case where a distribution tendency deviating from the predetermined distribution tendency of the error appears at a predetermined number of predictions.

That is, although the details will be described later, when the alarm unit 413 issues an alarm, the operator of the information processing system 1 can obtain an opportunity to verify the validity of the learning model 61.

The acquisition unit 414 (corresponding to the fourth acquisition unit) acquires the parameters P0 and the parameters P3 related to the learning model. Although the details will be described later, the parameter P0 acquired by the acquisition unit 414 is a parameter adjusted by the information processing apparatus 5. Further, the parameter P3 is a parameter optimized by the learning processing device 3.

The storage unit 415 stores the parameter P0 and the parameter P3 acquired by the acquisition unit 414. As a result, the parameter P0 is stored in the storage unit 415.

<Processing for predicting power demand>
FIG. 8 is a flowchart illustrating a flow of processing in which the demand forecasting device 4 predicts power demand. Note that FIG. 8 shows a process of predicting the power demand for one forecast target date. The process of predicting the power demand includes steps S201 to S206.

First, in step S201, the acquisition unit 410 acquires the parameter P0, the parameter P3, and the forecast value of the predictor.

Next, in step S202, the output unit 411 outputs the predicted value of the power demand from the learning model to which the parameters P0 and P3 are applied and the forecast value of the predictor acquired in step S201.

Next, in step S203, the acquisition unit 410 acquires the actual value of the predictor. That is, step S203 and subsequent steps are executed after the prediction target date.

Next, in step S204, the output unit 411 outputs the predicted value of the power demand from the learning model 61 and the actual value of the predictor acquired in step S203.

Next, in step S205, the acquisition unit 412 acquires the actual value of the electric power demand on the forecast target date announced by the electric power company.

Next, in step S206, the alarm unit 413 determines whether or not the error of the predicted value of the power demand output in step S204 with respect to the actual value of the power demand acquired in step S205 exceeds a predetermined threshold value. ..

FIG. 9 is a diagram illustrating the transition of the error. In FIG. 9, the horizontal axis is the forecast target date. The vertical axis is the error of the predicted value of power demand with respect to the actual value of power demand. If the error is inside the two dashed lines shown in FIG. 9, the absolute value of the error is less than or equal to the predetermined threshold T and is shown without fill. If the error is outside the two dashed lines shown in FIG. 9, the absolute value of the error exceeds a predetermined threshold T and is shown in fill.

When it is determined in step S206 that the absolute value of the error exceeds the threshold value T (S206: Y), in step S206, the alarm unit 413 issues an alarm and ends the prediction of the power demand on the forecast target date.

If it is determined in step S206 that the error does not exceed the threshold value (S206: Y), the prediction of the power demand on the prediction target date is terminated.

<Information processing device 5>
The information processing device 5 is a device for changing the learning model used by the demand forecasting device 4 to a different type of learning model and adjusting the parameter P0. The information processing apparatus 5 can further output the forecast result of the power demand by applying the adjusted parameter P0 to the learning model.

The operator of the information processing system 1 can better reproduce the learning model when, for example, the error between the actual value of the power demand and the predicted value of the power demand by the demand forecasting device 4 exceeds the allowable range. You may want to update to something. In such a case, the operator can adjust the parameter P0 by using the information processing apparatus 5.

[Hardware configuration of information processing device 5]
FIG. 10 is a diagram showing an example of the hardware of the information processing apparatus 5. The information processing device 5 exemplified includes a processor 500, a main storage device 501, an auxiliary storage device 502, an input device 503, an output device 504, and a communication device 505.

The configurations of the processor 500, the main storage device 501, the auxiliary storage device 502, the input device 503, the output device 504, and the communication device 505 of the information processing device 5 are the processor 300, the main storage device 301, and the auxiliary storage of the demand forecasting device 4, respectively. Since the configuration is the same as that of the device 302, the input device 303, the output device 304, and the communication device 305, the description of the common parts will be omitted.

The learning data 60 and the learning model information 7 are stored in the auxiliary storage device 502. The learning data is the same as that stored in the auxiliary storage device 302 in the learning processing device 3. Hereinafter, information 7 of the learning model will be described.

[Learning model information]
FIG. 11 is a diagram illustrating information 7 of the learning model of the present embodiment stored in the auxiliary storage device 502 described above. The auxiliary storage device 502 stores information 7 of the learning model for each of the plurality of different types of learning models.

