KR102143757B1 - Wind Power Generator incorporating therein AI Technology, including Deep Learning - Google Patents

Abstract

The present invention can provide an excellent control method of a wind power generator capable of producing more power than a mechanical and electronic control system in a wind power reality in our country where the wind direction and speed frequently change.
The present invention analyzes the state information of the wind turbine generator and controls the pitch angle, TSR, and the like using deep learning, thereby increasing the power produced by the wind turbine generator by 30%.

Description

Wind Power Generator incorporating therein AI Technology, including Deep Learning}

The present invention relates to a wind power generator using artificial intelligence, in particular, deep learning, and in particular, in order to increase the power produced by the wind power generator, the pitch of the wind power generator corresponding to the environmental variable or the power sale point, etc. It is about how to control using intelligence.

Sustainable renewable energy, which can lead economic growth while preserving the global environment, has been widely distributed in our daily lives. Renewable energy is an essential element of the new paradigm of'low carbon society', and it is demanding many changes in Korea's economic system based on fossil fuels. Since Korea imports more than 90% of its energy resources and produces 60% of its electric energy using fossil fuels, there is a need for improvement in many areas that require changes in the basic system.

A lot of research is being conducted to use solar power, tidal power, and geothermal heat, and power generation devices using wind power are also being spread in Korea. Wind power generation in Korea began to be supplied as an independent power source for remote islands after the oil shock in the 1970s. Since the 1990s, a 160kW-class pilot complex has been built, and research for demonstration and localization has been ongoing since then.

Wind power generation devices around the world have been concentrated in a small number of countries, mainly in Europe, such as the United States, Germany, and Spain, but since 2000, the market has been expanding to Asia such as China and India.

The first and foremost development goal targeted by manufacturers of wind power systems around the world is to reduce the cost of energy. Wind power generators are devices that convert wind kinetic energy into mechanical energy and then convert it back into electrical energy. It is a method of operating the power generation device after converting the windmill rotating by wind speed and the appropriate gear ratio. In countries where the four seasons are distinct, such as Korea, the change in wind speed as well as the direction of the wind is very large depending on the season, so it is most important to increase energy conversion efficiency through proper control of the wind speed generator.

Patent Document 1 relates to a method for controlling a wind power generator, comprising: a wind fan rotating unit having a wind fan rotating shaft; A gearbox having a wind fan connection shaft connected to the wind fan rotation shaft and a power generation device connection shaft; The control method of a wind power generator comprising a wind generator connected to the rotation shaft of the power generator, wherein the power generator connection shaft can be connected to and disconnected from the wind fan connection shaft by an electronic clutch, and the wind fan connection shaft and the In a state in which the power generation device connection shaft is released, a second power transmission unit for transmitting power from the first gear of the wind power fan connection shaft to the second gear of the power generation device connection shaft is provided, and the second power transmission unit is for the first gear. It is configured to be able to adjust the speed increase ratio of the second gear in a plurality of steps, and adjusts the speed increase ratio in a plurality of steps in response to the rotational speed of the wind fan connection shaft.

Patent Document 1 describes that it is possible to provide a control method of a wind power generation device with artificial intelligence, but only a control method for the mechanical configuration of the wind power generation device is described, which is specifically controlled using artificial intelligence. It does not describe the method.

Patent Document 2 relates to a wind turbine generator and a wind turbine generator, wherein the wind turbine generator comprises: a wind fan connection shaft connected to a wind fan rotation shaft at a front end of the speed reducer; A first speed sensor for sensing a rotation speed of the wind fan connection shaft; A first power generation device formed integrally with the wind fan connection shaft to generate electricity by rotation of the wind fan connection shaft; A first gear integrally formed with the wind fan connection shaft; A second power generation device for generating electricity by receiving power from the first gear; A first power transmission unit that is power-connected to the first gear to transmit power to the second power generation device; A power generation device connection shaft connected to the wind power generator at the rear end of the gearbox; An electronic clutch including a first disk connected to the wind fan connection shaft and a second disk connected to the power generation device connection shaft; A second gear integrally formed with the power generation device connection shaft; A second power transmission unit that is connected to the first gear and transmits power to the second gear; A second speed detection sensor for sensing the rotational speed of the power generation device connection shaft; A control unit for controlling the operation of parts; A first battery for charging electricity produced by the first power generation device and supplying electricity to the electronic clutch; And a second battery for charging electricity produced by the second power generation device and supplying electricity to the control unit.

