Verification of Prediction Method Based on Machine Learning under Wake Effect Using Real-Time Digital Simulator

Abstract

With the increase in the penetration rate of renewable energy sources, a machine-learning-based forecasting system has been introduced to the grid sector to improve the participation rate in the electricity market and reduce energy losses. In these studies, correlation analysis of mechanical and environmental variables, including geographical figures, is considered a crucial point to increase the prediction’s accuracy. Various models have been applied in terms of accuracy, speed calculation, and amount of data based on a mathematical model that can calculate the wake; however, it can be difficult to derive variables such as air density, roughness length, and the effect of turbulence on the structural characteristics of wind turbines. Furthermore, wake accuracy could decrease due to the excessive variables that come from the wake effect parameters. In this paper, we intend to conduct research to improve prediction accuracy by considering the wake effect of wind turbines using supervisory control and data acquisition (SCADA) data from the Dongbok wind farm. The wake divides the wind direction into four parts and then recognizes and predicts the affected wind turbine. The predicted result is the wake wind speed and its conversion to power generation by applying a power curve. We try to show the efficiency of machine learning by comparing the wake wind speed and the power generation in the wake model. This result shows the error rate using evaluation metrics of regression, such as mean squared error (MSE), root mean squared error (RMSE), and weighted absolute percentage error (WAPE), and attempts to verify power system impact and efficiency through a real-time digital simulator (RTDS).

1. Introduction

Widely adopted climate agreements are aimed at limiting carbon dioxide emissions that increase international environmental problems. The EU has agreed to increase the share of renewable energy to 32% of total energy consumption by 2030 [1]. According to the international renewable energy agency (IRENA), the total capacity of renewable energy installations in the EU was 609 GW, which represents 22% of the world’s renewable energy production capacity in 2020. Among them, wind power amounted to 207 GW, which constitutes 34% of all renewable energy installations in the EU [2]. As wind power has a small installation area and high generation efficiency, the penetration rate increases sharply compared to other renewable energy sources. However, wind energy is greatly affected by rapidly fluctuating wind speed, so it is difficult to balance supply and demand in its generation. Because of these features, along with the increase in the share of wind power facilities, the power system constantly has difficulties in maintaining a stable frequency and voltage within a certain range. Research aimed at forecasting incremental events in the field of renewable energy generation, securing power reserves, and enhancing the flexibility of the power system by accurately predicting wind energy has attracted attention [3].

Research into the application of forecasting and wind time series data analysis for the operation of the power grid is progressing step by step. The authors of reference [4] published a study on the processing and refinement of existing meteorological data in a specific area to make it suitable for the operation of a wind farm. These applications have evolved to be used by grid operators in a specific operating cycle by the short-term prediction of environmental variables [5,6]. SCADA-based management systems are also being enhanced to predict, analyze, and use real-time environmental variables [7]. The long–short-term memory model (LSTM), a relatively new technology based on machine learning, has excellent data storage capabilities. Thus, trends in input and output data can be expressed with high precision and are often used as time series forecasting models [8]. However, if rapidly changing variables are applied directly to the model without preprocessing, it is difficult to calculate the weight, and therefore, accuracy is reduced. Accordingly, dataset building, data preprocessing, and data mining prior to training have been evaluated to be important for improving prediction accuracy [9]. Convolutional neural networks (CNNs) can be used in data mining to efficiently extract features from data [10]. It has been rated as good at identifying the trend of time series data, even data consisting of nonlinear relationships. The use of the CNN-LSTM model combining CNN and LSTM is expected to increase the accuracy of the wake prediction.

In general, wind farms consider the wake effect of wind changes based on the efficiency of power generation [11,12]. The turbulence caused by the geographical features and the wake passing through the wind turbine is mixed, and the amount of power generated by the wind turbine installed at the rear changes [13,14]. Each wind turbine is exposed to single or multiple wakes depending on the wind direction, and the calculation process is complicated in many ways [15]. Commonly used wake models include the Jensen model and the eddy viscosity model [16,17,18,19]. In addition, various wake models have been researched for accurate wake calculations [20,21]. To apply the wake model, it is necessary to know the wake decay constant and the force of turbulence. The general wake decay constant is recommended to be 0.075 on land and 0.04 at sea, but the wake loss coefficient should be changed according to geographic characteristics [22]. Due to the above variables, inaccurate calculation values may be reflected in situations in which precise power generation is required by applying the wake model.

