/>

Transformer
Time
H₂
CH₄
C₂H₄
C₂H₆
C₂H₂
CO

MTR001
2013
8
0
7
1
5
156

2015
0
0
10.1
4.1
8.6
226.9

2015
0
0
0.7
1.3
2.2
36.2

2017
5.54
0
7.14
2.19
4.4
132.24

2018
5.9
0
15.8
1
8.5
248.6

2019
5.3
0
9.6
7.5
16.9
171.8

It is seen that the three kinds of gases including methane (CH₄), ethylene (C₂H₄), and ethane (C₂H₆) are reduced to 60% or less. The three kinds of gases may be used as the oil-filtering determination gas 250 to determine whether oil filtering is performed.

The oil-filtering determination gas 250 is exemplary, and another oil-filtering determination gas 250 may be used to determine whether oil filtering is performed. Such an embodiment may also fall within the scope of the present invention.

FIG. 3 is a conceptual diagram illustrating a method of predicting a state of a transformer on the basis of machine learning according to the exemplary embodiment of the present invention.

FIG. 3 shows a method of training a prediction model such as an LSTM model.

Referring to FIG. 3, training data may be separately transferred to generate a first prediction model (oil filtering) 300 and a second prediction model (oil non-filtering) 350.

Training data (oil filtering) including dissolved gas data and transformer state data of a case in which oil filtering is performed may be input so that the first prediction model (oil filtering) 300 may be generated. Training data (oil non-filtering) including dissolved gas data and transformer state data of a case in which oil filtering is not performed may be input so that the second prediction model (oil non-filtering) 350 may be generated.

According to the exemplary embodiment of the present invention, each prediction model may include the following sub-prediction models.

First, the first prediction model (oil filtering) 300 may include a plurality of sub-prediction models (e.g., a first sub-prediction model (oil filtering) 310, a second sub-prediction model (oil filtering) 320, a third sub-prediction model (oil filtering) 330, and a fourth sub-prediction model (oil filtering) 340. Each of the plurality of sub-prediction models (oil filtering) may be a prediction model for predicting each individual state of a transformer.

The second prediction model (oil non-filtering) 350 may include a plurality of sub-prediction models (e.g., a first sub-prediction model (oil non-filtering) 360, a second sub-prediction model (oil non-filtering) 370, a third sub-prediction model (oil non-filtering) 380, and a fourth sub-prediction model (oil non-filtering) 390. Each of the plurality of sub-prediction models (oil non-filtering) may be a prediction model for predicting each individual state of a transformer.

In the case of determining one of the four states of a transformer, the first to third sub-prediction models (oil filtering) 310 to 330 or the first to third sub-prediction models (oil non-filtering) 360 to 380 may be used, and the fourth sub-prediction model (oil filtering) 340 or the fourth sub-prediction model (oil non-filtering) 390 may not be included in the prediction models. The reason is that when a state of a transformer is not predicted on the basis of the third sub-prediction model (oil filtering) 330 or the third sub-prediction model (oil non-filtering) 380, the last transformer state on which a determination will be made is predicted as a subsequent transformer state without the fourth sub-prediction model (oil filtering) 340 or the fourth sub-prediction model (oil non-filtering) 390. In the exemplary embodiment of the present invention, it is assumed for convenience of description that the fourth sub-prediction model (oil filtering) 340 or the fourth sub-prediction model (oil non-filtering) 390 are present.

It may be assumed that transformer states are classified as the first state (normal), the second state (warning), the third state (critical), and the fourth state (fault).

Each sub-prediction model may be generated through individual training and used to predict a state of each transformer.

A method of training the sub-prediction models included in the first prediction model (oil filtering) 300 or the second prediction model (oil non-filtering) 350 will be described below with reference to FIGS. 4 and 5. A transformer state determined by each individual sub-prediction model is exemplary and may vary.