Patent Document 2 describes that it is possible to provide a gearbox for wind power generators with artificial intelligence, but only the mechanical configuration of the gearbox for wind power generators is described, and a specific method of using artificial intelligence is described. Not doing.

Patent Document 3 relates to a data processing and monitoring device of a wind turbine and a predictive management system of a wind power generation device, and is for a management and control system including a wind power generation area, a wind resistance, and a communication network. Patent Document 3 refers to artificial intelligence, but does not disclose a specific method for this.

For this problem, the pitch angle was adjusted using a conventional mechanical or electronic control system. However, when the direction and speed of the wind frequently change according to seasons or throughout the day, if this is controlled by using a simple mathematical model, the control is inevitably low in efficiency from an engineering point of view. In addition, if the electricity price changes with time every day, the generated electricity may be stored in the ESS and the electricity production value may vary depending on the time of sale. There is a limit to efficient control by applying these non-linear wind patterns or electricity transaction prices only with existing mathematical or numerical models, and no clear improvement measures have been suggested yet. In particular, there have been attempts to use artificial intelligence, but no specific solution has been proposed.

Republic of Korea Patent Publication No. 1537363 (2015.07.10.) Republic of Korea Patent Publication No. 1515157 (2015.04.20) US Patent Publication No. 2007-0140847 (2007.06.21)

The present invention is to solve the above problems, and to provide an excellent control method of a wind power generator capable of producing more power than a mechanical and electronic control system in a wind power reality in our country where the direction and speed of the wind frequently change. The purpose.

In order to solve the above problems, the present invention includes the steps of collecting state information of one or more wind power generators including wings, nacelles, and towers by a big data collection unit with sensors corresponding to each; Control of the wind power generator using deep learning, comprising: controlling, by a controller, the one or more wind power generators through deep learning based on the state information of the one or more wind power generators collected by the big data collector. Provides a way.

The state information of the at least one wind power generator includes wind speed, wind direction, nacelle direction, yaw error, air temperature, power generator temperature, power generator speed, and active power. Additionally, weather observation information, weather prediction information, or power transaction information of the power exchange may be added to the state information.

Prior to applying the state information of the one or more wind turbines to the deep learning, influence evaluation according to the state information may be performed, and only important state information may be selected and applied to the deep learning.

In the step of controlling the one or more wind turbines, controlling is at least one of a pitch angle, a tip speed ratio (TSR), and a yaw direction of the one or more wind turbines, and the additional control is an inverter Or at least one of a converter or an ESS (Energy Storage System).

The goal of improving through the deep learning may include power produced by each of the at least one wind turbine generator; Electricity production or electricity sales of the entire system including an ESS (Energy Storage System) linked to the at least one wind turbine generator; Alternatively, the accumulated sales amount is increased by controlling the sales amount, sales timing, and sales amount for the power of the entire system including the ESS (Energy Storage System) linked to the one or more wind power generators.

The deep learning learns through actor/critical reinforcement learning, and learning through reinforcement learning at this time uses deep learning, a modified DDPG that does not predict future rewards in a deep deterministic policy gradient (DDPG).

The deep learning is at least one of DDPG, Twin Delay DDPG, Advantage Actor-Critic, A3C, Soft Actor-Critic, CNN (Convolution Neural Network), DNN (Deep Neural Network), RNN, DBN, LSTM, GAN, and Softmax models. Meanwhile, the state information collected by the big data collection unit, the control method at this time, and information on the power generated at this time may be stored in a separate database format.

The present invention provides a wind power generator controlled by the control method of the wind power generator using the deep learning.

The present invention is to solve the above problems, and can provide an excellent control method of a wind power generator capable of producing more power than a mechanical and electronic control system in a wind power reality in our country where the wind direction and speed frequently change. have.

The present invention analyzes the state information of the wind turbine generator and controls the pitch angle, yaw direction, TSR, and the like using deep learning, thereby increasing the power produced by the wind turbine generator by 15%. In addition, it is possible to increase the final electricity sales amount by building artificial intelligence including the electricity transaction price and controlling the sales point.

