JP6687241B2 - Power generation power prediction device, server, computer program, and power generation power prediction method - Google Patents

Description

  The present invention relates to a power generation prediction device, a server, a computer program, and a power generation prediction method that provide power generation prediction data for photovoltaic power generation.

  In recent years, the introduction of renewable energy has been promoted to counter global warming and improve energy security. Photovoltaic power generation, which converts solar energy into electricity, is one of the renewable energies and has the advantage of not generating greenhouse gases, and its spread is expected.

  In solar power generation as well as in other forms of power generation, if it is not possible to balance the amount of power supply and the amount of power demand, it may not be possible to operate the power system in a stable manner, and power generation prediction is performed. This is very important. Therefore, by simulating a physical meteorological model related to the amount of water vapor in the atmosphere and clouds based on the weather forecast data, the light transmittance of the clouds is calculated, and the spectral data that reaches the ground at the predicted point is calculated. There is disclosed a power generation amount prediction system for a solar power generation device that calculates the amount of solar radiation based on data and predicts the power generation amount (see Patent Document 1).

JP, 2011-159199, A

  However, in the system of Patent Document 1, it depends on the accuracy of the weather forecast data. In addition, the distribution of the weather forecast data is generally updated after a certain time interval (for example, several hours). Therefore, in the conventional forecasting system, there is a possibility that the forecasting accuracy of the power generation amount may decrease depending on the accuracy of the weather forecast data and the time interval of distribution.

  The present invention has been made in view of the above circumstances, and provides a power generation prediction device capable of accurately predicting power generation, a server, a computer program for realizing the power generation power prediction device, and a power generation prediction method. The purpose is to provide.

  A power generation prediction apparatus according to an embodiment of the present invention is a power generation prediction apparatus that provides power generation prediction data for photovoltaic power generation, and includes time-series actually measured power generation related to power generation equipment installed at a plurality of points. A power generation information acquisition unit that acquires power generation information, a weather information acquisition unit that acquires weather information regarding the plurality of points, power generation information acquired by the power generation information acquisition unit, and weather information acquired by the weather information acquisition unit. A storage unit for storing, a prediction parameter determination unit that determines a prediction parameter including a prediction target point and a prediction target time, and power generation information and weather information stored in the storage unit and a prediction parameter determined by the prediction parameter determination unit And a prediction input data generation unit that generates prediction input data for a predetermined neural network model, and a prediction input data generation unit that generates the prediction input data. A prediction output data generation unit that generates prediction output data based on the input data for use and the neural network model, and a prediction target point and a prediction target time based on the prediction output data generated by the prediction output data generation unit. It is characterized by comprising a predicted generated power calculating section for calculating predicted generated power.

  A computer program according to an embodiment of the present invention is a computer program for causing a computer to provide power generation prediction data for photovoltaic power generation, and is a time series of a power generation facility installed at a plurality of points. A power generation information acquisition unit that acquires power generation information including measured power generation, a weather information acquisition unit that acquires weather information related to the plurality of points, a storage unit that stores the acquired power generation information and weather information, and a prediction target point And a prediction parameter determining unit that determines a prediction parameter including a prediction target time, and a prediction input that generates prediction input data to a predetermined neural network model based on the stored power generation information and weather information and the determined prediction parameter. Based on the data generator, the generated prediction input data and the neural network model. A predicted output data generation unit that generates predicted output data, and a predicted generated power calculation unit that calculates predicted generated power at the prediction target point and prediction target time based on the generated prediction output data. And

  A power generation prediction method according to an embodiment of the present invention is a power generation prediction method that provides power generation prediction data for photovoltaic power generation, and includes time-series actually measured power generation related to power generation facilities installed at a plurality of points. The power generation information acquisition unit acquires the power generation information, the weather information acquisition unit acquires the weather information related to the plurality of points, and the acquired power generation information and the weather information are stored in the storage unit, and the prediction target point and the prediction target time The prediction parameter determining unit determines the prediction parameter including, and based on the power generation information and the weather information stored in the storage unit and the determined prediction parameter, the prediction input data to the predetermined neural network model is input to the prediction input data. The prediction output data generation unit generates the prediction output data based on the prediction input data generated by the generation unit and the generated neural network model. The predictive power generated by the prediction target point and the prediction target time prediction generated power calculating unit and calculating on the basis of the generated predicted output data.

  The power generation information acquisition unit acquires power generation information including time-series actually measured power generation related to power generation facilities installed at a plurality of points. The power generation information can include, for example, facility information on the power generation facility, measurement unit information on a measurement unit that measures the power generated by the power generation facility, and position information on the power generation facility, in addition to the actually measured power generation.

  The weather information acquisition unit acquires weather information regarding a plurality of points. The weather information is provided by a weather forecast business operator or the like over time at each of a plurality of points, and includes past measured weather information and future weather information forecasts. The weather information is also time-series data similar to the measured power generation.

  The prediction parameter determination unit determines a prediction parameter including a prediction target point and a prediction target time. The prediction target point may be a point where the power generation facility is installed or a point where the power generation facility is installed in the future. The position of the power generation facility that predicts the generated power is the prediction target point, and the prediction target time specifies the time at which the generated power is predicted.

  The prediction input data generation unit generates prediction input data for a predetermined neural network model based on the power generation information and weather information stored in the storage unit and the prediction parameter determined by the prediction parameter determination unit. The neural network is composed of an input layer, an output layer and a plurality of intermediate layers. The prediction input data generation unit generates prediction input data in the vector format having the same size as the number of nodes in the input layer (size of the input layer).

  The prediction output data generation unit generates prediction output data based on the prediction input data generated by the prediction input data generation unit and the neural network model. The prediction output data generation unit is an output layer of the neural network, and generates prediction output data in vector format having the same size as the number of nodes in the output layer (size of output layer).

  The predicted generated power calculation unit calculates predicted generated power at the prediction target point and the prediction target time based on the predicted output data generated by the predicted output data generation unit. For example, the number of output layers is 1, 2, ... N, and the prediction output data generated by the prediction output data generation unit is y1, y2 ,. The predicted generated power calculation unit calculates predicted generated power based on the predicted output data y1, y2, ... Yk ,. For example, if yk = 1 and the other predicted output data is 0, the predicted generated power calculation unit calculates the predicted generated power based on yk.

  As described above, the time-series data of the measured power generation and the time-series data of the measured meteorological information at a plurality of points are used as the input data for prediction of the neural network. It is possible to accurately predict the generated power.

  In the generated power prediction apparatus according to the embodiment of the present invention, the prediction input data generation unit is a time-series measured generated power and each power generation facility for each power generation facility in a predetermined period before the prediction target time. The input data for prediction including the position information of is generated.

  The prediction input data generation unit generates prediction input data including time-series actually-measured generated power of each power generation facility and position information of each power generation facility in a predetermined period before the prediction target time. In addition, time means a time point or a time zone. It is assumed that the time is uniquely expressed as t1, t2, t3, ... For each unit time (for example, minutes) in time series. For example, assuming that the prediction target time is t1140, the predetermined period before the prediction target time t1140 can be a period from 80 steps before the unit time (for example, minutes) to 50 steps before, such as t1060 to t1090. . That is, the prediction input data includes time-series actually-measured generated power of each power generation facility at times t1060 to t1090. It should be noted that if the length of the period is increased, the number of pieces of prediction input data increases, so that the prediction accuracy of the generated power increases.

  With the above configuration, the time-series data of the actually-generated power generated by each power generation facility in the predetermined period is used to accurately predict the power generated at the prediction target point in the prediction target time after the predetermined period. can do.

  In the generated power predicting apparatus according to the embodiment of the present invention, the predictive input data generating unit may generate predictive input data including actually measured meteorological information in the predetermined period for the plurality of points. Characterize.

  The prediction input data generation unit generates prediction input data including actually measured weather information for a plurality of points in a predetermined period. For example, assuming that the prediction target time is t1140, the predetermined period before the prediction target time t1140 can be a period from 80 steps before the unit time (for example, minutes) to 50 steps before, such as t1060 to t1090. . That is, the input data for prediction includes meteorological information measured at times t1060 to t1090. It should be noted that if the length of the period is increased, the number of pieces of prediction input data increases, so that the prediction accuracy of the generated power increases.

  With the above configuration, the time-series data of the meteorological information actually measured at a plurality of points in a predetermined period is used to accurately predict the generated power at the prediction target point in the prediction target time after the predetermined period. can do.

  The generated power prediction apparatus according to the embodiment of the present invention is characterized in that the prediction input data generation unit generates prediction input data including weather information for prediction at the prediction target point and the prediction target time. To do.

  The prediction input data generation unit generates prediction input data including weather information for prediction at a prediction target point and a prediction target time. By including the weather information of the prediction at the prediction target point and the prediction target time in the prediction input data, it is possible to accurately predict the generated power.

  The generated power prediction apparatus according to the embodiment of the present invention is characterized in that the predicted output data generation unit generates predicted output data for each of the plurality of divided ranges of the generated power.

