Device and method for predicting power generation amount of virtual power plant based on machine learning operations

Abstract

A device for predicting a power generation amount of a virtual power plant (VPP) based on machine learning operations (MLOps) according to an embodiment of the present disclosure may comprise a processor for: executing a data collection pipeline configured to collect actual measurement data including power generation amount data and weather data on the basis of a first trigger signal; executing a data collection pipeline configured to collect forecast data including weather data on the basis of a second trigger signal; generating training data for training a power generation amount prediction model by pre-processing the collected actual measurement data; and generating input data of the power generation amount prediction model by pre-processing the collected forecast data.

Description

A device and method for predicting power generation in a virtual power plant based on machine learning operations.

[0001] The present invention relates to a power generation prediction device and a power generation prediction method of a virtual power plant based on machine learning operation.

[0002] Recently, the penetration of renewable energy-based distributed energy resources (DER) has been increasing worldwide. With the expansion of renewable energy, efforts are increasingly being made to build virtual power plants (VPPs), which integrate and operate distributed energy resources as a single power plant using information and communications technology (ICT) to address issues such as output uncertainty of renewable energy resources and the stability of power distribution systems.

[0003] Meanwhile, in order to efficiently operate the power grid linked to distributed energy and provide stable power supply services, a renewable energy power generation forecasting system was newly established that forecasts the power generation of renewable energy, such as solar and wind power plants with a capacity of 20 MW or more, and submits them to the power exchange, and sets the unit price based on the error rate with the actual power generation.

[0004] Accordingly, it is becoming increasingly important to improve the accuracy of power generation predictions along with the construction of virtual power plants.

[0005] The purpose of the present invention is to provide a power generation prediction device and a power generation prediction method of a virtual power plant based on machine learning operation that automatically collects data for power generation prediction more easily and accurately.

[0006] The purpose of the present invention is to provide a power generation prediction device and a power generation prediction method of a virtual power plant based on machine learning operation that more easily and accurately identifies the occurrence of missing data or prediction delay.

[0007] The purpose of the present invention is to provide a power generation prediction device and a power generation prediction method for a virtual power plant based on machine learning operation that distributes a power generation prediction model with a low prediction error rate more easily and accurately.

[0008] In accordance with one embodiment of the present invention, a power generation prediction device of a Virtual Power Plant (VPP) based on Machine Learning Operations (MLOps) may include a processor configured to execute a data collection pipeline set to collect actual data including power generation data and weather data based on a first trigger signal, execute a data collection pipeline set to collect forecast data including weather data based on a second trigger signal, preprocess the collected actual data to generate learning data for learning a power generation prediction model, and preprocess the collected forecast data to generate input data for the power generation prediction model.

[0009] The above processor can replace missing data among the irradiance items of the actual data with 0, and replace missing data among the remaining items of the actual data with actual data collected in the previous cycle.

[0010] The above processor can select irradiance data to be included in the learning data among the irradiance data of actual data received through two or more APIs (Application Programming Interfaces) based on the power plant location information.

[0011] The above processor can supplement missing data of the forecast data by using forecast data received through two or more APIs (Application Programming Interfaces).

[0012] The above processor can identify data to be included in the input data by applying a voting algorithm to items collected repeatedly among the above forecast data.

[0013] The processor can match the data ranges of the items collected in duplicate, categorize the data with the matched ranges into two or more categories, and identify one of the categories corresponding to the items collected in duplicate as data to be included in the input data based on a majority vote.

[0014] The processor may identify at least one additional data among the power generation efficiency of each power plant, the trigonometric ratio of the extreme angle, the trigonometric ratio of the azimuth, and the dew point using the actual data and the forecast data, and include the additional data in the learning data and the input data.

[0015] The above first trigger signal and second trigger signal may include power plant location information, collection API information, and announcement time information.

[0016] A method for predicting power generation of a Virtual Power Plant (VPP) based on Machine Learning Operations (MLOps) performed by a power generation prediction device according to one embodiment of the present invention may include: a step of executing a data collection pipeline set to collect actual data including power generation data and weather data based on a first trigger signal; a step of executing a data collection pipeline set to collect forecast data including weather data based on a second trigger signal; a step of preprocessing the collected actual data to generate learning data for learning a power generation prediction model; and a step of preprocessing the collected forecast data to generate input data for the power generation prediction model.

[0017] The step of generating the above learning data may include a step of replacing missing data among the irradiance items of the actual data with 0; and a step of replacing missing data among the remaining items of the actual data with actual data collected in the previous period.

[0018] The step of generating the above learning data may include a step of selecting the irradiance data to be included in the learning data from among the irradiance data of actual data received through two or more APIs (Application Programming Interfaces) based on the power plant location information.

[0019] The step of generating the above input data may include a step of supplementing missing data of the forecast data using forecast data received through two or more APIs (Application Programming Interfaces).

[0020] The step of generating the above input data may include a step of identifying data to be included in the input data by applying a voting algorithm to items collected in duplicate among the above forecast data.

[0021] The step of generating the above input data may include: a step of matching the data ranges of the items collected in duplicate; a step of categorizing the data with the matched ranges into two or more categories; and a step of identifying one of the categories corresponding to the items collected in duplicate as data to be included in the input data based on a majority vote.

[0022] The step of generating the above learning data and input data may include a step of identifying at least one additional data among the power generation efficiency of each power plant, the trigonometric ratio of the extreme angle, the trigonometric ratio of the azimuth, and the dew point using the actual data and the forecast data; and a step of including the additional data in the learning data and the input data.