1 is an architecture of an actor according to the present invention.
2 is an architecture of critic according to the present invention.
3 is a graph showing a comparison between a reward generated in a legacy system and a reward expected to be improved according to Example 1 of the present invention.
4 is a graph showing 200000 examples of wind speed, legacy system generated power, power generated by actors, power difference between legacy systems and actors, actual pitch values, and pitch values by actors according to Example 1 to be
5 is a general production power curve of the Pongsock power generation device in relation to wind speed.
6 is a graph showing a comparison between a reward generated in a legacy system and a reward expected to be improved according to Example 2 of the present invention.
7 is a graph showing 200000 examples of wind speed, legacy system generated power, power generated by actors, power difference between legacy systems and actors, actual pitch values, and pitch values by actors according to Example 2 to be.
8 is a graph showing a comparison between a reward generated in a legacy system and a reward expected to be improved according to Example 3 of the present invention.
9 is a graph showing 200000 examples of wind speed, legacy system generated power, power generated by actors, power difference between legacy systems and actors, actual pitch values, and pitch values by actors according to Example 3 to be.

Hereinafter, exemplary embodiments in which the present invention can be easily carried out by those of ordinary skill in the art to which the present invention pertains will be described in detail with reference to the accompanying drawings. However, when it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted in describing the operating principle of the preferred embodiment of the present invention in detail.

In addition, the same reference numerals are used for parts having similar functions and functions throughout the drawings. Throughout the specification, when a part is said to be connected to another part, this includes not only a case in which it is directly connected, but also a case in which it is connected indirectly with another element interposed therebetween. In addition, the inclusion of any component does not exclude other components unless specifically stated otherwise, means that other components may be further included.

Hereinafter, it will be described with reference to the embodiments of the present invention, but this is for a more easy understanding of the present invention, and the scope of the present invention is not limited thereto.

An example according to the present invention was carried out on the basis of data collected from four turbines. Each turbine has 692 parameters (data variables). Data for one year and eight months were used for the two turbines, and data for eight months were used for the other two turbines. The data collection period was 1 second, and sometimes missing data occurred. The missing data may affect the prediction result, but can be ignored because it corresponds to about 0.01% of the total data size.

In the examples of the present invention, the following variables were intensively analyzed.

Generated power, GenPowAct (Generator Active Power), SpeGen (Generator Speed), wind speed, wind direction, GenWndTemMax (Generator Temperature), DayEnePro (Energy Production per Day), PllAng (nacelle angle or yaw error), BlaPos (Blade Position or pitch, blade position or pitch), BlaPos003 (Third Blade Position or pitch, third blade position or pitch), pitch, nacelle temperature, ambient temperature

In one embodiment of the present invention, the goal to be improved is the power produced. For a given state information, an increased power value can be obtained by performing the best action using deep learning. For deep learning learning, reinforcement learning was performed with pitch as action, the remaining parameters as state information, and a reward to increase the power produced.

Wind speed

It was found that the wind speed according to the present invention mostly appears in the range of 0-20 m/s, and occasionally rises up to 60 m/s infrequently. In controlling the wind speed generator according to an embodiment of the present invention, the cut-in speed for data is about 3.5 m/s and the cut-off speed is about 20 m/s, so it is out of this range. In this case, care must be taken when building a critical network that does not generate power. Since these cut-in and cut-off speeds may differ for each wind speed generator, changes must be made to the device to be applied.

Wind direction

By comparing the wind direction and the direction of the nacelle, we can see how much wind energy the turbine can actually convert into kinetic energy. The nacelle is a part corresponding to the heart of the wind power generator, and is composed of all devices for converting the rotational force obtained by the rotor into electric energy.

It was found that most of the wind according to an embodiment of the present invention comes from 270 to 340 degrees or 50 to 70 degrees.

Nacelle direction

As mentioned in the wind direction, the greater the difference between the nacelle direction and the wind direction, the less kinetic energy is collected by the turbine blades, so the nacelle direction and the wind direction are highly related. It was found that the nacelle direction and the wind direction almost coincide.

Yaw error

Since the yaw error represents the difference between the wind direction and the nacelle direction, in the assumption for applying the method according to the present invention, the existence of the yaw error may act as a prediction error between the power generated in the wind power generator of the present application and other variables. , Since the yaw error according to the present invention is less than 5 degrees, it is predicted that the yaw error will not be significantly affected.

Air temperature:

This is important in that the higher the air temperature, the lower the air density and the higher the generator temperature. This is because less energy is transmitted to the power generation device due to the low air density, but when the power generation device temperature increases, the power generation device deceleration is a limiting condition for reasons of protection.

