KR102160884B1 - Transformer Design System with Machine Learning - Google Patents

Abstract

The present invention relates to a transformer design system applied with machine learning. In the present invention, after applying all possible values for values to be inputted by a designer, the best design is selected, but machine learning is applied thereto. In the present invention, of the transformer design, all possible designs can be made by applying machine learning, so that the best design to which machine learning is applied can be selected.

Description

Transformer Design System with Machine Learning {Transformer Design System with Machine Learning}

The present invention relates to a transformer design system to which machine learning is applied, and in more detail, in the transformer design, all possible designs can be applied by applying machine learning, so that the best design to which machine learning is applied can be selected. It works.

In general, the transformer design process is shown in FIG. 1 below. Once the user's requirements (efficiency, cost, weight, volume, etc.) are determined, enter the values (LV Turn, magnetic flux density, core cross-sectional area, winding design conductor, etc.) to be determined by the designer. Based on this value, other values ( V/T, number of turns, creepage distance, iron core window, core shape, weight, etc.) can be calculated. These calculated values become a design, check whether the design meets the user's requirements (derive the result), and if not, repeat this process until the desired design is obtained.

Transformer manufacturers design and manufacture transformers in the same process as above, and MS Execl is used to make this efficient.

As an example, in the prior art Patent Publication No. 10-2014-0085750 as a transformer design system,

A control unit that calculates the capacity of a design target transformer by a transformer capacity calculation program and calculates the capacity of the transformer based on inputted transformer design condition data;

A storage unit for storing the transformer capacity calculation program and providing the program to the control unit, and storing data related to the transformer capacity calculation under the control of the control unit;

An input unit for inputting commands and data applied according to manual manipulation to the control unit;

A transformer design system is disclosed, comprising a display unit for displaying a screen for calculating the transformer capacity under control of a controller.

In addition, Patent Application Publication No. 10-2007-0063204 discloses a method for designing a transformer core using DEAS;

The transformer designer inputs design specifications;

Calculating, by DEAS, the copper loss of the primary winding for a design parameter that is instantaneously selected for optimization;

Calculating the total copper loss after calculating the copper loss of the secondary winding;

Calculating a voltage fluctuation rate, a core iron loss, and a core increase temperature to calculate an objective function of the selected design parameter;

Calculating a cost function from the values;

Comparing a cost function value calculated by DEAS with existing searched values to find a next search point;

An optimal design method for a transformer core, characterized in that it comprises the step of outputting an optimum value after satisfying a given DEAS execution condition, is disclosed.

However, the conventional method may have a difference in the number of iterations according to the experience or knowledge of the design expert. And even if it is a design that satisfies the requirements, there is a disadvantage that it cannot be guaranteed as an optimal design.

Therefore, the present invention was conceived to solve the above problems, and in the transformer design, all possible designs can be made by applying machine learning, so that machine learning is applied to select the best design to which machine learning is applied. It is intended to provide a transformer design system.

The present invention relates to a transformer design system to which machine learning is applied. After applying all possible values for values to be entered by the designer, the best design is selected, but machine learning is applied.

Accordingly, the present invention has a remarkable effect of selecting the best design to which machine learning is applied, since it is possible to perform all possible designs by applying machine learning in transformer design.

1 is a general transformer design process diagram of the present invention
2 is a block diagram of neurons in machine learning of the present invention

The present invention relates to a transformer design system to which machine learning is applied. After applying all possible values for values to be entered by the designer, the best design is selected, but machine learning is applied.

In addition, the machine learning model is characterized by comprising an input layer, a hidden layer, and an output layer.

The present invention will be described in detail with reference to the accompanying drawings as follows. 1 is a schematic diagram of a general transformer design process of the present invention, and FIG. 2 is a block diagram of a neuron in machine learning of the present invention.

It is to apply all possible values to the values that the designer must enter. The table below shows the number of possible cases of each value in a 100kV pole transformer design.

