Neural Networks for Improving Wind Power Efficiency: A Review

Abstract

The demand for wind energy harvesting has grown significantly to mitigate the global challenges of climate change, energy security, and zero carbon emissions. Various methods to maximize wind power efficiency have been proposed. Notably, neural networks have shown large potential in improving wind power efficiency. In this paper, we provide a review of attempts to maximize wind power efficiency using neural networks. A total of three neural-network-based strategies are covered: (i) neural-network-based turbine control, (ii) neural-network-based wind farm control, and (iii) neural-network-based wind turbine blade design. In the first topic, we introduce neural networks that control the yaw of wind turbines based on wind prediction. Second, we discuss neural networks for improving the energy efficiency of wind farms. Last, we review neural networks to design turbine blades with superior aerodynamic performances.

1. Introduction

Concerns surrounding climate change continue to grow and related treaties have been signed. For example, in COP26 (26th UN Climate Change Conference of the Parties), most developed countries agreed to abolish traditional coal power generation and halt new construction. This agreement aims to control rises in temperature by achieving a “net zero” balance of carbon emission and absorption by 2050 [1]. Therefore, it is crucial to minimize, or eliminate if possible, carbon generation during energy harvesting processes.

Wind is a ubiquitous source of energy that can be harvested without carbon emission. Wind turbines (WT) are utilized to convert the kinetic energy of wind to usable forms of energy (e.g., electricity). Accordingly, the installation of WTs has significantly increased owing to their negligible carbon emission [2,3]. However, the number of WTs that can be installed is limited due to concerns related to species protection and geopolitical and supply risks [2]. Indeed, the high production and maintenance costs associated with WTs also present one of the main factors warranting consideration before installation [4]. Therefore, with limits in increasing the number of WTs, it is essential to maximize the zero-carbon energy generation from wind by further optimizing the energy efficiency of turbines.

The energy efficiency of WTs depends on the interactions between atmospheric flow and wind blades. Atmospheric flow—a turbulent boundary layer flow with moving boundaries and external energy sources—is highly non-linear. The interaction between turbulent flow and wind blades is also non-linear, and depends on factors such as wind direction, angle of attack, and blade geometry. Hence, designing WTs requires simulations or experiments that involve full interactions between turbulent atmospheric flow and turbine blades [5,6]. Therefore, significant challenges exist in optimizing the energy efficiency of WTs. Recently, several studies have employed machine learning (ML) models to optimize WT energy efficiency [7,8] or to maintain WTs [9,10,11] efficiently. In particular, neural networks (NN), a type of ML model, have shown significant potential to be utilized for wind energy harvesting. This is due to the state-of-the-art performance of NNs in learning non-linear patterns, such as the chaotic patterns in atmospheric flows. Studies on improving the energy efficiency of WTs using deep NNs can be categorized as follows (see Figure 1):

• NN-based single turbine control
• NN-based wind farm control
• NN-based wind blade design

We discuss the recent developments of deep NNs for improving the energy efficiency of WTs in the three categories above. This paper is organized as follows: NN-based turbine control is discussed in Section 2. Section 3 investigates approaches for NN-based wind farm control. NN-based blade design methods are introduced in Section 4, followed by concluding remarks in Section 5. All WTs covered in this paper are horizontal-axis upwind wind turbines (see Appendix A for more details about upwind wind turbines).

2. NN-Based Single Turbine Control

Atmospheric flows around WTs are turbulent, leading to chaotic changes in harvestable winds in space and time. Therefore, a WT must constantly respond to the surrounding atmospheric flow to stabilize power production and minimize fatigue damage [4,12]. Accordingly, modern WTs are designed with pitch- and yaw-angle control systems (Figure 2). A turbine pitch-control system controls the pitch of wind blades. A turbine yaw-control system controls the WT’s nacelle position (i.e., turbine face direction) to react to the chaotic surrounding flows properly [4].

