Analysis and comparison of turbulence models on wind turbine performance using SCADA data and machine learning technique

Abstract

Wind energy has grown significantly over the last decade. With this, various improvements in the design of the wind turbine are geared towards increasing the reliability of several components. Wind turbulence has a huge effect on the fatigue loading of wind turbines considered in the design. Several monitoring methodologies, such as turbulence intensity analysis, are used to identify wind turbulence. In this paper, a method based on machine learning techniques and data from Supervisory Control and Data Acquisition (SCADA) systems is described. Five machine learning models are generated and compared in this study with the use of the operational data from the SCADA of wind turbines in a single wind farm. Results showed that the model based on Linear Regression in terms of a quadratic hyperparameter has lesser errors compared to the other models that were generated. Each parameter used in the creation of the model affects its performance. Observations in the nacelle system also showed higher errors due to the relationship between rotor speed and the blade angle. The rotor performance is mostly influenced by wind turbulence as the variation in wind speeds and rotational speeds have a certain correlation. Based on the results, it can be concluded that the use of SCADA data in generating turbulence models provides key insights into the relation of the turbulence intensity to the various components. It can be used as the basis for developing turbulence monitoring models that could help improve the design and operation of wind turbines.

Keywords:

• wind turbine
• SCADA system
• wind turbulence
• machine learning
• wind turbine components

Public Interest Statement

The paper highlights the different turbulence models that can be used to monitor the behavior of the turbulence intensity of the wind turbine using SCADA data. Through this study, the SCADA data can be used as a reference for the generation and monitoring of turbulence using Machine Learning Techniques. Wind turbulence intensity, if properly monitored, could help in the improvement of the reliability of the wind turbines. Developing a data-driven turbulence model is a cost-effective and convenient method of modeling wind turbulence.

1. Introduction

At present, rapid population growth is a major contributor to the high percentage of energy demand, resulting in a scarcity of energy supply. The use of some renewable energy sources, such as wind energy, has a long history. Wind energy is one of the most promising markets in recent years, with very significant growth (Anup et al., 2020). Despite its popularity, its application has been limited due to a lack of knowledge in some fields. The effect of turbulence on wind turbine performance, for example, is an incomplete field of knowledge that requires more research.

Turbulence has a direct impact on turbine performance and structural design, so research in this area is crucial. The turbulence structure on turbine wakes must be known to predict the fatigue life of wind turbines grouped in a wind farm (Santolamazza et al., 2021). The turbulent atmospheric winds and the wakes created by upwind turbines are largely responsible for wind turbine performance and structural dynamics within a wind farm. The stability of the atmospheric boundary layer and the height of surface aerodynamic roughness have a significant impact on the mean wind profile and turbulence characteristics.

Supervisory control and data acquisition (SCADA) data is an industrial control system that can be used to monitor plant operations from a remote location or on-site (Liu & Corbita, 2021). SCADA systems are made up of both hardware and software that allow a user to control and monitor operations, so this would be a big help in terms of efficiency in correctly predicting failures, as well as the time it takes before the failure is predicted. A turbulence modeling based on SCADA data uses data already collected at the wind turbine controller. It’s a low-cost method of monitoring wind turbines for early detection of failures and performance issues. This study aims to compare machine learning techniques that are capable of using SCADA data to model turbulence intensity affecting the performance of wind turbines.

Wind turbines are devices that generate electricity by harnessing the power of the wind. Wind turbines have high operating and maintenance costs, which reduces their overall cost-effectiveness. Unscheduled maintenance due to unexpected failures is one of the biggest drivers of maintenance costs (Optis & Perr-Sauer, 2019). So, by checking its turbulence by validating wind turbulence models before they reach a catastrophic stage and eliminating unnecessary scheduled maintenance, it is possible to improve turbine reliability and lower maintenance costs. By analyzing the effects of different input and output parameters used in turbulence modeling, this research aims to evaluate, develop, and optimize the performance of wind turbulence models on wind turbines using SCADA data.