In this embodiment, a plurality of different types of learning models include a least squares method, a partial least squares method, a regression equation, a neural network, a support vector machine, a regression tree, and a random forest. Note that FIG. 11 shows some of these omitted.

The information 7 of the learning model is information including the parameter P0 relating to the learning model, the condition C1 (corresponding to the first condition), and the condition C2 (corresponding to the second condition). The condition C1 and the condition C2 are conditions for the parameter P0. The number and type of parameters P0 are unique to the learning model.

Here, the "parameter" includes a parameter P1 (corresponding to the first parameter) for determining the configuration of the learning model and a parameter P2 (corresponding to the second parameter) for learning the learning model. Hereinafter, unless otherwise specified, the parameter P1 and the parameter P2 are collectively referred to as "parameter P0 related to the learning model".

As shown in FIG. 11, when the learning model is a neural network, the parameter P0 includes the number of elements in the input layer, the number of hidden layers, the number of elements in the hidden layer, and the predictor used among the learning data (hereinafter, simply "prediction"). It is referred to as "factor"), learning period, kernel function, learning coefficient, number of learnings and moment term.

Here, in the present embodiment, at least a part is selected as the predictor from the maximum temperature, the minimum temperature, and the day of the week shown in the learning data 60 of FIG. Further, as the predictor, a predictor other than the predictor shown in FIG. 9 may be newly acquired and further selected from the newly acquired predictors. Examples of the predictors other than the predictors in FIG. 9 include the maximum discomfort index, the wind direction, and the like.

Further, in the present embodiment, as the learning period, at least a part of the period from January 1, 2020 to December 31, 2020 included in the learning data 60 is selected. Further, as the learning period, data relating to a period other than the period possessed by the learning data 60 in FIG. 9 may be newly acquired and further selected from the newly acquired period.

Here, the information regarding the number of elements of the input layer, the number of hidden layers, the number of elements of the hidden layer, and the predictors (for example, temperature, humidity, etc.) is the parameter P1. Further, the kernel function, the learning coefficient, the number of learning times and the moment term, and the learning period (the period of the learning data used for learning) are the parameters P2.

The condition C1 is a condition for determining whether or not the learning model can be learned. When the parameter P0 satisfies the condition C1, the learning model can be trained, and the learning model can be generated in principle. On the other hand, when the parameter does not satisfy the condition C1, the learning model cannot be learned, so that the learning model cannot be generated in principle.

The condition C1 will be specifically described with reference to FIGS. 11 and 12 by exemplifying a case where the learning model is a neural network. FIG. 12 is a diagram illustrating a configuration of a model of a neural network.

In the example of FIG. 12, the input layer is provided with three elements N, the hidden layer is provided with a large number of elements N, and the output layer is provided with one element N. The maximum temperature, the minimum temperature, and the day of the week are input to the three elements N of the input layer, respectively. Power demand is output to the element N of the output layer.

[Condition C1 for parameter P1]
When the learning model is a neural network, the condition C1 includes a condition that the value of the parameter P1 is 1 or more. In the present embodiment, the number of elements in the input layer, the number of hidden layers, and the number of elements in the hidden layer are all integers of 1 or more. By satisfying such conditions, it is possible in principle to learn a learning model.

[Condition C1 for parameter P2]
The condition C1 is a condition in which the number of predictors is 1 or more for the predictors. Further, the condition C1 is a condition that the number of data 60a in the learning data 60 used as the learning period is 1 or more for the learning period.

Further, the condition C1 includes a condition in which the value of the parameter P2 is a value larger than 0. In this embodiment, the learning coefficient is a positive real number, and the learning frequency is an integer of 1 or more. The moment term is a positive real number. By satisfying these conditions, it is possible in principle to learn a learning model.

The condition C2 is, for example, a condition preset by a designer or the like of the demand forecasting device 4 (hereinafter referred to as a designer or the like). The designer or the like can set the condition recommended for the parameter as the condition C2.

The conditions recommended for the parameters P1 and P2 are, for example, when the parameters P1 and P2 satisfy this condition, a highly stable prediction result can be obtained by the generated learning model, and the calculation cost is appropriate. It can be a condition judged by the designer or the like.

Here, the "highly stable prediction result" is, for example, a prediction result that is almost uniquely determined within the range of conditions recommended for the parameters P1 and P2.

Further, the "calculation cost" is the time, cost, etc. required for the calculation for generating the learning model or the calculation until the prediction result is output using the learning model.