[0023] According to one embodiment of the present invention, the data collection and preprocessing process required for learning a power generation prediction model can be automated.

[0024] According to one embodiment of the present invention, the data collection and preprocessing process required for power generation prediction using a model can be automated.

[0025] According to one embodiment of the present invention, the cornerstone of a new power business and a brokerage business for participation in the power market can be laid, and the foundation for a VPP solution service can be secured.

[0026] According to one embodiment of the present invention, a technology gap in machine learning operations (MLOps) can be filled by enabling a person in charge to quickly recognize when missing data or prediction delay occurs.

[0027] According to one embodiment of the present invention, in a system that must provide real-time prediction information, such as ultra-short-term predictions, a power brokerage business can be operated stably by taking follow-up measures when missing data or prediction delays occur.

[0028] According to one embodiment of the present invention, the process of generating and distributing a power generation prediction model can be automated.

[0029] According to one embodiment of the present invention, the prediction accuracy of a power generation prediction model can be upwardly standardized.

[0030] FIG. 1 is a schematic diagram illustrating a power generation prediction device of a virtual power plant based on machine learning operation according to one embodiment of the present invention.

[0031] FIG. 2 is a block diagram illustrating the configuration of a power generation prediction device according to one embodiment of the present invention.

[0032] FIG. 3 is a diagram illustrating an operation flow chart of a power generation prediction device according to a first embodiment of the present invention.

[0033] FIG. 4 is a diagram illustrating a list of data pipelines according to a second embodiment of the present invention.

[0034] FIG. 5 is a diagram illustrating a list of data pipelines according to a third embodiment of the present invention.

[0035] FIG. 6 is a diagram illustrating an operation flow chart of a power generation prediction device according to a fourth embodiment of the present invention.

[0036] FIG. 7 is a diagram illustrating an operation flow chart of a power generation prediction device according to a fifth embodiment of the present invention.

[0037] FIG. 8 is a diagram illustrating an operation flow chart of a power generation prediction device according to a sixth embodiment of the present invention.

[0038] FIG. 9 is a drawing illustrating the operation of a power generation prediction device according to one embodiment of the present invention.

[0039] FIG. 10 is a diagram illustrating the operation of a power generation prediction system according to one embodiment of the present invention.

[0040] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. The detailed description set forth below, together with the accompanying drawings, is intended to explain exemplary embodiments of the present invention and is not intended to represent the only embodiments in which the present invention may be practiced. In the drawings, portions irrelevant to the description may be omitted for clarity in describing the present invention, and the same reference numerals may be used throughout the specification for identical or similar components.

[0041] FIG. 1 is a schematic diagram illustrating a power generation prediction device based on machine learning operation of a virtual power plant according to one embodiment of the present invention.

[0042] A power generation prediction device (100) according to one embodiment of the present invention is a device that predicts the power generation of renewable energy resources using collected data, and can be implemented as a computer, a PLC (Programmable Logic Controller), a server, a smart phone, a tablet PC, a smart pad, a laptop, etc.

[0043] The power generation prediction device (100) can predict and manage the power generation of each resource through a virtual power plant constructed to operate different power systems that operate renewable energy-based distributed energy resources such as solar power generation, hydroelectric power generation, wind power generation, energy storage systems, and electric vehicles as a single power plant.

[0044] Power generation forecasting is required daily or hourly, and efforts are being made to improve prediction accuracy through repeated processes of data collection, data preprocessing, power generation forecasting model training, model deployment, and monitoring.

[0045] Meanwhile, the present invention introduces the concept of Machine Learning Operations (MLOps) to automate this process based on a virtual power plant. MLOps is a method for automating and managing the entire process, from development to deployment and operation of machine learning models. MLOps provides repeatable processes and enables continuous monitoring and improvement of model performance.

[0046] In the present invention, a method for automating data collection and preprocessing processes is proposed.

[0047] In the present invention, a method for generating notification information regarding data omission or prediction delay that occurs during the process of predicting power generation from data collection is proposed.

[0048] The present invention proposes a method for improving the prediction accuracy of a power generation prediction model by distributing a continuously updated model.

[0049] Hereinafter, the configuration and operation of a power generation prediction device (100) according to one embodiment of the present invention will be specifically described with reference to the drawings.

[0050] FIG. 2 is a block diagram illustrating the configuration of a power generation prediction device according to one embodiment of the present invention.

[0051] A power generation prediction device (100) according to one embodiment of the present invention includes an input unit (110), a communication unit (120), a display unit (130), a storage unit (140), and a processor (150).

[0052] The input unit (110) generates input data in response to user input of the power generation prediction device (100). For example, the user input may be user input required to build a data collection pipeline, such as setting a data collection period, setting items to be collected for each data, and setting a collection API (Application Programming Interface) for each data; user input setting a logic for preprocessing actual data and forecast data; user input setting a threshold for the time required for each step; user input setting a new model creation range and a new model verification range; and the like. In addition, the user input may be applied without limitation to automate the data collection and preprocessing process, generate notification information for data omissions and prediction delays that occur in the power generation prediction process from data collection, and increase the prediction accuracy of the power generation prediction model.

[0053] The input unit (110) includes at least one input means. The input unit (110) may include a keyboard, a key pad, a dome switch, a touch panel, a touch key, a mouse, a menu button, etc.

[0054] The communication unit (120) can perform communication with external devices such as a data provider server, API, power exchange server, operator terminal, business terminal, etc. to transmit and receive trigger signals, power generation data, weather data, learning data, input data, power plant location information, collection API information, announcement time information, additional data, power generation prediction model, time required for each step, threshold value for time required for each step, alarm information, new model creation range, new model verification range, calculation formula for prediction error rate, power generation prediction information, etc.