In this paper, we intend to use SCADA data from the Dongbok wind farm on Jeju Island. The Dongbok wind farm has 15 distributed wind turbines and severe turbulence due to a nearby quarry and landfill, as shown in Figure 1.

In general, wind farms have difficulty ensuring accurate values of on-site environmental data. Therefore, the wake model method is complicated to calculate the radius and speed of the wake, including turbulence. Because machine learning training is based on past causal relationships between variables, it has the advantage of being able to estimate the resulting value without including the exact present value.

Table 1. Comparison of the proposed method with other wake models.

	[16,17]	[18,19]	[20]	Proposed Method
Improved grid stability	✓	✓	✓	✓
Improved grid economic	✓	✓	✓	✓
Computational efficiency				✓
Instantaneous environment value not required (air density)				✓
Applicable to various system				✓

summarizes the comparison of the proposed method with previous studies. The machine learning model assumed the use of CNN-LSTM to construct a model, including data mining. CNN was used to extract data features efficiently, and LSTM was applied to continue training in chronological order. As a result, it was configured to capture trends of time series data consisting of nonlinear relationships. The model was used based on the sixth and fifth wind turbines in the Dongbok wind farm to predict the wake wind speed acting on nearby wind turbines in four different wind directions. The machine learning method proposed in this article was shown to be more accurate than the existing computational wake model. As a result of observing the point of common coupling (PCC) using RTDS, it was confirmed that the expected generation amount and the actual value had very similar values. Based on the research results, we believe that CNN-LSTM can be used to meet the individual forecast power generation requirements recommended by the International Energy Agency (IEA). Furthermore, it can be used in various forecasting systems and programs and is expected to have an impact on ensuring the stability of the distribution system linked to new and renewable energy. The contributions of the paper are as follows:

• The proposed method enables high-precision wake prediction, which can increase the forecast accuracy of wind farm power generation, reduce curtailments, and reduce energy economic losses.
• Because of the split wind directions to predict the wake of individual turbines, and even if some wind turbines are shut down for maintenance, the accuracy of the overall power generation remains the same.
• The actual wind farm is configured through RTDS, and there is a possibility that it can be used in the digital twin field through advanced work with machine learning in the future.

This paper is organized as follows: In Section 2, Jensen and eddy viscosity are used as computational wake models, and their explanations are given. In addition, the explanation of the variable decision process by correlation analysis is shown. In Section 3, the machine learning simulation process using CNN-LSTM and related contents continue to be explained in order to compare with computational wake models. Moreover, an explanation of regression evaluation metrics to confirm the error rate for the result value is continued. In Section 4, the general process of the simulation is presented, and the values of the results are divided into four directions, N, E, S, and W, and the predicted and actual values of the affected turbine are plotted. Then, the predicted wake effect error rate between the proposed method and Jensen and eddy viscosity is shown in a table quantified by applying MSE, RMSE, and WAPE regression metrics evaluation. The predicted wake was converted to power generation by applying the HJWT 2000 power curve. This was applied to the RTDS constituting the Dongbok wind farm, and the resulting values at the PCC are displayed on a graph. Finally, Section 5 consists of conclusions.

2. Model Optimization and Configuration

2.1. Wake Model

Wake is the flow behind an object that occurs when a fluid flows through it. Wind farms select the optimal location with regard to the wake for efficient generation by reducing the losses of individual wind turbines. Given the wind direction, the rear wind turbine induces a reduced generation due to the wake of the front wind turbine.

The six main types of energy loss that occur during the operation of a wind farm are the same as the loss factors in

Table 2. Loss factors of wind farm operation.

Loss Factors	Low (%)	Typical (%)	High (%)
Wake effects	3.0	6.7	15.0
Availability	2.0	6.0	10.0
Electrical	2.0	2.1	3.0
Performance	0.0	2.5	2.0
Environmental	1.0	2.6	6.0
Curtailments	0.0	0.0	5.0
Total losses	7.8	18.5	37.0

. The range of each loss is expressed as a percentage of AEP (annual energy production), with wake effects showing the highest percentage of losses [23]. Therefore, it is estimated that the wake derivation with high accuracy reduces losses in wind farms and contributes to optimal power generation.