As for the parameters P1 and P2, the condition C2 is set so that the condition C1 is automatically satisfied if the condition C2 is satisfied.

The condition C2 will be specifically described with reference to FIGS. 11 and 12 by exemplifying the case of a neural network.

[Condition C2 for parameter P1]
The condition C2 includes a condition that the number of elements in the input layer is equal to or less than a predetermined value. In this embodiment, the number of elements in the input layer is set to 1 or more and 200 or less. By satisfying such conditions, the calculation cost does not become enormous, and highly stable prediction results can be obtained.

The condition C2 also includes a condition that the value of the number of hidden layers is equal to or less than a predetermined value. In the present embodiment, the value of the number of hidden layers is set to 1 or more and 7 or less. By satisfying such conditions, the calculation cost does not become enormous, and highly stable prediction results can be obtained.

The condition C2 also includes a condition that the value of the number of elements in the hidden layer is equal to or less than a predetermined value indicating a predetermined multiple of the number of elements in the input layer. In the present embodiment, the value of the number of elements in the hidden layer is set to 1 or more and 10 times or less the number of elements in the input layer. By satisfying such conditions, the calculation cost does not become enormous, and highly stable prediction results can be obtained.

[Condition C2 for parameter P2]
Condition C2 is selected from the predictors shown in the training data with respect to the predictors, and is not particularly limited as long as the number of predictors is the same as the number of elements in the input layer.

Further, the condition C2 includes a condition in which the number of data 60a in the learning data 60 used as the learning period is within a predetermined range for the learning period. In this embodiment, the number of data 60a to be used is set to a value smaller than 1500. By satisfying such conditions, highly stable prediction results can be obtained.

The condition C2 also includes a condition in which the value of the learning coefficient is equal to or less than a predetermined value. In the present embodiment, the value of the learning coefficient is set to a value larger than 0 and smaller than 1. By satisfying such conditions, highly stable prediction results can be obtained.

The condition C2 also includes a condition in which the number of learnings is equal to or less than a predetermined value. In the present embodiment, the value of the number of learnings is set to a value smaller than 1000000. By satisfying such conditions, the calculation cost does not become enormous, and highly stable prediction results can be obtained.

The condition C2 also includes a condition in which the value of the moment term is equal to or less than a predetermined value. In the present embodiment, the value of the moment term is set to a value larger than 0 and smaller than 1. By satisfying such conditions, highly stable prediction results can be obtained.

[Functional block of information processing device 5]
FIG. 13 is a diagram showing a functional block of the information processing apparatus 5 of the present embodiment. In the information processing device 5 of the present embodiment, the processor 500 executes a predetermined program to provide an analysis unit 510, an acquisition unit 511, an input reception unit 512, an input reception unit 513, a determination unit 514, and the like. A determination unit 515, an alarm unit 516, a learning unit 517, an output unit 518, an acquisition unit 519, an output unit 520, an output unit 521, and a display control unit 522 are realized.

The analysis unit 510 analyzes the power demand predicted by the demand forecasting device 4, and calculates the degree of influence of the predictor. FIG. 14 is a diagram showing an example of the degree of influence of the predictor calculated by the analysis unit 510. From the calculation result by the analysis unit 510, the operator can consider the selection of the predictor when generating a new learning model.

The acquisition unit 511 acquires the learning data 60 stored in the auxiliary storage device 502.

The input receiving unit 512 (corresponding to the first input receiving unit) accepts the input of the selected learning model from a plurality of different types of learning models. The plurality of different types of learning models are learning models included in the information 7 (FIG. 11) of the learning model stored in the auxiliary storage device 502. The selected learning model is a learning model desired by the operator among a plurality of different types of learning models.

The input receiving unit 513 (corresponding to the second input receiving unit) accepts the input of the parameter P0. As described above, the parameter P0 includes the parameters P1 and P2.

At this time, the input receiving unit 513 accepts the input of the parameter P0 according to the learning model input via the input receiving unit 512. The parameter P0 is output to the determination units 514 and 515, which will be described later.

The determination unit 514 (corresponding to the first determination unit) determines whether or not the parameter P0 input via the input reception unit 513 satisfies the condition C1. Here, the condition C1 corresponds to the learning model input via the input reception unit 512. The determination result is output to the alarm unit 516, which will be described later.

