Apparatus and method for Virtual Power Plant Diagnosis

Abstract

The present invention relates to a virtual power plant diagnostic device and method. A virtual power plant diagnostic device according to the present invention comprises: a time series data collection unit that collects time series data including at least one of voltage, current, power, frequency, reactive power, active power, temperature, and humidity for a power generation device constituting a virtual power plant through an energy monitoring sensor and a temperature and humidity sensor; a token generation unit that converts the time series data into text-type data and tokenizes the time series data converted into text-type data to generate tokens; a prediction value generation unit that generates a prediction value by predicting a token after the last token among the tokens through a natural language processing-based artificial intelligence model; a prediction error calculation unit that calculates a prediction error by comparing additionally collected time series data with the prediction value; and a diagnostic unit that diagnoses the power generation device as faulty when the prediction error exceeds a preset error threshold.

Description

Technology Field

[0001] The present invention relates to a virtual power plant diagnostic device and method, and more specifically, to a virtual power plant diagnostic device and method for predicting the occurrence of abnormalities in a virtual power plant and diagnosing a virtual power plant. Background Technology

[0002] A virtual power plant refers to a system that utilizes information and communication technologies, such as artificial intelligence, smart grids, the Internet of Things, and cloud computing, to connect small-scale energy generation resources—including renewable energy sources like solar and wind power, energy storage devices, and electric vehicles—and integrates and controls them like a single centralized power plant.

[0003] Recently, there has been an increasing trend of small-scale power generation, storage, and utilization in homes and workplaces.

[0004] To maximize the utilization efficiency of these small-scale distributed resources, technology that detects early signs of anomalies in the power system is crucial.

[0005] In the operation of virtual power plants, proper prediction of anomalies can prevent accidents, reduce operating costs by predicting maintenance timing, and enhance the operational efficiency of the virtual power plant by providing a basis for forecasting power supply and demand.

[0006] Conventional anomaly detection technologies had a problem of insufficient precision because they did not reflect the contextual characteristics of the data.

[0007] Furthermore, conventional anomaly detection technologies had a problem of insufficient explanatory power because they failed to provide clear and explanatory grounds for the criteria for judging anomalies.

[0008] Consequently, there was a problem of reduced operational efficiency as the intervention of experts capable of providing explanations for abnormal phenomena in the power plant was essential for decision-making regarding the operation of the virtual power plant.

[0009] Furthermore, conventional anomaly detection technology had a problem in that it could not reflect the large amount of sensor data collected from the virtual power plant and the dynamic characteristics of the data.

[0010] In addition, conventional anomaly detection technology required the use of at least two artificial intelligence models for anomaly detection and prediction, which made system maintenance and updates difficult and complicated the analysis of the cause when a problem occurred. Prior art literature

[0011] Korean Registered Patent Publication No. 10-2141677 The problem to be solved

[0012] The present invention was devised to solve the problems described above, and the objective of the present invention is to provide a virtual power plant diagnostic device and method that diagnoses a virtual power plant by reflecting the contextual characteristics of data through a natural language processing-based artificial intelligence model.

[0013] Another objective of the present invention is to provide a virtual power plant diagnostic device and method that diagnoses a virtual power plant by providing clear grounds for anomaly judgment criteria through a natural language processing-based artificial intelligence model.

[0014] Another objective of the present invention is to provide a virtual power plant diagnostic device and method that diagnoses a virtual power plant by reflecting a large amount of sensor data collected from the virtual power plant and the dynamic characteristics of the data through a natural language processing-based artificial intelligence model.

[0015] Another objective of the present invention is to provide a virtual power plant diagnostic device and method that diagnoses a virtual power plant using a single artificial intelligence model through a natural language processing-based artificial intelligence model. means of solving the problem

[0016] For the above purpose, a virtual power plant diagnostic device according to one embodiment of the present invention comprises: a time series data collection unit that collects time series data including at least one of voltage, current, power, frequency, reactive power, active power, temperature, and humidity for a power generation device constituting a virtual power plant through an energy monitoring sensor and a temperature and humidity sensor; a token generation unit that converts the time series data into text-type data and tokenizes the time series data converted into text-type data to generate tokens; a prediction value generation unit that generates a prediction value by predicting a token after the last token among the tokens through a natural language processing-based artificial intelligence model; a prediction error calculation unit that calculates a prediction error by comparing additionally collected time series data with the prediction value; and a diagnostic unit that diagnoses the power generation device as faulty when the prediction error exceeds a preset error threshold.