  The predicted output data generation unit divides the range of generated power into a plurality of sections and generates predicted output data for each section. When the number of nodes in the output layer of the neural network is N, the range of generated power is divided into N sections. For example, when the power of 50 kW is divided into 50 sections, the generated power from 0 kW to 50 kW can be treated as output data in units of 1 kW. As a result, it is possible to predict the generated power in units of sections. The range of generated power is not limited to 50 kW. Also, the number of categories is not limited to 50.

  In the generated power prediction apparatus according to the embodiment of the present invention, the predicted generated power calculation unit calculates predicted power based on a representative value defined for each of the plurality of sections and predicted output data generated for each of the sections. Is calculated.

  The predicted generated power calculation unit calculates predicted generated power based on the representative value set for each of the plurality of sections and the predicted output data generated for each section. When the power of 50 kW is divided into 50 divisions, the division 1 is set to 0 kW or more and less than 1 kW, the division 2 is set to 2 kW or more and less than 3 kW, and the ranges of division 3 ... division 50 are similarly defined. The representative value can be the median value in each section. For example, the median value of category 1 is 0.5 kW, the median value of category 2 is 1.5 kW, and so on. The median value of the section i is represented by mi. Further, the prediction output data of the section 1 is y1, the prediction output data of the section 2 is y2, and similarly, the prediction output data of the section N is yN. The predicted output data of the section i is represented by yi.

  The predicted generated power calculation unit calculates the predicted generated power W by the formula W = {Σ (yi × mi)} / Σyi. For example, if the prediction output data y20 of the category 20 (range: 19 kW or more and less than 20 kW, median value: 19.5 kW) is 1 and all other prediction output data are 0, the predicted generated power W is 19.5 kW. Becomes As a result, it is possible to calculate the generated power according to the strength of the output data output from the output layer of the neural network (which can be interpreted as the probability for each category).

  A power generation prediction apparatus according to an embodiment of the present invention includes a learning parameter determination unit that determines a learning parameter including a learning target point and a learning target time, power generation information and weather information stored in the storage unit, and the learning parameter determination. Based on the learning parameter determined by the learning unit, learning input data generation unit for generating learning input data of the neural network model, and learning of the neural network model based on the learning parameter determined by the learning parameter determination unit Output data generating unit for generating output data for learning, learning input data generated by the input data generating unit for learning and learning output data generated by the output data generating unit for learning of the neural network model A learning processing unit that performs a learning process is provided.

  The learning parameter determination unit determines a learning parameter including a learning target point and a learning target time. The learning parameter is a prediction parameter used for learning and updating the neural network. The learning target point may be a point where the power generation facility is installed or a point where the power generation facility is installed in the future. The position of the power generation equipment that predicts the generated power is the learning target point, and the learning target time specifies the time at which the generated power is predicted.

  The learning input data generation unit generates learning input data of the neural network model based on the power generation information and the weather information stored in the storage unit and the learning parameter determined by the learning parameter determination unit. The input data for learning is input data used for learning and updating of the neural network, and the input data generation unit for learning uses a vector format learning of the same size as the number of nodes in the input layer (size of input layer). Input data for use.

  The learning output data generation unit generates learning output data of the neural network model based on the learning parameter determined by the learning parameter determination unit. The learning output data defines a distribution of output values expected as output data of the neural network. The learning output data generating unit generates learning output data in a vector format having the same size as the number of nodes in the output layer (size of output layer).

  The learning processing unit performs learning processing of the neural network model based on the learning input data generated by the learning input data generation unit and the learning output data generated by the learning output data generation unit. The output data output by the neural network based on the learning input data are y1, y2, ... YN. Further, output values expected as output data of the neural network are z1, z2, ... ZN. The learning processing unit optimizes the weight of the connection between the nodes of the neural network so that the error is minimized by the error function having the output data output by the neural network and the expected output value as variables.

  As described above, the learning parameters including the time-series data of the measured power generation and the time-series data of the measured meteorological information at the plurality of points, the learning target point and the learning target time are used as input data for learning of the neural network. Since the neural network is learned and updated, it is possible to accurately predict the generated power at any prediction target point and prediction target time.

  In the generated power prediction apparatus according to the embodiment of the present invention, the learning input data generation unit includes time-series actually-measured generated power and each power generation facility for each power generation facility in a predetermined period before the learning target time. The input data for learning including the position information of is generated.

  The learning input data generation unit generates learning input data including time-series actually-measured generated power of each power generation facility and position information of each power generation facility in a predetermined period before the learning target time. For example, assuming that the learning target time is t110, the predetermined period before the learning target time t110 can be a period from 80 steps to 50 steps before a unit time (for example, minutes) like t30 to t60. . That is, the learning input data includes the time-series actually-measured generated power of each power generation facility from time t30 to time t60. It should be noted that if the length of the period is increased, the number of input data for learning increases, so that the prediction accuracy of the generated power increases.

  With the above configuration, the time-series data of the actually-generated power generated by each power generation facility in the predetermined period is used to accurately predict the power generated at the prediction target point in the prediction target time after the predetermined period. The neural network for learning and updating can be performed.

  In the power generation predicting apparatus according to the embodiment of the present invention, the learning input data generating unit may generate learning input data including actually measured weather information for the plurality of points in the predetermined period. Characterize.

  The learning input data generation unit generates learning input data including actually measured weather information for a plurality of points in a predetermined period. For example, assuming that the learning target time is t110, the predetermined period before the learning target time t110 can be a period from 80 steps to 50 steps before a unit time (for example, minutes) like t30 to t60. . That is, the input data for learning includes meteorological information actually measured from time t30 to time t60. It should be noted that if the length of the period is increased, the number of input data for learning increases, so that the prediction accuracy of the generated power increases.

  With the above configuration, the time-series data of the meteorological information actually measured at a plurality of points in a predetermined period is used to accurately predict the generated power at the prediction target point in the prediction target time after the predetermined period. The neural network for learning and updating can be performed.

  The generated power predicting apparatus according to the embodiment of the present invention is characterized in that the learning input data generation unit generates learning input data including actual weather information at the learning target point and the learning target time. To do.

  The learning input data generation unit generates learning input data including actual measurement weather information at a learning target point and a learning target time. By including the actually measured weather information at the learning target point and the learning target time in the learning input data, the neural network for accurately predicting the generated power can be learned and updated.

  In the generated power prediction apparatus according to the embodiment of the present invention, the learning output data generation unit generates learning output data including actually measured generated power of a power generation facility installed at a point near the learning target point. It is characterized by

  The learning output data generation unit generates learning output data that includes actually measured generated power of a power generation facility installed near a learning target point. When the learning target point is a point where the power generation equipment is installed, the learning target point can be set as the relevant point. Further, when the learning target point is apart from any of the plurality of points where the power generation equipment is installed, the point closest to the learning target point among the plurality of points can be set as the learning target point. By including in the output data for learning the actually-measured generated power of the power generation equipment installed near the learning target point, it is possible to learn and update the neural network for accurately predicting the generated power.

  The power generation prediction apparatus according to the embodiment of the present invention is characterized in that the meteorological information includes at least one of weather, weather change, cloud amount, and solar radiation.

The weather information includes at least one of weather, weather change, cloud cover, and solar radiation. The weather is, for example, “fine”, “cloudy”, “finely cloudy”, “temporary clear”, or the like. The change of the weather is, for example, "after cloudy weather" or "after cloudy weather". Cloud amount is the ratio of clouds to the entire sky. The solar radiation is, for example, the solar radiation intensity per unit area (kW / m ² ), the amount of solar radiation per unit time / unit area (kWh / m ² ), and the like. As a result, it is possible to consider factors that affect the generated power of the photovoltaic power generation.

  The power generation prediction apparatus according to the embodiment of the present invention is characterized in that the power generation information includes a rated power generation capacity of the power generation facility.

  The power generation information includes the rated power generation capacity of the power generation facility. The rated power generation capacity is the rated value of the power generation output. As a result, the generated power can be accurately predicted regardless of the rated value of the power generation output of the power generation equipment.

  A server according to an embodiment of the present invention uses a power generation data acquisition unit that acquires power generation data related to power generation equipment from a company related to solar power generation, and a power generation data acquired by the power generation data acquisition unit. An output unit that outputs first information regarding power generation to a computer of the business operator under a predetermined first price condition, and an information provision request from the business operator under a second price condition that is higher than the first price condition. In this case, the power generation predicting apparatus according to the embodiment of the present invention is characterized by further comprising: an output unit configured to output the second information including the power generation prediction data to the computer of the business operator under the second price condition.

  The power generation data acquisition unit acquires power generation data regarding a power generation facility from a solar power generation company. The company is, for example, a power plant company or an O & M (Operation & Maintenance) company. The power generation data can be acquired from the computer of the business operator.

  The output unit outputs the first information regarding the photovoltaic power generation to the computer of the business operator under the predetermined first price condition based on the power generation data acquired by the power generation data acquisition unit. The first information relates to, for example, information about a calendar function (for example, a maintenance inspection schedule, a weather forecast, etc.), a maintenance inspection management function (output of a checklist, storage, management, etc.) or a power generation amount management function. Includes information related to basic services such as information. The first price condition may be free, for example.