[0055] To this end, the communication unit (120) can perform wireless communication such as 5G (5th generation communication), LTE-A (long term evolution-advanced), LTE (long term evolution), Wi-Fi (wireless fidelity), Bluetooth, or wired communication such as LAN (local area network), WAN (Wide Area Network), and power line communication.

[0056] The display unit (130) displays display data according to the operation of the power generation prediction device (100). The display unit (130) may display, for example, a screen for executing a data collection pipeline, a screen for preprocessing collected data, a screen for displaying power generation prediction information output from a power generation prediction model, a screen for displaying alarm information, a screen for displaying prediction error rate information, and a screen for receiving user input.

[0057] The display unit (130) includes a liquid crystal display (LCD), a light emitting diode (LED) display, an organic light emitting diode (OLED) display, a micro electro mechanical systems (MEMS) display, and an electronic paper display. The display unit (130) may be implemented as a touch screen by being combined with the input unit (110).

[0058] The storage unit (140) stores the operation programs of the power generation prediction device (100). The storage unit (140) includes a non-volatile storage that can preserve data (information) regardless of whether power is supplied, and a volatile memory that loads data to be processed by the processor (150) and cannot preserve data if power is not supplied. The storage includes a flash memory, a hard-disc drive (HDD), a solid-state drive (SSD), a read-only memory (ROM), etc., and the memory includes a buffer, a random access memory (RAM), etc.

[0059] The storage unit (140) can store power generation data, weather data, learning data, input data, power plant location information, collection API information, announcement time information, additional data, power generation prediction model, time required for each step, threshold value for time required for each step, alarm information, new model creation range, new model verification range, calculation formula for prediction error rate, power generation prediction information, etc.

[0060] The storage unit (140) can store calculation programs required in the process of executing a data collection pipeline, generating learning data / input data, preprocessing collected data, checking for missing data, measuring the time required for each step, generating alarm information, generating a power generation prediction model, calculating a prediction error rate, and distributing a power generation prediction model.

[0061] The processor (150) can control at least one other component (e.g., hardware or software component) of the power generation prediction device (100) by executing software such as a program, and can perform various data processing or calculations.

[0062] A processor (150) according to one embodiment of the present invention may execute a data collection pipeline set to collect actual data including power generation data and weather data based on a first trigger signal, execute a data collection pipeline set to collect forecast data including weather data based on a second trigger signal, preprocess the collected actual data to generate learning data for learning a power generation prediction model, and preprocess the collected forecast data to generate input data of the power generation prediction model.

[0063] The processor (150) can check whether there is missing data from actual data or forecast data for each API, measure the time required for each step of generating learning data for learning a power generation prediction model using the actual data, measure the time required for each step of obtaining power generation prediction information from the power generation prediction model using the forecast data, and generate alarm information when there is missing data in the actual data or forecast data or when the time required for each step exceeds a predefined time.

[0064] The processor (150) can generate a new power generation prediction model using actual data including power generation data and weather data collected within a new model generation range, calculate the prediction error rate of the power generation prediction model being distributed and the prediction error rate of the new power generation prediction model using actual data and forecast data collected within a new model verification range, and identify a power generation prediction model to be distributed based on the prediction error rate.

[0065] Meanwhile, the processor (150) may perform at least a portion of the data analysis, processing, and result information generation for performing the above operations using at least one of a machine learning, neural network, or deep learning algorithm as a rule-based or artificial intelligence (AI) algorithm. Examples of the neural network may include models such as a CNN (Convolutional Neural Network), a DNN (Deep Neural Network), and an RNN (Recurrent Neural Network).

[0066] Figure 3 is a flowchart illustrating the operation of a power generation prediction device according to a first embodiment of the present invention. This diagram describes a method for automating the data collection and preprocessing processes.

[0067] A processor (150) according to one embodiment of the present invention can execute a data collection pipeline set to collect actual data including power generation data and weather data based on a first trigger signal (S10).

[0068] The first trigger signal is a signal that initiates a data collection pipeline configured to collect actual data. In one embodiment, the first trigger signal may be generated to collect the previous day's actual data (power generation data and weather data) every day at 3:00 AM.

[0069] Power generation prediction models are generated and distributed for each power plant (a location that generates renewable energy resources, including but not limited to individual homes and large-scale power plants). Training data for model generation and input data for the model can also be generated for each model. To this end, trigger signals can be generated for each power generation prediction model. Specifically, the trigger signal includes information regarding which data to collect and from where. For example, this may include power plant location information, collection API information, and announcement timing information.

[0070] Power plant location information may include, for example, power plant latitude / longitude information, power plant ID, etc. Collection API information may include, for example, information regarding the API that collects actual data. Announcement time information may include, for example, information regarding whether actual data / forecast data is used, whether the data is a short-term forecast system or a very short-term forecast system. For example, this may include the forecast database time, the forecast date for forecast data, the past date for actual data, and the trigger signal call date.

[0071] In the case of the renewable energy power generation forecasting system described above, it is composed of a system that forecasts power generation for 24 hours from 0:00 to 24:00 on the trading day on the day before the transaction (hereinafter referred to as the short-term forecasting system) and a system that forecasts power generation for the next 24 hours based on the hourly forecasting time (hereinafter referred to as the ultra-short-term forecasting system).

