Review of Artificial Intelligence-Based Design Optimization of Wind Power Systems

Abstract

This paper reviews the applications of artificial intelligence (AI) in the design optimization of wind power systems, mainly including (1) wind farm layout optimization; (2) wind turbine design optimization; and (3) wind farm electrical system design optimization. Firstly, this paper introduces the general considerations in the optimal design of wind power systems and the AI methods commonly used for the optimal design of wind power systems. Then the applications of AI in the optimal design of wind farms are reviewed in detail. Finally, further research directions of using AI methods in the optimal design of wind power systems are recommended.

1. Introduction

Many countries are now concentrating on carbon neutrality in order to achieve the full implementation of the Paris Agreement. One of the best strategies to meet the carbon neutrality target, save the environment, and eliminate dependence on fossil fuels that are becoming more and more scarce is the broad use of renewable energy. As a type of renewable energy, wind energy is inexhaustible, and wind power technology is the most competitive, matured renewable energy technology, with large-scale development and good economic prospects. In order to lower the price of wind energy, efforts are still being made to build more dependable and economically viable WFs, even with the wind industry’s rapid expansion.

An American, Charles Brush, is considered to be the first person to use WT to generate electricity, which took place in the winter of 1887. At the same time, Professor James Blyth embarked on a similar experiment in July 1887 and finally obtained a British patent in 1891 [1]. Over the decades, wind power generation technology has undergone substantial advancements, leading to a consistent rise in the installed capacity. Figure 1 [2] illustrates the growth in global WT installations since 2010, highlighting the significant expansion of this renewable energy sector.

As wind farms continue to expand in scale, optimizing their design has become essential to enhance reliability and maximize profitability. However, the inherent randomness and intermittency of wind energy pose significant challenges to its widespread adoption. To address these issues, AI has emerged as a powerful tool, leveraging big data analysis, machine learning, and intelligent control to improve wind farm efficiency and reliability. Under the background of accelerating global carbon neutrality efforts, the wind power industry is paying more attention to efficiency optimization, making AI-driven solutions increasingly crucial.

This review’s main contribution lies in constructing a systematic framework for AI-based wind farm optimization, encompassing key aspects such as wind turbine layout, wind turbine component optimization, and wind farm collection system optimization. Additionally, it provides an in-depth analysis of the evolution of AI technologies, evaluating the applicability of various methods, including machine learning and evolutionary algorithms. Compared to existing reviews, this review’s main new contribution lies in presenting a comprehensive technical roadmap and exploring hybrid intelligent optimization methods, such as the combination of machine learning and evolutionary algorithms.

Despite AI’s promising potential in wind power optimization, challenges remain, including data acquisition, model interpretability, and computational costs. Future research should focus on integrating data-driven and physics-based models, advancing distributed intelligent optimization, and deepening AI’s application in intelligent operation and maintenance. These efforts will further promote the advancement of wind power technology and contribute to the efficient utilization of global renewable energy.

2. Basic Consideration of Wind Power System Design Optimization and Overview of Artificial Intelligence Applications

This paper systematically reviews 60 relevant studies published since 1993, focusing on strategies to reduce the Levelized Cost of Energy (LCOE) of wind energy. The review examines three key areas: wind farm layout optimization (WFLO), wind turbine component optimization (WTCO), and wind farm electrical collector system optimization (WFECSO). The reviewed literature is sourced primarily from ScienceDirect, IEEE Xplore, and SpringerLink databases. The selection process involves keyword-based searches, followed by classification according to the aforementioned three key areas. A flowchart summarizing the article selection process is presented in Figure 2.