Input value range change Number Number of all cases Min Max LV Turn 26 46 2 11



2,083,143,744

Winding design 0.5 One 0.1 6 2 3.6 0.1 18 200 400 5 41 2 3.6 0.1 18 65 100 One 36 Iron core cross section 13000 14000 100 11 One 2 0.2 6

Table 1 shows the number of possible cases of each value in the design of a 100kV pole transformer. For LV Turn, the range is from 26 to 46 companies and varies by 2, so the number of possible cases is 11. Multiplying all the cases of each value is the number of all possible cases, which is about 2.08×09 for a 100kV pole transformer.

In the present invention, design is made for all possible cases. The advantage is that you can try all possible designs, so you can choose the best one.

As another embodiment of the present invention, machine learning is applied to the transformer design.

The machine learning model was defined as follows. The input layer was composed of one neuron, the hidden layer was composed of 5 layers and each layer was composed of 100 neurons, and the output layer was composed of 4 neurons.

Here, the input is the user's requirement, the efficiency, and the output is the values that the designer has to decide. Therefore, when a designer inputs a requirement, the model provides design values that fit that requirement. (In order to simplify the explanation here, the requirements are efficiency, and the inputs by the designer are simplified from 8 to 4 (total area, LV Trun, Lv thickness, Hv thickness).)

For learning data, the designs obtained in Method-1 were used. Structures and examples of learning data are shown in Tables 2 to 4 below.

Item range Input Efficiency Real number with two decimal places
Output Core area Real between 1.0 and 2.0 Lv turn Even numbers from 26 to 46 Lv thickness Real between 0.5 and 1.0 Hv thickness Real between 2.0 and 3.6

Table 2 shows the structure and example of learning data

Input Output Efficiency Core area Lv turn Lv thickness Hv thickness 99.54 1.56 36 0.54 3.01 99.52 1.46 38 2.12 5.0 99.53 1.46 38 2.62 4.94 99.51 1.46 38 2.79 4.42 99.52 1.55 36 2.66 4.11 99.53 1.55 36 2.64 4.42 99.51 1.37 36 2.18 4.36

Table 3 shows the structure of the training data and the model was trained with 100 examples of training data, and the results of the trained model are as follows.

Input Output Efficiency Core area Lv turn Lv thickness Hv thickness 99.54 1.38 36 2.55 4.83 99.53 1.47 38 2.53 4.76 99.52 1.47 38 2.50 4.66 99.51 1.65 38 2.49 4.70 99.50 1.62 42 2.52 4.83

In Table 4, in order to find out the structure of the training data and the accuracy of the results provided by the pre-trained model, the efficiency was calculated using the results and compared with the input efficiency. As shown in the table, it can be seen that this model provides accurate values.

Efficiency as input (A) Efficiency from output (B) Difference
|(AB)| 99.54 99.54 0.0 99.53 99.53 0.0 99.52 99.52 0.0 99.51 99.53 0.02 99.50 99.51 0.01

Table 5 is a comparison table of the efficiency of learning data. In general, there are n inputs and one or m outputs. And it is characterized by handling real numbers with two decimal places. So, in applying machine learning to the transformer design, it provided inaccurate values that could not be used at first, but through model tuning (activation function, number of layers, number of neurons, application of CNN, RNN, etc.) , I created a model that gives the correct values.

In the present invention, since all possible designs can be tried, the best design can be selected.

Therefore, the present invention has a remarkable effect of selecting the best design, as all possible designs can be tried in transformer design.

Claims (2)

After applying all possible values for the values that the designer needs to enter, select the best design, but in the transformer design system to which machine learning is applied,
The transformer is a pole transformer,
The machine learning model is composed of an input layer, a hidden layer, and an output layer, and the input layer is composed of one neuron, and the hidden layer is composed of five layers, each layer having 100 It is composed of four neurons, and the output layer is composed of four neurons,
The input layer is the efficiency that is the user's requirement, and the output layer is the values that the designer needs to determine, so when the designer inputs the requirements, the machine learning model provides design values that meet the requirements. The things to be done are the core area, LV Turn, Lv thickness, Hv thickness of the pole transformer.
The efficiency of the learning data of the machine learning model is one input and multiple outputs, and a transformer design system with machine learning, characterized in that it handles real numbers of two decimal places. delete

Images