A turbine pitch-control system generally aims to prevent WT damage caused by large aerodynamic loads due to excessively high-speed winds with large turbulent fluctuations [4,12,13,14,15,16]. This can be achieved by controlling the pitch angle of wind blades, leading to a change in angle of attacks [12]. Newer pitch-control systems employ electrical or hydraulic drive systems with a proportional–integral–derivative (PID) controller or a proportional–integral (PI) controller [13,14]. Since NN-based pitch-control systems generally aim to prevent turbine damage, they are not further reviewed in this paper. In contrast to pitch-control systems, turbine yaw-control systems aim to maximize energy production. Controlling the yaw angle is one of the most effective methods for maximizing wind energy efficiency. Generally, the maximum wind energy can be harvested when the turbine’s nacelle position is aligned with the surrounding wind velocity vector. One approach for this could be to align the position continuously with the guidance of sensors. However, conventional wind turbines rotate with a slow angular velocity of almost $\sim 0.5°$/s [17,18,19]. Due to the slow angular velocity, changes in the surrounding wind direction can occur during rotation, making it challenging to maintain the desired alignment. Continuous movement would also lead to faster wear. In addition, according to Yang et al. [20], the wind data measured by a turbine’s sensor differs from the true wind conditions. Therefore, it is suggested to use future wind directions for improving wind turbine performance. However, forecasting wind velocity presents a significant challenge due to the turbulent characteristics of atmospheric flow. Recently, researchers have attempted to control the yaw angle using NNs for wind forecasting. Details of such methods are discussed below.

The idea to control yaw angles of WTs based on wind prediction was proposed by Hure et al. [17]. The authors developed a yaw controller using information from short-term wind predictions and reported an improvement in power generation. The wind prediction was performed using a multi-layer perceptron (MLP) type NN developed by Kani and Ardehali [21] and Ðalto et al. [22]. Consequently, researchers have actively employed MLPs for predicting wind [23,24,25]. Figure 3 provides an example of such an MLP. In this example, the input layer has three neurons ($I_1$, $I_2$, $I_3$), each hidden layer has four neurons ($X_1$, …, $X_4$; $Y_1$, …, $Y_4$), and the final layer has two neurons ($O_1$, $O_2$).

NNs are reported to display state-of-the-art performance in learning non-linear data. They can be considered non-linear functions that transform input information into predictions. Let $X_i$ and $Y_j$ be two connected layers with $i\in \{1,\dots ,m\}$ and $j\in \{1,\dots ,n\}$ neurons, with $X\in {\mathbb{R}}^n$ and $Y\in {\mathbb{R}}^m$. The information in X is then transported to the next layer Y by a matrix calculation with non-linear activation ($f(·)$) as:

$${\displaystyle Y=f(WX+b),}$$

(1)

where W is the weight matrix connecting the layers X and Y, and b is a bias vector. The process is illustrated in Figure 4 for the computation of $X_1$ from Figure 3. Note that $W\in {\mathbb{R}}^{n\times m}$ and $b\in {\mathbb{R}}^n$. The objective of NNs is to learn the best weights W and biases b that best model the mapping between input and output data. The activation function ($f(·)$) enables NNs to learn non-linear relationships between input and output data [26]. Note that a leaky-Rectified linear unit (ReLU) activation function, a variation of the ReLU activation function [27], is one of the most representative activation functions, and is defined as

$$f\left(x\right)=max(\alpha x,x),$$

(2)

where x is an arbitrary tensor, $\alpha$ is typically a value between $0.01<\alpha <0.1$ [28,29]. The input information is transported to the output layer by repeating the process in Equation (1). The final layer of the NN outputs the prediction. Subsequently, the errors between the model prediction and ground truth data (i.e., labeled data) are computed to evaluate the model prediction. Generally, the mean squared error (MSE) is utilized to evaluate errors in regression results. The MSE is the mean value of squared errors defined as:

$${\displaystyle MSE=\frac{1}{n}\sum _{i=1}^n{(\widetilde{y_i}-y_i)}^2,}$$

(3)

where $\tilde{y}$ and y are the predicted and labeled values, respectively. The weights and biases are updated with respect to a minimization of the calculated error.