One of the wake models, Jensen, presents a relatively simple computational process with a single wake model [24]. Regardless of the strength of the surrounding turbulence, the wake behind the wind turbine is linear. The above model is applied to calculate the wake for a distance of at least three times the rotor diameter. The Jensen model formula is as follows [25]. The wake diameter is represented by ${\mathrm{D}}_{\mathrm{w}}$ in Equation (1).

$${\mathrm{D}}_{\mathrm{w}}=D\left(2ks\right)$$

(1)

$D$ is the rotor diameter, and k is the wake decay constant, which is related to height, turbulence, and atmospheric stability but is commonly used as 0.075 on land and 0.038 at sea. $s$ is the relative distance to the rear of the rotor in $x/D$.

$$U_{def}=\frac{1-\sqrt{1-C_T}}{{\left(1+2ks\right)}^2}$$

(2)

$U_{def}$ is the speed decay, and $C_T$ is the thrust coefficient. This can be expressed by Equation (3).

$$C_T=\frac{T}{\frac{1}{2}\rho U_{\infty }^2{A}_d}=4a\left(1-a\right)$$

(3)

T is the thrust exerted on the disk by the pressure difference generated between the front and rear of the disk. A is the axial induction coefficient, which determines the speed drop in the rotor plane [26]. Equation (3) is equal to Equation (4) when expressed as the ratio of the wind speed $(U_d)$ passing through the disk to the free wind speed $(U_{\infty })$.

$$\frac{U_d}{U_{\infty }}=\left(1-2a\right)=\sqrt{1-C_T}$$

(4)

As a result, the wind speed for a one-dimensional linear wake can simply be expressed as Equation (5).

$$U_w=U_{\infty }\left(1-U_{def}\right)$$

(5)

The eddy viscosity differs from the Jensen model because it takes turbulence into account in the calculation process. Suppose the wake begins at a distance of at least twice the rotor diameter. The velocity distribution at a distance x from the rotor uses the boundary condition of the Gaussian distribution. The Gaussian distribution formula is the same as Equation (6).

$$1-\frac{U_w}{U_{\infty }}=D_{m}exp\left[-3.56{\left(\frac{r}{B_w}\right)}^2\right]$$

(6)

$r$ represents the radius of the rotor, and $B_w$ represents the wake width, as calculated by Equation (7). The wind turbine may be affected by single or multiple wakes, and the width of the wakes must be calculated to determine whether they are affected.

$$B_w=\sqrt{\frac{3.56C_T}{8D_m\left(1-0.5D_m\right)}}$$

(7)

The initial speed reduction rate $D_{mi}$ of the wake central axis at the 2D distance of the rotor diameter can be calculated from Equation (8).

$$D_{mi}=C_T-0.05-\left(16C_T-0.5\right)I/1000$$

(8)

$I$ is the turbulence intensity [27].

2.2. Data Construction

Model composition is important but must be preceded by an understanding of the data composition that is applicable to a particular model. Therefore, data preprocessing is an important part of data construction. The data for model training was obtained via the wind farm SCADA system. The data includes relevant environmental variables and wind turbine details that affect wind speed. Time series data are based on a 15 min unit timescale against the flow of time. The model is to be composed of CNN-LSTM in order to solve the problem of long-term dependencies with data mining.

When creating a forecasting system with the use of time series data, it should take care to ensure that the input to the model does not adversely affect the output because of data preprocessing [28]. After going through imperfect data handling through data preprocessing, it is necessary to apply normalization and input data to the prediction system. Normalization means adjusting the scale differences between the variables used as input so that the individual data are dimensioned to the same unit.

In this paper, we used the most commonly used Min–Max Scaler. If there are multiple data variables (features) to be taught, the value of each data variable with different maximum comment values is converted to a value ranging from 0 to 1 or −1 to 1, and the scale is adjusted before training. Below is the corresponding Equation (9).

$$Min-MaxScaler\left(x\right)=\frac{x-x_{min}}{x_{max}-x_{min}}$$

(9)

The changing wind conditions after passing the wind turbine include conditions for the air. Therefore, data affecting wind speed, such as air density, pressure, and humidity, were derived from the correlation analysis. This process is performed because prediction accuracy may be degraded when the progress of learning involves unnecessary information. In addition, it is possible to reduce the loss of time and memory wastage that occurs during the learning process with huge amounts of data. The correlation analysis using the heatmap shown in Figure 2 only presents the correlation between wind turbine No. 6 and the neighboring turbine No. 9. Wind speed between No. 6 and No. 9 is the highest correlation among the different variables confirmed.