1.1. Wind turbulence

Turbulence is measured in the wind energy industry using a metric called turbulence intensity, which is the standard deviation of horizontal wind speed divided by the average wind speed over a given time, usually 10 minutes (Letzgus et al., 2022). The turbulence intensity will be high if the wind fluctuates rapidly. Stable winds, on the other hand, have less turbulence. In this paper, the researcher will discuss the engineering turbulence model by using turbine SCADA data.

Turbulence Intensity depends on upstream terrain, atmospheric stability, and wakes from neighbor turbines. Turbulence intensity causes the same material damage as the turbulence from all directions and it is defined in analogy to the equivalent load range (Balduzzi et al., 2020). The turbulence intensity is defined as the ratio of the standard deviation of the wind velocity in the average wind direction, divided by the average wind velocity. It is assumed that the wind turbine creates an additional turbulent kinetic energy that should be added to the ambient one.

The implicit assumption behind the effective turbulence intensity concept is that load ranges at a fixed wind speed are proportional to turbulence intensity and that typical weighting makes differences in the shape of load spectra of minor importance (Porté-Agel et al., 2020). The wake effects must be included in the model because wake turbulence is formulated as a combination of background turbulence and added turbulence as shown in equation 1(1) ${\mathit{I}}_{\mathit{eff}}\left({\mathit{V}}_{\mathit{hub}}\right)=\left[\frac{1}{2\mathit{\pi }}{\int }_0^{2\mathit{\pi }}\mathit{\rho }(\mathit{\theta }|{\mathit{V}}_{\mathit{hub}}){\mathit{I}}^{\mathit{m}}]\mathit{\hspace{0.17em}}({\mathit{V}}_{\mathit{hub}}\mathit{\theta }\right)\mathit{d\theta }{]}^{\frac{1}{\mathit{m}}}$(1) .

(1) ${\mathit{I}}_{\mathit{eff}}\left({\mathit{V}}_{\mathit{hub}}\right)=\left[\frac{1}{2\mathit{\pi }}{\int }_0^{2\mathit{\pi }}\mathit{\rho }(\mathit{\theta }|{\mathit{V}}_{\mathit{hub}}){\mathit{I}}^{\mathit{m}}]\mathit{\hspace{0.17em}}({\mathit{V}}_{\mathit{hub}}\mathit{\theta }\right)\mathit{d\theta }{]}^{\frac{1}{\mathit{m}}}$(1)

Where,

p—the probability density function of wind direction.

I—the turbulence intensity combined of ambient and wake flow from wind direction θ, and

m—the Wöhler (SN-curve) exponent for the considered material.

The implicit assumption behind the effective turbulence intensity concept is that load ranges at a fixed wind speed are proportional to turbulence intensity and that typical weighting with a high Wöhler exponent makes differences in the shape of load spectra of minor importance.

1.2. Machine learning

Machine learning is a category of artificial intelligence wherein software applications are trained using data to perform something. It is very useful in creating applications that can be used to solve repetitive problems. Mathematical modeling is very useful in generating machine-learning techniques.

1.2.1. Linear regression

Linear regression is an algorithm in machine learning which is more focused on supervised learning. It models a target prediction value that is based on independent variables. The main function of this model is for determining the relationship between variables and forecasting. This can be modeled using equation 2(2) $\mathit{y}=\mathit{\hspace{0.17em}}{\mathit{\theta }}_1+{\mathit{\theta }}_2\mathit{x}$(2) .

(2) $\mathit{y}=\mathit{\hspace{0.17em}}{\mathit{\theta }}_1+{\mathit{\theta }}_2\mathit{x}$(2)

Where,

x—input data

y—response variable

θ₁ and θ₂ -coefficients of the best regression fit line

1.2.2. Stepwise linear regression

Stepwise linear regression is an iterative method that examines the statistical significance of the parameter based on a linear regression model. This can be done through a forward selection approach, backward elimination, and bidirectional elimination. The method used equation 3(3) ${\mathit{b}}_{\mathit{j},\mathit{std}}=\mathit{\hspace{0.17em}}{\mathit{b}}_{\mathit{j}}\mathit{\hspace{0.17em}}\left(\frac{{\mathit{S}}_{\mathit{xj}}}{{\mathit{S}}_{\mathit{y}}}\right)$(3) as the basis of the standardization between the variables.