More recently, owing to the success of long short-term memory (LSTM) NNs, some studies have used LSTMs for predicting the wind direction [30,31]. An LSTM is proposed to compensate for a recurrent neural network (RNN)’s difficulty in learning long-term time-series data.

An RNN predicts sequential values by inputting the previous time step’s hidden state ($h_{t-1}$) and the current time step’s input ($i_t$) to a gate with a hyperbolic tangent function (see Figure 5). The current time step’s output ($h_t$) functions as the next time step’s hidden state. This process can be seen as having a short-term memory of past values.

Figure 6 provides a detailed illustration of an LSTM that is known to outperform RNNs in learning sequential correlations for the sake of the long- and short-term memories. Unlike RNNs, an LSTM contains three gates:

• Forget gate: The forget gate calculates the element-wise product of the $sigmoid$ ($1/(1+e^{-x})$) values of the current input ($i_t$) and previous hidden state ($h_{t-1}$). The forget gate affects the current cell state ($C_t$).
• Input gate: The input gate computes the element-wise product of the $sigmoid$ and $tanh$ values of $i_t$ and $h_{t-1}$. The result of the input gate is used together with the result of the forget gate to compute $C_t$.
• Output gate: The output gate performs the element-wise product between the $tanh$ value of $C_t$ and $sigmoid$ values of $h_{t-1}$ and $i_t$. As a result, the current hidden state ($h_t$) can be calculated.

The cell states and hidden states in an LSTM cell function as long- and short-term memories, respectively.

Using an LSTM, Chen et al. [18] successfully controlled the yaw angle of a wind turbine. Their LSTM model was trained with wind data acquired by a light detection and ranging (LiDAR) system. The input of their model was the wind LiDAR data, while the outputs were the desired yaw angle and actuation time that can cope with future wind. Using the LSTM model, energy efficiency improved by approximately $3.5\%$ and a reduction of 5.3% in travel distance in yaw direction were reported.

In addition to LSTM, convolutional neural networks (CNN) are also well known for their capability of learning spatio-temporal data [32,33,34,35]. Few studies have integrated yaw-control systems with recent CNNs to advance wind prediction. In this paper, we also review CNN-based wind prediction methods that could improve turbine yaw-control systems.

CNNs with different dimensionalities have been exploited for wind prediction, such as one-, two-, and three-dimensional CNNs. First, one-dimensional (1D) CNNs are capable of learning one-dimensional time-sequential data (Figure 7a). For example, Harbola and Coors [36] developed a 1D CNN that predicts temporal wind data in the Netherlands and in the German city of Stuttgart. The authors pre-processed the wind velocity data by categorizing it into 11 classes with different ranges (lower range, upper range):

$$\begin{array}{c}\{(\mu -k_1\sigma ,\mu +k_1\sigma ),(\mu +k_1\sigma ,\mu +k_2\sigma ),(\mu +k_2\sigma ,\mu +k_3\sigma ),(\mu +k_3\sigma ,\mu +k_4\sigma ),(\mu +k_4\sigma ,\mu +k_5\sigma ),\\ (\mu +k_5\sigma ,\infty ),(\mu -k_2\sigma ,\mu -k_1\sigma ),(\mu -k_3\sigma ,\mu -k_2\sigma ),(\mu -k_4\sigma ,\mu -k_3\sigma ),(\mu -k_5\sigma ,\mu -k_4\sigma ),(-\infty ,\mu -k_5\sigma )\},\end{array}$$

where $\mu$ and $\sigma$ are the mean and standard deviation of the wind velocity data, and $(k_1,k_2,k_3,k_4,k_5)=(0.15,0.45,0.65,0.95,1.25)$.