[0017] Preferably, the device further includes a text data collection unit that collects text data including at least one of an accident report and a maintenance log related to a short circuit and discharge of the power generation device.

[0018] Preferably, the diagnostic unit generates a diagnostic basis, which is the basis for diagnosis of the power generation device, through a natural language processing-based artificial intelligence model that uses the time series data and the text data as training data.

[0019] Preferably, the diagnostic unit generates a diagnostic basis corresponding to the time series data based on a pattern between the time series data and the text data.

[0020] Preferably, the token generation unit comprises a feature extraction unit that extracts statistical features including at least one of a time-series trend, a peak, and an average from the time-series data;

[0021] It includes a repetition period calculation unit that performs matching for times when the statistical features are similar in the time series data and calculates a repetition period based on the difference between the matched times; and a token conversion unit that converts the statistical features and the repetition period into data in the form of strings and then converts them into tokens.

[0022] And, a virtual power plant diagnostic method according to one embodiment of the present invention comprises: a step of collecting time series data including at least one of voltage, current, power, frequency, reactive power, active power, temperature, and humidity for a power generation device constituting a virtual power plant through an energy monitoring sensor and a temperature and humidity sensor; a step of converting the time series data into text-type data and tokenizing the time series data converted into text-type data to generate tokens; a step of generating a predicted value by predicting a token after the last token among the tokens through a natural language processing-based artificial intelligence model; a step of calculating a prediction error by comparing additionally collected time series data with the predicted value; and a step of diagnosing the power generation device as faulty when the prediction error exceeds a preset error threshold.

[0023] Preferably, after the diagnostic step, the method further includes a step of collecting text data including at least one of an accident report and a maintenance log related to a short circuit and discharge of the power generation device.

[0024] Preferably, after the step of collecting the text data, the method further includes the step of generating a diagnostic basis, which is the basis for diagnosis of the power generation device, through the natural language processing-based artificial intelligence model using the time series data and the text data as training data.

[0025] Preferably, the step of generating the diagnostic basis generates a diagnostic basis corresponding to the time series data based on a pattern between the time series data and the text data.

[0026] Preferably, the step of generating the token comprises: a process of extracting a statistical feature including at least one of an hourly trend, a peak, and an average from the time series data; a process of performing matching for times in the time series data where the statistical feature is similar and calculating a repetition period based on the difference between the matched times; and a process of converting the statistical feature and the repetition period into data in the form of a string and then converting them into the token. Effects of the invention

[0027] According to the present invention, since it reflects the contextual characteristics of the data, it has the effect of accurately diagnosing a virtual power plant.

[0028] In addition, according to the present invention, it has the effect of providing clear grounds for the criteria for determining abnormal phenomena and diagnosing a virtual power plant.

[0029] In addition, according to the present invention, since it reflects a large amount of sensor data collected from a virtual power plant and the dynamic characteristics of the data, it has the effect of accurately diagnosing the virtual power plant.

[0030] In addition, according to the present invention, since a single artificial intelligence model is used, it has the effect of diagnosing a virtual power plant through a system that is simpler to maintain and update. Brief explanation of the drawing

[0031] FIG. 1 is a configuration diagram of a virtual power plant diagnostic device according to one embodiment of the present invention. FIG. 2 is a flowchart of a virtual power plant diagnosis method according to one embodiment of the present invention. Figure 3 is a detailed configuration diagram of the time series data collection unit of Figure 1. Figure 4 is a detailed configuration diagram of the token generation unit of Figure 1. FIG. 5 is a flowchart of a learning method for a virtual power plant diagnostic device according to one embodiment of the present invention. Specific details for implementing the invention

[0032] Hereinafter, embodiments disclosed in this specification will be described in detail with reference to the attached drawings. Identical or similar components regardless of drawing symbols are assigned the same reference number, and redundant descriptions thereof will be omitted. The suffixes "module" and "part" used for components in the following description are assigned or used interchangeably solely for the ease of drafting the specification and do not inherently possess distinct meanings or roles. Furthermore, in describing the embodiments disclosed in this specification, if it is determined that a detailed description of related prior art could obscure the essence of the embodiments disclosed in this specification, such detailed description will be omitted. Additionally, the attached drawings are intended only to facilitate understanding of the embodiments disclosed in this specification; the technical concept disclosed in this specification is not limited by the attached drawings, and it should be understood that they include all modifications, equivalents, and substitutions that fall within the spirit and technical scope of the invention.

