KR20220132824A - Distribution facility condition monitoring system and method - Google Patents

Abstract

According to the present invention, a deep learning based system for monitoring abnormal points in power distribution facilities includes: a LSTM time series data learning model based on measurement data accumulated as big data; a data prediction unit which predicts future measurement data based on the data learning model; a distribution information management unit which calculates and stores distribution information for the measurement data accumulated as the big data; a reliability information calculation unit for deriving distribution information of the data predicted by the data prediction unit and calculating cross-correlation coefficient information for the distribution information stored in the distribution information management unit; and a model construction unit for re-learning the data learning model. Therefore, it is possible to quickly and accurately detect a failure of a distribution facility at a low cost by utilizing manager's experience or expertise.

Description

DISTRIBUTION FACILITY CONDITION MONITORING SYSTEM AND METHOD

The present invention relates to a system and method for monitoring an abnormal location in a power distribution facility that can detect and estimate a change in condition of a distribution facility based on deep learning.

Existing processing and underground power distribution facility diagnosis relies on instantaneous diagnosis, so-called time-based maintenance in which a person carries equipment such as ultrasonic detection equipment, partial discharge measurement equipment, and current detector to inspect all facilities regardless of suspicious locations. (Time Based Maintenance), and even that, it was difficult to directly access the overhead power distribution facility, so it was only at the level of checking with the naked eye.

For this purpose, there is a safety problem because the distribution power actually climbs or opens the distribution facility for direct inspection, and it is inefficient because it takes a lot of time and money.

As machine learning has been applied to various fields, there have been attempts to make the above-described power distribution check task into machine learning. For example, as a conventional method for diagnosing partial discharge, a partial discharge diagnosis method using a noise removal sensor has been proposed. In this case, it is advantageous to be implemented with a communication system that can transmit a sensor-based signal for inspection, such as IoT technology.

In the above method, a partial discharge type of sensor data is classified using a convolutional neural network (CNN) based on data received from the sensor data. Then, the model is trained and the pattern is identified using the image accumulated for 1 minute of the PRPD image.

However, as a classification model for the purpose of diagnosing partial discharge types, the CNN model requires a lot of learning data for each partial discharge type for learning and has a large amount of learning. In particular, prior to diagnosing the partial discharge type, a process of accumulating sufficient data by detecting the occurrence of partial discharge with high reliability should be preceded.

As such, even in diagnosing only partial discharge using machine learning techniques, it was difficult to directly apply it to the actual field because of the considerable prerequisites and the burden of prior data accumulation.

Republic of Korea Publication No. 10-2020-0130528

An object of the present invention is to provide a system and method for monitoring an abnormal location in a power distribution facility based on deep learning that can quickly and accurately detect a failure of a distribution facility at a low cost by utilizing the experience or expertise of a manager.

In accordance with an aspect of the present invention, there is provided a system for monitoring an abnormal state of a distribution facility, comprising: an LSTM time series data learning model based on measurement data accumulated as big data; a data prediction unit for predicting future measurement data based on the data learning model; a distribution information management unit for calculating and storing distribution information on the measurement data accumulated as the big data; a reliability information calculation unit for deriving distribution information of the data predicted by the data prediction unit and calculating cross-correlation coefficient information with respect to the distribution information stored in the distribution information management unit; and a model configuration unit for re-learning the data learning model.

Here, the distribution information management unit may derive a quartile as distribution information for the accumulated measurement data and retain it in the form of a DB.

Here, a measurement data acquisition unit that receives measurement values from various sensors installed on power facilities or distribution lines through an assigned communication channel and accumulates related big data in an internal storage means or an external DB server may include more.

Here, the cross-correlation coefficient information may be visually provided to the distribution facility manager, and may further include a user interface for receiving the manager's evaluation.

Here, the model configuration unit may re-learn the data learning model by reflecting the expert determination of the manager through the user interface.