Specifically, a sequence, in which the individual sub-prediction models included in each of the first prediction model (oil filtering) 300 and the second prediction model (oil non-filtering) 350 are arranged to sequentially determine a transformer state, may vary. For example, a sequence in which the first sub-prediction model (oil filtering) 310 to the fourth sub-prediction model (oil filtering) 340 included in the first prediction model (oil filtering) 300 are arranged may be adaptively changed in consideration of prediction accuracy. Likewise, a sequence in which the first sub-prediction model (oil non-filtering) 360 to the fourth sub-prediction model (oil non-filtering) 390 included in the second prediction model (oil non-filtering) 350 are arranged may be adaptively changed.

FIG. 4 is a conceptual diagram illustrating a method of generating a first prediction model (oil filtering) according to an exemplary embodiment of the present invention.

FIG. 4 shows a method of generating a plurality of sub-prediction models included in the first prediction model (oil filtering). The number of sub-prediction models may be exemplary.

Referring to FIG. 4, the plurality of sub-prediction models may include a first sub-prediction model (oil filtering), a second sub-prediction model (oil filtering), a third sub-prediction model (oil filtering), and a fourth sub-prediction model (oil filtering). Each of the sub-prediction models (oil filtering) may be a model which is individually trained to distinguish one state from other states. Each of the plurality of sub-prediction models may be an LSTM model which performs predictions in time series.

The first sub-prediction model (oil filtering) 400 may be a model for distinguishing the third state (critical) from others, the first state (normal), the second state (warning), and the fourth state (fault). To train the first sub-prediction model (oil filtering) 400, a first oil-filtering data set 410 and a second oil-filtering data set 420 may be input as training data. Specifically, the first oil-filtering data set 410 may include dissolved gas data and transformer state data with which a transformer is determined to be in the third state (critical), and the second oil-filtering data set 420 may include dissolved gas data and transformer state data with which a transformer is determined to be in the first state (normal), the second state (warning), or the fourth state (fault). The first oil-filtering data set 410 and the second oil-filtering data set 420 may include dissolved gas data and transformer state data of a case in which oil filtering is performed.

Specifically, the first oil-filtering data set 410 may be annual dissolved gas data of a transformer, which is determined to be in the third state (critical) in the fifth year, during a period from the first year to the fourth year and annual transformer state data of the transformer during a period from the second year to the fifth year. The second oil-filtering data set 420 may be annual dissolved gas data of a transformer, which is determined to be in the first state (normal), the second state (warning), or the fourth state (fault) in the fifth year, during a period from the first year to the fourth year and annual transformer state data of the transformer during a period from the second year to the fifth year.

Dissolved gas data of a specific year is input to each of the n layers as an input value, and transformer state data of the subsequent year is input as an output value so that the n layers may be modeled.

The first oil-filtering data set 410 and the second oil-filtering data set 420, which are separately labeled and correspond to the case in which oil filtering is performed, may be input to the first sub-prediction model (oil filtering) 400.

As a detailed example, {training data of the year 2013, training data of the year 2014, training data of the year 2015, training data of the year 2016, training data of the year 2017} included in the first oil-filtering data set 410 may be labeled with first information and input to the first sub-prediction model (oil filtering) 400. The training data may include annual dissolved gas data and annual transformer state data of a transformer which is determined to be in the third state (critical) in 2017.

Dissolved gas data of the year 2013 may be input to x_t-3, transformer state data of the year 2014 may be input to y_t-3, dissolved gas data of the year 2014 may be input to x_t-2, transformer state data of the year 2015 may be input to y_t-2, dissolved gas data of the year 2015 may be input to x_t-1, transformer state data of the year 2016 may be input to y_t-1, dissolved gas data of the year 2016 may be input to x_t, and transformer state data of the year 2017 may be input to y_t. In the first oil-filtering data set 410, the transformer state data of the year 2017 in y_tmay be the third state (critical).