Generator temperature

Generator speed is mostly affected by air temperature and water cooling. Most wind turbines use water cooling for cooling.

Generator speed

The generator speed (the rotational speed of the turbine) is directly correlated with the power produced, but may be limited by certain limits to protect the generator and turbine in situations such as typhoons with very high wind speeds. It was found that the generator speed control system is controlled by selecting one of three main speeds according to wind speed and other factors.

pitch

Pitch is the most important control variable to maximize production power. The pitch should be adjusted according to wind speed, generator speed and temperature.

Production power

The power produced is a goal to improve value by predicting the most appropriate behavior using actors as well as predicting using critical.

Deep learning

In the wind power generator according to an embodiment of the present invention, a problem is that there is no actual wind power generator for simulation or a simulator close thereto, so that it is necessary to directly learn from the existing accumulated data. During deep learning learning, there is a possibility that the real wind turbine may be damaged due to various output values of the actor. Therefore, the learning method according to an embodiment of the present invention will be applied in an expensive wind power generator. The learning method according to the present invention cannot directly apply the DDPG algorithm.

Deep learning is learned through actor/critical reinforcement learning, and learning through reinforcement learning at this time is a deep deterministic policy gradient (DDPG), but in an embodiment according to the present invention, deep learning, a modified DDPG that does not predict future rewards. Was used. In other words, by ignoring future rewards and applying certain actions to a given state, it focuses only on immediate rewards that an action can create. The specific algorithm for DDPG is described in https://arxiv.org/pdf/1709.05077.pdf, so a detailed description is omitted.

Actor architecture

The actor according to the present invention is designed to be accurate and flexible to accommodate various state inputs and to predict various possible actions. 1 shows the architecture of an actor according to the present invention.

Actors are learned by reflecting the results (rewards) of actions calculated from the critic network. Because the actors are trained by considering trends in order to maximize the outcome of the critic while learning the critic, learning of the actor is unstable and difficult to converge. In the present invention, this problem is solved by adding the Ornstein-Uhlenbeck process noise to the behavior before calculating the critic's tendency.

The hyper parameters for the actor network are as follows:

Learning rate: 1x10 ^-4

Maximum pitch: 90 (inferred from data)

Initializer: xavier

Critical architecture:

Critique takes an input of (state + response) and predicts the power produced (Q1). The second output (Q2) is used to calculate the draft of the actor. The ultimate goal is to maximize the power produced by predicted behavior from an actor in a given state. So Q2 should output the maximum possible value. In one embodiment according to the present invention, when Q2=Q1 ² was the most effective. 2 shows a critic architecture according to the present invention.

In order to simulate the power that can be generated according to the wind speed, the wind speed coefficient was set as follows.

Wind speed coefficient = 1/(1+e ^{2(-0.3 wind speed} + ^2.3)) -1/(1+e ^{40(-0.22 wind speed} + ^4.5) )

These values are only best suited for a given wind turbine type, and small modifications are required when using other types of turbines. The wind speed coefficient can also prevent damage to the turbine due to irrational determination of the actor by always setting the output to a realistic range.

Critical network hyper parameters are as follows.

Learning rate: 1x10 ^-2

MAXPOWER: 2700000

Initializer: xavier

Data statistics

The impact evaluation was conducted according to the state information, and important state information was selected, and the entire data was collected as follows.

In the table above, Wind Speed is the wind speed, Wind Direction is the wind direction, PIIAng is the nacelle angle, Air Temperature is the ambient temperature, Generator Temperature is the temperature of the generator, Nacelle Temperature is the temperature of the nacelle, and Generator Speed is the speed of the generator (turbine), Pitch means pitch, and Power means produced power. Y-axis count is the number of data generated, mean, std, min, 25%, 50%, 75%, max is the mean, standard deviation, minimum value, 25%, 50%, 75%, maximum value of each data set Represents.

The above data is over 43 million, and it takes too long to train the critical and actor of the reinforcement learning model. In the present invention, out of 43 million data (seconds) of data, continuous data of 1 million seconds having a distribution similar to that of the entire data were selected for an example. The table below is the statistical results for data of 1 million seconds.

Example 1

The important inputs in Example 1 were used as follows.