[0072] Currently, the power exchange requests power generation forecast information twice (at 10:00 AM and 5:00 PM) on the day before a transaction for the short-term forecast system, and hourly for the ultra-short-term forecast system (note that the number of forecasts and the timing of forecasts for the short-term / ultra-short-term forecast systems are merely examples and may change according to policy). As such, since the timing and frequency of submission of power generation forecast information to the power exchange are determined, the processor (150) can set the timing and frequency of trigger signal generation based on these.

[0073] Actual data refers to actual measured past data, which may include power generation data and meteorological data. The processor (150) can build a data collection pipeline to collect actual data based on the first trigger signal. Specifically, the processor (150) can build a data collection pipeline to collect actual data through APIs from servers such as a power exchange, the Korea Meteorological Administration, or an external meteorological center, and store the collected data in a desired location. Similar to the trigger signal, the data collection pipeline can be individually generated based on the power plant location, collection API, and announcement time.

[0074] The power generation data may include, for example, the power generation of the power plant, and may also include information such as the collective power resource name, unique region name, generator type, generator ID, whether discharged / charged, date, and time.

[0075] Meteorological data may include, for example, temperature, precipitation, wind speed, humidity, solar irradiance, snow cover, and cloud cover in the power plant area, and may also include information such as collective resource name, unique region name, latitude, longitude, date, and time. Among the meteorological data, solar irradiance is a key piece of information for power generation forecasting, and can be collected from two or more meteorological centers (i.e., two or more APIs). In addition to solar irradiance, information such as azimuth angle and zenith angle can also be collected.

[0076] A processor (150) according to one embodiment of the present invention can execute a data collection pipeline set to collect forecast data including weather data based on a second trigger signal (S20).

[0077] The second trigger signal is a signal to execute a data collection pipeline configured to collect forecast data. In one embodiment, for short-term forecasts, the second trigger signal may be generated to collect the most recently released forecast data (e.g., 8:00 a.m., 2:00 p.m.) based on the trigger signal generation time, twice a day (e.g., 8:30 a.m., 2:30 p.m.). For ultra-short-term forecasts, the second trigger signal may be generated to collect forecast data released a certain amount of time prior to the trigger signal generation time (e.g., 3:30 hours prior to the hour) at regular intervals (e.g., every 50 minutes).

[0078] Forecast data is data predicting weather changes and may include weather data. The processor (150) can build a data collection pipeline to collect forecast data based on a second trigger signal. Specifically, the processor (150) can build a data collection pipeline to collect forecast data through an API from a server such as the Korea Meteorological Administration or an external weather center and store the data in a desired location. Similar to the trigger signal, the data collection pipeline can be individually generated based on the power plant location, collection API, and announcement time.

[0079] Weather data may include, for example, temperature, precipitation, wind speed, humidity, snow cover, cloud cover, solar irradiance, azimuth angle, zenith angle, etc., and may also include information such as collective resource name, small-scale generator name, latitude, longitude, forecast announcement date (base_date), forecast announcement time (base_time), API date / time, and forecast date / time. At this time, the forecast data may include different forecasts depending on the forecasting technology of each weather center, and may be received through two or more APIs.

[0080] A processor (150) according to one embodiment of the present invention can preprocess collected actual data to generate learning data for learning a power generation prediction model (S30).

[0081] Below, various methods for preprocessing real-world data are described, and the order of preprocessing is not limited to the order in which they are described.

[0082] The processor (150) can replace missing data among the irradiance items of the actual data with 0, and replace missing data among the remaining items of the actual data with actual data collected in the previous cycle.

[0083] The processor (150) can select irradiance data to be included in the learning data from among the actual data received through two or more APIs based on the power plant location information. As described above, irradiance is important for power generation prediction and is greatly influenced by geography, so more accurate data can be selected from among the irradiance data acquired by multiple APIs. For example, consider a case where irradiance data is acquired from meteorological centers A and B. For a power plant geographically close to meteorological center A, the irradiance data collected from meteorological center A is used. At this time, it is possible to pre-set which API from which irradiance data will be used for each power plant.

[0084] The processor (150) can identify at least one additional data among the power generation efficiency (pv_efficiency) of each power plant, the trigonometric ratio of the extreme angle (cosine, sine), the trigonometric ratio of the azimuth angle (cosine, sine), and the dew point using the actual measurement data, and include the additional data in the learning data.

[0085] In addition, the processor (150) can extract actual data for a certain time period (e.g., 5:00 to 20:00) or remove actual data for an output limited time (outlier).

[0086] A processor (150) according to one embodiment of the present invention can preprocess collected forecast data to generate input data of a power generation prediction model (S40).

[0087] Below, we describe various methods for preprocessing forecast data, and similarly, the preprocessing order is not restricted by the order in which it is described.

[0088] The processor (150) can supplement missing data in the forecast data by using the forecast data received through two or more APIs.

[0089] The processor (150) can identify data to be included in the input data by applying a voting algorithm to items collected repeatedly among the forecast data. When collecting forecast data through two or more APIs, data may be collected repeatedly for each item. At this time, for items with high feature importance among the parameters of the power generation prediction model, utilizing data collected through multiple APIs rather than data collected through a single API may help improve prediction accuracy. For example, among the multiple items included in the forecast data, items such as irradiance, azimuth, total cloud cover, humidity, and temperature may have somewhat higher importance. At this time, the method for calculating importance is not limited to any one, and items identified as having high importance may vary depending on the power generation prediction model. Meanwhile, the voting algorithm may be employed in various ways and is not limited to any one.

[0090] The processor (150) can match the data ranges for each item collected in duplicate, categorize the data ranges into two or more values, and identify data to be included in the input data based on a majority vote.