Likewise, {training data of the year 2013, training data of the year 2014, training data of the year 2015, training data of the year 2016, training data of the year 2017} included in the second oil-filtering data set 420 may be labeled with second information and input to the first sub-prediction model (oil filtering) 400. The training data may include annual dissolved gas data and annual transformer state data of a transformer which is determined to be in the first state (normal), the second state (warning), or the fourth state (fault) in 2017.

Dissolved gas data of the year 2013 may be input to x_t-3, transformer state data of the year 2014 may be input to y_t-3, dissolved gas data of the year 2014 may be input to x_t-2, transformer state data of the year 2015 may be input to y_t-2, dissolved gas data of the year 2015 may be input to x_t-1, transformer state data of the year 2016 may be input to y_t-1, dissolved gas data of the year 2016 may be input to x_t, and transformer state data of the year 2017 may be input to y_t. In the second oil-filtering data set 420, the transformer state data of the year 2017 in y_tmay be the first state (normal), the second state (warning), or the fourth state (fault).

In this way, the first sub-prediction model (oil filtering) 400 may learn weights for distinguishing the third state (critical) from the other states on the basis of the first oil-filtering data set 410 labeled with the first information and the second oil-filtering data set 420 labeled with the second information.

Likewise, the second sub-prediction model (oil filtering) may learn weights for distinguishing the fourth state (fault) from the other states (the first state (normal), the second state (warning), and the third state (critical)) on the basis of a third oil-filtering data set labeled with the first information and a fourth oil-filtering data set labeled with the second information.

The third sub-prediction model (oil filtering) may learn weights for distinguishing the first state (normal) from the other states (the second state (warning), the third state (critical), and the fourth state (fault)) on the basis of a fifth oil-filtering data set labeled with the first information and a sixth oil-filtering data set labeled with the second information.

The fourth sub-prediction model (oil filtering) may learn weights for distinguishing the second state (warning) from the other states (the first state (normal), the third state (critical), and the fourth state (fault)) on the basis of a seventh oil-filtering data set labeled with the first information and an eighth oil-filtering data set labeled with the second information.

As described above, an arrangement sequence of sub-prediction models may be set differently in a prediction model depending on accuracy in determination, and such an embodiment may also fall within the scope of the present invention.

FIG. 5 is a conceptual diagram illustrating a method of generating a second prediction model (oil non-filtering) according to an exemplary embodiment of the present invention.

FIG. 5 shows a method of generating a plurality of sub-prediction models included in the second prediction model (oil non-filtering). The number of sub-prediction models may be exemplary.

Referring to FIG. 5, the plurality of sub-prediction models may include a first sub-prediction model (oil non-filtering), a second sub-prediction model (oil non-filtering), a third sub-prediction model (oil non-filtering), and a fourth sub-prediction model (oil non-filtering). Each of the sub-prediction models (oil non-filtering) may be a model which is individually trained to distinguish one state from other states. Each of the plurality of sub-prediction models may be an LSTM model which performs predictions in time series.

The first sub-prediction model (oil non-filtering) 500 may be a model for distinguishing the first state (normal) from others, the second state (warning), the third state (critical), and the fourth state (fault). To train the first sub-prediction model (oil non-filtering) 500, a first oil non-filtering data set 510 and a second oil non-filtering data set 520 may be input as training data. Specifically, the first oil non-filtering data set 510 may include dissolved gas data and transformer state data with which a transformer is determined to be in the first state (normal), and the second oil non-filtering data set 520 may include dissolved gas data and transformer state data with which a transformer is determined to be in the second state, the third state, or the fourth state. The first oil non-filtering data set 510 and the second oil non-filtering data set 520 may include dissolved gas data and transformer state data of a case in which oil filtering is not performed.