Since 2007, with the rapid advancement of artificial intelligence (AI), research on AI-driven wind farm optimization has grown significantly. There have been many scholars who have made contributions in this area [3,4,5,6,7]. Reference [3] considers the application of artificial intelligence to wind power generation in various aspects, including the optimal design of wind farm controllers. It is shown that artificial neural network-based controllers are superior to traditional PID controllers in wind farm control. Reference [4] mainly analyzes the application of artificial intelligence algorithms in the design and optimization of offshore wind farm towers, which describes the application of sensors and digital twin technology in wind power operation and maintenance. Reference [5] analyzes the application of AI algorithms in the whole life of offshore wind turbines, including site selection, structural optimization, operation and maintenance, etc., as well as describing the advantages and disadvantages of AI in these aspects. Reference [6] discusses the application of artificial intelligence in the power collection system of offshore wind farms. The article points out the problems with existing algorithms in solving problems and the advantages of multimodal algorithms that have emerged in recent research in identifying multiple optimal solutions in different dimensions, providing assistance for the future application of artificial intelligence in WFECSO. Reference [7] searched several scientific databases to analyze the application of AI in the design, operation, and maintenance of wind farms, including optimization of blade aerodynamics, tower structure and foundation design, siting, environmental impact assessment, etc. The paper describes the use of AI to identify birds near wind turbines and to make advance preventive measures by combining convolutional neural network and computer vision techniques.

Compared with the existing research, this paper mainly elaborates on the depth of AI algorithms in wind farm optimization design from three aspects, which not only includes a single algorithm but also discusses the innovation of fusion algorithms; the specific comparison is shown in

Table 1. A comparison table of the advantages of this article over existing literature.

References	Scope of the Study	Technical Feature	Data Support
This review paper	Three aspects: WFLO, WTCO and WFECSO	In-depth analysis of single and innovative algorithms, comparing the strengths and weaknesses of various AI algorithms	Quantitative analysis based on 69 papers, including multi-dimensional statistics such as time trend (1993–2022), country distribution, sea–land differences, etc.
Reference [3]	Focus on wind farm controller optimization	Traditional PID vs. ANN Controller	No AI application in the optimal design of wind farms
Reference [4]	Focus on offshore wind turbine tower design	Discusses the application of a few artificial intelligence methods to the optimal design of wind power generation	No other aspects than tower design
Reference [5]	Analyzing full life-cycle applications	Isolated applications of technology	No AI application in the optimal design of wind farms
Reference [6]	WFECSO	The algorithms are divided into two categories: deterministic algorithms and heuristic algorithms	In terms of WFECSO, 18 papers were discussed.
Reference [7]	Covering many aspects of site selection, blade design, etc., but lack of system integration	Multi-Technology Simple Stacking	No AI application in the optimal design of wind farms

.

This study analyzes the publication trends by country, revealing that Denmark, China, Spain, and the United States have made the most substantial contributions in this field. The paper summarizes the core findings of each study, evaluates the impact of various optimization methods on wind farm economics, and discusses research trends based on chronological and geographical distributions. The analysis highlights AI as a dominant tool in wind farm optimization, with future research expected to emphasize multi-objective optimization, intelligent algorithms, and comprehensive cost models.

By examining existing research and identifying key areas for future exploration, this review provides valuable insights into the ongoing development of wind farm optimization, aiding researchers and industry professionals in advancing cost-effective and efficient wind energy solutions.

2.1. General Considerations of Wind Power Design Optimization

In wind power generation, the Levelized Cost of Energy (LCOE) is a key economic performance indicator. It represents the cost per unit of electricity generated over the lifetime of a wind farm, accounting for both the total costs and the total energy output in present value terms. In other words, LCOE reflects the average cost of producing one unit of electricity, considering capital investment, operation and maintenance, and other associated expenses throughout the wind project’s operational life. The expression for calculating LCOE is given by Equation (1):

$$\mathrm{L}\mathrm{C}\mathrm{O}\mathrm{E}={\displaystyle \frac{{\sum }_{\mathrm{t}=1}^{\mathrm{n}}\frac{{\mathrm{C}}_{\mathrm{t}-\mathrm{o}\mathrm{u}\mathrm{t}}}{{\left(1+\mathrm{r}\right)}^{\mathrm{t}}}}{{\sum }_{\mathrm{t}=1}^{\mathrm{n}}\frac{{\mathrm{A}\mathrm{E}\mathrm{P}}_{\mathrm{t}}}{{\left(1+\mathrm{r}\right)}^{\mathrm{t}}}}}$$