[0033] Terms including ordinal numbers, such as first, second, etc., may be used to describe various components, but said components are not limited by said terms. These terms are used solely for the purpose of distinguishing one component from another.

[0034] When it is stated that one component is "connected" or "connected" to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. On the other hand, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between.

[0036] Hereinafter, a virtual power plant diagnostic device and method according to the present invention will be described in detail with reference to FIGS. 1 to 5.

[0037] FIG. 1 is a configuration diagram of a virtual power plant diagnostic device according to one embodiment of the present invention, and FIG. 2 is a flowchart of a virtual power plant diagnostic method according to one embodiment of the present invention.

[0038] Referring to FIG. 1, a virtual power plant diagnostic device (100) according to one embodiment of the present invention may include a time series data collection unit (110), a token generation unit (120), a prediction value generation unit (130), a prediction error calculation unit (140), a diagnostic unit (150), a text data collection unit (160), etc.

[0039] Figure 3 is a detailed configuration diagram of the time series data collection unit of Figure 1.

[0040] Referring to FIG. 3, the time series data collection unit (110) collects time series data including at least one of voltage, current, power, frequency, reactive power, active power, temperature, and humidity for a power generation device constituting a virtual power plant through an energy monitoring sensor (111) and a temperature and humidity sensor (112) (see step S210 of FIG. 2).

[0041] For example, the time series data collection unit (110) can measure interrelated electrical variables such as voltage, current, power consumption, and energy from an energy monitoring sensor when power is supplied to a load such as a PC or refrigerator, and specifically can measure AC (RMS) voltage, current, and power (single phase). In addition, the temperature and humidity sensor can measure the surrounding air through a humidity capacitor and a temperature transistor and output the measured value as a digital signal.

[0042] For example, the energy monitoring sensor can be implemented with the PZEM-004T, and the temperature and humidity sensor can be implemented with the DHT22.

[0043] Figure 4 is a detailed configuration diagram of the token generation unit of Figure 1.

[0044] Referring to FIG. 4, the token generation unit (120) converts the time series data into text-type data and tokenizes the time series data converted into text-type data to generate tokens (refer to step S220 of FIG. 2).

[0045] For example, the token generation unit (120) may include a feature extraction unit (121) that extracts a statistical feature including at least one of a time-by-time trend, a peak, and an average from the time series data, a repetition cycle calculation unit (122) that performs matching for each time point where the statistical feature is similar in the time series data and calculates a repetition cycle based on the difference between the matched times points, and a token conversion unit (123) that converts the statistical feature and the repetition cycle into data in the form of a string and then converts them into the token.

[0046] The above token conversion unit (123) can generate at least one of an increase, decrease, maximum value, and minimum value as data in the form of a string based on the above statistical characteristics and then convert it into the above token.

[0047] The prediction value generation unit (130) generates a prediction value by predicting the token after the last token among the above tokens through a natural language processing-based artificial intelligence model (see step S230 of FIG. 2).

[0048] For example, the prediction value generation unit (130) can process the numeric data of the time series data into a string data form and output a time series data prediction value through the next token prediction.

[0049] Here, the time-series data collected from the virtual power plant has a sequential structure in which the meaning changes according to time or order.

[0050] To predict such time series data, pretrained foundation models (PFMs) for time series data are required.

[0051] However, while a large-scale training dataset is required for the pre-training of foundation models, existing time-series data presents a problem in that they cannot be integrated because there are differences in sample size, number of features, and segmentation type among individual datasets, as well as differences in statistical characteristics and scale across domains such as manufacturing, finance, and medicine.

[0052] Consequently, there was a problem in that a foundation model for time series forecasting could not be built due to the absence of a large-scale training dataset.

[0053] Meanwhile, when using a natural language processing-based artificial intelligence model pre-trained on a large, systematically organized set of text data, it becomes possible to capture arrangements and patterns among data with the aforementioned sequential structure.