The determination unit 515 (corresponding to the second determination unit) determines whether or not the parameter P0 input via the input reception unit 513 satisfies the condition C2. Here, the condition C2 corresponds to the learning model input via the input reception unit 512. The determination result is output to the alarm unit 516, which will be described later.

The alarm unit 516 (corresponding to the first alarm unit) issues an alarm when the determination unit 514 determines that at least one of the parameters P0 does not satisfy the condition C1.

The alarm unit 516 further issues an alarm when the determination unit 515 determines that at least one of the parameters P0 does not satisfy the condition C2.

When the determination units 514 and 515 determine that each of the parameters P0 satisfies the condition C1, the learning unit 517 learns the learning model selected by using the parameter P0. By this learning, the parameter P3 is optimized. The procedure for generating the learning model is the same as the procedure for the learning unit 311 described above to generate the learning model.

The output unit 518 outputs the parameter P0 adjusted by the above procedure to the learning processing device 3 and the demand forecasting device 4.

The acquisition unit 519 (corresponding to the first acquisition unit) sets the prediction factor (corresponding to the first prediction factor) on the prediction target date (corresponding to the first prediction target date) in the same manner as the acquisition unit 410 of the above-mentioned demand forecasting device 4. Get it by the method of. The acquisition unit 519 further acquires the forecast value or the actual value of the predictor on the forecast target date.

In the present embodiment, the forecast target date is a predetermined period in the past. In addition, the predictor is the actual value on the forecast target date.

The output unit 520 (corresponding to the first output unit) outputs a predicted value of the power demand from the prediction factor acquired by the acquisition unit 519 by using the learning model generated by the learning unit 517 learning.

The output unit 521 outputs an error of the predicted value of the power demand output by the output unit 520 with respect to the actual value of the power demand.

The display control unit 522 generates a selection screen or the like when the input reception unit 512 or 513 accepts an input such as a learning model, and a pop-up screen for displaying an alarm issued by the alarm unit 516, and displays the pop-up screen on the output device 504. Details of these screens will be described later.

<Processing until parameter P0 is saved>
FIG. 15 is a flowchart illustrating a flow of processing until the information processing apparatus 5 saves the parameter P0. The process up to saving the parameter P0 includes steps S301 to S311.

First, in step S301, the acquisition unit 511 acquires the learning data 60. The learning data 60 shown in FIG. 3 is an example of the learning data acquired here by the acquisition unit 511.

Next, in step S302, the input receiving unit 512 receives the input of the learning model. At this time, the input receiving unit 512 accepts the input of the learning model via the learning model selection screen 8. FIG. 16 is a learning model selection screen 8 as an example. The learning model selection screen 8 is generated by the display control unit 522 and displayed on the output device 504.

FIG. 16 lists a plurality of selectable learning models. The operator can select one of the plurality of learning models displayed in the list with a mouse or the like. Based on the learning model selected here, a new learning model is generated in a later step.

FIG. 16 shows an aspect in which a neural network is selected as a learning model. When the setting button B1 is selected in this state, the neural network is input as a learning model.

Next, in step S303, the input receiving unit 513 receives the input of the parameter P0. At this time, the input receiving unit 513 accepts the input of the parameter P0 via the parameter setting screen 9. FIG. 17 is a parameter setting screen 9 as an example at this time.

On the parameter setting screen 9, the type of the parameter P0 and the input fields 9a to 9f for inputting the respective values are displayed.

In step S303, the operator displays the number of elements of the input layer, the number of hidden layers, the number of elements in the hidden layer, the predictor, and the learning period in the input fields 9a to 9i of the parameter setting screen 9 in the input fields 9a to 9i. , Kernel function, learning coefficient, number of learnings, and moment term.

Here, the number of elements in the input layer, the number of hidden layers, the number of elements in the hidden layer, and the predictor are the parameters P1. The learning period, the kernel function, the learning coefficient, the number of learnings, and the moment term are parameters P2.

FIG. 18 is an example of a state in which the operator has completed the input of the parameter P0 in step S303. When the operator completes the input of the parameter P0, the operator selects the setting button B2.

Next, in step S304, the determination unit 514 determines whether or not each of the parameters P0 input via the parameter setting screen 9 satisfies the condition C1.

Here, in FIG. 18, 0 is input as the value of the number of hidden layers, which is one of the parameters P0. The value of the number of such hidden layers does not satisfy the condition C1 as described above. Therefore, the determination unit 514 determines that the condition C1 is not satisfied.