[0091] For example, let's assume that the cloud cover item is applied to a voting algorithm. The data range for the cloud cover item may vary across APIs. For example, for weather center A, cloud cover may be expressed in tertiles from 1 to 3, while for weather center B, cloud cover may be expressed in deciles from 0 to 10. Therefore, the cloud cover for weather center A can be converted into deciles to match the two data sets (or vice versa).

[0092] Afterwards, if the data is 6 or higher, it can be categorized as 'clear', and if it is less than 6, it can be categorized as 'cloudy'. This categorization can also be done by setting various reference values ​​or categorizing values. For example, the total cloud cover can be categorized as 'clear' (8-10), 'moderate' (4-7), and 'cloudy' (0-3).

[0093] The processor (150) can determine which data to include in the input data based on the categorized values ​​by majority vote. For example, the total cloudiness items are collected from each of the weather centers A, B, and C, and as a result of categorizing them according to the above process, it can be confirmed that A is forecasted as clear, B as clear, and C as cloudy. Accordingly, the processor (150) can use the total cloudiness data of the weather centers A and / or B that are predicted as clear. Various methods of utilization can also be adopted, for example, the total cloudiness data of the weather centers A and B can be averaged and used.

[0094] In addition, the processor (150) can identify at least one additional data among the power generation efficiency of each power plant, the trigonometric ratio of the extreme angle (cosine, sine), the trigonometric ratio of the azimuth (cosine, sine), and the dew point using the forecast data, and include the additional data in the input data.

[0095] According to one embodiment of the present invention, the data collection and preprocessing process required for learning a power generation prediction model can be automated.

[0096] According to one embodiment of the present invention, the data collection and preprocessing process required for power generation prediction using a model can be automated.

[0097] According to one embodiment of the present invention, the cornerstone of a new power business and a brokerage business for participation in the power market can be laid, and the foundation for a VPP solution service can be secured.

[0098] FIG. 4 is a diagram illustrating a list of data pipelines according to a second embodiment of the present invention. Specifically, FIG. 4 illustrates an example of a list of data pipelines configured for collecting real-world data.

[0099] Referring to Figure 4, the data pipeline list shows the API, path and pipeline name, description of the corresponding pipeline, trigger point, notes, etc.

[0100] APIs may include the Korea Meteorological Administration (KMA), external meteorological centers, and the Korea Power Exchange (KPX).

[0101] For example, you can see that a route is set up to receive real-time data from the Korea Meteorological Administration every day at 3:00 AM. However, it's clear that the API type, number, trigger point, and specific route are not limited to any one specific API.

[0102] Figure 5 is a diagram illustrating a list of data pipelines according to a third embodiment of the present invention. Specifically, Figure 5 illustrates an example of a list of data pipelines configured for forecast data collection. Likewise, it should be understood that the API type, number, trigger point, specific path, etc. are not limited to any one specific type.

[0103] Figure 6 is a flowchart illustrating the operation flow of a power generation prediction device according to a fourth embodiment of the present invention. Figure 6 describes a method for generating notification information regarding data omissions and prediction delays that may occur during the power generation prediction process, from data collection to prediction.

[0104] A processor (150) according to one embodiment of the present invention can check whether there is missing data from actual data or forecast data by API (Application Programming Interface) (S610).

[0105] There are various ways to collect actual and forecast data. For example, as previously described with reference to FIG. 3, the processor (150) may collect data through a data collection pipeline for actual and forecast data. In this case, in an automated data collection process using machine learning operations (MLOps), the power generation prediction device (100) may continuously collect actual and forecast data through a defined data collection pipeline and accumulate them in a separate database.

[0106] The processor (150) can retrieve actual data or forecast data accumulated in a database for learning a power generation prediction model or for power generation prediction. This process is called data synchronization.

[0107] Some or all of the actual data and forecast data may be missing for various reasons, such as unstable network connection between the API and the power generation prediction device (100) or missing data during the data synchronization process.

[0108] A processor (150) according to one embodiment of the present invention can measure the time required for each step of generating learning data for learning a power generation prediction model using actual data (S620).

[0109] The processor (150) can generate training data for training a power generation prediction model based on a third trigger signal. The third trigger signal is a signal for generating training data using collected actual data. In one embodiment, the third trigger signal may be generated to generate training data using collected data at 3:45 AM every day.

[0110] Specifically, the step of generating learning data may include 1) a step of synchronizing the learning data from a database that has collected the learning data according to a third trigger signal, 2) a step of matching the format of the learning data received from two or more APIs, 3) a step of checking whether the learning data contains missing data, and 4) a step of preprocessing the learning data to generate learning data. According to one embodiment of the present invention, the processor (150) may provide a repeatable process through machine learning operations (MLOps) for the process of generating learning data.

[0111] 2) Data format matching is a step that matches the formats of real-world data received from different APIs, and is a step distinct from the preprocessing step 4). The preprocessing step 4) can be identically adapted from the content described above with reference to Figure 3.

[0112] The processor (150) can measure the time required for each of the above steps.

[0113] A processor (150) according to one embodiment of the present invention can measure the time required for each step of obtaining power generation prediction information from a power generation prediction model using forecast data (S630).

[0114] The processor (150) can generate input data for a power generation prediction model based on a fourth trigger signal. The fourth trigger signal is a signal for generating input data using collected forecast data. In one embodiment, the fourth trigger signal can be generated at 8:45 AM and 3:45 PM for short-term forecasts, and can be generated every hour for ultra-short-term forecasts.