Here, the model configuration unit may re-learning by reflecting the error between the actual measurement data and the prediction data through the LSTM.

Here, the user interface may display a measured value or a predicted value on the screen together with quartile information, which is a combination of values that divide the probability distribution into quarters with respect to the accumulated measurement big data.

Here, the user interface may set a time interval for quartile information, which is a reference for configuring the screen according to an instruction of an administrator.

Here, the LSTM time series data learning model is a time series data analysis method, by decomposing data trends and cycle data, respectively, to detect abnormal measurements, and STR (Seasonal-Trend-Residual) additive model The seasonal factor of can be used instead of the periodic factor.

In accordance with another aspect of the present invention, there is provided a method for monitoring an abnormal state of a power distribution facility, the method comprising: preparing an LSTM time series data learning model to which measurement data of a power system is applied; collecting measurement data of the power system as big data; predicting future measurement data using the LSTM time series data learning model trained with the collected measurement data; calculating quantile information as distribution information on the measurement data accumulated with the big data and converting it into a DB; providing information comparing the quantile information with predicted measurement data or actual measurement data to a manager; and retraining the LSTM time series data learning model according to the mutually compared information.

Here, in the DB forming step, quartile information may be calculated for the measured values belonging to a specified time interval, and the calculated quartile information may be stored in the DB.

Here, in the step of collecting the measurement data as big data, measurement values may be input from various measurement devices installed in the power distribution facility through the IoT sensor network.

Here, the method may further include the step of receiving an administrator's decision through a user interface, and in the re-learning of the learning model, the data learning model may be re-trained by reflecting the administrator's decision.

Here, in the step of providing the mutually compared information to the manager, the measured value or the predicted value may be displayed on the screen together with the quartile information, which is a combination of values that divide the probability distribution into quarters with respect to the accumulated measured big data.

Here, by using the information collected in the step of collecting as the big data, the step of forming the DB and the step of predicting the future measurement data may be performed in parallel.

When the deep learning-based power distribution facility status abnormality monitoring system and/or method according to the spirit of the present invention having the above configuration is implemented, the failure of the distribution facility is quickly and accurately detected at a low cost by utilizing the manager's experience or expertise There are advantages to being able to

The deep learning-based power distribution facility status abnormality monitoring system and/or method of the present invention has the advantage of reducing power saving costs by precisely detecting and promptly recovering the faulty section of the power system, particularly including the underground section.

The deep learning-based power distribution facility status abnormality monitoring system and/or method of the present invention has the advantage of reducing instantaneous, inspection, and diagnosis costs through real-time online status monitoring.

1 is a block diagram illustrating a system for monitoring an abnormal state of a power distribution facility according to an embodiment of the present invention.
2 is a graph showing aspects according to time series of measurement signals;
3 is a DB of the accumulated wiring equipment measurement big data into quartiles (quartile deviation coefficient), and displays the measured values together with quartile information, which is a combination of values that divide the probability distribution into quarters on an arbitrary random variable axis. An example screen of the form.
4 is a flowchart illustrating a method for monitoring an abnormal state of a power distribution facility according to the spirit of the present invention.
FIG. 5 is a conceptual diagram illustrating a method for monitoring an abnormal state of a distribution facility performed in the system for monitoring an abnormal state of a distribution facility of FIG. 1 , focusing on a data flow generated according to each step; FIG.

In describing the present invention, terms such as first, second, etc. may be used to describe various components, but the components may not be limited by the terms. The terms are only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component.

When a component is referred to as being connected or connected to another component, it may be directly connected or connected to the other component, but it can be understood that other components may exist in between. .

The terms used herein are used only to describe specific embodiments, and are not intended to limit the present invention. The singular expression may include the plural expression unless the context clearly dictates otherwise.