(3) ${\mathit{b}}_{\mathit{j},\mathit{std}}=\mathit{\hspace{0.17em}}{\mathit{b}}_{\mathit{j}}\mathit{\hspace{0.17em}}\left(\frac{{\mathit{S}}_{\mathit{xj}}}{{\mathit{S}}_{\mathit{y}}}\right)$(3)

Where,

S_xj and S_y—standard deviation of the variables

1.2.3. Fine tree

A fine tree is a decision tree in machine learning that is used to create a tree-like model of decisions. The algorithm is represented as a tree and it is used as a model to determine the decisions that could be used to solve specific problems. Figure 1 shows an example of how a fine tree is being modeled.

Figure 1. Fine Tree Model.

1.2.4. Linear support vector machine

Support vector machine is a type of machine learning model that can be used for classification and regression problems. Linear support vector machine focused on sets of data that are linearly separable. Figure 2 shows a classification problem that is solved using a support vector machine.

Figure 2. Support Vector Machine Classification.

1.3. Related studies

The discovery of wind-power-generated electricity dates to the end of the last century and has encountered many ups and downs in its more than 100-year history (Gambuzza & Ganapathisubramani, 2021). Wind energy extraction returned in the latter half of the twentieth century, thanks to a better understanding of aerodynamics and advances in materials, particularly polymers (Balduzzi et al., 2020). Wind energy devices, also known as wind turbines, are now used to generate electricity. A wind turbine is a device that converts the kinetic energy of the wind into mechanical energy in the form of a rotating shaft.

The SCADA system is a type of industrial measurement and control system that consists of a central host or master, one or more field data acquisition and control or remote units, and a collection of standards and/or custom software used to monitor and control field data elements that are located remotely. This type of system is characterized by open-loop control and relies heavily on long-distance communications. Closed-loop control and short-distance communication may also be present. Furthermore, SCADA systems generate huge amounts of data, which are frequently ignored even though important insights may be concealed within them (Meneveau, 2019).

According to an article in (Rezaeiha et al., 2019), the SCADA system gathers data on over 100 parameters and stores it every ten minutes. Because the wind farm’s turbines are all identical, one was chosen at random for an in-depth investigation of wind turbulence modeling. The approach given in this research, however, can be applied to any turbine. Turbulence modeling is a construction and application of a mathematical model to predict the effects of turbulence. Turbulence modeling is used to forecast the physical behavior of the turbulent flow generated in a system. A model’s accuracy, simplicity, and computational efficiency should all be considered (Ricci et al., 2020).

The turbulence characteristics of wind farm inflow have a significant impact on energy production and wind farm lifetime. The most common method is to estimate turbulence intensity, but this is not always available, and turbulence varies across the wind farm’s length. The energy losses in a wind farm caused by the effects of wind turbine wakes can often range from 10% to 20% (Dong et al., 2021). So, as the explanation for this, wind turbine wakes also increase turbulence, which can cause downstream wind turbines to become fatigued sooner. As a result, to estimate wind farm annual energy production and wind turbine loads, reliable and practical modeling of the influence of wind turbine wakes in wind farms is required. The atmospheric turbulence and wind turbine wakes are calculated using the chosen model which was developed to estimate the wind speed using the operational supervisory control and data acquisition (SCADA) system. It was created for use in real-time wind farm calculations that are required to perform control strategies, optimize the performance, evaluated the efficiency, and adhere to the cost.

SCADA systems are presently ubiquitous in most modern wind turbines. Because it relies on interpreting SCADA data, wind turbine conditioning and monitoring employing SCADA data analysis are both cost-effective (data collection and sensor networks are already in place) and dependable (J. H. Liu et al., 2022). The link between multiple signals can be studied and the wind turbine’s health components can be determined using SCADA data. An article in (Nybø et al., 2022) presents a wind turbine conditioning and monitoring based on SCADA that uses normal behavior models and fuzzy logic.

Turbulence is a key inflow characteristic that has been studied in several types of research. According to an article in (Martini et al., 2022), turbulence’s influence on a conventional power curve can be divided into two main components. The normal 10-minute averaging of power and wind speed data has an effect first. According to an article in (Grinderslev et al., 2021), when treating the power curve as a transfer function between wind speed and power, its non-linearity leads the produced power to be dependent on both the variance and the mean of the wind speed.