-Action: Peach

-Status information: WinSpe (Wind Speed, wind speed), WinDir (Wind Direction, wind direction), PllAng, EnvTem (Environment Temperature or ambient temperature), GenWndTemMax (Generator Temperature), NacTem (Nacelle Temperature), SpeGen (Generator Speed), BlaPos003 (Third Blade Position or pitch).

-Reward (goal): The production power is a goal to improve the value by predicting the most appropriate behavior using an actor, as well as making predictions using critical.

In the embodiment, it is intended to predict the status information at time t and the reward by the actor.

To ensure stability of actor learning, actor training was not started until the critic average absolute error (MAE) was less than 7%.

3 is a graph showing a comparison between a reward generated in a legacy system and a reward expected to be improved according to Example 1 of the present invention. In FIG. 3, the x-axis is described as time from 200 to 750, and the y-axis is from 7.0X10 ⁷ to 1.3X10 ⁸ as rewards. In FIG. 3, a straight line represents a legacy system reward, a red waveform represents an actual value, and an orange waveform represents a reward expected to be improved according to the present invention. In Fig. 3, the critic took about 250 epoxies (1 hour and 14 minutes) to stabilize. Before that, the action of the actor (orange line) is not shown in the graph. The table below summarizes the results according to Example 1.

As shown in the table above in Fig. 3, the reinforcement learning agent does not appear to have any significant improvement over the legacy system. It is understood that this is because data at 1-second intervals are used in Example 1. Since it takes time to move the blade to the desired pitch, it is interpreted that this result is shown because it has an effect after a certain time, not the time of the actual state information. In Example 1 according to the present invention, it takes about 12 seconds to change the pitch from 0 degrees to 90 degrees.

4 is a graph showing 200000 examples of wind speed, legacy system generated power, power generated by actors, power difference between legacy systems and actors, actual pitch values, and pitch values by actors. In Figure 4, the x-axis is the order of 200000 examples, and in graph order, wind speed, legacy system generated power, power generated by actors, power difference generated by legacy system and actor, actual pitch value, and pitch value by actor. Show. Due to the stability problem of the system due to the limitation of wind speed, the reinforcement learning system according to the present invention does not always show the best movement, especially when the wind speed is 8m/s or more.

5 is a general production power curve of the Pongsock power generation device in relation to wind speed. As can be seen from FIG. 5, even if the wind speed increases, there is a limit to the possible generated power. When applying these limitations to deep learning, better predictions and singularities must be avoided.

Example 2

Embodiment 2 is the same as Embodiment 1 except that the compensation (t+1) for the combination of state information/actors of time t is predicted.

6 is a graph showing a comparison between a reward generated in a legacy system and a reward expected to be improved according to Example 2 of the present invention. In FIG. 6, the x-axis is described as time from 500 to 4.0K (4000), and the y-axis is from 6.0X10 ⁷ to 1.8X10 ⁸ as a reward. In FIG. 6, a straight line represents a legacy system reward, a red waveform represents an actual value, and an orange waveform represents a reward expected to be improved according to the present invention. In Fig. 6, the critic took about 400 epoxies (1 hour and 42 minutes) to stabilize. Before that, the action of the actor (orange line) is not shown in the graph. The table below summarizes the results according to Example 2.

As can be seen from Fig. 6 and the table above, the reinforcement learning agent according to Example 2 achieved significant improvement over the legacy system. Since it is reasonable for the action of time t to give a result at time t+1, it is believed that this performance improvement is possible.

7 is a graph showing 200000 examples of wind speed, legacy system generated power, power generated by actors, power difference between legacy systems and actors, actual pitch values, and pitch values by actors according to Example 2 to be. In Fig. 7, the x-axis is the order of 200000 examples, and in graph order, the wind speed, the power generated by the legacy system, the power generated by the actor, the difference between the power generated by the legacy system and the actor, the actual pitch value, and the pitch value by the actor. Show. Example 2 overall shows better results compared to legacy systems. Sometimes when wind speeds are above 10m/s, the reinforcement learning system makes some bad decisions and performance is slightly degraded. When looking at the actual wind speed results of the system according to the present invention, this problem is unlikely to have a significant impact on the overall result of an increase in energy production of 23.8% because more than 90% of the wind speed is less than 10 m/s.

Example 3

Embodiment 3 is the same as Embodiment 1 except that the compensation (t+2) for the combination of state information/actors of time t is predicted.