In this specification, the terms include or include are intended to designate that a feature, number, step, operation, component, part, or combination thereof described in the specification exists, and includes one or more other features or numbers, It may be understood that the possibility of the presence or addition of steps, operations, components, parts, or combinations thereof is not precluded in advance.

In addition, shapes and sizes of elements in the drawings may be exaggerated for clearer description.

The present invention uses Long Short-Term Memory models (LSTM) and quantile cross-correlation analysis, which are a kind of deep learning-based time series data learning technique RNN, which are recently utilized in various fields, as a manager with rich expertise and/or experience. We present a support system that allows us to identify measurement outliers in distribution facilities and verify data.

Wireless diagnostic sensors in the Internet of Things data collection environment are attached to power distribution facilities (poles, ground switchgear, ground transformer, pole transformer), and each measurement signal (temperature, humidity, current, 3-axis vibration, phase current, leakage current, UV light) at regular intervals , ultrasound, etc.) to the server. The wireless diagnostic sensor may transmit, for example, a measurement signal to the server through an IoT sensor network.

In the present invention, the LSTM technique is used to estimate future data based on pre-measured and accumulated time series data. LSTM is a type of deep learning-based artificial neural network, unlike general artificial neural networks, iteratively learns by itself so that the information obtained in the previous step continuously affects it, learns using time series data, and estimates the next value.

In order to verify the advantages of the present invention, the correlation and similarity of measurement data of the IoT-based Mussen sensor for estimating the state of the pole transformer is analyzed in a way that analyzes time series data, for example, state measurement data for a month Through the task of estimating data for the next two days by learning, the measured values with high dependence on temperature were distinguished, and factors with trend and periodicity were removed to determine the exact abnormal (failure) value.

According to the main idea of the present invention, it is possible to reduce the size of data by configuring quantiles for each predetermined time for efficient operation and storage of generated/accumulated big data. The quantile is a standard number when dividing the entire data included in the range into a specific number. In the case of real-time data whose form of measurement data is outside the normal distribution or the degree of variance is large, the average alone cannot represent the given data. It is a representative value used.

The size of the quantile varies depending on how many total data are divided.

In order to compare LSTM-estimated data and quantiles of measured data for each location in the DB, after changing the estimated data to the same quantile, the cross-correlation coefficient is calculated to detect outliers and the distribution and estimated data of measured data so far. Compare the changes in their distribution.

Next, the data estimated by the LSTM is updated and re-learned through comparison with the actually measured data.

1 is a block diagram illustrating a system for monitoring an abnormal state of a power distribution facility according to an embodiment of the present invention.

The illustrated power distribution facility status abnormality monitoring system includes: an LSTM time series data learning model 130 based on measurement data accumulated as big data; a data prediction unit 140 for predicting future measurement data based on the data learning model; a quantile information management unit 150 for calculating and storing quantiles as distribution information on the measurement data accumulated with the big data; a reliability information calculation unit 160 for deriving a quantile as distribution information of the data predicted by the data prediction unit and calculating cross-correlation coefficient information for distribution information based on an actual measurement value stored in the distribution information management unit; and a model configuration unit 120 for re-learning the data learning model by reflecting the error between the actual measurement data and the prediction data through the LSTM.

As shown, a measurement data acquisition unit 110 for collecting measurement data that can be accumulated as big data; and a user interface 180 that visually provides the cross-correlation coefficient information to the distribution facility manager and receives the manager's evaluation.

The measurement data acquisition unit 110 receives measurement values from various sensors installed on power facilities or distribution lines, such as partial discharge sensors, through an assigned communication channel (eg, IoT sensor network), and receives internal storage means or Relevant big data can be accumulated by storing it in an external DB server.

According to implementation, the LSTM time series data learning model 130 is a time series data analysis method, by decomposing a trend data and a cycle data of data, respectively, to detect an anomaly measured value, and STR (Seasonal-Trend) -Residual) The seasonal factor of the additive model can be used instead of the periodic factor.