When the determination unit 514 determines that the condition C1 is not satisfied (step S304: N), the process proceeds to step S305, and the alarm unit 516 issues an alarm.

FIG. 19 is an example of an alarm issued by the alarm unit 516 after the determination unit 514 determines that the condition C1 is not satisfied. In this example, a pop-up screen 90 indicating an alarm is displayed.

When the operator selects the "OK" button B2 on the pop-up screen 90, the process returns to step S303, and the input receiving unit 513 accepts the input of the parameter P0 again. The operator can re-enter an appropriate number of hidden layers via the parameter setting screen 9.

If the determination unit 514 determines in step S304 that the condition C1 is satisfied (step S304: Y), the process proceeds to step S306.

In step S306, the determination unit 515 determines whether or not each of the parameters P0 input via the parameter setting screen 9 satisfies the condition C2.

In FIG. 18, 5000 is input as the value of the number of elements of the hidden layer, which is one of the parameters P0. The value of the number of such hidden layers does not satisfy the condition C2. Therefore, the determination unit 515 determines that the condition C2 is not satisfied.

When the determination unit 515 determines that the condition C2 is not satisfied (step S306: N), the process proceeds to step S307, and the alarm unit 516 issues an alarm.

FIG. 20 is an example of an alarm issued by the alarm unit 516 after the determination unit 515 determines that the condition C2 is not satisfied. In this example, a pop-up screen 91 showing an alarm is displayed.

On the pop-up screen 91, a message is displayed stating that the value of the number of elements in the hidden layer is too large at 5000 and that the recommended value is 30 or less.

Next, in step S308, the alarm unit 516 accepts an input as to whether or not to continue. On the pop-up screen 91, a message asking whether to continue is displayed.

When the operator selects the "No" button B5 displayed on the pop-up screen 91, the process returns to step S303, and the input receiving unit 513 accepts the input of the parameter P0 again. The operator can re-input the value of the number of elements of the preferable hidden layer via the parameter setting screen 9.

In step S306, if the determination unit 515 determines that the condition C2 is satisfied (step S306: Y), or if the operator selects the "Yes" button B4 in step S308 (step S308: Y), the process proceeds to step S309. move on.

In step S309, the learning unit 517 performs learning based on the learning model and the input parameter P0.

Next, in step S310, the learning unit 517 acquires the learning model obtained by the learning in step S309.

Next, in step S311, the output unit 520 outputs the predicted value of the power demand from the learning model acquired in step S310 and the actual value of the predictor.

Next, in step S312, the output unit 521 outputs an error of the predicted value of the power demand output in step S311 with respect to the actual value of the power demand.

FIG. 21 is a diagram illustrating the transition of the error. In FIG. 21, the horizontal axis is the prediction target date, and the prediction target date in this example is the same as the prediction target date in the example of FIG. The vertical axis is the error of the predicted value of power demand with respect to the actual value of power demand. The error shown in FIG. 21 is smaller than that in FIG. 9 due to the new adjusted parameter P0 being used in the demand forecast.

The operator can refer to the error shown in FIG. 21 and consider whether or not to adopt the parameter P0 adjusted in the procedure of steps S302 to S308 as a new parameter P0.

Next, in step S313, the output unit 521 receives an input as to whether or not to continue. In step S313, when the operator selects not to continue via an input screen (not shown) (step S313: N), the process returns to step S303, and the input receiving unit 513 accepts the input of the parameter P0 again.

When the operator selects to continue via the input screen in step S313 (step S313: Y), the process proceeds to step S314.

Next, in step S314, the output unit 518 outputs the parameter P0 related to the obtained learning model to the auxiliary storage device 302 and the auxiliary storage device 402. The output unit 518 may further output the parameter P3 to the auxiliary storage device 302 and the auxiliary storage device 402.

Through the above processing, the information processing apparatus 5 can generate and acquire a new learning model. Then, the demand forecast can be made by using the new trained model.

<Processing until parameter P3 is stored>
FIG. 22 is a flowchart illustrating a processing flow until the learning processing device 3 and the demand forecasting device 4 store the parameters P0 and P3. The process until the parameters P0 and P3 are stored includes steps S401 to S408. Note that step S401 is processing by the information processing apparatus 5, steps S402 to S404 are processing by the learning processing apparatus 3, and steps S405 to S408 are processing by the demand forecasting apparatus 4.

First, in step S401, the output unit 518 of the information processing apparatus 5 outputs the parameter P0 related to the learning model to the learning processing apparatus 3 and the demand forecasting apparatus 4.