(1)

In Formula (1), C_t-out represents the total cash outflow in year t (including capital expenditures, operation and maintenance expenses, etc.) and AEP_t represents annual electricity production in year t. Our discount rate is r. The project lifetime, measured in years, is n. AEP is a significant index for evaluating the economy of a wind power project and is calculated by dividing the present value of all WF costs by the present value of all wind energy output. It is calculated as

$$\mathrm{A}\mathrm{E}\mathrm{P}={\sum }_{\mathrm{i}=1}^{\mathrm{N}}\left(\mathrm{P}\left({\mathrm{v}}_{\mathrm{i}}\right)\times \mathrm{T}\left({\mathrm{v}}_{\mathrm{i}}\right)\right)$$

(2)

where $\mathrm{P}\left({\mathrm{v}}_{\mathrm{i}}\right)$ represents the power generation (in kilowatts, kW) at a specific wind speed ${\mathrm{v}}_{\mathrm{i}}$. $\mathrm{T}\left({\mathrm{v}}_{\mathrm{i}}\right)$ represents the time proportion of wind speed at ${\mathrm{v}}_{\mathrm{i}}$. in one year. N is the number of wind speed intervals.

The optimal design of the wind farm aims to effectively reduce the cost required for investment and operation and maximize AEP, thereby reducing the LCOE of the entire wind farm. In recent years, LCOE has been greatly reduced [8], as shown in

Table 2. Changes in LCOE from 2010 to 2021 [3].

Changes in LCOE	2010	2021	Change in %
Onshore wind	0.102	0.033	−68%
Offshore wind	0.188	0.075	−60%

.

When measuring the economic benefits of wind turbines, the cost of energy (COE) is often used as an evaluation index. The concept of COE is similar to that of LCOE, but LCOE provides a more comprehensive perspective, considering all costs and outputs throughout the entire lifecycle of a generation project, suitable for evaluating and comparing the economics of long-term projects. COE refers to the direct cost of generating electricity, usually focusing on the cost during the operational phase, and may not include all capital costs, financing costs, or other fixed costs. COE is calculated by dividing the present value of all wind farm costs by the present value of the total wind energy output, as expressed in Equation (3):

$$\mathrm{COE}={\displaystyle \frac{\mathrm{Operation} \mathrm{and} \mathrm{Maintenance} \mathrm{Costs}}{\mathrm{Electricity} \mathrm{Generation}}}$$

(3)

The Net Present Value (NPV) is a fundamental financial metric used to evaluate the profitability of investment projects by calculating the difference between the present value of expected cash inflows and the present value of cash outflows over the project’s operational period. It incorporates the time value of money, ensuring that future financial returns and costs are appropriately discounted to their present value. A positive NPV indicates that a project is expected to generate financial gains beyond its investment costs, while a negative NPV suggests the opposite. Although the Levelized Cost of Energy (LCOE), discussed in previous sections, is also based on discounted cash flows, it focuses specifically on the average cost per unit of generated electricity over the project’s life, serving as a cost-efficiency indicator rather than a direct profitability measure. In contrast, NPV provides a direct assessment of the project’s overall financial return. The NPV is expressed in Equation (4):

$$\mathrm{NPV}={{\displaystyle \sum }}_{\mathrm{t}=1}^{\mathrm{T}}{\displaystyle \frac{{\mathrm{C}}_{\mathrm{t}-\mathrm{in}}}{{\left(1+\mathrm{r}\right)}^{\mathrm{t}}}}-\mathrm{I}.$$

(4)

where r is the discount rate, ${\mathrm{C}}_{\mathrm{t}-\mathrm{in}}$ represents the net cash inflow in year t, I is the initial investment, and T is the project lifetime in years.