Their CNNs aim to predict the class of wind velocity on a future occasion. They developed a 1D CNN that predicts the future wind speed and velocity based on the past 25-h wind velocity data within 30-min time-intervals. This 1D CNN model successfully classified the range of the future wind speed and direction at 96.8%, 99.7%, respectively, for Stuttgart, and 98.8%, 99.4%, respectively, for the Netherlands.

However, the underlying physics of atmospheric flow is not one-dimensional. It is a function related to both space and time. Therefore, physically, the inclusion of spatial information should improve the performance of wind prediction. Two-and three-dimensional CNNs can learn spatial correlations in wind data. Hong and Satriani [37] used a two-dimensional (2D) CNN model to predict wind at a target site in Taiwan, using additional wind data from six other regions in South Korea, China, and the Philippines. In total, 78 h of past wind data at the seven sites, which can be rearranged to a $78\times 7$ array, were used as the input of their 2D CNN model (Figure 7b). Their model showed a lower root mean square error (RMSE), normalized mean square error (NMSE), and mean absolute percentage error (MAPE) compared to an LSTM and an MLP. The errors are defined as:

$$\begin{array}{cccc}\hfill \mathrm{root}\mathrm{mean}\mathrm{square}\mathrm{error}:\hfill & \hfill RMSE\hfill & \hfill =\hfill & \hfill \sqrt{MSE},\hfill \\ \hfill \mathrm{normalized}\mathrm{mean}\mathrm{square}\mathrm{error}:\hfill & \hfill NMSE\hfill & \hfill =\hfill & \hfill MSE/{\sigma }^2,\hfill \\ \hfill \mathrm{mean}\mathrm{absolute}\mathrm{percentage}\mathrm{error}:\hfill & \hfill MAPE\hfill & \hfill =\hfill & \hfill 100/n\sum _{i=1}^n|(\widetilde{y_i}-y_i)/\widetilde{y_i}|,\hfill \end{array}$$

(4)

where y is the ground truth value, $\tilde{y}$ is the predicted value, n is the number of data, and ${\sigma }^2$ is the variance of the ground truth data.

Wu et al. [38] proposed a novel wind prediction method based on a graph neural network (GNN) with a Transformer. GNN is a model capable of learning unstructured data in graph structures [39]. A Transformer is an NN model derived from attention mechanisms proposed by Vaswani et al. [40], which have shown state-of-the-art performance in learning sequential data. The developed GNN model with a Transformer was reported to outperform a CNN-LSTM model and a 2D-CNN model for predicting wind speed.

Table 1. NN-based wind prediction and single WT control methods.

References	NN Model	Purpose	Data Source	Sampling Rate	Quantitative Performance
Kani and Ardehali (2011) [21]	MLP	Wind direction prediction	-	2.5 s	$MAPE=3.144\%$ (wind speed prediction)
Dzulfikri et al. (2020) [24]	MLP	Wind direction prediction	Measured wind data	3 h	$MAPE=0.4\%$ (wind direction prediction)
Zhang et al. (2022) [25]	MLP	Wind speed on the WT surface prediction	Measured wind	-	$MAPE=1.479\%$ (wind speed prediction)
Delgado and Fahim (2020) [30]	LSTM	Wind speed and direction prediction	Measured wind data in Turkey	10 min	R-squared $>0.902$ (wind speed prediction) and R-squared $>0.366$ (wind direction prediction)
Chen et al. (2020) [18]	LSTM	Yaw angle control	Simulated wind	10 s	power increases by 3.5% (at the wind speed of 8 m/s)
Harbola and Coors (2019) [36]	1D CNN	Dominant wind speed and direction classification	Measured wind data in Germany and Netherlands	30 min	98.8% and 99.7% accurate in dominant wind speed and direction classification
Hong and Satriani (2020) [37]	2D CNN	Dominant wind speed and direction classification	Measured and simulated wind data in China and S.Korea and Taiwan and Philippines	1 h	$MAPE=13.840\%$ for best CNN.
Wu et al. (2022) [38]	GNN & Transformer	Wind speed and direction prediction	Measured wind data in Denmark and Netherlands	1 h	Mean absolute error $=1.244$ (wind speed [m/s] prediction in Denmark dataset after 6 h)

summarizes the studies of predicting wind information using NNs. The information on the employed NN models, the purpose of the study, the data source and sampling rate, and the quantitative performance of studies introduced in the current section are included in the table.

