The digital twin structure for wind power prediction is shown in Figure 2, which mainly includes four parts: entity layer, data layer, virtual layer, and decision-making layer.

The SVM algorithm is used to predict wind power and to train and model the historical data of active power and its corresponding variable data and the test data are used to test the prediction model. First, the sensor is used to collect the real-time meteorological data of the wind turbine environment. After normalization, it is tested as the input of the power prediction model, and the preliminary results are obtained. The SVM algorithm has good generalization ability and can well-solve the types of problems in non-linear pattern recognition. On this basis, the characteristics of the GA algorithm are used to optimize the parameter finding problem of the SVM.

On this basis, the digital twin is used as the skeleton of the power prediction model to connect wind farm, twin data, and the prediction model as a whole; the prediction model and twin data are interconnected in real time to enhance the timeliness and reliability of wind power prediction of the wind farm. Meanwhile, the sensor compensation mechanism is introduced to compensate the prediction deviation of the wind farm caused by real-time change of meteorological data so as to further improve the wind power accuracy. The flow of the digital twin wind power prediction system is shown in Figure 3.

With the continuous updating of twin data, the compensation of the meteorological database for prediction power will be more accurate, and the prediction accuracy will be continuously improved within a certain range.

Whether the data are selected correctly directly affects the experimental results. Due to factors such as system failure, wind power curtailment policy, and environmental impact of collection, the data of wind farms will be missing and abnormal. Wind farm output will be affected by environmental factors such as wind speed, temperature, and wind direction, but if all influencing factors are taken into account, it will interfere with the model. Therefore, it is necessary to select several key variables to predict wind power.

The wind power is calculated as follows: p = 1 2 A ∗ V 3 ∗ C p ∗ D ∗ φ , (1) where p is wind power, A is the sweeping area, V is the wind speed, Cp is the conversion rate of wind energy, D is the air density, and φ is the power factor.

According to Eq. 1, the sweeping area is related to the fan blade radius and wind direction angle, and the wind energy conversion rate varies with the technology of the fan manufacturer; therefore, in the wind speed prediction for a single fan, the most direct influencing factors are wind speed and wind direction.

To ensure data continuity, the missing data need to be filled by linear interpolation. Linear interpolation is often used in the fields of statistics and computer science.

The linear interpolation formula is shown in Eq. 2: P i + m = P i + m P i + n − P i n , 0 < m < n , (2) where P i , P i + m , and P i + n represent the values at times i, i + m, and i + n of the same time series, respectively.

The collected wind speed, power, and other data will have noise due to the impact of the environment, which affects the reliability of the data, in order to make it complex with the real field environment. Therefore, the data noise reduction process is carried out. The wind speed and power data have good continuity and smoothness, so the wavelet soft threshold denoising method is used for denoising. If the absolute value of the wavelet coefficient is less than the given threshold, it is set to 0; when it is greater than the threshold, the threshold is subtracted from the wavelet coefficient, and the formula is shown in Eq. 3: ω λ = sgn ω ω − λ ω ≥ λ 0 ω < λ . (3)

The determination of a reasonable threshold can separate the signal and noise, so it can denoise effectively.

The next step is data normalization. It is to enlarge or narrow the relevant data to a relative interval according to a certain proportion so that the indicators are in the same order of magnitude. Data normalization is done to process the data, and the formula is shown in Eq. 4: y = x − 1 2 x max + x min 1 2 x max + x min , (4) where x _min is the minimum value of the original data, x _max is the maximum value of the original data, x is the original data, and y is the normalized target value.

SVR transforms the non-linear problem into a high-dimensional linear solvable problem through the mapping of kernel function. The task of the homing machine is to find the parameter ω , b so that the error between the objective function f x i = ω ⋅ x i + b and the actual value is within an acceptable range (Chen, Y et al., 2021).

Suppose there is a training sample x i , y i , i = 1 , 2 … , I ; x i ∈ R n ; y i ∈ R . A nonlinear mapping φ ⋅ is used to map the sample space to the feature space φ x i and construct linear regression function in high-dimensional feature space: y x = ω φ x + b , where ω is the weight vector, ω ∈ R k , b ∈ R . At this time, the problem of nonlinear regression in the original sample space is transformed into the problem of linear regression in the high-dimensional feature space. It can be expressed as an optimization function: min ⁡ J ω , ξ = 1 2 ω T ω + C ∑ i = 1 l ζ i + ζ i *   s . t .   y i − y x i ≤ ε + ζ i y x i − y i ≤ ε + ζ i * , (5) where ε is a constant, C is a penalty factor, and ζ i and ζ i * are relaxation factors.