[0054] In one embodiment of the present invention, the weights of the natural language processing-based artificial intelligence model used can be implemented through fine-tuning to output predicted values.

[0055] Accordingly, the prediction value generation unit (130) can generate an answer through a natural language processing-based artificial intelligence model that takes a template-based prompt and time series data as input.

[0056] For example, the context template can be implemented with the start and end dates of the analysis interval, the power generation device being analyzed, and the average power produced by the power generation device from the start date to the end date, and the question template can be implemented with the predicted value of the average power produced by the power generation device on the day following the end date.

[0057] Accordingly, the token generated by the token generation unit can be placed in the start and end dates within the context template, the identification number of the power generation device to be analyzed, and the average power produced per date variable.

[0058] On the other hand, the collected time series data is continuous in chronological order for each variable, and only a single scalar value is stored for each time point. Since these single scalar values ​​cannot provide the trend information necessary for forecasting, the contextual information inherent in the time series data cannot be reflected in the prediction.

[0059] Generally, since time series data is non-stationary data whose distribution changes over time, it is necessary to reflect the local context of the time series data through the repetition period calculated by the repetition period calculation unit (122).

[0060] Accordingly, the prediction value generation unit (130) can generate a prediction token by predicting the token after the last token among the tokens, and can generate the prediction value by converting the prediction token into time series data.

[0061] The prediction error calculation unit (140) calculates a prediction error by comparing the additionally collected time series data with the prediction value (see step S240 of FIG. 2).

[0062] For example, the prediction error calculation unit (140) can calculate the difference between the predicted value calculated based on the normal operation of the virtual power plant and the sensor data actually collected from the energy monitoring sensor.

[0063] The diagnostic unit (150) diagnoses the power generation device as faulty when the prediction error exceeds a preset error threshold (see step S250 of FIG. 2).

[0064] For example, the diagnostic unit (150) can diagnose the abnormal signs of the virtual power plant when a large prediction error is calculated.

[0065] The text data collection unit (160) collects text data including at least one of an accident report and a maintenance log related to a short circuit and discharge of the power generation device (see step S260 of FIG. 2).

[0066] For example, the text data collection unit (160) can collect the title of the report, the date and time of the accident, the name and identification number of the power generation device where the accident occurred, the details of the accident circumstances and cause analysis, the details of the measures and responses taken after the accident, the results of the device inspection after the accident and the necessary repairs, the future preventive measures and inspection plan, and the name and signature of the author.

[0067] The diagnostic unit (150) generates a diagnostic basis, which is the basis for diagnosis of the power generation device, through the natural language processing-based artificial intelligence model that uses the time series data and the text data as training data (see step S270 of FIG. 2).

[0068] For example, the diagnostic unit (150) can output the circumstances and causes of the accident corresponding to the input time series data based on the collected text data and the pattern of the time series data collected at the time the accident occurred.

[0069] For example, from the sensor, it is possible to predict the possibility of a short circuit due to changes in current and voltage patterns, such as insulation performance due to temperature and humidity, leakage current, voltage drop, and voltage spike, or to determine whether a short circuit has occurred.

[0070] Accordingly, the diagnostic unit (150) can generate a diagnostic basis by generating text data corresponding to the pattern of the time series data based on the pattern between the time series data and the text data. Here, the diagnostic basis refers to an explanation that humans can understand regarding the prediction.

[0071] Accordingly, when applying the virtual power plant diagnostic method according to one embodiment of the present invention, the reliability, efficiency, and transparency of operation can be improved.

[0072] FIG. 5 is a flowchart of a learning method for a virtual power plant diagnostic device according to one embodiment of the present invention.

[0073] Referring to FIG. 5, time-series streaming data is collected from a sensor (refer to step S510 of FIG. 5).

[0074] Then, time-series streaming data collected from the sensor is converted into tokens for a Large Language Model (LLM) (see step S520 in Fig. 5).

[0075] Then, after inputting into a large-scale language model, it is detected whether the newly input streaming data corresponds to an outlier based on the predicted value derived from the previously collected data (see step S530 of Fig. 5).

[0076] First, if the difference between newly input sensor data and existing predicted values ​​exceeds a certain threshold, it can be detected as an outlier.