In general, a typical modern wind farm has wake losses of 10–20 percent (Nybø et al., 2021). As a result, estimating how much power a wind farm can produce for a given wind inflow is one part of the problem of wind farm wake. The development of wind turbine wakes in offshore wind farms is influenced by atmospheric stability, which is an important meteorological variable. Several studies have investigated this impact in the last few years. According to an article in (Somoano et al., 2021), a more intelligent wind farm operational strategy based on wake models could reduce equivalent fatigue loads on turbines while significantly increasing power output. In this study, the wake is not considered in the turbulence models since the wind farm is arranged in a single row.

Machine learning is becoming a research trend nowadays with the development of artificial intelligence in a variety of applications. A machine learning approach in a study in (Y. Liu et al., 2021) makes use of the technique to predict energy yields in wind turbines based on the degradation of some of its components. In this study, energy yield was successfully predicted using machine learning techniques.

Another study that utilizes machine learning in the field of wind turbines involves the prediction of wind power in the wind farm (Alom et al., 2021). The study utilizes the use of an artificial neural network in the wake models to predict the wind power in the entire wind farm.

This paper addresses the main gap in this literature by using the operational data of SCADA to generate turbulence models. The current trend in the use of SCADA data coupled with the integration of machine learning techniques is the key method being used in this study. Table 1 shows the summary of the models used by related literature and their corresponding details.

Table 1. Summary of Related Works

Turbulence Model	Brief Methodology	Highlights	Limitations
Kaimal and Mann Models (Dong et al., 2021)	Investigate the adoption of lidar-assisted control in wind turbine design.	Lidar measurement coherence is lower for the Mann turbulence model compared to the Kaimal Model.	The algorithm is not capable of using SCADA data in the generation of the model.
Eddy-viscosity turbulence model (Rezaeiha et al., 2019)	A comparative analysis of commonly used eddy-viscosity turbulence models.	Results showed that there is a high sensitivity of simulation results to the turbulence model.	The algorithm is not capable of prediction-based turbulence modeling.
Spalart-Allmaras and k-w SST model (Alom et al., 2021)	Presents a numerical study that assessed the effect of surface roughness distribution.	Both turbulence models gave comparable results except for the high angles of attack of the drag.	The algorithm is not capable of using SCADA data for model generation.

To address the limitations of the previous studies, particularly on the non-employment of the SCADA data and the capability of a prediction-based model, the researchers used algorithms that are capable of using the SCADA data and prediction-based modeling and performed a comparative analysis of the models.

2. Materials and methods

To achieve the aim of this research study, a certain framework is being followed in the conduct of the data analysis. To understand more about this framework, the researcher shows in Figure 3, the conceptual framework which illustrates the step-by-step process required to achieve the results of turbulence on a wind turbine farm.

Figure 3. Research Framework of the Study.

The first part of this study is to gain data from historical SCADA data from a wind farm. After extracting the SCADA data, the data will undergo pre-processing to get the data that is important to the turbulence modeling and ignore some data that is not needed during the analysis. When the necessary data is gathered, it will be subjected to parameter classifications, where the parameters related to the turbulence intensity model will be discussed. Following the classification of the parameters, the researcher will perform performance calculations using statistical analysis and analyze the effects of every parameter on the wind turbulence model, obtaining all the data and displaying the results in graphs.

2.1. Data extraction

Data extraction is the process of gathering or retrieving various types of data from a variety of sources, many of which are unstructured or poorly organized. Data extraction allows for the consolidation, processing, and refinement of data before it is stored in a centralized location and transformed. The SCADA system’s user interface could only provide historical data for the last two years.

In extracting the data, it is important to know the information regarding the wind turbine farm. As shown in Table 2, those are the information about the wind turbine farm and SCADA system would be useful for data extraction.