For Eq. 5, Lagrange multipliers a i and a i * are introduced to transform it into a dual problem, and the optimization function is given, see Eq. 6: min ⁡ I a , a * = 1 2 ∑ i , j = 1 l a i * − a i a j * − a j K x i − x j − ε ∑ i = 1 l a i * + a i − ∑ i = 1 l a i * − a i y j s . t . ∑ i = 1 J a i * − a i = 0 0 ≤ a i , a j * ≤ C / l , i = 1 , ⋯ , I , (6) where K x i , x j = φ x i φ x j is a kernel function.

Finally, the regression function is Eq. 7: f x = ∑ i = 1 l a i K x , x i + b . (7)

By training the sample data and adjusting the C value and gamma value most suitable for the model, the prediction model can be obtained. With the passage of time, the sample database will be increasingly larger, the twin data will be increasingly perfect, and the prediction accuracy will be continuously improved. Digital twinning is based on a large amount of data and has high requirements for data. It should not only have timeliness but also have reliability. Therefore, the focus of digital twinning should be on the process of data and modeling and simulation. When using MATLAB software for modeling and simulation, it is necessary to test the model many times and compare the uncertain quantities in the model again and again until the expected results are obtained.

Because of the meteorological data changes in real time, the prediction error is increased; the prediction results of the SVM model optimized using the genetic algorithm should be compensated and corrected. The paper sets up the wind power error compensation link. The new data are constructed as the objective function, and the actual output power and predicted power of similar meteorological days searched in the historical database are used to correct the initial prediction results. The compensation process is shown in Figure 4.

The meteorological data and measured data of a wind farm in Chagan, Inner Mongolia Autonomous Region, are selected for simulation analysis. The accuracy of the digital twin prediction model is verified, and it is proved that this prediction method has reference significance.

The sensors collected data on wind speed, wind direction, and temperature every 5 min, 24 h a day from April 2020 to May 2020 and uploaded the collected data to the database every hour. All the data in the database, except for the test day, are used as the prediction model, and a day in the later stage of the study is selected as the test day for verification. First, the data of the detection day are input into the prediction model of the GA-SVM neural network to obtain the preliminary prediction results. Then, the corresponding meteorological data are collected into the previous database for analysis and comparison; the actual power under similar conditions collected by similar meteorology is used to correct the preliminary prediction results and obtain the digital twin-predicted value.

This simulation includes two processes, one is to predict the power of a single fan in a wind farm in Chagan, Inner Mongolia; Next, the power of a wind farm in Chagan, Inner Mongolia is predicted. This study intercepts actual power from April 1 to April 2 in the training database. In the first process, the output power of a single fan of a wind farm in Chagan, Inner Mongolia, is predicted, and the results are shown in Figure 5.

According to Figure 5, the digital twin-predicted power curve and GA-SVM-predicted power curve are roughly consistent with the expected output curve, but there are still deviations. The prediction effect of digital twins in some time periods is not as good as that of the SVM model optimized using the genetic algorithm, which may be due to the drastic change in weather in the corresponding time period. It may be that the established historical database is not updated in time, resulting in the difference of the compensation effect, or it may be that similar meteorological data are not searched. In most of the time, the change in weather is relatively stable and has a good compensation effect. For a single wind turbine, the predicted value of the digital twin is closer to the expected output and has higher prediction accuracy.

Some common error evaluation indexes mainly include E _MAE , E _RMSE , average percentage error, accuracy, and qualification rate (Wang, S et al., 2017). The short-term power errors required by the national energy administration to be reported to the dispatching department are root mean square errors. ERMSE and EMAE were chosen as the evaluation indexes of the prediction model (Huang, L.2011). The specific formulas are as follows: E R M S E = ∑ i = 1 n p a − p b 2 n × 100 % , (8) E M A E = 1 n ∑ i = 1 n p a − p b , (9) where p _a is the actual value, p _b is the predicted value, and n is the number of predicted samples.

The E _RMSE and E _MAE of the two prediction models are calculated; it is shown in the following table.

The error comparison of prediction results in Table 1 shows that for a single typhoon generator set, the E _RMSE and E _MAE of the predicted value of the digital twin are higher than the predicted value of the GA-SVM algorithm. It is concluded that in the prediction of single-fan output power, the prediction accuracy of the digital twin is higher as a whole.

In the second process, the power prediction results of a wind farm in Chagan, Inner Mongolia, are shown in Figures 6–8.

It can be seen from Figures 6–8 that the digital twin-predicted power curve and GA-SVM-predicted power curve are roughly consistent with the expected output curve, and the numerical twin-predicted power curve is closer to the expected output curve.

The errors can be compared according to Eqs 8, 9, and it is shown in the following table.

The prediction results in Table 2 show that for the power prediction of the whole wind farm, the predicted value of the digital twin under different wind speeds is closer to the expected output, and the accuracy is higher than that of the SVM model optimized using the genetic algorithm.