Next, in step S402, the acquisition unit 313 of the learning processing device 3 acquires the parameter P0 output by the output unit 518.

Next, in step S403, the storage unit 314 of the learning processing device 3 stores the parameter P0 acquired by the acquisition unit 313.

Next, in step S404, the learning unit 311 of the learning processing device 3 optimizes the parameter P3 by performing learning based on the parameter P0 stored in step S403 and the learning data 60.

Further, in step S405, the acquisition unit 414 of the demand forecasting device 4 acquires the parameter P0 output by the learning unit 517.

Next, in step S406, the storage unit 415 stores the parameter P0 acquired by the acquisition unit 414.

Next, in step S407, the acquisition unit 414 of the demand forecasting device 4 acquires the parameter P3 optimized in step S404.

Next, in step S408, the storage unit 415 stores the parameter P3 acquired by the acquisition unit 313.

By the above processing, the parameter P0 adjusted by the information processing device 5 can be stored in the learning processing device 3 and the demand forecasting device 4, and the demand forecasting device 4 is the parameter adjusted by the information processing device 5. The parameter P3 optimized by P0 and the learning processing apparatus 3 can be stored.

The above embodiment is for facilitating the understanding of the present invention, and is not for limiting the interpretation of the present invention. Further, the present invention can be changed or improved without departing from the spirit thereof, and it goes without saying that the present invention includes an equivalent thereof.

For example, in the present embodiment, the mode in which the learning processing device 3, the demand forecasting device 4, and the information processing device 5 are separately provided in the information processing system 1 is shown, but the present invention is not limited thereto. For example, the learning processing device 3 and the demand forecasting device 4 may be integrally provided in the information processing system 1. Alternatively, the learning processing device 3, the demand forecasting device 4, and the information processing device 5 may be integrally provided in the information processing system 1.

Even in such cases, according to the information processing system 1, it is possible to easily adjust the parameters for determining the configuration of the learning model and the parameters for learning the learning model to appropriate values. Become.

== Summary ==
As described above, the information processing apparatus 5 of the above embodiment has an input receiving unit 512 that receives input of a selected learning model from a plurality of different types of learning models, and an input receiving unit that receives input of a parameter P0 related to the selected learning model. When the determination unit 515 and the determination unit 515 for determining whether the selected learning model can be learned or not satisfy the condition C1 and the parameter P0 satisfy the condition C1. It includes a learning unit 517 that learns a learning model selected by using the parameter P0.

According to such a configuration, the operator does not set an inappropriate value when adjusting the parameter P0 related to the learning model, even if he / she lacks the expertise and experience in designing the learning model. Therefore, according to such a configuration, the parameter P0 can be easily adjusted to an appropriate value.

The information processing apparatus 5 further includes an alarm unit 516 that issues an alarm when the parameter P0 does not satisfy the condition C1. According to such a configuration, the operator can recognize this when an inappropriate parameter P0 is input.

The information processing apparatus 5 includes a determination unit 515 for determining whether or not the parameter P0 satisfies the condition C2 for learning the selected learning model, and the alarm unit 516 does not satisfy the condition C2 for the parameter P0. If so, issue an alarm.

According to such a configuration, the operator can recognize this when the non-recommended parameter P0 is input. This allows the operator to reconsider the selection of parameter P0. Therefore, according to such a configuration, the parameter P0 related to the learning model can be easily adjusted to a more appropriate value.

In the information processing apparatus 5, the parameter P0 includes a parameter P1 for determining the configuration of the selected learning model and a parameter P2 for learning the selected learning model. According to such a configuration, the parameter P1 and the parameter P2 can be easily adjusted to more appropriate values.

In the information processing apparatus 5, the selected learning model is a neural network, the parameter P1 includes at least one of the number of hidden layers and the number of elements in the hidden layers, and the condition C1 is the value of the parameter P1. Includes a condition of 1 or more. According to such a configuration, when the learning model by the neural network is generated, the value of the inappropriate parameter P1 is not set.

In the information processing apparatus 5, the parameter P2 includes at least one of the learning coefficient, the moment term, and the number of learnings, and the condition C1 includes a condition in which the value of the parameter P2 is larger than 0. According to such a configuration, when the learning model by the neural network is generated, the value of the inappropriate parameter P2 is not set.