Table 2. General Information of Wind Turbine Farm

Location	Taiwan
Number of Wind Turbine Heads	Thirteen (13)
Nominal Power Capacity of Wind Turbine	2 MW
Transmission Connection	Direct-Drive
Number of Parameters Used	10
Downloaded Data Interval	10 minutes averaged

2.2. Data preprocessing

Data preprocessing involves the filtering of data values to exclude erroneous data recorded from the SCADA. This is performed to overcome one challenge of using SCADA data which is the presence of erroneous values. Table 3 shows the filters that were used in the study and their notation details. These filters were carried out using the Matlab software using the notation as indicated.

Table 3. Data Filter Used in the Study

Data Filter	Notation
Missing Data Filter	delete ${\mathit{x}}_{\mathit{t}}$ if $\left\{{\mathit{x}}_{\mathit{t}}|{\mathit{x}}_{\mathit{ti}}=\mathit{not}\mathit{available}\right\}$
General Boundary Filter	delete ${\mathit{x}}_{\mathit{t}}$ if $\{\left({\mathit{x}}_{\mathit{t}}|{\mathit{x}}_{\mathit{ti}}\mathit{\hspace{0.17em}}>\mathit{upper}\mathit{bound}\right)$ or $\left({\mathit{x}}_{\mathit{t}}|{\mathit{x}}_{\mathit{ti}}\mathit{\hspace{0.17em}}<\mathit{lower}\mathit{bound}\right)\}$
Efficiency Filter	delete ${\mathit{x}}_{\mathit{t}}$ if $\{\left({\mathit{x}}_{\mathit{t}}|{\mathit{x}}_{\mathit{ti}}\mathit{\hspace{0.17em}}>\mathit{upper}\mathit{limit}\right)$ or $\left({\mathit{x}}_{\mathit{t}}|{\mathit{x}}_{\mathit{ti}}\mathit{\hspace{0.17em}}<\mathit{lower}\mathit{limit}\right)\}$

2.3. Classification of input and output parameters

Wind turbulence is determined by several parameters, including, airfoils, blade pitch angle, rotor speed, rotor radius, nacelle parts, and wind speed. The SCADA system is used to select parameters that are relevant to the analysis of the main components of turbulence intensity. They are very useful if they reduce the model’s complexity. They incorporate better predictions and insights into the relevance of the variables in terms of turbulence intensity. Several output parameters are chosen from among the parameters for each module to serve as output models for the model generation.

Different combinations of input parameters from the SCADA system can be used to generate output models. This is accomplished by combining the filtered data from all of the wind farm’s turbine heads for each parameter. These parameters were chosen to represent the wind turbine’s turbulence intensity. The researcher combines all possible input parameters for each input parameter, and the output model is the result. The performance will be evaluated based on which input parameters correlate to the output models’ turbulence intensity. The researcher also divided its parameters into three sections of the wind turbine SCADA system used in this study: the generator system, the nacelle system, and the impeller/rotor system. The study’s parameters are divided and fitted into each section with wind speed data in every system, as shown in Table 4.

Table 4. Selected Parameters

Generator System	Nacelle System	Rotor System
Wind Speed Generator Speed Generator Voltage Generator Current	Wind Speed Resultant Force Wind Direction Atmospheric Pressure	Wind Speed Rotor Speed Rotor Angle Blade Rotor Angle

To determine which of the variables made available by SCADA systems are ideal for developing a turbine monitoring system through its components, bibliographic research is required for this step. The use of a combination of data approaches to determine the relationship between these quantities is common.

The proposed models’ various inputs and outputs are shown. Their decision was based on a review of the scientific literature to determine the variables that could be used. Generally, wind speed is the main parameter of the three systems to assess the wind turbulence model. As presented in (Castorrini et al., 2022), with the addition of wind speed, generator voltage, generator current, and generator speed are related in the wind turbine. The voltage generated normally increases as the speed increases, and the two have a direct link, therefore all parameters in the generator system are directly proportional. There is also a link between the nacelle system parameter and the rotor system parameter. The nacelle system’s parameters, which include wind speed, resultant force, wind direction, and atmospheric pressure, are used to evaluate the turbulence model and determine its correlation. Because of the relationship between rotor speed and angle, the data obtained in the nacelle system can be used to determine wind speed analysis in a plane of a rotor of a wind turbine.