[0077] For example, when implementing a large-scale language model using sensor data from a power generation unit that previously failed and accident and maintenance reports written at the time as training data, the training data can be configured based on sensor data from the power generation unit in the section where no failure occurred.

[0078] Meanwhile, when implementing a large-scale language model, it can be implemented to detect the corresponding pattern by learning sensor data from the power generation device where the failure occurred.

[0079] For example, even if the prediction error is below a certain threshold value, if the newly input sensor data matches the pattern of the sensor data collected at the time of an accident, it can be implemented so that the type of accident and anomaly indication is output through a large-scale language model.

[0080] If there is an error between the newly input sensor data and the existing predicted value, accident and maintenance keywords corresponding to the newly input sensor data can be extracted through a large-scale language model using the sensor data of the power generation device that previously failed and the accident and maintenance report written at that time as training data (see step S540 of Fig. 5).

[0081] Meanwhile, instead of extracting existing accident and maintenance keywords, it can be implemented so that only extracted statistical features, such as time-based trends, peaks, and averages extracted from the feature extraction unit (121), are output.

[0082] Then, the diagnostic unit (150) generates a prompt to provide an explanation of the outlier and outputs an answer regarding the basis for the diagnosis (see step S550 of FIG. 5).

[0083] For example, explanatory information can be provided to non-experts by synthesizing the contents of sensor data collection time intervals where abnormal signs were detected, prediction errors, statistical characteristics of the sensor data within those time intervals, and accident and maintenance reports.

[0084] If no outliers are detected in the streaming data, the streaming data can be retrieved and input as training data for the model (see step S560 in Fig. 5).

[0085] Then, some of the input training data is used as training data to reset the weights (see step S570 in Fig. 5).

[0086] In addition, a universally suitable model can be selected by using some of the input training data as validation data (see step S580 of Fig. 5).

[0087] When a suitable model is selected, outliers are detected using that model instead of the existing model (see step S590 of FIG. 5).

[0088] Accordingly, when utilizing the virtual power plant diagnostic method according to one embodiment of the present invention, individuals operating small-scale distributed resources, such as small solar panels and electric vehicles constituting virtual power plants which are increasing due to recent eco-friendly policies, can economically obtain information without hiring experts.

[0089] In addition, in power systems where maintenance is critical—such as balancing power production and consumption, maintaining constant voltage values, and preventing safety accidents—the operational efficiency of a virtual power plant can be enhanced by predicting trends through forecast values ​​and controlling each power generation unit via communication devices.

[0090] In addition, while conventional methods were limited to using inaccurate prediction models or statistical techniques due to the difficulty of constructing foundation models caused by limitations in the configuration of time-series learning datasets, one embodiment of the present invention utilizes a natural language processing model with guaranteed performance in pattern learning to fine-tune time-series numerical data and maintenance data preprocessed into text-based tokens as training data.

[0092] In the specification of this disclosure (particularly in the claims), the use of the term "above" and similar descriptive terms may be in both singular and plural. Furthermore, where a range is described in this disclosure, it is to include an invention to which individual values ​​belonging to said range are applied (unless otherwise stated), as is equivalent to describing each individual value constituting said range in the detailed description of the invention.

[0093] Unless explicitly stated otherwise, the steps constituting the method according to the present disclosure may be performed in a suitable order. The present disclosure is not necessarily limited by the order in which the steps are described. The use of all examples or exemplary terms (e.g., etc.) in the present disclosure is merely for the purpose of describing the present disclosure in detail, and the scope of the present disclosure is not limited by such examples or exemplary terms unless limited by the claims. Furthermore, a person skilled in the art will understand that various modifications, combinations, and changes may be made according to design conditions and factors within the scope of the claims or equivalents to which they are added.