The data prediction unit 140 estimates an LSTM time series data learning model based on the accumulated measurement big data and future measurement data based thereon. To this end, the data prediction unit 140 uses the LSTM time series data learning model 130 in performing data prediction, and may be implemented in the form of a SW module integrated with the LSTM time series data learning model 130 . have.

The LSTM time series data learning model 130 and the data prediction unit 140 apply measurement values from various sensors as inputs and apply future expected measurement values of various sensors as outputs, the model The configuration unit 120 evaluates the prediction of the data prediction unit 140 using the LSTM time series data learning model 130 in the past with the measurement value of the currently measured specific sensor, and the LSTM time series data learning model 130 . can be relearned.

The quantile information management unit 150 compresses the wiring equipment measurement big data accumulated by the measurement data acquisition unit 110 for quick real-time analysis and HW burden reduction, more specifically, relative scatter information, by applying quantile information. (summary) do. In summary, the quantile information management unit 150 calculates and stores quantiles for the measurement data.

2 is a graph illustrating aspects according to time series of measurement signals.

It can be seen that the items shown in FIG. 2 are temperature, ultrasound, and ultraviolet light, and abnormal data detection is possible through time series data analysis.

3 is a DB of the accumulated wiring equipment measurement big data into quartiles (quartile deviation coefficient), and displays the measured values together with quartile information, which is a combination of values that divide the probability distribution into quarters on an arbitrary random variable axis. This is an example screen of the form.

As shown, the position of the measured value of the corresponding parameter measured by the corresponding sensor is relatively displayed along with the quartile information for each sensor and each measurement parameter. In the illustrated case, it corresponds to an abnormal (failure) state value and is displayed in the form of a singular value higher than the maximum value.

To this end, the quantile information management unit 150 determines a time section to be analyzed as a quartile, classifies values belonging to the determined time section among the accumulated wiring equipment measurement big data, and as quartile information for each sensor And, for each measurement parameter, the maximum value, the minimum value, the median value, the first quartile, and the third quartile may be calculated and stored.

According to the implementation, not only the value measured immediately before by the corresponding sensor, but also the expected value for the corresponding parameter of the corresponding sensor (that is, the predicted value of the data prediction unit 140) is displayed together with the quartile information as shown. can do.

In this case, in addition to the quartile information shown based on the big data accumulated for a relatively long period of time, the quartile information of the modeling reflection data derived based on the recently accumulated big data of a relatively short period used to generate the corresponding prediction value is added. can be displayed

The reliability information calculating unit 160 is for generating an evaluation for re-learning of the LSTM time series data learning model 130 based on quantile information, and the distribution (quantiles) of predicted data through the LSTM and actual measured values. Based on the distribution of data based on , a cross-correlation coefficient capable of comparing the quantiles of two groups may be calculated, and this may be provided to the model configuration unit 120 . Then, the model configuration unit 120 may perform re-learning according to the cross-correlation coefficient.

The user interface 180 is for receiving information provided by a manager as a multi-experienced person and/or a technical expert in the power system. For example, the user interface 180 is a screen similar to that shown in FIG. 2 , and displays the predicted measurement value and quartile information together, so that the administrator with respect to the predicted measurement value has his/her expertise and/or experience can be used to evaluate its adequacy.

In addition, in determining the abnormal (failure) state value for the predicted value or the measured value, the administrator's expert confirmation (determination) through the user interface 180 may be reflected.

According to implementation, the administrator may set a time interval for quartile information as a reference for composing the screen of FIG. 2 through the user interface 180 . In this case, the administrator primarily observes the measured (or predicted) value displayed along with the quartile information according to the time interval set as a default, and if you want to compare it in a wider range, select the time interval for the quartile information. Adjust widely to look at the measured (or predicted) value with quartile information in a wide time, or adjust the time interval for the quartile information to a recently narrower measurement (or prediction) with quartile information in a narrow recent time ) can be viewed.