2.2. Overview of Artificial Intelligence Applications in Design Optimization of Wind Power System

In this paper, a total of 69 relevant publications in the literature are reviewed. These papers mainly explore how to reduce the LCOE of wind farms from three aspects: WFLO, WTCO and WFECSO. Figure 3 illustrates the distribution of the reviewed articles across these three aspects. It can be observed that the majority of recent research focuses mainly on WFLO.

The literature discussed in this paper mainly comes from Science Direct, IEEE Xplore and Springer Link databases from 1993 to the present. After 2007, there are more and more studies on the use of artificial intelligence to optimize the design of wind farms. The number of articles published each year is shown in Figure 4, which also shows the number of documents published in different countries each year. In all articles, there are 59 journal articles and 10 conference publications, respectively. This analysis reveals that China (16 articles), Denmark (13 articles), the United States (10 articles) and Spain (7 articles) are leading in terms of research publications in this field. The distribution of articles from different countries is illustrated in Figure 5, where the color represents publications in three optimization aspects: WFLO, WTCO, and WFECSO, respectively. In the reviewed literature, more research related to WFLO is focused on onshore wind farms, while more research related to WFECSO is centered on offshore wind farms, as depicted in Figure 6. WTCO-related research does not significantly distinguish between onshore and offshore wind farms, as also shown in Figure 6.

3. Application of Artificial Intelligence Methods in the Optimization Design of Wind Power Systems

Various AI methods have been investigated for design optimization of wind power systems. Those proposed AI methods may be summarized as shown in Figure 7. All the following algorithms are used in the summarized articles and do not represent all AI methods. This paper divides the following methods into three categories, machine learning, heuristic algorithms and other algorithms, among which neural networks usually belong to the category of supervised learning in the optimization design of wind farms. Section 3.1, Section 3.2 and Section 3.3 will discuss the AI algorithms in WFLO, WTCO and WFECSO, respectively.

3.1. Application of Artificial Intelligence in Wind Farm Layout Optimization

WFLO primarily investigates WT layout optimization in WFs. Its primary focus is on WFs’ wake impact, with the goal of achieving maximum AEP and minimum PC. Studies [9] indicate that the wake of the upstream WTs will have a major effect on the performance of the downstream WTs, resulting in a 10–20% drop in the power generation of the downstream WTs. At present, models of wake loss include the Jensen model, Larsen model, Frandsen model, Gaussian model and so on. In early work, the empirical estimation method was often used to estimate the influence of wake decay characteristics on the efficiency of WFs, but this method cannot provide any physical insights into the flow process [10]. Subsequently, numerical calculation methods were employed to compute the wake model. Differential equations were utilized to calculate parameters such as wake width and turbulent viscosity, allowing for the simulation of the wake velocity field. However, the calculation process is complicated, and the considered scenario is often simple. It is very difficult to calculate the wake model under complex and changeable wind conditions [10]. Numerous AI-based techniques have been put out to maximize power output and optimize wind power plant designs by lowering wake losses. These algorithms include genetic algorithm (GA), particle swarm optimization (PSO), simulated annealing arithmetic (SAA), coral reef optimization (CRO), etc. In these algorithms, GA has the characteristics of randomness, simple search process, and easy combination with other algorithms. It is better to deal with discontinuously changing variables. PSO is more efficient than GA and is better at dealing with continuously changing variables. It offers a more practical means of WFLO. There are also some derivative algorithms, such as multi-population genetic algorithm (MPGA), adaptive particle swarm optimization (APSO), etc.

Because there are many GA and PSO-based algorithms in studies on WFLO, the WFLO aspect is divided into three parts, GA, PSO and other algorithms, as follows.

3.1.1. Application of Genetic Algorithm in Wind Farm Layout Optimization

GA was first proposed by American John Holland in the 1970s [11]. GA is an algorithm that mimics the natural evolution process to obtain the best answer. Its search process is simple, random and easily combined with other algorithms. It easily deals with discontinuous variables.