  When the output unit receives the information provision request from the business operator under the second price condition that is higher than the first price condition, the output unit outputs the second information including the power generation prediction data by the power generation prediction device according to the second embodiment. Output to the operator's computer under the price condition. The second price condition can be charged, for example. The information provision request includes, for example, requesting a paid optional function (intelligence function).

  With the above configuration, it is possible to obtain the power generation data on the power generation equipment from the business while providing the basic service on the solar power generation to the business related to the solar power generation free of charge. Can be collected as so-called big data. The collected big data (power generation data) can be used to learn and update the neural network of the power generation prediction apparatus, and the prediction accuracy of the generated power of the power generation prediction apparatus can be improved. Further, for a business operator who wants to use the pay option, it is possible to provide a service related to power generation prediction whose accuracy is improved by machine learning.

  According to the present invention, it is possible to accurately predict the generated power.

It is a block diagram showing an example of composition of a power plant and a power generation monitoring system of this embodiment. It is explanatory drawing which shows an example of the power generation information of a power station. It is a block diagram which shows an example of a structure of the electric power generation prediction apparatus of this Embodiment. It is an explanatory view showing an example of power generation information. It is explanatory drawing which shows an example of power station detailed information. It is explanatory drawing which shows an example of weather condition information. It is explanatory drawing which shows typically the relationship with the time change of the weather in the position where several power stations were installed, and the electric power generation of each power station. It is explanatory drawing which shows typically an example of a structure of the neural network model part of this Embodiment. It is a flow chart which shows an example of a procedure of learning processing and update processing of a neural network model part by a power generation prediction device of this embodiment. It is an explanatory view showing an example of generated power information and weather information in each power station. It is explanatory drawing which shows a mode that the prediction parameter for learning is determined. It is explanatory drawing which shows an example of the measured data of the power station of the conditions close to the prediction parameter for learning. It is explanatory drawing which shows an example of the time series one-dimensional vector of the generated electric power information and weather information for every power station and measurement unit. It is explanatory drawing which shows an example of the generation method of the input data for learning by the electric power generation prediction apparatus of this Embodiment. It is explanatory drawing which shows an example of the input data for learning which the power generation prediction apparatus of this Embodiment produces | generates. It is explanatory drawing which shows an example of the output data for learning which the power generation prediction apparatus of this Embodiment produces | generates. It is explanatory drawing which shows the relationship of the data for learning used for learning and update of the neural network model part by the power generation prediction apparatus of this Embodiment. It is explanatory drawing which shows an example of the class of the output data of the neural network model part of this Embodiment. It is explanatory drawing which shows an example of the output value of the output data of the neural network model part of this Embodiment. It is explanatory drawing which shows an example of distribution of the output value expected as output data of the neural network model part of this Embodiment. It is explanatory drawing which shows typically an example of the hierarchical structure of the output layer for calculation of a softmax function. It is a flow chart which shows an example of a procedure of prediction processing of generated power by a power generation prediction device of this embodiment. It is an explanatory view showing an example of generated power information and weather information in each power station. It is explanatory drawing which shows the mode of determination of a prediction parameter. It is explanatory drawing which shows an example of the time series one-dimensional vector of the generated electric power information and weather information for every power station and measurement unit. It is explanatory drawing which shows an example of the generation method of the input data for a prediction by the electric power generation prediction apparatus of this Embodiment. It is explanatory drawing which shows an example of the input data for a prediction which the power generation prediction apparatus of this Embodiment produces | generates. It is explanatory drawing which shows the relationship of the prediction data used for the prediction of the electric power generation by the electric power generation prediction apparatus of this Embodiment. It is explanatory drawing which shows an example of the output value of the output data of the neural network model part of this Embodiment. It is a block diagram which shows an example of a structure of the server of this Embodiment. It is explanatory drawing which shows the 1st example of the 1st information regarding solar power generation. It is explanatory drawing which shows the 2nd example of the 1st information regarding solar power generation. It is explanatory drawing which shows the 3rd example of the 1st information regarding solar power generation. It is a flow chart which shows an example of a processing procedure of a server of this embodiment. It is explanatory drawing which shows an example of the character string of this Embodiment. It is explanatory drawing which shows an example of conversion from a character string to a numerical value by the power generation prediction apparatus of this Embodiment. It is explanatory drawing which shows an example of the numerical conversion of this Embodiment.

  Hereinafter, the present invention will be described based on the drawings showing the embodiments. FIG. 1 is a block diagram showing an example of the configurations of a power plant 10 and a power generation monitoring system 20 according to this embodiment. The power plant 10 and the power generation monitoring system 20 are connected by a communication line 30 such as the Internet. Further, the power generation monitoring system 20 and a power generation prediction device described later are also connected by a communication line 31 such as the Internet. Although only one power plant 10 and one power generation monitoring system 20 are illustrated in FIG. 1, the number of power plants 10 and one power generation monitoring system 20 is not limited to one. Alternatively, one power generation monitoring system 20 may be connected to a plurality of power stations 10.

  The power plant 10 is a photovoltaic power plant as a power generation facility, and includes a photovoltaic power generation panel 11, an inverter 12, a current collector 13, a measurement unit 14, a communication unit 15, and the like. The inverter 12 converts the direct current output from the photovoltaic power generation panel 11 into alternating current. The photovoltaic power generation panel 11 and the inverter 12 are provided for each of a plurality of systems. The current collector 13 collects the electric power obtained in each system. The measurement unit 14 measures the generated power information collected by the current collector 13 and converts it into a numerical value or a character string. The equipment to be measured by the measurement unit 14 is the photovoltaic power generation panel 11 and the inverter 12 arranged in the preceding stage of the current collector 13. Each power plant 10 includes one measurement unit 14. The communication unit 15 transmits information such as the measurement unit information and the generated power information output by the measurement unit 14 to the power generation monitoring system 20.

  The power generation monitoring system 20 includes a communication unit 21, a processing unit 22, a database 23, a processing unit 24, a communication unit 25, and the like. The communication unit 21 receives the information transmitted by the communication unit 15 of the power plant 10. The processing unit 22 performs a process of registering the information received by the communication unit 21 in the database 23. The database 23 stores the power generation information of the power plant.

  FIG. 2 is an explanatory diagram showing an example of power generation information of a power plant. In the example of FIG. 2, the power generation information of two power stations is illustrated. As shown in FIG. 2, the power generation information includes time information, measurement unit information, generated power information (measured generated power), and the like. The power generation information is information in which time information, measurement unit information, generated power information, and the like are regularly or irregularly measured and recorded at regular intervals.

  The time information is a numerical value or a character string indicating a uniquely determined time point or time zone. For example, the character string “January 2016” (year month), the character string “May 23, 2016 am” (date and time zone), the character string “2016-05-23T1 11:00:01” (date and second) Unit time), or a numerical value "16101" (the number of milliseconds that has elapsed from 0:00:00 on January 1, 2016). In the example of FIG. 2, the time information is represented by the symbols t1 and t2.

  The measurement unit information is a numerical value or a character string that uniquely indicates the measurement unit 14. In the example of FIG. 2, the measurement unit information is represented by symbols u1 and u2.

  The generated power information is a DC / AC instantaneous generated power, a numerical value or a character string indicating the generated power amount or the strength of the generated power. For example, the numerical value "2" (the number of pulses per unit time of the pulse type watt hour meter), the numerical value "15.0" (measured value in the electric power meter which outputs a kW unit), the numerical value "20.01" (kWh), The character string “less than 10 kW” or the character string “10 kWh or more and less than 20 kWh” is included. In the example of FIG. 2, the generated power information is represented by symbols p1-1, p1-2, p2-1, p2-2, and the like.

  The processing unit 24 reads the power generation information from the database 23. The communication unit 25 transmits the power generation information read by the processing unit 24 to the power generation prediction device.

  FIG. 3 is a block diagram showing an example of the configuration of the power generation prediction device 50 according to the present embodiment. The power generation prediction device 50 is connected to a plurality of power generation monitoring systems 20, 20, ..., A weather forecast business operator 40 (more precisely, a server of the weather forecast business operator) via a communication line such as the Internet.

  As illustrated in FIG. 3, the power generation prediction device 50 includes a communication unit 51, a registration unit 52, a database 53 as a storage unit, a processing unit 54, a communication unit 55 that communicates with an external server, and the like. The processing unit 54 also includes an input data generation unit 541, a parameter determination unit 542, a neural network model unit 543, a predicted generated power calculation unit 544, a learning processing unit 545, a learning data generation unit 546, and the like.

  The database 53 stores power generation information 53a, power station detailed information 53b, weather condition information 53c, and the like. Details of the power generation information 53a, the power station detailed information 53b, and the weather condition information 53c will be described later.

  The communication unit 51 has a function as a power generation information acquisition unit, and acquires power generation information including time-series actually measured power generation related to the power plant 10 (power generation facility) installed at a plurality of points. In addition, the communication unit 51 has a function as a weather information acquisition unit and acquires weather information regarding a plurality of points.

  The registration unit 52 registers various information acquired (received) by the communication unit 51 in the database 53.