In the information processing apparatus 5, the parameter P1 includes the number of hidden layers, and the condition C2 includes a condition that the value of the number of hidden layers is equal to or less than the first value. With such a configuration, it is possible to prevent setting a value for the number of hidden layers, which is not recommended, when generating a learning model by a neural network.

In the information processing apparatus 5, the parameter P1 includes the number of elements in the hidden layer, and the condition C2 is a value indicating that the value of the number of elements in the hidden layer is a predetermined multiple of the number of elements in the input layer (corresponding to the second value). Includes the following conditions. With such a configuration, it is possible to prevent setting the number of elements in the hidden layer, which is not recommended, when the learning model by the neural network is generated.

In the information processing apparatus 5, the parameter P2 includes a learning coefficient, and the condition C2 includes a condition in which the value of the learning coefficient is equal to or less than the third value. With such a configuration, it is possible to prevent setting a value of a learning coefficient that is not recommended when generating a learning model by a neural network.

In the information processing apparatus 5, the parameter P2 includes a moment term, and the condition C2 includes a condition in which the value of the moment term is equal to or less than the fourth value. With such a configuration, it is possible to prevent setting a value of a moment term that is not recommended when generating a learning model by a neural network.

In the information processing apparatus 5, the parameter P2 includes the number of learnings, and the condition C2 includes a condition in which the number of learnings is equal to or less than the fifth value. With such a configuration, it is possible to prevent setting an unrecommended learning count value when generating a learning model by a neural network.

The information processing apparatus 5 uses the acquisition unit 519 for acquiring the prediction factor on the prediction target date and the learning model learned by the learning unit 517 to obtain the predicted value of the demand of the predetermined target on the prediction target date from the prediction factor. Further includes an output unit 520 for output. With such a configuration, the validity of the learning model learned by the learning unit 517 can be verified.

The information processing system 1 uses an information processing apparatus 5, an acquisition unit 410 that acquires a prediction factor on a prediction target date, and a predetermined learning model to output a predicted value of demand of a predetermined target on the prediction target date. The error between the unit 411, the acquisition unit 412 that acquires the actual value of the demand of the predetermined target on the forecast target date, and the predicted value of the demand of the predetermined target and the actual value of the demand of the predetermined target is a predetermined condition. It is provided with a demand forecasting device 4 including an alarm unit 413 that issues an alarm when the condition is satisfied. With such a configuration, the validity of the learning model used in the demand forecasting device 4 can be verified.

In the information processing system 1, the predetermined condition is that when the error exceeds the predetermined value, the number of times the error is positive or negative exceeds the predetermined number of times in the predetermined number of predictions, the predetermined number of predictions. In the case where the variation of the error exceeds a predetermined value, or when a distribution tendency deviating from the predetermined distribution tendency of the error appears in a predetermined number of predictions. With such a configuration, the degree of freedom of the conditions for the alarm unit 413 to issue an alarm is increased.

In the information processing system 1, the predictor is an actual value on the prediction target date. With such a configuration, only the error caused by the learning model can be extracted.

In the information processing system 1, the demand forecasting device 4 further includes an acquisition unit 414 for acquiring the parameter P0 related to the learning model generated by learning by the learning unit 517, and a storage unit 415 for storing the parameter P0. According to such a configuration, the demand forecasting device 4 can perform demand forecasting after applying the parameter P0 to the learning model.

1: Information processing system 2: Sensor 3: Learning processing device 300: Processor 301: Main memory device 302: Auxiliary storage device 303: Input device 304: Output device 305: Communication device 310: Acquisition unit 311: Learning unit 312: Storage unit 313: Acquisition unit 314: Storage unit 4: Demand forecasting device 400: Processor 401: Main storage device 402: Auxiliary storage device 403: Input device 404: Output device 405: Communication device 410: Acquisition unit 411: Output unit 412: Acquisition unit 413: Alarm unit 414: Acquisition unit 415: Storage unit 5: Information processing device 500: Processor 501: Main storage device 502: Auxiliary storage device 503: Input device 504: Output device 505: Communication device 510: Analysis unit 511: Acquisition unit 512: Input reception unit 513: Input reception unit 514: Judgment unit 515: Judgment unit 516: Alarm unit 517: Learning unit 518: Output unit 519: Acquisition unit 520: Output unit 521: Output unit 522: Display control unit 60: Learning Data 61: Learning model 7: Learning model information 8: Learning model selection screen 9: Parameter setting screen 90: Pop-up screen 91: Pop-up screen

Claims (15)