Specifically, the first oil non-filtering data set 510 may be annual dissolved gas data of a transformer, which is determined to be in the first state (normal) in the fifth year, during a period from the first year to the fourth year and annual transformer state data of the transformer during a period from the second year to the fifth year. The second oil non-filtering data set 520 may be annual dissolved gas data of a period from the first year to the fourth year of a transformer, which is determined to be in the second state (warning), the third state (critical), or the fourth state (fault) in the fifth year, and annual transformer state data of the transformer during a period from the second year to the fifth year.

The training data may include results of collecting information on the six kinds of dissolved gases and component ratio information of the six kinds of dissolved gases n times and results of making a diagnosis n times, and n layers may be trained on the basis of the training data. Dissolved gas data of a specific year is input to each of the n layers as an input value, and transformer state data of the subsequent year is input as an output value so that the n layers may be modeled.

The first oil non-filtering data set 510 and the second oil non-filtering data set 520 which are separately labeled and correspond to the case in which oil filtering is not performed may be input to the first sub-prediction model (oil non-filtering) 500.

As a detailed example, {training data of the year 2013, training data of the year 2014, training data of the year 2015, training data of the year 2016, training data of the year 2017} included in the first oil non-filtering data set 510 may be labeled with first information and input to the first sub-prediction model (oil non-filtering) 500. The training data may include annual dissolved gas data and annual transformer state data of a transformer which is determined to be in the first state (normal) in 2017.

Dissolved gas data of the year 2013 may be input to x_t-3, transformer state data of the year 2014 may be input to y_t-3, dissolved gas data of the year 2014 may be input to x_t-2, transformer state data of the year 2015 may be input to y_t-2, dissolved gas data of the year 2015 may be input to x_t-1, transformer state data of the year 2016 may be input to y_t-1, dissolved gas data of the year 2016 may be input to x_t, and transformer state data of the year 2017 may be input to y_t. In the first oil non-filtering data set 510, the transformer state data of the year 2017 in y_tmay be the first state (normal).

Likewise, {training data of the year 2013, training data of the year 2014, training data of the year 2015, training data of the year 2016, training data of the year 2017} included in the second oil non-filtering data set 520 may be labeled with second information and input to the first sub-prediction model (oil non-filtering) 500. The training data may include annual dissolved gas data and annual transformer state data of a transformer which is determined to be in the second state (warning), the third state (critical), or the fourth state (fault) in 2017.

Dissolved gas data of the year 2013 may be input to x_t-3, transformer state data of the year 2014 may be input to y_t-3, dissolved gas data of the year 2014 may be input to x_t-2, transformer state data of the year 2015 may be input to y_t-2, dissolved gas data of the year 2015 may be input to x_t-1, transformer state data of the year 2016 may be input to y_t-1, dissolved gas data of the year 2016 may be input to x_t, and transformer state data of the year 2017 may be input to y_t. In the second oil non-filtering data set 520, the transformer state data of the year 2017 in y_tmay be the second state (warning), the third state (critical), or the fourth state (fault).

In this way, the first sub-prediction model (oil non-filtering) 500 may learn weights for distinguishing the first state (normal) from the other states on the basis of the first oil non-filtering data set 510 labeled with the first information and the second oil non-filtering data set 520 labeled with the second information.

Likewise, the second sub-prediction model (oil non-filtering) may learn weights for distinguishing the second state (warning) from the other states (the first state (normal), the third state (critical), and the fourth state (fault)) on the basis of a third oil non-filtering data set labeled with the first information and a fourth oil non-filtering data set labeled with the second information.

The third sub-prediction model (oil non-filtering) may learn weights for distinguishing the third state (critical) from the other states (the first state (normal), the second state (warning), and the fourth state (fault)) on the basis of a fifth oil non-filtering data set labeled with the first information and a sixth on-oil-filtering data set labeled with the second information.

The fourth sub-prediction model (oil non-filtering) may learn weights for distinguishing the fourth state (fault) from the other states (the first state (normal), the second state (warning), and the third state (critical)) on the basis of a seventh oil non-filtering data set labeled with the first information and an eighth oil non-filtering data set labeled with the second information.