  FIG. 4 is an explanatory diagram showing an example of the power generation information 53a. As shown in FIG. 4, the power generation information includes power station information, time information, measurement unit information, generated power information, and the like. The power generation information is information in which information such as power station information, time information, measurement unit information, and generated power information is regularly and irregularly acquired and recorded at regular intervals. The power generation information is information in which power station information is added to the power generation information of each power station 10 transmitted by each power generation monitoring system 20. The power plant information is a numerical value or a character string that uniquely indicates the power plant 10. In the example of FIG. 4, the power plant information is represented by the symbols s1 and s2.

  FIG. 5 is an explanatory diagram showing an example of the power station detailed information 53b. As shown in FIG. 5, the power station detailed information includes power station information, measurement unit information, facility information, position information, and the like. The equipment information is information about equipment such as the photovoltaic panel 11 and the inverter 12. The facility information includes, for example, information such as the number of photovoltaic panels 11, manufacturer, model, rated capacity and total rated capacity, and number of inverters 12, manufacturer, model, rated capacity and total rated capacity. The facility information can be divided into a facility capacity and a facility type value. The equipment information is a numerical value or a character string that uniquely indicates the equipment. In the example of FIG. 5, the equipment information is represented by reference signs e1, e2, and the like.

  The position information is information in which the position of the power plant is represented by a numerical value or a character string. The position information may be represented by latitude and longitude, or may be represented by a character string indicating a region such as a city, a county, or a town. In the example of FIG. 5, the position information is represented by symbols l1, l2, and the like.

  FIG. 6 is an explanatory diagram showing an example of the weather condition information 53c. As shown in FIG. 6, the weather condition information 53c includes position information, time information, weather information, and the like. The weather information is information in which the weather at a specific time and position is represented by a numerical value or a character string. The weather information includes at least one of weather, weather change, cloud amount, and solar radiation at a weather station or power plant 10 whose position is near. The weather, transition of the weather, cloud amount and solar radiation are also collectively referred to as weather. The weather information includes forecasts of past measured weather information and future weather information.

  In the example of FIG. 6, the weather information in the position information 11 and the time information t1 is represented by w1-1. Similarly, the weather information in the position information 11 and the time information t2 is represented by w1-2. The weather information in the position information 12 and the time information t1 is represented by w2-1. Similarly, the weather information in the position information 12 and the time information t2 is represented by w2-2. Hereinafter, a specific example will be described.

  FIG. 7 is an explanatory diagram schematically showing the relationship between the temporal change of weather at the position where a plurality of power plants 10 are installed and the power generated by each power plant 10. The upper diagram shows the positions of multiple power plants and the state of temporal changes in the weather, and the lower diagram shows an example of temporal changes in generated power. For the sake of convenience, two power plants s1 and s2 (indicated by circled numbers) will be described. The positions of the two power plants s1 and s2 are designated as l1 and l2, respectively.

  The diagram on the left side of the upper row shows the weather at the position l1 of the power plant s1 by w1-1 and the weather at the position l2 of the power plant s2 by w2-1, which shows the state of the weather at time t1. The weather w1-1 is “cloudy” and the weather w2-1 is “clear”.

  The figure on the right side of the upper row shows the weather condition at the position l1 of the power plant s1 by w1-2 and the weather condition at the position l2 of the power plant s2 by w2-2, which shows the state of the weather at the time t2 after the time t1. Indicate. Meteorological w1-2 is “fine and sometimes cloudy”, and weather w2-2 is “temporary and fine”.

  As shown in the upper diagram, at time t1, the weather at the position of the power plant s1 is "cloudy", so the generated power of the power plant s1 decreases at time t1 and its vicinity. Further, at time t2, the weather at the position of the power station s2 is "temporarily cloudy", so the power generated by the power station s2 decreases at and near time t2.

  As described above, the weather information can be represented by a character string indicating the weather (for example, “cloudy”, “fine”, “fine cloudy”, “temporary clear”, etc.). Further, the weather information can be expressed by a code value indicating the weather. For example, "cloudy", "fine", "partly cloudy" and "temporary cloudy" can be described as "3", "1", "2" and "4", respectively. Further, the weather can be represented by the value of the solar radiation intensity measured at each power plant 10. For example, "cloudy", "fine", "partially cloudy" and "temporary cloudy" are described as "0.5", "0.9", "0.8" and "0.6", respectively. You can also In addition, the weather can be represented by a value of cloud amount (a ratio of clouds to the entire sky in the sky) observed at each weather station (for example, an integer value of 11 levels from 0 to 10). The cloud amount can also be indicated by a value that is inversely proportional to the insolation intensity. It can also be written as "4".

  The weather information can be given the same format for each measurement unit 14, but may be given a different format for each measurement unit 14. For example, to the measurement unit 14 whose measurement unit information is u1 (position information is l1), the weather observation value is given as weather information every three hours, and the measurement unit information whose measurement unit information is u2 (position information is l2). A measured value of solar radiation intensity can be given to 14 as weather information.

  The parameter determination unit 542 functions as a prediction parameter determination unit when predicting the generated power. That is, the parameter determination unit 542 determines the prediction parameter including the prediction target point and the prediction target time. The prediction target point may be the point where the power plant 10 is installed, or the point where the power plant 10 is installed in the future. The position of the power plant 10 that predicts the generated power is the prediction target point, and the prediction target time specifies the time at which the generated power is predicted.

  The parameter determining unit 542 functions as a learning parameter determining unit when learning and updating the neural network model unit 543. That is, the parameter determination unit 542 determines the learning parameter including the learning target point and the learning target time. The learning parameter is a prediction parameter used for learning and updating of the neural network model unit 543. The learning target point may be the point where the power plant 10 is installed, or the point where the power plant 10 is installed in the future. The position of the power plant 10 that predicts the generated power is the learning target point, and the learning target time specifies the time at which the generated power is predicted.

  The input data generation unit 541 functions as a prediction input data generation unit when predicting the generated power. That is, the input data generation unit 541 generates prediction input data to the neural network model unit 543 based on the power generation information and weather information stored in the database 53 and the prediction parameter determined by the parameter determination unit 542. The input data generation unit 541 generates vector input prediction data having the same size as the number of nodes in the input layer (size of the input layer) of the neural network model unit 543.

  FIG. 8 is an explanatory diagram schematically showing an example of the configuration of the neural network model unit 543 of this embodiment. As shown in FIG. 8, the neural network model unit 543 includes an input layer, an output layer, and a plurality of intermediate layers. Although two intermediate layers are shown in FIG. 8 for convenience, the number of intermediate layers is not limited to two and may be three or more. Further, in the example of FIG. 8, for convenience, the number of nodes in the input layer (layer size) is 5, the number of nodes in the intermediate layer is 3, and the number of nodes in the output layer is 4, but the layer size is It is not limited to the example of FIG.

  One or a plurality of nodes (neurons) exist in the input layer, the output layer, and the intermediate layer, and the nodes in each layer are unidirectionally coupled with desired weights to the nodes existing in the layers before and after. A vector having the same number of components as the number of nodes in the input layer (in the example of FIG. 8, only five are shown for convenience) is given as input data of the learned neural network model unit 543. In FIG. 8, actual measurement values and prediction parameters are shown as input data. The measured value includes, for example, the above-mentioned position information, facility information, time information, weather information, generated power information, and the like. Further, the prediction parameters include position information, time information, weather information, facility information, and the like.

  When the data given to each node of the input layer is input to the first hidden layer and given, the output of the hidden layer is calculated using the weight and the activation function, and the calculated value is passed to the next hidden layer. It is given to the subsequent layers (lower layers) one after another until the output of the output layer is obtained in the same manner. Note that all the weights that connect the nodes are calculated by the learning algorithm described later.

  The output layer of the neural network model unit 543 has a function as a prediction output data generation unit and generates prediction output data. The predicted output data is vector format data having a component having the same size as the number of nodes in the output layer (size of the output layer).

  The output layer of the neural network model unit 543 generates predicted output data for each of a plurality of classes (also referred to as divisions) of the generated power range. When the number of nodes in the output layer of the neural network model unit 543 is N, the range of generated power is divided into N sections. For example, when the power of 50 kW is divided into 50 sections, the generated power from 0 kW to 50 kW can be treated as output data in units of 1 kW. Thereby, the generated power can be predicted in units of classes (classes). The range of generated power is not limited to 50 kW. Also, the number of classes is not limited to 50. In the example of FIG. 8, the output value for each class is represented by y_i (i = 1 to N). The output value of the class is also called the output strength and can be interpreted as the probability of being classified into each class.

  The predicted generated power calculation unit 544 calculates predicted generated power at the prediction target point and the prediction target time based on the prediction output data output by the output layer of the neural network model unit 543. For example, the number of output layers is 1, 2, ... N, and the predicted output data output by the output layer is y1, y2 ,. The predicted generated power calculation unit 544 calculates the predicted generated power based on the predicted output data y1, y2, ... Yk ,. For example, when yk = 1 and the other predicted output data is 0, the predicted generated power calculation unit 544 calculates the predicted generated power based on yk.