A heatmap is an important process of data analysis, allowing one to visualize it and check for the existence of a linear relationship between variables. In this paper, the variables are determined by analyzing the correlation between them and the output values of the wind power generators through data heat maps utilizing the Seaborn library. The variables measured by the power generator and the environmental variables, excluding the output variable, were found to be appropriate for the calculation of the output value. A high correlation coefficient indicates a high linear relationship, while a low one indicates nonindependent data, possibly a nonlinear relationship, and requires each variable to be considered. Figure 2 shows the heatmap for each variable. The analysis of the heat map shows that the wind speed has the most linear relationship with the output result. Temperature and humidity with negative correlation coefficients imply a linear but inverse relationship. In this regard, it can be seen that the output of the wake effect wind speed decreases when the temperature and humidity are high. The influence of environmental variables measured by the meteorological tower on the wake wind speed of No. 9 was also found. Due to the above connection, environmental variables are included in the training dataset. Using additional variables allows for the analysis of their impact on the results and is considered important in constructing datasets to improve the accuracy of future forecasting. Random selection of variables without going through such a process is likely to negatively affect the prediction results. Because wind speed is a rapidly changing variable, it is very difficult to predict. Thus, to improve forecast accuracy, a weather variable is set that affects wind speed in the input data set. Therefore, heat maps and correlations are necessary processes to select and refine the variables that influence the result.

3. Machine Learning Model

3.1. Prediction Algorithm

The predictive model uses the CNN-LSTM model, which combines CNN and LSTM. CNN is a neural network of complex products and is mainly used for image classification, but it is a machine learning method recently used as a kind of data mining [29]. CNN acts as a feature extraction function that learns input data, analyzes patterns, filters the data, and extracts features [30]. The convolution layer and the max pooling layer are developed by stacking multiple times, as shown in Figure 3.

The CNN layer ensures spatial features of the data by the filter performing the composing operation and extracting the maximum value using the max pooling layer [31]. Therefore, complex datasets can be organized effectively. Then, to apply it to the LSTM, a model suitable for the given time series data, using the flatten function, which converts the 2D array output in CNN to 1D. By applying the dimensionally transformed data to the LSTM, we extracted the features of the data and constructed a CNN-LSTM model that reflects the temporal trend. LSTMs, which are currently in use, are a type of RNN possessing the ability to learn when working with long-term time series data. The basic LSTM structure is the same, but the environment variables have been added to the input values in order to increase the amount of learning. The added environmental variables were temperature, humidity, and air pressure, and they were configured to have the same duration and timescale as the variables measured by the existing wind turbine. Cell state, forget gate, and input and output gates are part of the LSTM. The forget gate is determined by the sigmoid layer whether to save the information from the cell state and is represented by Equation (10).

$$f_t=\sigma \left(W_f · \left[h_{t-1,} x_t\right]+b_f\right)$$

(10)

The input gate determines whether the incoming data are stored in a cell state. A decision has to be made whether the value is to be saved by the sigmoid layer, and then a new vector will be created from the tanh layer. Then, it is updated to the cell state. This process is expressed in Equations (11)–(13).

$$i_t=\sigma \left(W_i · \left[h_{t-1,} x_t\right]+b_i\right)$$

(11)

$$\widetilde{C_t}=tanh\left(W_C · \left[h_{t-1,} x_t\right]+b_C\right)$$

(12)

$$C_t=f_t\times C_{t-1}+i_t\times \widetilde{C_t}$$

(13)

The output gate selectively exports information through the sigmoid layer, as in Equation (14), and the tanh layer, as in Equation (15).

$$o_t=\sigma \left(W_o · \left[h_{t-1,} x_t\right]+b_o\right)$$

(14)

$$h_t=o_t\times tanh\left(C_t\right)$$

(15)

The predictive system algorithm proposed in this paper uses composite stacking and max pooling to construct CNN, and the CNN activation function is specified as the ReLU value. Functions such as sigmoid and tanh cause backpropagation problems. As the design of the neural network deepens, the signal begins to disappear in backpropagation, which is called a “vanishing gradient.” However, since ReLU has a constant gradient for positive inputs, it avoids the problem of signal loss [32]. Moreover, taking into account the characteristics of the time series data, a sliding window algorithm in the CNN-LSTM model was applied in order to improve the accuracy over time [33]. Figure 4 shows the derivation method by CNN-LSTM.