References [12,13,14,15,16] considered the optimization effect under different wind conditions. By linking the wake superposition-based WF simulation model with the genetic search code, reference [12] optimized it. By examining the outcomes of a few straightforward applications, the viability of this approach is demonstrated. Three wind conditions (single direction, steady wind speed and changing direction, and changing wind speed and changing direction) are used to optimize the number and placement of WTs. Energy optimization and installation cost minimization are the objectives. Reference [13] used GA to determine the optimal WT design with the maximum output capacity. Three experimental cases with different wind conditions were used to demonstrate the effectiveness of GA in WFLO. The three cases studied in reference [14] and reference [13] are the same, and GA is also used, but reference [14] proposed a new coding method; under the wind conditions of changing directions and fluctuating speeds of 8, 12 and 17 m/s, the optimization result is the best. After optimization, there are 28 WTs and 32,261 kW of power is produced annually. At this time, the efficiency is 91%. The efficiency has increased by 4.38%, the total power has grown by 223 kW, and there are 11 fewer WTs than in reference [13]. The Weibull function in [15] describes the likelihood of the wind speed distribution, and the WT speed–power curve is used to estimate the power generation of the WT. The optimization problem is described utilizing the increased wind and models of WTs in order to minimize the energy cost of WF. Binary coded GA is used to create the optimum model framework. The best WF layout was found using a local search (LS) approach based on the binary real-coded genetic algorithm (BRCGA) in reference [16]. Accompanying the suitable wake interaction modeling is the robust single wake model collection. Each WT’s position is represented via a binary matrix. The power produced by every WT is represented by the matrix element. Better setup and increased power productivity can be attained with the suggested technique.

References [17,18,19] used NPV as the optimization objective function; NPV is a metric used to assess an investment’s profitability. Reference [17] used GA to optimize the NPV of WF investment. The economic conditions such as initial investment, power generation value and investment period are considered. These variables are influenced by the kind, quantity, height, and arrangement of WTs inside WFs. Reference [18] optimized WF setup using GA. The full WF’s expense model was used in the optimization process to determine the best beginning investment and the annual NPV during the lifetime of WFs. The suggested algorithm computes the yearly revenue from the sales of net power generation while accounting for the wake decay effect’s production loss on a single WT. In the face of the current large-scale computation problem, GA can no longer be solved well, and it easily falls into ‘precocity’. Some improved GA derivative algorithms have emerged. Reference [19] uses sequential optimization of IGA to tackle the optimal location problem of WTs in huge-scale OWFs by dividing the available marine plots into smaller suitable-size areas. The purpose is to maximize the economic benefits of the project. The optimization results increased the NPV by 1.02%.

The wake effect’s influence on electricity production increases as WFs become bigger. The wake loss and power output will be impacted by the WT configuration. The reduction in wake loss can be significantly impacted by GA-based layout improvement. Novel GA-based coding was presented in reference [20], which made use of six years’ worth of wind speed and direction data collected at three-hour intervals in a specific region of Iran. This research examines three optimization scenarios: varying population and generation sizes, varying effective distances, and varying longitudinal (x) and latitudinal (y) distances. The findings demonstrate that increasing x between WTs reduces the wake effect on WT and increases efficiency. In this instance, x = 372.8 and y = 186.4 yielded better optimization outcomes than x = 186.4 and y = 372.8. In the former, there are 56 WTs, which is 8 more than in the latter. The latter was 1.15% lower than the effective rate of 89.54%. References [21,22] both address not only the layout of wind farms but also the topology of cable routing. In both studies, genetic algorithms (GA) are used to optimize wind turbine placement. The key difference lies in the approach to cable connection: the former employs the GeoSteiner algorithm to determine the shortest cable routing scheme and focuses on minimizing costs, while the latter uses the ant colony optimization (ACO) algorithm to identify the connection topology, with an emphasis on improving grid efficiency.