[0115] Specifically, the step of obtaining power generation prediction information may include 1) a step of synchronizing forecast data from a database that has collected forecast data according to a fourth trigger signal, 2) a step of matching the format of forecast data received from two or more APIs, 3) a step of checking whether forecast data contains missing data, 4) a step of preprocessing forecast data to generate input data of a power generation prediction model, and 5) a step of inputting the input data into the power generation prediction model to obtain power generation prediction information. According to one embodiment of the present invention, the processor (150) may provide a repeatable process through machine learning operations (MLOps) for the process of obtaining power generation prediction information.

[0116] Similarly, data format matching in 2) is a step that matches the formats of real-world data received from different APIs, and is a step distinct from the preprocessing in 4). The preprocessing step in 4) can be identically adapted from the content described above with reference to Figure 3.

[0117] A processor (150) according to one embodiment of the present invention can generate alarm information when there is missing data in actual data or forecast data or when the time required for each step exceeds a predefined time (S640).

[0118] The predefined time serves as a threshold for the time required for each step, and the processor (150) can set the predefined time for each step. The time can be set to the same for each step, or different for each step. For example, in the process of generating learning data, step 1) can be set to 2 minutes, step 2) to 3 minutes, step 3) to 3 minutes, and step 4) to 5 minutes.

[0119] In addition, the processor (150) may generate alarm information if the time required for the entire step exceeds a predefined time. For example, alarm information may be generated if the entire step of generating learning data exceeds 15 minutes.

[0120] Alarm information generated when missing data exists may include information about whether it is a learning data generation process / power generation forecast information acquisition process, whether it is a short-term forecast / ultra-short-term forecast in the case of a power generation forecast information acquisition process, the API where the missing data occurred, and the missing data items.

[0121] Alarm information generated when the time required for a step exceeds a predefined time may include information about the time required for the step, the step that exceeded the predefined time, etc.

[0122] In addition, alarm information may include the number of attempts, start time, duration, power plant information, and contact information.

[0123] If alarm information is generated during the process of acquiring power generation prediction information, the processor (150) may use power generation prediction information acquired in the previous cycle. In this case, the alarm information may be applied regardless of the occurrence of missing data or prediction delay.

[0124] When alarm information is generated due to missing data in actual data or forecast data, the processor (150) can identify the items of missing data and replace the items of missing data with actual data or forecast data of the previous cycle.

[0125] According to one embodiment of the present invention, a technology gap in machine learning operations (MLOps) can be filled by enabling a person in charge to quickly recognize when missing data or prediction delay occurs.

[0126] According to one embodiment of the present invention, in a system that must provide real-time prediction information, such as ultra-short-term prediction, a power brokerage business can be operated stably by taking follow-up measures when missing data or prediction delay occurs.

[0127] Fig. 7 is a diagram illustrating an operational flow chart of a power generation prediction device according to a fifth embodiment of the present invention. Fig. 7 is a diagram schematically illustrating the process of obtaining power generation prediction information described above with reference to Fig. 6.

[0128] The step of obtaining power generation forecast information may include 1) a step of synchronizing forecast data from a database that collected forecast data according to a fourth trigger signal (Sync), 2) a step of matching the format of forecast data received from two or more APIs (Clean), 3) a step of checking whether forecast data contains missing data (Check), 4) a step of preprocessing forecast data to generate input data for a power generation forecast model (Make), and 5) a step of inputting input data into a power generation forecast model to obtain power generation forecast information (Predict).

[0129] Referring to Figure 7, in the synchronization (Sync) step, it can be seen that the forecast data collected from the Korea Meteorological Administration (KMA) is synchronized (701), the forecast data collected from the first external meteorological center (API 1) is synchronized (702), and the forecast data collected from the second external meteorological center (API 2) is synchronized (703). At each step, the success / failure and the required time can be confirmed. For example, it can be seen that the step of synchronizing the forecast data collected from the Korea Meteorological Administration (KMA) (701) was successfully completed and took 2 minutes and 6 seconds.

[0130] After the synchronization step, a step (Clean) (711, 712, 713) for matching the format of forecast data received from two or more APIs can be performed.

[0131] Thereafter, steps (Check) (721, 722, 723) for checking whether the forecast data contains missing data may be performed. At this time, it may be determined that missing data occurs (Failed) in the forecast data collected from the Korea Meteorological Administration (KMA). In this case, the processor (150) may generate alarm information, as described above with reference to FIG. 6.

[0132] Thereafter, a step (Make) (731) of preprocessing the forecast data to generate input data for a power generation prediction model can be performed. The process of preprocessing the forecast data to generate input data can be borrowed from the content described above with reference to FIG. 3.

[0133] Meanwhile, if alarm information is generated, as described above with reference to FIG. 6, the processor (150) may replace missing data with forecast data received from another API, or terminate the forecasting process and transmit the power generation forecast information collected in the previous cycle to the power exchange. As an extension of the step (Make) (731) of generating input data, a step (Check) (732) of verifying the generated input data may be performed.

[0134] Afterwards, a step (Predict) (741) of obtaining power generation prediction information by inputting input data into a power generation prediction model can be performed, and a step (Save) (751) of saving the obtained power generation prediction information can be performed.

[0135] Fig. 8 is a flowchart illustrating the operation of a power generation prediction device according to a sixth embodiment of the present invention. Fig. 8 describes a method for improving the prediction accuracy of a power generation prediction model.

[0136] A processor (150) according to one embodiment of the present invention can create a new power generation prediction model using actual data including power generation data and weather data collected within a new model generation range (S810).