4 is a flowchart illustrating a method for monitoring an abnormal state of a power distribution facility according to the spirit of the present invention.

The illustrated method for monitoring an abnormal state of a power distribution facility includes the steps of preparing an LSTM time series data learning model to which measurement data of a power system is applied (S10); Collecting measurement data of the power system as big data (S110); predicting future measurement data using the LSTM time series data learning model trained with the collected measurement data (S140); Calculating quantile information as distribution information on the measurement data accumulated with the big data and converting it into a DB (S150); providing information comparing the quantile information with predicted measurement data or actual measurement data to a manager (S160); and retraining the LSTM time series data learning model according to the mutually compared information (S190).

In the step of preparing the LSTM time series data learning model (S10), the LSTM time series data learning model 130 adapted to the distribution facility to be monitored is stored in the storage device of the distribution facility status abnormality monitoring system of FIG. 1 . can be done in this way.

In the step of collecting the measurement data as big data (S110), the measurement values for the corresponding measurement parameters (eg, temperature, ultrasonic waves, ultraviolet rays) are transmitted from various measurement equipment installed in the distribution facility to be monitored through a predetermined communication channel ( For example, it can be performed by receiving input through an IoT sensor network).

Predicting the future measurement data ( S140 ) may be performed by applying the LSTM time series data learning model by the data prediction unit 140 of FIG. 1 .

The DB forming step ( S150 ) is performed by the quantile information management unit 150 of FIG. 1 , and may be performed in parallel with the step ( S140 ) of predicting the future measurement data.

In the DB forming step ( S150 ), quartile information may be calculated for the measured values belonging to a time period designated in advance by an administrator or a default unit time period, and the calculated quartile information may be stored in the DB. The quartile information stored in the DB has a small amount of data (that is, it can be viewed as a kind of compressed information), so the transmission burden is small when necessary, and the processing burden in subsequent processing such as analysis is also small. In the case of an implementation to reduce the DB itself, the original measured values are not stored in the DB at all, but can be stored in the DB only in the form of quartile information calculated from the measured values belonging to a predetermined unit time section.

According to implementation, a time interval for calculating the quartile information may be set according to an instruction from an administrator through the user interface 180 of FIG. 1 .

The step S160 may be performed in such a way that the information generated by the reliability information calculating unit 160 of FIG. 1 is provided to the manager through the user interface 180 on a screen. For example, as shown in FIG. 3 , the measured value may be displayed together with quartile information. In this case, when the measured value is out of a preset area, it can be expressed as a singular value as shown.

The step S190 may be performed by the model configuration unit 120 of FIG. 1 , and depending on the implementation, the administrator's evaluation result of the predicted value of the LSTM time series data learning model 130 may be reflected.

FIG. 5 is a conceptual diagram illustrating a method for monitoring an abnormal state of a distribution facility performed in the system for monitoring an abnormal state of a distribution facility of FIG. 1 , focusing on the flow of data generated according to each step rather than the flow of time.

As shown, using the information (data) collected in the step (S110) of collecting as the big data, the step of making the DB (S150) and the step of predicting the future measurement data (S140) are performed in parallel can be found

On the other hand, the step of providing the manager with information comparing the quantile information and the predicted measurement data or the actual measurement data (S160), the step of forming the DB (S150) and predicting the future measurement data performed in parallel It can be seen that the method of comparing the respective results of (S140) is performed. In addition, it can be seen that the performance result of step S160 is used in the re-learning process (S190) for the LSTM time series data learning model prepared in step S10.

Those skilled in the art to which the present invention pertains should understand that the present invention may be embodied in other specific forms without changing the technical spirit or essential characteristics thereof, so the embodiments described above are illustrative in all respects and not restrictive. only do The scope of the present invention is indicated by the following claims rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present invention. .