  As described above, the time-series data of the measured power generation and the time-series data of the measured meteorological information at a plurality of points are used as the prediction input data of the neural network model unit 543. It is possible to accurately predict the generated power in time.

  Next, the operation of the power generation prediction device 50 of this embodiment will be described. First, the learning process and the updating process of the neural network model unit 543 will be described, and then the prediction process of the generated power by the power generation prediction device 50 will be described.

  FIG. 9 is a flowchart showing an example of a procedure of learning processing and updating processing of the neural network model unit 543 by the power generation prediction device 50 of the present embodiment. Note that, in the following, for convenience of explanation, the main body of processing will be described as the processing unit 54. The processing unit 54 reads the generated power information and weather information at each power plant 10 from the database 53 (S11). In addition, in the following description, the generated power information and the weather information read from the database 53 are also referred to as actual measurement data.

  FIG. 10 is an explanatory diagram showing an example of generated power information and weather information at each power station 10. FIG. 10 shows generated power information and weather information at each power plant 10 read by the processing unit 54 in step S11. In the example of FIG. 10, each of the five power stations 10 designated by the symbols s1 to s5 has one measurement unit, and 120 pieces of generated power information from time t1 to t120 are recorded. The time information is uniquely determined. That is, the time ti is a time that has elapsed by unit time × i with the time t0 as a base point. For example, if the unit time is minutes and the base time t0 is 12:00 on May 10, 2016, the time t1 is 12:01 on May 10, 2016, and the time t120 is May 10, 2016. It will be 14:00. In the example of FIG. 10, 120 pieces of time information from time t1 to t120 are shown, but the number of time information may be larger. Assuming that the unit time is minutes, one day is 1440 minutes, so that there are 1440 pieces of time information corresponding to one day. Also, the time information corresponding to one year (365 days) is 525600. In addition, since the weather changes gently, the same weather (for example, “clear” or “cloudy”) may occur over a plurality of consecutive times. Also, the records at each power plant 10 may be missing. If it is missing, an alternative value indicating that it is missing may be given.

  The processing unit 54 (parameter determination unit 542) determines the prediction parameter for learning (S12) and specifies the actually measured weather information at the prediction target position and the prediction target time for learning (S13).

  FIG. 11 is an explanatory diagram showing how a prediction parameter for learning is determined. In the example of FIG. 11, the prediction target time t110, the prediction target position l2, and the prediction target facility information e1 are determined as the prediction parameters for learning. Then, the weather information w2-110 at the prediction target time t110 and the prediction target position 12 is specified from the actual measurement data read in step S11.

  The processing unit 54 identifies the actual measurement data of the power plant 10 under the condition close to the prediction parameter for learning (S14).

  FIG. 12 is an explanatory diagram showing an example of the actual measurement data of the power plant under the condition close to the learning prediction parameter. As shown in FIG. 12, since the prediction target time of the learning prediction parameter is t110, the prediction target position is l2, and the prediction target facility information is e1, the actual measurement data of the power plant s2 is specified as a power plant with similar conditions. It

  The processing unit 54 (learning data generation unit 546) generates learning input data based on the prediction target time for learning (S15). The learning data generation unit 546 transforms the generated power information and the weather information for each of the power plant 10 and the measurement unit 14 into a time-series one-dimensional vector as a preprocess for generating learning input data.

  FIG. 13 is an explanatory diagram showing an example of a time-series one-dimensional vector of generated power information and weather information for each power plant 10 and measurement unit 14. In the example shown in FIG. 13, each power plant is transformed into a one-dimensional vector having power plant information, measurement unit information, generated power information from time t1 to t120, and weather information from time t1 to t120 as components. .

  FIG. 14 is an explanatory diagram showing an example of a method for generating learning input data by the power generation prediction device 50 according to the present embodiment. As input data for learning, actual measurement data under conditions close to the prediction parameters for learning are prepared. In the process of predicting the generated power, prediction is performed for future time, so the time information of the input data for learning is limited to the time information past the prediction target time. For example, as shown in FIG. 14, the measured data (power generation) from the unit time 80 to 50 steps before the prediction target time t110 (see FIG. 11) (from the time t30 to the time t60 when the prediction target time is t110) Power information and weather information) are prepared. That is, in order to predict the power generation information of the prediction target position 12 and the prediction target facility information e1 at time t110, the actual measurement data of each power plant from time t30 to time t60 is used. The period from time t30 to time t60 is an example, and is not limited to the above example. Further, in the above example, the last time of the period (time t30 to time t60) is set to time t60 with respect to the prediction target time t110, but the time t60 may be predicted at what time. It can be set appropriately depending on whether to perform.

  FIG. 15 is an explanatory diagram showing an example of the learning input data generated by the power generation prediction device 50 of the present embodiment. As illustrated in FIG. 15, the learning data generation unit 546 expands the actually measured data and the prediction parameter into a one-dimensional vector to generate input data for learning.

  The processing unit 54 (learning data generation unit 546) generates learning output data based on the actual measurement data identified in step S14 (S16).

  FIG. 16 is an explanatory diagram showing an example of learning output data generated by the power generation prediction device 50 according to the present embodiment. As illustrated in FIG. 16, the learning data generation unit 546 adopts the prediction target time t110, the prediction target position l2, and the generated power information of the power plant s2 close to the prediction target facility information e1 as learning data.

  FIG. 17 is an explanatory diagram showing the relationship of learning data used for learning and updating of the neural network model unit 543 by the power generation prediction device 50 of this embodiment. In FIG. 17, the horizontal axis represents time, and the vertical axis represents a set of actually measured data. In the figure, the curves indicated by reference numerals s1 to s5 represent a collection of measured data of each of the five power stations 10 (that is, the measured data such as generated power information and meteorological information are collectively represented by one curve for convenience. ). In addition, in FIG. 17, a thin line curve indicates existing measurement data (data that is actually measured and exists), and a thick line curve indicates data used for learning among the existing measurement data. It should be noted that the curve shown in FIG. 17 simply shows how the actual measurement data changes with the passage of time, and does not show the actual change of the actual measurement data.

  The forecast of generated power is made at a future time. Therefore, first, the prediction target time t110 and the prediction target position 12 are determined from the past measured data collected in advance. Then, the actual measurement data from time t30 to time t60, which is further past the prediction target time t110, is adopted as the learning input data. This actual measurement data includes power generation power information and weather information from time t30 to time t60. Further, when predicting the generated power, the forecast of the weather information at the forecast time (future weather information) is also used as the input data, so that the weather information (measured value) at the forecast target time t110 and the forecast target position l2 is also obtained. Used as input data for learning. Furthermore, since the generated power information at the prediction target time t110 and the predicted target position 12 corresponds to the predicted value when the generated power is predicted, the generated power information at the prediction target time t110 and the predicted target position 12 is used for learning. Used as output data.

  When repeating the learning process and the updating process, for example, the following two processes can be considered. In the case of predicting the generated power of the day based on the actual measurement data of the previous day, as the first process, the actual measurement data from time t30 to time t60 is not changed and the prediction target time t110 is incremented by a unit time in the future. The process can be repeated by changing the direction. In addition, when predicting the generated power in the near future of the day (for example, after 30 minutes, 1 hour, 2 hours, etc.) based on the actual measurement data of the day, as the second process, the prediction target time is set. The processing can be repeated by changing t110 by the unit time in the future direction and changing the time (t30, t60) of the actual measurement data by the unit time in the future direction.

  The processing unit 54 (learning processing unit 545) learns and updates the neural network model unit 543 (S17). The learning algorithm will be described below.

  FIG. 18 is an explanatory diagram showing an example of the class of output data of the neural network model unit 543 of this embodiment. As shown in FIG. 18, it is divided into classes (i) from 1 to 50, class 1 has a power range of 0 kW or more and less than 1 kW, class 2 has a power range of 1 kW or more and less than 2 kW, and so on. The power range of 50 is 49 kW or more and less than 50 kW. Further, the median value (m_i) is determined as the representative value of each class. For example, in class 1, the median is 0.5 kW. Thereby, the generated power from 0 kW to 50 kW can be handled as output data in units of 1 kW. The power range, the number of classes, and the like can be appropriately changed according to the power generated by the power plant 10.

  FIG. 19 is an explanatory diagram showing an example of output values of the output data of the neural network model unit 543 of this embodiment. As shown in FIG. 19, the output value of the output data of the neural network model unit 543 is represented by y_i (i = 1 ... N) for each class (i).

  FIG. 20 is an explanatory diagram showing an example of a distribution of output values expected as output data of the neural network model unit 543 of this embodiment. The expected result for output data is when only the class to which the learning data belongs is output. Therefore, when the class to which the learning data belongs is j, only the expected output value y_j is set to 1 and the output values of the other classes are set to 0. For example, as shown in FIG. 20, when the learning output data is 1.8 kW, the learning output data belongs to class 2, so the output value z_2 of class 2 is set to 1 and the other output values are set to 0. And

  The learning processing unit 545 calculates an error by an error function having the output value of the neural network model unit 543 and the expected output value as variables, and the neural network model unit 545 uses the error back propagation method to minimize the error. The weight of the connection between the nodes 543 is optimized.