3.2. Regression Evaluation Metrics

Regression metrics for evaluation is a method of advanced training with data and calculating the difference between the actual value and the predicted value when obtaining a forecast result. Prediction accuracy can be improved by making corrections to the data and hyperparameters applied in the model by evaluating the regression metrics. This study uses the MSE, RMSE, and WAPE methods to evaluate the regression metrics. The formula for each regression metrics evaluation is the same as the formulas shown in Equations (16)–(18).

$$MSE=\frac{1}{n}{\displaystyle \sum }_{i=1}^n{\left(y_i-t_i\right)}^2$$

(16)

$$RMSE=\sqrt{\frac{1}{n}{\displaystyle \sum }_{i=1}^n{\left(y_i-t_i\right)}^2}$$

(17)

$$WAPE=\frac{{{\displaystyle \sum }}_{i=1}^n\left|y_i-t_i\right|}{{{\displaystyle \sum }}_{i=1}^n\left|y_i\right|}$$

(18)

4. Simulation

4.1. Case Design

Figure 5 shows the entire simulation process in this study, from data acquisition, data preprocessing, and machine learning model application to RTDS application. A machine learning model was applied using SCADA time series data between 24 February 2020 and 30 June 2022 from the Dongbok wind farm located in Dongbok-ri, Gwaneup, Jeju. The simulation algorithm is shown in Figure 6. The Dongbok wind farm with the HJWT2000 model name consists of 15 wind turbines configured as shown in Figure 1.

In this paper, we have established a wake effect analysis to assess the neighbor wind turbine performance when the wind passes through turbines No. 5 and 6. SCADA data secured wind speed, wind direction, and blade pitch angle data measured by each wind turbine and wind speed data measured by weather towers. We also constructed a dataset that included environmental variables, such as humidity, atmospheric pressure, and temperature. The dataset comprises seven items as 15 min units of data, and 80% of the training data is divided into 20% of the test data to perform wake prediction in 15 min units. Therefore, the test data consists of approximately 5 months of January 2022 data.

The technical specifications of the HJWT2000 are shown in

Table 3. HJWT 2000 technical specifications [34].

Classification	HJWT 2000
Rated power	2000 kW
Start-up wind speed	3.5 m/s
Rated wind speed	12 m/s
Static wind speed	25 m/s
Average annual wind speed	8.5 m/s
Turbulence intensity	18%
Rotor diameter	86 m
Hub height	85 m

. The meteorological tower in the wind farm is equipped with anemometers (80 m, 76 m, 50 m, 35 m), wind vanes (76.5 m, 74 m, 35 m), thermometers, hygrometers, and atmospheric pressure gauges (75 m). The wind speed measured by the meteorological tower was defined as the free wind speed for calculation using the wake model.

To use the wake model, we need the distances between the wind turbines we want to calculate. Therefore, the distances from each wind turbine are shown in

Table 4. Standard distance for wind turbine numbers 5 and 6 (meters).

No.	5th	6th
1st	950	780
2nd	810	700
3rd	570	550
4th	520	610
5th	-	610
6th	430	-
7th	680	430
8th	450	300
9th	290	360
10th	270	500
11th	460	720
12th	640	820
13th	770	870
14th	820	830
15th	1000	940

, with wind turbines No. 5 and 6 as a reference. The distance between the wind turbines was confirmed to consist of 2D, 3D, or more of the rotor diameter, which is a distance considered suitable for the Jensen and eddy velocity model calculation.

The analysis results of all wind data are shown in Figure 7, where (a),(b) are graphs of wind direction/speed based on wind turbines (No. 5 and 6). The wind speed was classified into units of 5 m/s and subdivided according to the wind direction. The most common analysis result of the analysis is a wind speed of 5 m/s or less, with the main direction of wind power generator No. 5 being WNW and wind power generator No. 6 being NW.