2.4. Model generation

Different combinations of input parameters from the SCADA system can be used to generate output models. This is accomplished by combining the filtered data from all the wind farm’s turbine heads for each parameter. The researcher uses linear regression to generate predictive models based on input-output combinations. The MAE and MSE were used as linear regression performance measures. The researcher used five different models to assess and find the lowest value that can be used to analyze and interpret data: Fine Tree in Regression Tree, Linear Regression, Linear SVM, Linear Regression with a Hyperparameter in terms of Quadratic and Stepwise Linear Regression. A pseudo-code for the algorithm is shown in Figure 4.

Figure 4. Pseudo Code of the Algorithm.

An article in (Aboelezz et al., 2022) investigated the application of decision trees in the identification of desirable features. The decision tree is dependent on the features’ information gain. The research investigates which statistical attribute delivers the most data. Decision trees, a family of well-developed machine-learning algorithms, are well-known for sharing these qualities. The main benefit of decision tree algorithms is that the generated set of classification rules may be easily applied in many systems using fuzzy logic.

One of the strategies for predicting quantitative values is linear regression. Simple linear regression assumes that the input and output parameters have a simple connection and that the relationship is linear (Purohit et al., 2022). Linear SVM is a supervised learning algorithm used in classification and regression. It constructs a linear discriminant function separating instances as widely as possible (Ti et al., 2021).

Similar to linear regression, quadratic regression has a constant term, linear terms, and quadratic terms that incorporate interactions. A linear regression model with a quadratic term, as employed in the study in (Yang et al., 2022), can be utilized to determine a link between independent input process parameters and output data. The input parameters affect the output parameters, and they can be connected to the wind turbine mechanism to determine the various causes of system problems.

2.5. Performance analysis and measures

The airfoils, blade pitch angle, rotor speed, rotor radius, nacelle parts, and wind speed. Are all variables that can affect wind turbine performance in terms of turbulence. The mathematical results of the experimental parameters are examined.

The performance measures used in this study are the mean absolute error and the mean square error. The researcher will use MATLAB as a computing tool for calculation. These values can be used to analyze the predictive model’s performance, which in this case includes a comparison of wind turbine heads while also comparing different input combinations for each output model. The performance of different wind turbines is also compared for each output model using the model with the best input combination.

3. Results

3.1. Model performance

The MAE and MSE were used as performance measures for the linear regression. To assess and find the lowest value that can be used to analyze and interpret data, the researcher used five different models namely: Fine Tree in Regression Tree, Linear Regression, Linear SVM, Linear Regression with a Hyperparameter in terms of Quadratic and Stepwise Linear Regression. The effectiveness of the models is compared using data from the SCADA system. The models are then separated based on the results with the provided data, and the model with the fewest errors is analyzed in the MAE of the Generator as an example shown in Figure 5.

Figure 5. MAE of Turbine Head 5 Generator in Different Models.

It is shown in Figure 3 that among all the models, the Fine Tree in the Regression Tree and Linear Regression with a Hyperparameter in terms of Quadratic, had the fewest error rates of 0.054229 and 0.066813 respectively. The performance of the models shows that a fine tree algorithm could provide a better turbulence model owing to the nature of the algorithm (Cappugi et al., 2021). The nonlinearity of the SCADA data is one reason why the fine tree algorithm was more effective.

As shown in Figure 6, indicates that these two models (Linear Regression and Linear Regression Quadratic) perform better than the other models in terms of MSE based on data related to the generator output parameter in addition to wind speed. The data that are observed are used in linear regression to compare and relate data sets for a specific output parameter.

Figure 6. MSE of Turbine Head 5 Generator in Different Models.

3.2. Wind turbulence analysis of wind turbines

In recent years, the importance of meteorological influences on turbine siting and wind resource assessment has changed due to the increasing size of turbines and the erection of entire wind parks offshore and in difficult terrain. A study in (Purohit et al., 2022) stated that the velocity deficit and extra turbulence caused by any upstream wind turbine affect the flow towards a given wind turbine in a cluster of wind turbines. In general, interaction results in a complicated flow structure in and around the wind farm. Regions with multiple velocities wakes and multiple turbulence wakes, for example, may be identified depending on the wind direction and the layout of the wind farm.