[0137] The renewable energy generation forecasting system applies a differential unit price based on the forecast error rate for each time period. For example, if the forecast error rate is 6% or less, the unit price may be 4 won / kWh, and if it exceeds 6% but is less than 8%, the unit price may be 3 won / kWh. Therefore, the accuracy of the generation forecasting model directly affects the profits from electricity trading. Therefore, continuously updating the model to improve its accuracy is crucial.

[0138] To this end, power generation prediction models are being optimized for each power plant, and new power generation prediction models are being continuously created by reflecting the latest data.

[0139] The new model generation range may be the period of time required to collect actual data needed to create a new power generation forecast model. For example, assume that new power generation forecast models are generated on the 8th, 18th, and 28th of each month, and that the new model generation range is one year from the model generation date. In this case, the new model generation range for a new power generation forecast model generated on August 8, 2024, is from August 8, 2023, to August 7, 2024. Furthermore, the new model generation range for a new power generation forecast model generated on August 18, 2024, is from August 18, 2023, to August 17, 2024.

[0140] A power generation prediction model can be trained to derive predicted power generation data based on input forecasted weather data by utilizing the correlation between actual and measured weather data. Various techniques exist for generating power generation prediction models, including the gradient boosting algorithm.

[0141] A processor (150) according to one embodiment of the present invention can calculate the prediction error rate of a power generation prediction model being distributed and the prediction error rate of a new power generation prediction model using actual data and forecast data collected within a new model verification range (S820).

[0142] Meanwhile, in machine learning operations (MLOps), model deployment can be operated by comparing the performance difference between the currently deployed model and the newly created model and updating it to a model with better performance.

[0143] At this time, the prediction error rate calculation can be performed within the new model validation range. The new model validation range may be the collection period of actual data and forecast data required to validate the new power generation prediction model and compare its performance differences with existing power generation prediction models. For example, assume that new power generation prediction models are generated on the 8th, 18th, and 28th of each month, and the new model validation range is one month from the model creation date. In this case, the new model validation range for the new power generation prediction model generated on August 8, 2024, is from July 8, 2024, to August 7, 2024. Additionally, the new model validation range for the new power generation prediction model generated on August 18, 2024, is from July 18, 2024, to August 17, 2024.

[0144] In addition, the scope of new model creation and the scope of new model verification may be the same and are not limited to either one.

[0145] The processor (150) can obtain power generation prediction information of a power generation prediction model currently being distributed (hereinafter referred to as a first power generation prediction model) and a newly created power generation prediction model (hereinafter referred to as a second power generation prediction model) using the forecast data collected within the new model verification range.

[0146] As previously described with reference to FIG. 3, the processor (150) can preprocess forecast data to generate input data. The processor (150) can obtain power generation prediction information by inputting input data to the first power generation prediction model and the second power generation prediction model, respectively. By comparing the prediction results of the two models with the actual power generation, the prediction error rate of each model can be calculated.

[0147] Specifically, the processor (150) can compare the power generation prediction information of the first power generation prediction model and the power generation prediction information of the second power generation prediction model with the actual data (power generation data) collected within the new model verification range and calculate the prediction error rate of the first power generation prediction model and the prediction error rate of the second power generation prediction model, respectively.

[0148] For example, the processor (150) can calculate the prediction error rate using the following mathematical expression 1.

[0149]

[0150] At this time, according to the renewable energy power generation prediction system, if the utilization rate is less than 10%, it may be excluded from the calculation of the prediction error rate, but this is not limited to this.

[0151] A processor (150) according to one embodiment of the present invention can identify a power generation prediction model to be distributed based on a prediction error rate (S830).

[0152] The processor (150) can distribute the first power generation prediction model when the prediction error rate of the first power generation prediction model being distributed is lower than the prediction error rate of the newly generated second power generation prediction model.

[0153] Conversely, the processor (150) may distribute the second power generation prediction model if the prediction error rate of the first power generation prediction model is higher than the prediction error rate of the second power generation prediction model.

[0154] According to one embodiment of the present invention, the process of generating and distributing a power generation prediction model can be automated.

[0155] According to one embodiment of the present invention, the prediction accuracy of a power generation prediction model can be upwardly standardized.

[0156] FIG. 9 is a drawing illustrating the operation of a power generation prediction device according to one embodiment of the present invention.

[0157] The power generation forecasting process can be broadly divided into data collection, model creation, and power generation forecasting. As described with reference to the drawings above, data collection may include collecting and preprocessing measured data, including power generation data and weather data (history), and forecast data, including weather data (forecast).

[0158] The power generation prediction device (100) can create a power generation prediction model using actual measurement data. At this time, the power generation prediction device (100) can remove outliers from the actual measurement data and evaluate / update the model.

[0159] Thereafter, the power generation prediction device (100) can manage the model by comparing the prediction error rate between the power generation prediction model being distributed and the newly generated power generation prediction model, or can manage data automatically collected according to a trigger signal. In addition, the power generation prediction device (100) can obtain power generation prediction information corresponding to the forecast data by using the power generation prediction model being distributed. The power generation prediction device (100) can submit the power generation prediction information obtained by prediction type (short-term prediction / ultra-short-term prediction, etc.) to the power exchange according to the power generation prediction system. Alternatively, the power generation prediction device (100) can transmit the power generation prediction information to the operator terminal of the in-house operation team that utilizes the power generation prediction information.

[0160] Additionally, the power generation prediction device (100) can transmit alarm information to the operator terminal when missing data or delayed prediction occurs throughout the entire process.