A first input receiving unit that accepts input of a selected learning model from a plurality of different types of learning models,
A second input receiving unit that accepts input of parameters related to the selected learning model, and
A first determination unit for determining whether or not the parameter satisfies the first condition for determining whether or not the selected learning model can be learned.
When the parameter satisfies the first condition, a learning unit that learns the selected learning model using the parameter and a learning unit.
When the parameter does not satisfy the first condition, a first alarm unit that issues an alarm and
A second determination unit that determines whether or not the parameter satisfies the second condition for learning the selected learning model.
Equipped with
The first alarm unit issues an alarm when the parameter does not satisfy the second condition.
Information processing device. The information processing apparatus according to claim 1.
The parameter includes a first parameter for defining the configuration of the selected learning model and a second parameter for learning the selected learning model.
Information processing device. The information processing apparatus according to claim 2.
The selected learning model is a neural network.
The first parameter includes at least one of the number of hidden layers and the number of elements in the hidden layer.
The first condition includes a condition that the value of the first parameter is 1 or more.
Information processing device. The information processing apparatus according to claim 3.
The second parameter includes at least one of a learning coefficient, a moment term, and the number of learnings.
The first condition includes a condition in which the value of the second parameter is greater than 0.
Information processing device. The information processing apparatus according to claim 3 or 4.
The first parameter includes the number of hidden layers.
The second condition is an information processing apparatus including a condition that the value of the number of hidden layers is equal to or less than the first value. The information processing apparatus according to any one of claims 3 to 5.
The first parameter includes the number of elements in the hidden layer.
The second condition includes a condition that the value of the number of elements in the hidden layer is equal to or less than the second value indicating a predetermined multiple of the number of elements in the input layer.
Information processing device. The information processing apparatus according to any one of claims 4 to 6.
The second parameter includes a learning coefficient and includes.
The second condition includes a condition in which the value of the learning coefficient is equal to or less than the third value.
Information processing device. The information processing apparatus according to any one of claims 4 to 7.
The second parameter includes a moment term.
The second condition includes a condition in which the value of the moment term is equal to or less than the fourth value.
Information processing device. The information processing apparatus according to any one of claims 4 to 8.
The second parameter includes the number of learnings and includes
The second condition includes a condition in which the number of learnings is equal to or less than the fifth value.
Information processing device. The information processing apparatus according to any one of claims 1 to 9.
The first acquisition unit that acquires the first prediction factor on the first prediction target date,
An information processing device further comprising a first output unit that outputs a predicted value of demand of a predetermined target on the first prediction target date from the first predictor using the learning model learned by the learning unit. The information processing apparatus according to any one of claims 1 to 10.
The second acquisition unit that acquires the second predictor on the second forecast target date,
Using a predetermined learning model, a second output unit that outputs a predicted value of the demand of the predetermined target on the second forecast target date, and a second output unit.
The third acquisition unit that acquires the actual value of the demand of the predetermined target on the second forecast target date, and the third acquisition unit.
Information processing including a demand forecasting device including a second alarm unit that issues an alarm when an error between the predicted value of the demand of the predetermined target and the actual value of the demand of the predetermined target satisfies a predetermined condition. system. The information processing system according to claim 11.
The predetermined conditions are
If the error exceeds a predetermined value,
When the number of times the error is either positive or negative exceeds the predetermined number of times in the predetermined number of predictions.
When the variation of the error exceeds the predetermined value in the predetermined number of predictions,
When a distribution tendency that deviates from the predetermined distribution tendency of the error appears at a predetermined number of predictions,
An information processing system that is one of the following. The information processing system according to claim 11 or 12.
The second predictor is an information processing system which is an actual value on the second prediction target date. The information processing system according to any one of claims 11 to 13.
The demand forecaster is
A fourth acquisition unit that acquires the parameters generated by learning by the learning unit, and
An information processing system further including a storage unit for storing the parameters. A step that accepts input from a selected learning model from multiple different types of learning models,
A step that accepts input of parameters related to the selected learning model,
A step of determining whether or not the parameter satisfies the first condition for determining whether or not the selected learning model can be trained.
When the parameter satisfies the first condition, a step of learning the selected learning model using the parameter and a step of learning the selected learning model.
If the parameter does not meet the first condition, a step of issuing an alarm and
A step of determining whether or not the parameter satisfies the second condition for learning the selected learning model.
If the parameter does not meet the second condition, a step of issuing an alarm and
Information processing methods including.

Images