8 is a graph showing a comparison between a reward generated in a legacy system and a reward expected to be improved according to Example 3 of the present invention. In FIG. 8, the x-axis is described as time from 1.4K (1400) to 4.0K (4000), and the y-axis is described from 2.0X10 ⁷ to 1.6X10 ⁸ as rewards. In FIG. 8, a straight line represents a legacy system reward, a red waveform represents an actual value, and an orange waveform represents a reward expected to be improved according to the present invention. In Fig. 8, the critic took about 1300 epoxies (5 hours and 37 minutes) to stabilize. Before that, the action of the actor (orange line) is not shown in the graph. The table below summarizes the results according to Example 3.

9 is a graph showing 200000 examples of wind speed, legacy system generated power, power generated by actors, power difference between legacy systems and actors, actual pitch values, and pitch values by actors according to Example 3 to be. In Figure 9, the x-axis is the order of 200000 examples, and in graph order, wind speed, legacy system generated power, power generated by actors, power difference generated by legacy systems and actors, actual pitch value, and pitch value by actor. Show. Example 3 overall shows better results compared to the legacy system and Examples 1 and 2. The MAE of the critic in Example 3 is slightly higher than in the previous example. In the case of data at 1-second intervals using both the reward (t+1) and the reward (t+2), the amount of power produced is higher than that of using only the reward (t) with a similar MAE ratio.

In theory, the reinforcement learning system according to Example 3 shows about 8% better results than the reinforcement learning system according to Example 2. Upon checking the pitch value of the actor according to Example 3, it was found that the reinforcement learning system sometimes changes the pitch value too quickly (up to 4-6 seconds). The turbine used in practice requires about 12 seconds to completely change the angle. On the other hand, in Example 2, this case rarely occurred except for a few times during the last 3 hours.

Claims (14)

Collecting, by a big data collection unit, state information of one or more wind power generators including wings, nacelles, and towers with sensors corresponding to each;
Controlling, by a controller, the at least one wind power generator by applying to deep learning based on the state information of the at least one wind power generator collected by the big data collection unit;
In the control method of a wind power generator using deep learning comprising a ,
The deep learning is learned through actor/critical reinforcement learning,
The learning through the actor/critical reinforcement learning is a modified DDPG that does not predict future rewards in a deep deterministic policy gradient (DDPG), a method of controlling a wind power generator using deep learning. The method of claim 1,
The state information of the one or more wind power generators includes wind speed, wind direction, nacelle direction, yaw error, air temperature, power generator temperature, power generator speed, and reactive power. The method of claim 1,
A method of controlling a wind power generator using deep learning in which observation information on weather, prediction information on weather, or power transaction information of a power exchange is added to the state information. The method of claim 1,
Prior to applying the state information of the one or more wind power generators to the deep learning, influence evaluation according to the state information is performed to select only important state information and apply it to deep learning. The method of claim 1,
Controlling in the step of controlling the at least one wind power generator is at least one of a pitch angle and a tip speed ratio (TSR) of the wind power generator. The method of claim 5,
In the step of controlling the one or more wind power generators, the additional control is at least one of an inverter, a converter, or an ESS (Energe Storage System). The method according to any one of claims 1 to 3,
The goal to be improved through the deep learning is a control method of a wind power generator using deep learning, which is power produced by each of the one or more wind power generators. The method according to any one of claims 1 to 3,
The goal to be improved through the deep learning is a control method of a wind power generator using deep learning, which is the amount of power production or power sales of the entire system including an ESS (Energy Storage System) linked to the one or more wind power generators. The method according to any one of claims 1 to 3,
The goal of improving through the deep learning is to increase the accumulated sales by controlling the sales amount, sales timing, and sales amount of the entire system including the ESS (Energy Storage System) linked to the one or more wind power generators. Control method of wind power generation device using in-deep learning. delete delete The method of claim 1,
The deep learning is at least one of DDPG, Twin Delay DDPG, Advantage Actor-Critic, A3C, Soft Actor-Critic, CNN (Convolution Neural Network), DNN (Deep Neural Network), RNN, DBN, LSTM, GAN, and Softmax models. A method of controlling a wind power generator using deep learning. The method of claim 1,
The state information collected by the big data collection unit, the control method at this time, and information on the power generated at this time are stored in a separate database format. A wind power generator controlled by the control method of the wind power generator using deep learning according to any one of claims 1 to 6.

Images