  As a specific example, a case where a softmax function is used as the activation function of the output layer and a cross entropy function is used as the error function will be described below. The output value y_i of the neural network model unit 543 is obtained by the softmax function shown in Expression (1).

  FIG. 21 is an explanatory diagram schematically showing an example of the hierarchical structure of the output layer for calculating the softmax function. With the softmax function, the output strength of each class in the output layer can be interpreted as the probability of being classified into each class. This is because the softmax function normalizes the output strength of each class in the output layer to a range of 0 to 1, and the sum of the output strengths of the classes becomes 1.

  The probability p (z | y) that the expected output value is obtained can be expressed by Expression (2). In addition, in Formula (2), code | symbol ^ shows a power operation. As the error function, the probability is maximized and the error is minimized by using the cross entropy function E (y, z) obtained by multiplying the logarithmic value of the probability by -1 as shown in Expression (3). That is mathematically equivalent. ^

  The processing unit 54 determines whether or not to end the processing (S18), and when the processing is not to be ended (NO in S18), the prediction parameter for learning is changed (S19), and the processing from step S13 onward is continued, When the process is to be ended (YES in S18), the process is ended.

  In this way, the prediction parameter for learning is changed, the generation of learning data, and the learning and updating of the neural network model unit 543 are repeated. It is possible to use a mini-batch method in which learning data corresponding to a plurality of prediction parameters for learning are created in advance and learning and updating of the neural network model unit 543 are collectively performed.

  As described above, according to this embodiment, the learning parameters including the time-series data of the measured power generation and the time-series data of the measured meteorological information at the plurality of points, the learning target point, and the learning target time are set to the neural network. Since the neural network model unit 543 is learned and updated using the input data for learning of the model unit 543, it is possible to accurately predict the generated power at any prediction target point and prediction target time.

  Further, the learning data generation unit 546 generates learning input data including time-series actually-measured generated power of each power station 10 and position information of each power station 10 in a predetermined period before the learning target time. For example, assuming that the learning target time is t110, the predetermined period before the learning target time t110 can be a period from 80 steps to 50 steps before a unit time (for example, minutes) like t30 to t60. . That is, the learning input data includes time-series actually-measured generated power of each power plant 10 from time t30 to time t60. It should be noted that if the length of the period is increased, the number of input data for learning increases, so that the prediction accuracy of the generated power increases.

  With the above-described configuration, the time-series data of the actually-generated power at each power plant 10 in the predetermined period is used to accurately generate the power at the prediction target point in the prediction target time after the predetermined period. The neural network model unit 543 for prediction can be learned and updated.

  Further, the learning data generation unit 546 generates learning input data including actually measured weather information for a plurality of points in a predetermined period. For example, assuming that the learning target time is t110, the predetermined period before the learning target time t110 can be a period from 80 steps to 50 steps before a unit time (for example, minutes) like t30 to t60. . That is, the input data for learning includes meteorological information actually measured from time t30 to time t60. It should be noted that if the length of the period is increased, the number of input data for learning increases, so that the prediction accuracy of the generated power increases.

  With the above configuration, the time-series data of the meteorological information actually measured at a plurality of points in a predetermined period is used to accurately predict the generated power at the prediction target point in the prediction target time after the predetermined period. It is possible to perform learning and updating of the neural network model unit 543 for doing so.

  In addition, the learning data generation unit 546 generates learning input data that includes actual measurement weather information at the learning target point and the learning target time. By including the meteorological information measured at the learning target point and the learning target time in the learning input data, the neural network model unit 543 for accurately predicting the generated power can be learned and updated.

  In addition, the learning data generation unit 546 generates learning output data including the actually-measured generated power of the power plant 10 installed near the learning target point. When the learning target point is the point where the power plant 10 is installed, the learning target point can be set as the relevant point. Further, when the learning target point is apart from any of the plurality of points where the power plant 10 is installed, the point closest to the learning target point among the plurality of points can be set as the learning target point. By including the actually generated power of the power plant 10 installed near the learning target point in the output data for learning, the neural network model unit 543 for accurately predicting the generated power can be learned and updated. it can.

  Next, the prediction process of the generated power by the power generation prediction device 50 will be described.

  FIG. 22 is a flowchart showing an example of the procedure of the prediction process of the generated power by the power generation prediction device 50 of this embodiment. Note that, in the following, for convenience of explanation, the main body of processing will be described as the processing unit 54. The processing unit 54 reads the generated power information and weather information at each power plant 10 from the database 53 (S31).

  FIG. 23 is an explanatory diagram showing an example of generated power information and weather information at each power station 10. FIG. 23 shows generated power information and weather information at each power plant 10 read by the processing unit 54 in step S31. In the example of FIG. 23, each of the five power stations 10 designated by the symbols s1 to s5 has one measurement unit, and 120 pieces of generated power information from time t1001 to t1120 are recorded. Similar to the learning stage, the time ti is a time elapsed from the time t0 as a base point by unit time × i. Note that in FIG. 23, the initial time of the time information is set to t1001 in order to indicate that time has elapsed from the learning stage.

  The processing unit 54 determines a prediction parameter (S32) and acquires a forecast of weather information at the prediction target position and the prediction target time (S33).

  FIG. 24 is an explanatory diagram showing how the prediction parameters are determined. In the example of FIG. 24, the prediction target time t1140, the prediction target position 12 and the prediction target facility information e1 are determined as the prediction parameters. Then, the weather information w2-1140 (forecast value) at the prediction target time t1140 and the prediction target position l2 is acquired.

  The processing unit 54 (input data generation unit 541) generates prediction input data based on the prediction target time (S34). The input data generation unit 541 transforms the generated power information and the weather information for each of the power plant 10 and the measurement unit 14 into a time-series one-dimensional vector as a preprocess for generating the input data for prediction.

  FIG. 25 is an explanatory diagram showing an example of a time-series one-dimensional vector of generated power information and weather information for each power station 10 and measurement unit 14. In the example shown in FIG. 25, each power plant is transformed into a one-dimensional vector having power plant information, measurement unit information, generated power information from time t1001 to t1120, and weather information from time t1001 to t1120 as components. .

  FIG. 26 is an explanatory diagram showing an example of a method of generating prediction input data by the power generation prediction device 50 of this embodiment. As the input data for prediction, actual measurement data under conditions close to the prediction parameters are prepared. For example, as shown in FIG. 26, the measured data (power generation from the unit time 80 to 50 steps before the prediction target time t1140 (see FIG. 24) (from the time t1060 to the time t1090 when the prediction target time is t1140) Power information and weather information) are prepared. That is, in order to predict the power generation information of the prediction target position 12 and the prediction target facility information e1 at time t1140, the actual measurement data of each power plant 10 at time t1060 to time t1090 is used. The period from time t1060 to time t1090 is an example, and is not limited to the above example. Further, in the above example, the last time of the period (from time t1060 to time t1090) is set to time t1090 with respect to the prediction target time t1140. It can be set appropriately depending on whether to perform.

  FIG. 27 is an explanatory diagram showing an example of the prediction input data generated by the power generation prediction device 50 of the present embodiment. As shown in FIG. 27, the input data generation unit 541 expands the actually measured data and the prediction parameter into a one-dimensional vector to generate prediction input data.

  FIG. 28 is an explanatory diagram showing the relationship of the prediction data used for predicting the generated power by the power generation prediction device 50 of the present embodiment. In FIG. 28, the horizontal axis represents time and the vertical axis represents a set of actually measured data. In the figure, the curves indicated by reference numerals s1 to s5 represent a collection of measured data of each of the five power stations 10 (that is, the measured data such as generated power information and meteorological information are collectively represented by one curve for convenience. ). Further, in FIG. 28, a thin line curve indicates existing measurement data (data that is actually measured and exists), and a thick curve indicates data used for learning among the existing measurement data. Note that the curve shown in FIG. 28 merely shows how the actually measured data changes with the passage of time, and does not show the actual change of actually measured data.

  The current time is set to t1120. There is a time lag associated with the process of collecting actual measurement data. In order to represent the time lag, in FIG. 28, the latest time (the right end of the curve indicated by the thin line) of the actual measurement data of each power plant 10 varies and the current time t1120 is not reached. At time t1090, the actual measurement data of each power plant 10 exists. The forecast of generated power is made at a future time. First, the prediction target time t1140 and the prediction target position 12 are determined. Then, the actual measurement data from the time t1060 to the time t1090, which is earlier than the prediction target time t1140, is adopted as the input data for prediction. This actual measurement data includes power generation power information and weather information from time t1060 to time t1090. Further, the forecast of the weather information at the prediction target time t1140 and the prediction target position l2 is also adopted as the prediction input data. By supplying the prediction input data to the input layer of the learned neural network model unit 543, the predicted generated power at the prediction target time t1140 and the prediction target position 12 can be obtained.

  When predicting the generated power, for example, the following two types of processing can be considered. In the case of predicting the generated power of the day based on the actual measurement data of the previous day, as the first process, the actual measurement data from time t1060 to time t1090 is not changed, and the prediction target time t1140 is incremented by a unit time in the future. The process can be repeated by changing the direction. In addition, when predicting the generated power in the near future of the day (for example, after 30 minutes, 1 hour, 2 hours, etc.) based on the actual measurement data of the day, as the second process, the prediction target time is set. The process can be repeated by changing t1140 by the unit of time in the future direction and changing the time (t1060 and t1090) of the actual measurement data by the unit of time in the future direction.