[0161] According to one embodiment of the present invention, everything from data collection to preprocessing, model learning, model update, version management, and prediction and transmission can be automatically performed based on machine learning operations (MLOps).

[0162] South Korea operates its systems by dividing them into demand-side resource markets (demand-side VPPs) and small-scale power brokerage markets (supply-side VPPs), restricting participation to specific technologies and power sources (tech-specific). In this context, in the short term, we should focus on developing an integrated VPP operation platform for supply-side VPP resources. In the mid- to long-term, we need to develop a platform capable of operating converged resources, including demand-side VPP resources such as DR and electric vehicles. As described in this invention, by implementing an automated process based on machine learning operations (MLOps), we expect to build a stable resource portfolio and secure diverse business opportunities.

[0163] FIG. 10 is a diagram illustrating the operation of a power generation prediction system according to one embodiment of the present invention.

[0164] The power generation prediction device (100) can build a data process to perform the process from data collection to model distribution and monitoring to power generation prediction on a cloud-based platform.

[0165] In the data collection stage, the power generation prediction device (100) can connect to APIs and databases (storage) that collect actual data (history) and forecast data (forecast).

[0166] In the data storage and processing step, the power generation prediction device (100) stores all collected data in a database.

[0167] In the model learning and verification stage, the power generation prediction device (100) can obtain the necessary actual data and forecast data by synchronizing with the database, and perform preprocessing and model learning / generation.

[0168] In the model distribution and monitoring stage, the power generation prediction device (100) can obtain power generation prediction information using the model at a set time according to the power generation prediction system.

Claims

1. In a power generation prediction device of a virtual power plant (VPP) based on machine learning operations (MLOps), Execute a data collection pipeline set to collect actual data including power generation data and weather data based on the first trigger signal, Execute a data collection pipeline set to collect forecast data including weather data based on a second trigger signal, Preprocess the collected actual data to generate learning data for learning the power generation prediction model, A power generation prediction device including a processor that preprocesses the collected forecast data to generate input data for the power generation prediction model.

2. In paragraph 1, The above processor, Missing data among the solar radiation items of the above measured data are replaced with 0, A power generation prediction device that replaces missing data among the remaining items of the above measured data with measured data collected in the previous cycle.

3. In paragraph 1, The above processor, A power generation prediction device that selects irradiance data to be included in learning data from among irradiance data of actual data received through two or more APIs (Application Programming Interfaces) based on power plant location information.

4. In paragraph 1, The above processor, A power generation prediction device that supplements missing data in the forecast data by using forecast data received through two or more APIs (Application Programming Interfaces).

5. In paragraph 1, The above processor, A power generation prediction device that identifies data to be included in the input data by applying a voting algorithm to items collected repeatedly among the above forecast data.

6. In paragraph 5, The above processor, Match the data range of the items collected in duplicate above, Categorize the data with matching ranges into categories with two or more distinct categories, A power generation prediction device that identifies one of the categories corresponding to the above-mentioned duplicated collected items as data to be included in the input data based on a majority vote.

7. In paragraph 1, The above processor, Using the above measured data and forecast data, at least one additional data among the power generation efficiency, trigonometric ratio of extreme angle, trigonometric ratio of azimuth, and dew point of each power plant is identified, A power generation prediction device that includes the above additional data in the above learning data and the above input data.

8. In paragraph 1, The above first trigger signal and second trigger signal are, A power generation prediction device that includes power plant location information, collected API information, and announcement time information.

9. A method for predicting power generation of a virtual power plant (VPP) based on machine learning operations (MLOps) performed by a power generation prediction device, A step of executing a data collection pipeline set to collect actual data including power generation data and meteorological data based on a first trigger signal; A step of executing a data collection pipeline set to collect forecast data including weather data based on a second trigger signal; A step of preprocessing the collected actual data to generate learning data for learning a power generation prediction model; A power generation prediction method comprising a step of preprocessing the collected forecast data to generate input data for the power generation prediction model.

10. In paragraph 9, The steps for generating the above learning data are: A step of replacing missing data among the solar radiation items of the above measured data with 0; A method for predicting power generation, comprising a step of replacing missing data among the remaining items of the above measured data with measured data collected in the previous cycle.

11. In paragraph 9, The steps for generating the above learning data are: A method for predicting power generation, comprising a step of selecting irradiance data to be included in learning data among irradiance data of actual measurement data received through two or more APIs (Application Programming Interfaces) based on power plant location information.

12. In paragraph 9, The steps for generating the above input data are: A power generation prediction method comprising a step of supplementing missing data of the forecast data using forecast data received through two or more APIs (Application Programming Interfaces).

13. In paragraph 9, The steps for generating the above input data are: A power generation prediction method comprising a step of identifying data to be included in the input data by applying a voting algorithm to items collected in duplicate among the above forecast data.

14. In paragraph 13, The steps for generating the above input data are: A step of matching the data range of the items collected in duplicate; A step of categorizing data with matching ranges into two or more distinct categories; A method for predicting power generation, comprising a step of identifying one of the categories corresponding to the above-mentioned duplicated collected items as data to be included in the input data based on a majority vote.

15. In paragraph 9, The steps of generating the above learning data and input data are: A step of identifying at least one additional data among the power generation efficiency, trigonometric ratio of extreme angle, trigonometric ratio of azimuth, and dew point of each power plant using the above measured data and forecast data; A method for predicting power generation, comprising the step of including the additional data in the learning data and the input data.

16. In paragraph 9, The above first trigger signal and second trigger signal are, A method for predicting power generation, characterized by including power plant location information, collected API information, and announcement time information.

Images