In this paper, we classified the data based on the wind direction to advance the wake effect prediction. Based on the wind direction measured by wind power generators No. 5 and 6, it was divided into four directions, N, E, S, and W, and the wind power generators affected by the wind direction were grouped as shown in

Table 5. Grouping of wind turbines by wind direction.

Direction	No. 5	No. 6
N (315~45°)	1, 2, 3, 4, 13, 14, 15	1, 2, 3, 13, 14, 15
E (45~135°)	11, 12	4, 5, 10, 11, 12
S (135~225°)	9, 10	8, 9
W (225~315°)	6, 7, 8	7

for learning.

4.2. Case Study

The total number of data was approximately 84,600, and in the case of missing values, the relevant data were excluded. Figure 8 shows the wind turbines affected by N, E, S, and W wake with reference to the No. 5 wind turbine. It can be seen that the number of time data for the four turbines shown in the graph is not the same because the number of data is divided by wind direction. Therefore, the number of predicted data corresponding to each wind power generator does not match.

Figure 9 is based on the No. 6 wind turbine. Similar to Figure 8, the number of forecast data for each wind turbine is different for grouping by wind direction, and in the cases of wind turbines No. 1 and No. 9, it can be concluded that there is relatively much missing data and little forecasting data.

Table 6. Comparison of wake error caused by wind turbine No. 5.

	MSE			RMSE			WAPE
WT	N.O. Jensen	Eddy Viscosity	CNN LSTM	N.O. Jensen	Eddy Viscosity	CNN LSTM	N.O. Jensen	Eddy Viscosity	CNN LSTM
1	2.870	3.710	1.821	1.694	1.926	1.350	0.233	0.265	0.175
2	2.542	3.557	2.053	1.594	1.886	1.433	0.196	0.233	0.166
3	2.701	3.988	2.236	1.643	1.997	1.495	0.240	0.293	0.201
4	1.893	3.307	1.646	1.376	1.818	1.283	0.193	0.255	0.166
6	3.430	6.102	1.504	1.852	2.470	1.226	0.207	0.268	0.135
7	3.968	5.562	1.721	1.992	2.358	1.312	0.221	0.260	0.149
8	2.771	5.000	1.568	1.665	2.236	1.252	0.201	0.268	0.152
9	5.864	13.554	2.932	2.422	3.682	1.712	0.303	0.453	0.204
10	4.912	12.718	3.167	2.216	3.566	1.780	0.293	0.464	0.221
11	3.185	6.916	0.838	1.785	2.630	0.915	0.315	0.466	0.143
12	3.953	6.492	1.631	1.988	2.548	1.277	0.345	0.450	0.214
13	2.108	3.045	1.836	1.452	1.745	1.355	0.187	0.224	0.168
14	2.696	3.579	2.135	1.642	1.892	1.461	0.224	0.258	0.188
15	3.414	4.329	2.186	1.848	2.081	1.478	0.180	0.203	0.135

and

Table 7. Comparison of wake error caused by No. 6 wind turbine.

	MSE			RMSE			WAPE
WT	N.O. Jensen	Eddy Viscosity	CNN LSTM	N.O. Jensen	Eddy Viscosity	CNN LSTM	N.O. Jensen	Eddy Viscosity	CNN LSTM
1	2.636	3.710	1.735	1.624	1.926	1.317	0.223	0.265	0.164
2	2.367	3.557	1.882	1.539	1.886	1.372	0.189	0.233	0.153
3	2.803	3.988	2.191	1.674	1.997	1.480	0.243	0.293	0.199
4	3.681	6.149	0.996	1.919	2.480	0.998	0.345	0.448	0.163
5	4.380	8.102	1.164	2.093	2.846	1.079	0.352	0.482	0.171
7	2.617	5.562	1.074	1.618	2.358	1.036	0.176	0.260	0.107
8	6.046	14.627	3.210	2.459	3.825	1.792	0.312	0.481	0.215
9	6.511	14.113	2.995	2.552	3.757	1.731	0.326	0.474	0.206
10	4.980	8.424	1.046	2.232	2.902	1.023	0.375	0.490	0.160
11	4.701	6.916	1.078	2.168	2.630	1.038	0.383	0.466	0.172
12	4.688	6.492	1.903	2.165	2.548	1.379	0.380	0.450	0.225
13	2.204	3.045	1.754	1.484	1.745	1.324	0.192	0.224	0.160
14	2.739	3.579	2.041	1.655	1.892	1.429	0.226	0.258	0.183
15	3.332	4.329	2.210	1.825	2.081	1.487	0.178	0.203	0.134