3. NN-Based Wind Farm Control

A wind farm is a power plant composed of arrays of WTs. The total power output of a wind farm is lower than the summation of power outputs of stand-alone WTs. This is because WTs at downstream locations produce less wind energy by encountering wake flow with less momentum due to the upstream turbines [41,42] (Figure 8a). The computational fluid dynamics (CFD) results in Figure 8b, which were generated with the framework developed in [43,44,45], provide an example of such wake effects for a wind farm. The velocity field of an offshore wind farm is shown with the easterly wind flowing from right to left. The wind farm is located in the North Sea and contains 370 neighboring WTs. It is clearly visible that the wind speed declines and that wake flow significantly affects wind power generation. Therefore, it is crucial to include wake effects when controlling all individual turbines in a wind farm.

The most representative method to embed wake effects in yaw angle control systems is reinforcement learning (RL) [46,47,48]. RL is used to learn a sequence of actions that leads to an optimal state in an environment concerning a pre-defined optimization criterion. RL is an iterative mechanism where a single agent or multiple agents take an action based on a state provided by an environment. The environment then evaluates the action(s), provides feedback in the form of a reward, and passes the new state to the agent(s). A schematic diagram of controlling a wind farm using RL is shown in Figure 9. The optimum action(s) of changing the yaw angle of each WT based on a given state of wind flow could be learned with respect to the total power output. This can be achieved through trial and error explorations performed by an agent or agents, who are rewarded by a quantity related to energy production.

Zhao et al. [47] used an RL algorithm based on the deep deterministic policy gradient (DDPG) approach to control wind turbines in a wind farm. The DDPG is an RL algorithm for environments with continuous action space, such as yaw angles of WTs. For a discrete action space, each action could be evaluated with a so-called Q-value. For a continuous action space, however, the complete space cannot exhaustively be evaluated with this approach. Instead, a gradient-based learning rule is formulated that takes advantage of a Q-function that is differentiable with respect to the yaw angles chosen by an agent or multiple agents. For a more detailed explanation of DDPG, readers may refer to the following reference [49]. In [47], an algebraic wake model was used to model the wake effect and control the wind farm through knowledge-assisted learning DDPG (KA-DDPG). KA-DDPG constrains the learning of agents to physical rules and domain knowledge. Because of this constraint, an agent of KA-DDPG becomes capable of avoiding dangerous or low-reward actions. They reported $10\%$ improvement compared to a control algorithm (e.g., greedy algorithm) which does not involve any future consequences of an action of WTs. Similarly, Dong et al. [48] controlled WTs in a wind farm using a variant of DDPG. They reported a $15\%$ improvement in power production. It is also worth noting that they calculated the flow fields in a wind farm using high-fidelity simulations. Thus, no algebraic wake model was used.

Predicting the changes of flow fields due to wakes in wind farms using simulations is highly expensive, while algebraic wake models are inexpensive but inherently contain uncertainty. This is due to the challenging nature of accurately modeling turbulence and geometric effects. Therefore, developing an accurate wake model is important to improve the yaw control of wind farms. Recently, researchers have attempted to model wake by using NNs. For example, Li et al. [50] predicted unsteady turbine wakes using a CNN with two types of inputs. The first input is velocity field data over a certain time interval from the past. The second input is data of inflow velocity and yaw angle of each WT in the wind farm from the same past time interval. Their CNN model predicts the velocity fields at a future time based on these two inputs. The predictions under diverse conditions were compared with the result of large-eddy simulations (LES). LES computes turbulent flow by modeling sub-grid scale turbulence. Their CNN model showed an average error of 3.7% in predicting the wind velocity in a wind farm. Bleeg [51] developed a graphical neural network (GNN) to model wake effects. Reynolds-averaged Navier–Stokes (RANS) simulations were employed to generate training data. They developed a GNN that predicts the decay rate of wind speed in downstream locations relative to the front row of WTs. They reported that the developed GNN was capable of predicting the velocity decay rate with a small error (0.0028 in Mean absolute error).