Figure 7 shows the wind speed in December 2021, and analyzes the mean power is dependent on the mean wind speed and any factors like wind turbine spacing, turbulence intensity, and stability conditions. Although the analysis reveals some general trends in the relationship between wind speed turbulence and stability, it should be noted that variability within each class is typically greater than differences between classes.

Figure 7. Wind Speed Data in December 2021.

Turbulence in the atmosphere can have a substantial impact on turbine output and the loads that induce wind turbine component fatigue. Only by collecting turbulence measurements can the impact of turbulence be understood. A study in (Castorrini et al., 2022) presents three-dimensional turbulence at altitudes within, above, and below a turbine rotor disk as relevant measurements for understanding turbulence’s impact on wind energy. Atmospheric turbulence has various consequences on wind energy: it affects the generator’s individual turbines’ power performance, it affects turbine loads and fatigue, and it determines how wind turbines will impact their nearby environment through wake effects.

The processes of measuring the rotor’s rotational speed, the blades’ inclination angle, and the power generated by the wind turbine’s conversion machine, which is the parameter employed in the study, form the method of determining the wind speed in a plane of a rotor of a wind turbine. Developing a wind turbine dynamic regression model relating the rotational speed of the rotor to the wind speed in the rotor plane, the inclination angle of the wind turbine blades, and the power provided by the wind turbine.

Because of the link between rotor speed and angle, the data observed in the nacelle system is higher. The MSE of the trained models is larger. Because rotor operation is prone to vibrations, wind turbulence in the rotor is also higher. The rotor of the wind turbine is directly connected to the rotating blades, resulting in more inaccuracies than other sections of the wind turbine. Because the results are seen to be larger in this area of the wind turbine, possible reasons for failures due to wind turbulence can be located in the nacelle system.

3.3. Performance of wind turbines

Enhancing parameter combinations is an essential part of standardizing the SCADA system’s condition monitoring process. The MSE performance of the various input combinations is analyzed using linear regression models. To put it another way, the parameters of the models are analyzed concerning other parameters.

The collected data is analyzed with a smaller set of errors using the MSE in the SCADA system as shown in Figure 8. The parameters inside the generator are used to observe wind turbulence in all turbines of the generator part. These data sets are correlated with one another, and wind turbulence has been observed to affect the wind turbine generator. In the study in (Nybø et al., 2021), generator voltage, generator current, and generator speed are related in the wind turbine with the addition of wind speed. Wind speed is caused by the production of wind turbulence due to atmospheric stability. As you get closer to a wind turbine generator, the wind speed drops, and turbulence rises in anticipation of the flow-disrupting and energy-extracting object. Furthermore, there is proportionality in the increase between the generator voltages to that of the increase in the generator speed by the wind speed making the model perform better for those combinations with generator voltage and generator current. The data shows that all of the wind turbine heads in the generator have almost no MSE except for the H07, which has 0.16346 because the generator system of the H07 has inconsistent data and is always damaged in the year 2021.

Figure 8. MSE of all Wind Turbine Heads in the Generator.

The rotor response is lagging in more convective and turbulent conditions as the turbine responds more quickly to drops in wind speed. Rotor blades are used to simulate the flow near a wind turbine. According to an article in (Nybø et al., 2022), over the rotor disk, the wind is assumed to have constant turbulence intensity and a non-uniform mean wind speed that varies vertically according to the shear exponent. As a result, both the mean wind speed and the amount of turbulent wind speed fluctuations around the mean are vertically non-uniform in the resulting wind. Turbulence is produced by a wind turbine at a scale of the rotor size, rotor angle, and the correlating frequency or rotor speed that has been used as a parameter in this model. As seen in Figure 9, all the wind turbine heads have a considerable amount of MSE having the lowest error of 1.6454. Each parameter affects the wind turbulence of a wind turbine and provides information about it by giving its MSE. Wind turbulence in the rotor has an impact on the wind turbine’s performance.

Figure 9. MSE of all Wind Turbine Heads in the Rotor.