show the error rates of the predicted values and actual values corresponding to 14 units based on wind power generators No. 5 and 6, respectively. Machine-learning-based CNN-LSTM has the lowest error rate and highest wake prediction accuracy, and the error rate increases in the order of N.O. Jensen and eddy viscosity. Eddy viscosity is a method of calculating the wake, including surrounding turbulence. However, this paper estimates that the error rate increased considering only the turbulence corresponding to the wind power generator at once.

In addition, by using time series data on a 15 min timescale and dividing the data according to wind direction, the amount of data was reduced, and the time series characteristics became inconsistent, which is believed to have affected the forecast accuracy.

To verify the effect of machine learning prediction, we intend to analyze the impact on the power system using RTDS. By applying RTDS, it is possible to understand the system’s influence on the PCC by reflecting line or power current losses rather than calculating and summing up the wind turbines individually. The system diagram of the Dongbok wind farm was built using the RSCAD program. For the construction of the Dongbok wind farm, the 15 wind turbines consisted of individual PQ models, and the server was configured to exchange data using Modbus.

RTDS data is entered in real time, so data in seconds is required. Therefore, the prediction of wind speed in seconds was performed under the same conditions as in the previous model to provide the predicted data, and then a power curve was applied. The forecast generation data obtained by applying the power curve is transferred to an RTDS wind turbine configured as a PQ model, and the result is displayed. The result is a simulation of one wind turbine and is shown in Figure 10. It can be seen that the expected generation is applied, similar to the actual power generation. When applied directly to a real system, the wake can be predicted with less error, which is expected to contribute to system stability. Through this process, changes at the PCC can be identified if the predicted value is applied to an actual power system. In addition, the results shown in the figure below are judged as the basis for the correctness of the system, reflecting the loss and characteristics of tidal currents when predicted values by machine learning are applied.

5. Conclusions

In this paper, the N.O. Jensen and eddy viscosity wake model and the CNN-LSTM composite machine learning model were compared and validated for power generation and wake wind speed. A simulation of the Dongbok wind farm compared three methods based on two wind turbines. When calculated as the average of the regression index evaluation MSE, the machine learning value was 1.877, the Jensen model 3.571, and the eddy viscosity model 6.23. The mean of RMSE was 1.351 for machine learning, 1.863 for the Jensen model, and 2.418 for the eddy viscosity model. Finally, the average of WAPE was 0.172 for machine learning, 0.259 for the Jensen model, and 0.335 for the eddy viscosity model. The prediction accuracy is highest in order of CNN-LSTM, Jensen, and eddy viscosity, and the accuracy of eddy viscosity is judged to be low because only one turbulence is considered for each of the No. 5 and No. 6 wind turbines. CNN-LSTM improved forecast accuracy by adding environmental variables related to wind speed through correlation analysis. It was confirmed that there are very few environmental variables that can be considered when performing simulations. The presence of additional environmental variable data is considered very useful for correlation analysis and the derivation of results. In addition, the SCADA data of the wind farm had blank data due to communication errors, so it was supplemented, but it is considered that the error rate can be reduced by using the generative adversarial network (GAN) method in the future. Based on the prediction results, a power curve was applied and replaced with power generation. As a result of its application to the RTDS constituting a wind farm, the actual and expected generation were very similar, and the effect was verified. Even with realistic scaled-down modeling to account for the wake of the wind farm, it has the drawback of not being able to accurately measure environmental variables because it cannot reproduce the climate environment of the site. The wake effect has many variables that affect wind speed, including turbulence, and may increase the error rate when applying a wake model based on mathematical calculations to a wind farm. This can lead to curtailment due to power supply and demand imbalances, which can lead to lower participation rates compared to renewable installed capacity. In addition to the field of prediction, machine learning is also usefully used to build analytics and strategies with data [35]. Because RTDS takes into account internal power flow and controls in systems containing renewable energy sources, it exhibits a very similar operating regime to real systems. Through the advancement of machine learning and RTDS linkage in the future, various simulations will be possible when used in the digital twin field, and it is judged to be useful in deriving proposals for problem-solving.