Table 2. NN-based wind farm control and modeling methods.

References	Method	Purpose	Data Source	Quantitative Performance
Zhao et al. (2020) [47]	RL (KA-DDPG)	WT Yaw control	Wind farm model based on 2D CFD simulation	$10\%$ improvement in power generation
Dong et al. (2021) [48]	RL (DDPG)	WT Yaw control	Wind farm simulation	$15\%$ improvement in power generation
Li et al. (2022) [50]	CNN	Freestream wind speed prediction	Wind farm simulation	$3.7\%$ error
Bleeg (2020) [51]	GNN	Wind speed decay in wind farms	RANS simulation	Mean absolute error $=0.0028$ for a non-dimensionalized wind speed.

summarizes the studies of yaw control and wind speed prediction in wind farms using NNs. The information on the employed machine learning method, the purpose of the study, the data source, and the quantitative performance of studies introduced in the current section are included in this table.

4. NN-Based Turbine Blade Design

WTs convert the rotational energy of a rotor to a usable energy source. The rotation of rotors occurs due to the aerodynamic lift force generated on the turbine blade surfaces. The key to maximizing wind energy efficiency is to maximize the lift-to-drag ratio of a turbine blade. This ratio describes how effectively a blade can generate lift (i.e., rotation) while experiencing energy loss from drag. It is highly dependent on the shapes of the turbine blades and the flow conditions. Optimizing the design of a turbine blade aims to improve the lift-to-drag ratio at flow conditions in an operating WT. However, such an optimization process involves numerous design elements and conditions. In addition, for design robustness, flow analysis on the blade considering these factors should be repeated for various flow conditions. Therefore, the optimization process becomes costly. The increased computational costs can be overcome by employing NNs for efficiently designing optimal turbine blades.

Wen et al. [52] proposed a design method that combines a genetic algorithm and a back-propagation (GABP) NN to optimize the design of airfoils with reduced computational cost and complexity. Their NN was trained to predict a WT blade’s lift coefficient and lift-to-drag ratio, whose shape is represented by Bessel polynomials. The NN is reported to have a $90\%$ accuracy in predicting an arbitrary turbine blade’s lift coefficient and lift-to-drag ratio. Based on this network, the authors were able to design a blade with improved aerodynamic properties. Lalonde et al. [53] developed and compared six different NNs which function as surrogate models of an aerodynamic blade. These surrogate models are helpful because an NN connects input and output data without time-consuming computations such as CFD simulations. For example, they used the time-independent MLP with a full set of input data architecture with the following aerodynamic input data: wind speed, blade geometry, and blade deflected shape. Consequently, the CNN model predicted the aerodynamic load on the blade with a normalized root mean squared error (NRMSE) of 0.66%, where NRMSE is defined as:

$$NRMSE(\%)=\frac{RMSE}{\overline{o}}\times 100$$

(5)

where $\overline{o}$ is the average of the ground truth values.

Jasa et al. [54] proposed an approach for an airfoil design that utilizes an invertible neural network (INN), as illustrated in Figure 10. An INN processes inputs to outputs and vice versa. The INN was trained with data acquired from two-dimensional numerical simulations of airfoils. The inputs of their INN are airfoil shapes and operating conditions, whereas the outputs are aerodynamic performances and structural considerations. The INN’s invertible nature allows the creation of an airfoil design that fully satisfies the desired aerodynamic performances and structural considerations. Their optimized blade is reported to increase the annual energy production by 3.4% compared to non-optimized airfoil.