  The processing unit 54 generates predicted output data (S35).

  FIG. 29 is an explanatory diagram showing an example of output values of the output data of the neural network model unit 543 of this embodiment. As shown in FIG. 29, the output value of the output data of the neural network model unit 543 is represented by y_i (i = 1 ... N) for each class (i).

  The processing unit 54 (predicted generated power calculation unit 544) calculates predicted generated power (S36). The predicted generated power W can be calculated by the equation (4).

  The predicted generated power calculation unit 544 calculates the predicted generated power based on the median value (representative value) determined for each of a plurality of classes (classes) and the predicted output data generated for each class. For example, if the predicted output data y10 of the class 10 (range: 9 kW or more and less than 10 kW, median: 9.5 kW) is 1 and all other predicted output data are 0, the predicted generated power W is 9.5 kW. Becomes Thereby, it is possible to calculate the generated electric power according to the strength of the output data output from the output layer of the neural network model unit 543.

  Note that the method for calculating the predicted generated power is not limited to the method according to the above equation (4). For example, in the output values y_i (i = 1 ... N) for each class (i), the representative value (for example, the median value) of the class having the maximum output value can be calculated as the predicted generated power.

  Further, in the output value y_i (i = 1 ... N) for each class (i), when there are two or more classes having the maximum output value, the one having the largest class value (the highest power) The representative value (eg, median) of the class may be calculated as the predicted generated power, or the representative value (eg, median) of the class with the smaller class value (smaller power) may be used as the predicted generated power. It may be calculated. Whether to adopt the larger class value or the smaller class value may be appropriately determined according to the purpose of predicting the generated power.

  The processing unit 54 determines whether or not to end the processing (S37), and when the processing is not to be ended (NO in S37), the processing of step S32 and subsequent steps is continued, and when the processing is ended (YES in S37), the processing To finish.

  As described above, according to the present embodiment, the input data generation unit 541 determines the time-series actually-generated power and the position of each power plant 10 for each power plant 10 in the predetermined period before the prediction target time. Generate prediction input data containing information. For example, assuming that the prediction target time is t1140, the predetermined period before the prediction target time t1140 can be a period from 80 steps before the unit time (for example, minutes) to 50 steps before, such as t1060 to t1090. . That is, the input data for prediction includes the time-series actually-measured generated power of each power plant 10 at times t1060 to t1090. It should be noted that if the length of the period is increased, the number of pieces of prediction input data increases, so that the prediction accuracy of the generated power increases.

  With the above-described configuration, the time-series data of the actually-generated power at each power plant 10 in the predetermined period is used to accurately generate the power at the prediction target point in the prediction target time after the predetermined period. Can be predicted.

  In addition, the input data generation unit 541 generates prediction input data including actually measured weather information for a plurality of points in a predetermined period. For example, assuming that the prediction target time is t1140, the predetermined period before the prediction target time t1140 can be a period from 80 steps before the unit time (for example, minutes) to 50 steps before, such as t1060 to t1090. . That is, the input data for prediction includes meteorological information measured at times t1060 to t1090. It should be noted that if the length of the period is increased, the number of pieces of prediction input data increases, so that the prediction accuracy of the generated power increases.

  With the above configuration, the time-series data of the meteorological information actually measured at a plurality of points in a predetermined period is used to accurately predict the generated power at the prediction target point in the prediction target time after the predetermined period. can do.

  The input data generation unit 541 also generates prediction input data including weather information for prediction at the prediction target point and the prediction target time. By including the weather information of the prediction at the prediction target point and the prediction target time in the prediction input data, it is possible to accurately predict the generated power.

As described above, the weather information includes at least one of weather, change of weather, cloud cover, and solar radiation. The weather is, for example, “fine”, “cloudy”, “finely cloudy”, “temporary clear”, or the like. The change of the weather is, for example, "after cloudy weather" or "after cloudy weather". Cloud amount is the ratio of clouds to the entire sky. The solar radiation is, for example, the solar radiation intensity per unit area (kW / m ² ), the amount of solar radiation per unit time / unit area (kWh / m ² ), and the like. The weather, transition of the weather, cloud amount and solar radiation are also collectively referred to as weather. As a result, it is possible to consider factors that affect the generated power of the photovoltaic power generation.

  The power generation information also includes the rated power generation capacity of the power plant. The rated power generation capacity is the rated value of the power generation output. As a result, the generated power can be accurately predicted regardless of the rated value of the power output of the power plant.

  The power generation prediction device 50 of the present embodiment can also be realized using a general-purpose computer including a CPU (processor), a GPU (graphics processing unit), a RAM, and the like. That is, as shown in FIG. 9 and FIG. 22, a computer program that defines the procedure of each process is loaded into the RAM provided in the computer, and the CPU (processor) executes the computer program to predict power generation on the computer. The device 50 can be realized. Further, the power generation prediction device 50 can also be configured by one or a plurality of servers.

  In the above-described embodiment, the unit time is described as minutes, but the unit time is not limited to minutes. For example, the unit time may be 5 minutes, 10 minutes, 30 minutes, 1 hour, or the like.

  Next, a method of providing a predetermined service to a power plant operator or an O & M (Operation & Maintenance) operator using the power generation prediction device 50 of the present embodiment will be described.

  FIG. 30 is a block diagram showing an example of the configuration of the server 100 of this embodiment. As shown in FIG. 30, the server 100 is connected to a power plant operator 200 and an O & M operator 300 (specifically, a computer owned by each operator) via a communication line 32 such as the Internet. Further, the server 100 is connected to the power generation prediction device 50.

  The server 100 includes a control unit 101, a communication unit 102, a storage unit 103, a first information generation unit 104, a second information generation unit 105, and the like.

  The communication unit 102 has a function as a power generation data acquisition unit, and acquires power generation data on power generation equipment from a solar power generation business operator (power station business operator 200 or O & M business operator 300). The power generation data can be acquired from the computer of the business operator. The power generation data includes, for example, the power generation information illustrated in FIG. The power generation data acquired by the communication unit 102 is stored in the storage unit 103. The communication unit 102 also has a function as an output unit that outputs first information and second information described below.

  The first information generation unit 104 generates the first information regarding solar power generation based on the power generation data acquired by the communication unit 102, that is, the power generation data stored in the storage unit 103.

  FIG. 31: is explanatory drawing which shows the 1st example of the 1st information regarding photovoltaic power generation. As shown in FIG. 31, the first information is information relating to a calendar function (for example, a maintenance inspection schedule, a weather forecast, etc. are described).

  FIG. 32: is explanatory drawing which shows the 2nd example of the 1st information regarding solar power generation. As shown in FIG. 32, the first information is information regarding the maintenance and inspection management function (output of inspection table, storage, management, etc.).

  FIG. 33 is an explanatory diagram showing a third example of the first information regarding solar power generation. As shown in FIG. 33, the first information is information regarding the management function of the power generation amount or information for providing the power generation record.

  The second information generation unit 105 generates second information including power generation prediction data by the power generation prediction device 50 according to the present embodiment.

  Next, the operation of the server 100 of this embodiment will be described. FIG. 34 is a flowchart showing an example of the processing procedure of the server 100 of this embodiment. Hereinafter, for convenience, the main body of processing will be described as the control unit 101. The control unit 101 acquires power generation data regarding power generation equipment from the power plant operator 200 and / or the O & M operator 300 (S51).

  The control unit 101 outputs the first information generated by the first information generation unit 104 to the computer of the power plant operator 200 and / or the O & M operator 300 under a predetermined first price condition (S52). The first price condition may be free, for example.

  The control unit 101 determines whether there is a request for an optional function from the power plant operator 200 or the O & M operator 300 (S53). The request for the optional function is, for example, a request for providing information under a second price condition that is higher than the first price condition. The second price condition can be charged, for example.

  When there is a request for the optional function (YES in S53), the control unit 101 outputs the power generation prediction data to the computer of the power plant operator 200 or the O & M operator 300 that has made the request under the second price condition ( S54), and the process ends. When there is no request for the optional function (NO in S53), the control unit 101 ends the process without performing the process of step S54.

  With the above configuration, it is possible to obtain the power generation data on the power generation equipment from the business while providing the basic service on the solar power generation to the business related to the solar power generation free of charge. Can be collected as so-called big data. The collected big data (power generation data) can be used to learn and update the neural network of the power generation prediction device 50, and the power generation prediction accuracy of the power generation prediction device 50 can be improved. Further, for a business operator who wants to use the pay option, it is possible to provide a service related to power generation prediction whose accuracy is improved by machine learning.