At lower wind speeds, an article in (Ricci et al., 2020) stated that mixing in the atmosphere during more convective conditions, as well as turbine interaction, may cause additional motion, exaggerating nacelle blockage effects and causing underestimation. Rotor efficiency may influence flow induction and thus the wind speeds measured on the back of the nacelle during more turbulent conditions. Furthermore, if the turbine and rotor efficiencies are lower during convective and more turbulent periods, less momentum may pass through the rotor and along the nacelle. That explains why the results in Figure 10 show that almost the entire wind turbine heads have a bigger value of MSE.

Figure 10. MSE of all Wind Turbine Heads in the Nacelle.

4. Discussion

The quantification results indicate that in terms of the model performance, fine tree and linear regression in terms of quadratic models outperform other models since they have less error. The data on the performance of wind turbine heads in the generator demonstrates that, except for the H07, all of the wind turbine heads have nearly no MSE. In terms of the rotor, all the wind turbine heads have a considerable amount of MSE having the lowest error of 1.6454. For the nacelle, since the turbine and rotor efficiencies are lower during convective and more turbulent periods, less momentum may pass through the rotor and along the nacelle which is why almost all the wind turbine heads have a massive value of MSE.

The performance of the various machine learning techniques used was dependent on the behavior of the data used in the generation of the models. The linear regression model has the highest error value since SCADA data are not linear. This is also evident in the linear SVM and the stepwise linear regression since the linearity of the data is focused on the behavior of the data. The fine tree model has the lowest error value since the data is modeled based on various decision levels generated in the model.

The results presented are an indicator that the SCADA data provided was used efficiently and highly useful in assessing the performance of the wind turbine. This exemplifies that, despite the high operating and maintenance costs of wind turbines, it is possible to improve turbine reliability and lower maintenance costs by improving the design and reliability with a more reliable turbulence model. Improving parameter combinations is an important step toward standardizing the SCADA system’s condition monitoring process. Linear regression models are used to examine the MSE performance of various input combinations.

4.1. Recommendation

The use of healthy data in the system is also a factor in the analysis and the production of more accurate outcomes. The methodological approach was tested in a real-world case study involving wind turbines, and it proved successful in detecting anomalous behaviors before the emergence of turbulence. As a result, the developed system has shown to be a vital support for wind farm maintenance, offering extra information to update the present maintenance policy based on time-based scheduled inspections and alarms, which frequently do not allow enough time to prevent critical scenarios. The progression of mentioned methodology toward wind turbulence diagnostics is recommended for this approach.

Wind turbulence that happened should be correlated with specific patterns on the control charts, creating the foundations for further automation of the wind turbulence diagnostic, in addition to recognizing the abnormal behavior. To accomplish so, professionals and maintenance people are needed. This has interesting implications, especially in terms of maintenance job scheduling optimization. One recommendation made was to consider the frequency of the wind turbine if the wind turbine will generate in different nations, such as the Philippines. Experimenting with the frequency of the wind turbine could be part of the future scope for testing the wind turbine’s s durability and detecting portions that are easily damaged.

Author contributions

Conceptualization, Jui-Hung Liu, Jien-Chen Chen, and Nelson T. Corbita Jr.; Data curation, Jui-Hung Liu; Formal analysis, Jien-Chen Chen and Nelson T. Corbita Jr.; Investigation, Jui-Hung Liu; Methodology, Jui-Hung Liu and Nelson T. Corbita Jr.; Project administration, Jui-Hung Liu and Jien-Chen Chen; Validation, Nelson T. Corbita Jr.; Writing – original draft, Nelson T. Corbita Jr.; Writing – review & editing, Jui-Hung Liu. All authors have read and agreed to the published version of the manuscript.

Acknowledgements

The support from Taipower Company was highly appreciated.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

The authors received no direct funding for this research.

Notes on contributors

Jui-Hung Liu

Jui-Hung Liu and his team have been focused on large wind farm operations and maintenance for over 15 years. Based on Taiwan’s special wind conditions for the wind turbines, the team keeps developing monitoring systems, control strategies, and intelligent algorithms to improve the turbines’ operation and performance. The present paper utilizes the SCADA data and relates to the special turbulent condition in that wind farm to show the importance of a correct turbulent model.