[0094] Accordingly, the scope of the present disclosure should not be limited to the embodiments described above, and all scopes equivalent to or equivalently modified from the claims set forth below, as well as the claims set forth below, shall be considered to fall within the scope of the scope of the present disclosure. Explanation of the symbols

[0095] 110: Time series data collection unit 120: Token generation section 130: Prediction value generation unit 140: Prediction error calculation unit 150: Diagnostic Department

Claims

Claim 1 A virtual power plant diagnostic device comprising: a time series data collection unit that collects time series data including at least one of voltage, current, power, frequency, reactive power, active power, temperature, and humidity for a power generation device constituting a virtual power plant through an energy monitoring sensor and a temperature and humidity sensor; a token generation unit that converts the time series data into text-type data and tokenizes the time series data converted into text-type data to generate tokens; a prediction value generation unit that generates a prediction value by predicting a token after the last token among the tokens through a natural language processing-based artificial intelligence model; a prediction error calculation unit that calculates a prediction error by comparing additionally collected time series data with the prediction value; and a diagnostic unit that diagnoses the power generation device as faulty when the prediction error exceeds a preset error threshold. Claim 2 A virtual power plant diagnostic device according to claim 1, wherein the virtual power plant diagnostic device further comprises a text data collection unit that collects text data including at least one of an accident report and a maintenance log related to a short circuit and discharge of the power plant. Claim 3 In paragraph 2, the diagnostic unit generates a diagnostic basis, which is the basis for diagnosis of the power generation device, through a natural language processing-based artificial intelligence model that uses the time series data and the text data as training data, in a virtual power plant diagnostic device. Claim 4 In paragraph 3, the diagnostic unit generates a diagnostic basis corresponding to the time series data based on a pattern between the time series data and the text data, a virtual power plant diagnostic device. Claim 5 In claim 4, the token generation unit comprises: a feature extraction unit that extracts a statistical feature including at least one of a time-by-hour trend, a peak, and an average from the time series data; a repetition period calculation unit that performs matching for each time point where the statistical feature is similar in the time series data and calculates a repetition period based on the difference between the matched times points; and a token conversion unit that converts the statistical feature and the repetition period into data in the form of a string and then converts them into the token, thereby forming a virtual power plant diagnostic device. Claim 6 In claim 5, the token conversion unit generates at least one of an increase, a decrease, a maximum value, and a minimum value as data in the form of a string based on the statistical features and then converts it into the token, a virtual power plant diagnostic device. Claim 7 In claim 6, the prediction value generation unit generates a prediction token by predicting a token after the last token among the tokens, and generates the prediction value by converting the prediction token into time-series data, a virtual power plant diagnostic device. Claim 8 A virtual power plant diagnostic method comprising: collecting time series data including at least one of voltage, current, power, frequency, reactive power, active power, temperature, and humidity for a power generation device constituting a virtual power plant through an energy monitoring sensor and a temperature and humidity sensor; converting the time series data into text-type data and tokenizing the time series data converted into text-type data to generate tokens; generating a predicted value by predicting a token after the last token among the tokens through a natural language processing-based artificial intelligence model; calculating a prediction error by comparing additionally collected time series data with the predicted value; and diagnosing the power generation device as faulty when the prediction error exceeds a preset error threshold. Claim 9 A virtual power plant diagnostic method according to claim 8, further comprising, after the diagnostic step, a step of collecting text data including at least one of an accident report and a maintenance log related to a short circuit and discharge of the power generation device. Claim 10 A virtual power plant diagnosis method according to claim 9, further comprising, after the step of collecting the text data, the step of generating a diagnostic basis which is the basis for diagnosis of the power generation device through the natural language processing-based artificial intelligence model using the time series data and the text data as training data. Claim 11 In claim 10, the step of generating the diagnostic basis comprises generating a diagnostic basis corresponding to the time series data based on a pattern between the time series data and the text data, a virtual power plant diagnostic method. Claim 12 A virtual power plant diagnostic method according to claim 11, wherein the step of generating the token comprises: a process of extracting a statistical feature including at least one of a time-series trend, a peak, and an average from the time series data; a process of performing matching for time points in the time series data where the statistical feature is similar and calculating a repetition period based on the difference between the matched time points; and a process of converting the statistical feature and the repetition period into data in the form of a string and then converting them into the token. Claim 13 In claim 12, the process of converting into the above token is a virtual power plant diagnostic method in which at least one of an increase, decrease, maximum value, and minimum value is generated as data in the form of a string based on the above statistical characteristics and then converted into the above token. Claim 14 In claim 13, the step of generating the predicted value comprises generating a predicted token by predicting a token after the last token among the tokens, and generating the predicted value by converting the predicted token into time-series data. A virtual power plant diagnostic method.

Images