110: measurement data acquisition unit
120: model component part
130: LSTM time series data learning model
140: data prediction unit
150: quantile information management unit
160: reliability information calculation unit
180 : User Interface

Claims (15)

LSTM time series data learning model based on measurement data accumulated as big data;
a data prediction unit for predicting future measurement data based on the data learning model;
a distribution information management unit for calculating and storing distribution information on the measurement data accumulated as the big data;
a reliability information calculation unit for deriving distribution information of the data predicted by the data prediction unit and calculating cross-correlation coefficient information with respect to the distribution information stored in the distribution information management unit; and
A model construction unit for re-learning the data learning model
A monitoring system for abnormal conditions in distribution facilities, including a.
According to claim 1,
The distribution information management unit,
A distribution facility status abnormality monitoring system that derives quartiles as distribution information for the accumulated measurement data and holds them in the form of a DB.
According to claim 1,
Measurement data acquisition unit that receives measurement values from various sensors installed on power facilities or distribution lines through an assigned communication channel, and accumulates related big data by storing it in an internal storage means or an external DB server
Distribution facility status abnormality monitoring system further comprising a.
According to claim 1,
A user interface that visually provides the cross-correlation coefficient information to the distribution facility manager and receives the manager's evaluation
Distribution facility status abnormality monitoring system further comprising a.
5. The method of claim 4,
The model component,
A distribution facility status abnormality monitoring system for re-learning the data learning model by reflecting the expert judgment of the manager through the user interface.
5. The method of claim 4,
The model component,
A monitoring system for abnormal conditions in distribution facilities that re-learns by reflecting the error between the actual measured data and the predicted data through LSTM.
5. The method of claim 4,
The user interface is
For the accumulated measurement big data, it is a distribution facility abnormality monitoring system that displays the measured or predicted values along with the quartile information, which is a combination of values that divide the probability distribution into quarters.
8. The method of claim 7,
The user interface is
A distribution facility status abnormal location monitoring system that sets a time section for quartile information, which is a standard for composing the screen according to an administrator's instruction.
According to claim 1,
The LSTM time series data learning model is
As a time series data analysis method, anomalies are detected by decomposing data trend data and cycle data, respectively, and seasonal factors of STR (Seasonal-Trend-Residual) additive model can be used instead of periodic factors. A monitoring system for abnormal conditions in distribution facilities.
preparing an LSTM time series data learning model to which the measurement data of the power system is applied;
collecting measurement data of the power system as big data;
predicting future measurement data using the LSTM time series data learning model trained with the collected measurement data;
calculating quantile information as distribution information on the measurement data accumulated with the big data and converting it into a DB;
providing information comparing the quantile information with predicted measurement data or actual measurement data to a manager; and
retraining the LSTM time series data learning model according to the mutually compared information
A method for monitoring an abnormal state of a distribution facility comprising a.
11. The method of claim 10,
In the step of forming the DB,
A method for monitoring abnormalities in distribution facilities by calculating quartile information for measured values belonging to a specified time period and storing the calculated quartile information in a DB.
11. The method of claim 10,
In the step of collecting measurement data as big data,
A method for monitoring abnormalities in distribution facilities by receiving measurement values from various measuring devices installed in distribution facilities through IoT sensor network.
11. The method of claim 10,
Further comprising the step of receiving the decision of the administrator through the user interface,
In the step of re-learning the learning model,
A method for monitoring an abnormal state of a power distribution facility for re-learning the data learning model by reflecting the decision of the manager.
11. The method of claim 10,
In the step of providing mutually compared information to the manager,
For the accumulated measurement big data, it is a method of monitoring abnormalities in the distribution facility that displays the measured or predicted values on the screen together with the quartile information, which is a combination of values that divide the probability distribution into quarters.
11. The method of claim 10,
By using the information collected in the step of collecting as the big data, the step of forming the DB and the step of predicting the future measurement data are performed in parallel to the monitoring method of the abnormal location of the power distribution facility.

Images