Oh [55] developed a neural network for predicting and optimizing the aerodynamic performances of WT airfoils. The developed NN demonstrated a superior optimization performance compared to conventional design methods, such as the response surface method (RSM) [56]. They found their NN suitable for learning the integrated influences of airfoil geometry and flow conditions on aerodynamic performances. Their optimized airfoil using NN showed 18.560% and 8.194% increases in the lift-to-drag ratio compared to an unoptimized airfoil and an airfoil optimized through RSM, respectively.

Jia et al. [57] demonstrated a tunable WT blade using RL. The twist angle distribution of the turbine blades was tuned using RL to maximize aerodynamic performances. Based on the power coefficient ($c_p$) of a blade with seven controllable actuators. they stated an average improvement of 12.9% in the high-speed wind regime (9∼13 m/s), but showed an average reduction in performance of -11.3% at low speed (5∼8 m/s), compared to blade control using a conventional genetic algorithm (GA).

Table 3. NN-based turbine blade design methods.

References	Method	Purpose	Data Source	Quantitative Performance
Wen et al. (2019) [52]	MLP + GA	WT blade’s lift coefficient & lift-to-drag ratio prediction	Experimental airfoil database	90% accuracy (maximum lift-drag ratio and the maximum lift coefficient prediction)
Lalonde et al. (2020) [53]	CNN	Aerodynamic load on the blade prediction	Wind turbine simulation	$NRMSE=0.66\%$ (Aerodynamic load prediction)
Jasa et al. (2022) [54]	INN	Airfoil design optimization	CFD simulation	Increase in Annual energy production by $3.4\%$
Oh (2020) [55]	MLP	Creation of surrogate model and Design optimization model	Airfoil simulation	18.560% and 8.194% increase over baseline and RSM (lift-to-drag ratio)
Jia et al. (2021) [57]	RL	Creation of surrogate model and Finding optimal blade twist angle distribution	Wind turbine simulation	Average improvement of 12.9% compared to GA (at the high-speed wind regime 9∼13 m/s)

summarizes the studies on designing WT blades using NNs. The information on the employed machine learning method, the purpose of the study, the data source, and the quantitative performance of studies introduced in the current section are included in this table.

5. Conclusions

WTs generate wind power from the kinetic energy of atmospheric flow. NNs are suitable for efficiently modeling the effects of atmospheric flow to wind power generation since NNs are capable of modeling non-linear functions. We investigated studies that have successfully improved power generation efficiency using NNs. To the best of our knowledge, we are the first to provide a review of neural-network-based methods to improve WT power generation in categories that cover wind turbines to wind farms and from control methods to blade design strategies.

First, studies focused on increasing the efficiency of a WT through the control of yaw angles were reviewed. In these studies, various types of NNs, e.g., MLP, LSTM, or CNN, were used to predict wind directions. Maximum power generation of a WT generally occurs when the nacelle direction and the wind direction are parallel. Based on accurate and fast wind predictions, the yaw of a WT can be controlled to achieve an optimal position and increase power generation.

Second, studies aimed at improving the energy efficiency of wind farms were reviewed. Due to the momentum loss at upstream WTs’ wake, the total power generation of a wind farm is inevitably lower than the sum of the maximum power generation of individual WTs. Typically, quantifying such interactions, e.g., wake flows, is computationally expensive. This renders the analysis of a wind farm arrangement difficult. We reviewed studies that employ NNs to overcome this limitation. A number of studies employed RL to control wind farms in a way that minimizes the negative effects of interactions between WTs and maximizes total power output. Other studies concentrate on modeling complex wake effects by developing NN-based surrogate models.

Finally, NN- and RL-based studies that optimized WT blade designs to improve the aerodynamic performances of WTs were reviewed. An optimization of turbine blades normally requires iterative CFD computations. Since CFD is a time-consuming method that requires a large amount of computational resources, NN-based surrogate models of CFD have been actively proposed. The developed surrogate models show a small margin of error and only take a fraction of the time needed to conduct simulations, leading to blade designs with improved wind power generation efficiency.