  In the present specification, the character string is not limited to a simple character string. The significance of the character string will be described below by taking the facility information as an example. FIG. 35 is an explanatory diagram showing an example of the character string of this embodiment. In FIG. 35, the upper diagram shows an example of the facility information, and the lower diagram shows an example of a character string related to the facility information of the upper diagram. As shown in the upper diagram, the facility information e1 is such that the type of the photovoltaic panel 11 is a1, the manufacturer is b1, the rated capacity is 100, the number of units is 6, and the total rated capacity is 600. Is a3, the manufacturer is b1, the rated capacity is 500, the number of units is 1, and the total rated capacity is 500. Equipment information e2 and e3 are also shown in the figure. In this case, the equipment information e1 can be represented by a character string as shown in the lower diagram. The same applies to the facility information e2 and e3. That is, as shown in FIG. 35, the character string may have a format such as a definition statement or a structural statement in the program code.

  Further, in the power generation prediction device 50 of the present embodiment, the information stored in the database in the form of character strings such as facility information and weather information is converted into numerical values according to a predetermined conversion rule, and the converted numerical values are converted into numerical values. It can be input to the input layer of the neural network model unit 543. The data input to the input layer of the neural network model unit 543 may be a character string. Next, conversion from a character string to a numerical value will be described.

  FIG. 36 is an explanatory diagram showing an example of conversion from a character string to a numerical value by the power generation prediction device 50 of this embodiment. In FIG. 36, the upper diagram is an example of a conversion table for converting a character string into a numerical value. As shown in the upper diagram of FIG. 36, for example, the models a1, a2, and a3 are converted into 1010, 1020, and 1030 (referred to as model values), respectively. Further, the manufacturers b1 and b2 are converted into 101 and 102 (referred to as manufacturer numbers), respectively. The lower diagram of FIG. 36 shows an example of equipment information converted into numerical values. The equipment information shown in the lower part of FIG. 36 is obtained by converting the equipment information in the character string illustrated in FIG. 35 into numerical values.

  In the present embodiment, for example, when the number of elements of the one-dimensional vector of the input data for learning illustrated in FIG. 14 and the one-dimensional vector of the input data for prediction illustrated in FIG. 26 is M, for example. In the examples of FIGS. 14 and 26, the number of elements as the facility information is one, but the facility information shown in FIGS. 14 and 26 is converted into a one-dimensional vector of 10 items as shown in FIG. 35, for example. Each element can be converted into a unitary vector having M + 9 elements. As a result, the number of input data increases, so that the prediction accuracy of the generated power can be improved.

  FIG. 37 is an explanatory diagram showing an example of numerical conversion according to the present embodiment. For information that can be divided into multiple items, any item can be extracted from it, and the value of the extracted item can be converted and integrated by a predetermined conversion formula to be the value of that information. . For example, as shown in the diagram on the left side of FIG. 37, when a conversion formula for adding the value obtained by dividing the model value of the inverter model by 10,000 as the integer part of the total inverter rated capacity of the facility information is used, The numerical value of can be converted into one numerical value as shown in the right figure. As shown in the diagram on the right side, since the equipment information values of the equipment information e1 and e2 are the same value, the equipment information e1 and e2 can be regarded as the same equipment.

10 power plant 20 power generation monitoring system 50 power generation prediction device 51 communication unit 52 registration unit 53 database 54 processing unit 541 input data generation unit 542 parameter determination unit 543 neural network model unit 544 predicted generation power calculation unit 545 learning processing unit 546 learning data generation Part 55 Communication part 100 Server 101 Control part 102 Communication part 103 Storage part 104 1st information generation part 105 2nd information generation part 200 Power generation company 300 O & M company

Claims (10)

A power generation prediction device that provides power generation prediction data for photovoltaic power generation,
A power generation information acquisition unit that acquires power generation information including time-series actually measured generated power related to power generation facilities installed at a plurality of points,
A meteorological information acquisition unit that acquires meteorological information related to the plurality of points,
A storage unit that stores the power generation information acquired by the power generation information acquisition unit and the weather information acquired by the weather information acquisition unit,
A prediction parameter determination unit that determines a prediction parameter including a prediction target point and a prediction target time,
A prediction input data generation unit that generates prediction input data to a predetermined neural network model, based on the power generation information and weather information stored in the storage unit, and the prediction parameter determined by the prediction parameter determination unit,
A prediction output data generation unit that generates prediction output data based on the prediction input data generated by the prediction input data generation unit and the neural network model;
And a predicted generated power calculation unit that calculates predicted generated power at the prediction target point and the prediction target time based on the predicted output data generated by the predicted output data generation unit ,
Furthermore, the prediction input data generation unit,
Generating prediction input data including time-series actually-generated power and position information of each power generation facility relating to each power generation facility in a predetermined period before the prediction target time,
Generating prediction input data including actual weather information in the predetermined period related to the plurality of points,
Generating prediction apparatus characterized that you generate prediction input data including weather information for prediction in the prediction target point and the prediction target time. The predicted output data generation unit,
Generating prediction apparatus according to claim 1, characterized in that generating the predicted output data range of generated power for each said section, which is divided into a plurality of sections. The predicted power generation calculation unit,
The power generation prediction apparatus according to claim 2, wherein the predicted power generation power is calculated based on the representative value determined for each of the plurality of sections and the predicted output data generated for each of the sections. A learning parameter determination unit that determines a learning parameter including a learning target point and a learning target time,
A learning input data generation unit that generates learning input data of the neural network model based on the power generation information and the weather information stored in the storage unit and the learning parameter determined by the learning parameter determination unit,
A learning output data generation unit that generates learning output data of the neural network model based on the learning parameter determined by the learning parameter determination unit;
A learning processing unit that performs learning processing of the neural network model based on the learning input data generated by the learning input data generation unit and the learning output data generated by the learning output data generation unit ;
Further, the learning input data generation unit is
Generate learning input data including time-series actually-generated power and position information of each power generation facility related to each power generation facility in a predetermined period before the learning target time,
Generate learning input data including actual weather information in the predetermined period of time related to the plurality of points,
Generating prediction apparatus according to any one of claims 1 to 3, characterized that you generate input data for learning including weather information measured in the learning target point and the learning target time. The learning output data generation unit,
The power generation prediction device according to claim 4 , wherein the power generation prediction device generates learning output data including actually generated power generated by a power generation facility installed at a point near the learning target point. The weather information is
The power generation prediction apparatus according to any one of claims 1 to 5 , wherein the power generation prediction apparatus includes at least one of weather, weather change, cloud amount, and solar radiation. The power generation information is
The power generation prediction apparatus according to any one of claims 1 to 6 , comprising a rated power generation capacity of the power generation facility. A power generation data acquisition unit that acquires power generation data related to power generation equipment from a solar power generation company,
An output unit that outputs the first information related to solar power generation to the computer of the business operator under a predetermined first price condition based on the power generation data acquired by the power generation data acquisition unit;
The power generation prediction data by the power generation prediction device according to any one of claims 1 to 7 , when an information provision request is received from the business operator under a second price condition that is higher than the first price condition. An output unit that outputs the second information including the second information to the computer of the business under the second price condition. A computer program for causing a computer to provide power generation prediction data for photovoltaic power generation,
Computer,
A power generation information acquisition unit that acquires power generation information including time-series actually measured generated power related to power generation facilities installed at a plurality of points,
A meteorological information acquisition unit that acquires meteorological information related to the plurality of points,
A storage unit that stores the acquired power generation information and weather information,
A prediction parameter determination unit that determines a prediction parameter including a prediction target point and a prediction target time,
A prediction input data generation unit that generates prediction input data to a predetermined neural network model based on the stored power generation information and weather information and the determined prediction parameter,
A prediction output data generation unit that generates prediction output data based on the generated prediction input data and the neural network model;
Based on the generated predicted output data, function as a predicted generated power calculation unit that calculates predicted generated power at the prediction target point and the prediction target time ,
Furthermore, the prediction input data generation unit,
Generating prediction input data including time-series actually-generated power and position information of each power generation facility relating to each power generation facility in a predetermined period before the prediction target time,
Generating prediction input data including actual weather information in the predetermined period related to the plurality of points,
Computer program characterized that you generate prediction input data including weather information for prediction in the prediction target point and the prediction target time. A predicted power generation method of providing power generating predicted data of the solar power,
The power generation information acquisition unit acquires power generation information including time-series measured power generation related to power generation equipment installed at a plurality of points,
The meteorological information acquisition unit acquires meteorological information related to the plurality of points,
Store the acquired power generation information and weather information in the storage unit,
The prediction parameter determining unit determines the prediction parameter including the prediction target point and the prediction target time,
Based on the power generation information and weather information stored in the storage unit and the determined prediction parameters, the prediction input data generation unit generates prediction input data to a predetermined neural network model,
Based on the generated prediction input data and the neural network model, the prediction output data is generated by the prediction output data generation unit,
The predicted generated power calculation unit calculates the predicted generated power at the prediction target point and the prediction target time based on the generated predicted output data ,
Furthermore, the prediction input data generation unit,
Generating prediction input data including time-series actually-generated power and position information of each power generation facility relating to each power generation facility in a predetermined period before the prediction target time,
Generating prediction input data including actual weather information in the predetermined period related to the plurality of points,
Predicted power generation method characterized by generating a predictive input data including weather information for prediction in the prediction target point and the prediction target time.

Images