The Jensen wake model was employed in reference [23] to determine the loss resulting from every WT wake. The findings demonstrate that an optimized architecture can increase WF’s power output even in the case of unchanged WT counts and WF acreage. When calculating the velocity loss, however, the Jensen wake model is not entirely compatible with field measurement and computational fluid dynamics simulation results. As a result, reference [24] employed GA along with the WFLO approach of the Gaussian wake model to lower WF’s annual energy cost. In every scenario that was examined, this approach produced greater annual power generation and reduced calculation times.

Applying GA’s binary string to the infinite layout method will increase the calculation time. To this end, reference [25] proposed the multi-objective genetic algorithm (MOGA) and used new boundary constraints to increase the area of WFs so as to obtain the initial layout in MOGA. Then, WTs outside the original area are moved to the nearest edge to increase the possibility of WT edge positioning. Considering a variety of factors and the historical wind conditions in a certain area, the efficiency of WFs is increased, and the unit power generation cost is reduced. However, the increase in the length and width of WFs may affect the final result. Therefore, this problem needs further study.

MPGA for resolving the optimal configuration problem of WTs is proposed in [26], aiming to address the issue that GA is prone to local extrema while handling complex optimization issues. Consider the following different wind conditions (a): the wind speed is 12 m/s, and the wind direction is constant; (b): the wind speed is 12 m/s, and the wind direction is changing; (c): the wind speeds are 8 m/s, 12 m/s, 17 m/s, and the wind direction is changing. In terms of both solution speed and quality, the MPGA outperforms the single-population GA method when compared to the regular GA. It can find the best solution more quickly and for the least amount of money.

In reference [27], a new bi-criteria identification and relocation mechanism was added to GA and a variety of IGAs. The computational complexity is not increased by the addition of this mechanism. Numerous experiments confirm the efficacy of this new technique, which significantly lowers the wake loss. Reference [25] used a new method combining fixed-point selection (DPS) and GA, which better reduces the unit power cost and the minimum wake effect and gives a reasonable turbine spacing.

This paper conducted a statistical analysis of the articles reviewed in Section 3.1.1 and compiled the relevant authors and optimization characteristics of the articles in

Table 3. Application of GA in WFLO.

Authors and Literature	Optimization Characteristics
	Wake Model	Data Sources	Optimization Objectives
G. Mosetti et al. [12]	—	—	COE
S.A. Grady et al. [13]	Jensen	—	The cost per unit of energy
Alireza Emami et al. [14]	Jensen	—	The cost per unit of energy
Chunqiu Wan et al. [15]	Jensen	—	The cost per unit of energy
Ali M. Abdelsalam et al. [16]	Gaussian	—	The cost per unit of energy
Jose Calero Baron et al. [17]	—	—	NPV
Javier Serrano Gonzalez et al. [18]	—	—	NPV
Javier Serrano Gonzalez et al. [19]	—	—	NPV
Majid Khanali et al. [20]	Jensen	Real wind data in Kahrizak, Tehran, Iran	The cost per unit of energy
Yan Wu et al. [21]	Jensen	The air volume data come from an enterprise and are collected once an hour in Hailar, China, for one year.	AEB
Wu, Y.K et al. [22]	—	—	AEP
Rabia Shakoor et al. [23]	Jensen	—	The capitalized cost
Leandro Parada et al. [24]	Gaussian	—	COE
Ying Chen et al. [25]	Jensen	Monthly actual wind data in Corpus Christi, Texas	The cost per unit of energy
Xiaoxia Gao et al. [26]	Jensen	Wind data of real offshore wind farms in the southeastern waters of Hong Kong for nearly 20 years (1992~2011)	LCOE
Feng Liu et al. [27]	Gaussian	Actual data of two wind farms	COE

.

In summary, the performance of a single genetic algorithm in simple wind conditions and small-scale WFLO is relatively reliable, but it may be easy to fall into local optima in complex scenarios (such as multi-speed wind conditions, large-scale wind farms), and needs to be combined with improved strategies (such as MPGA, IGA) by hybrid methods. The improved genetic algorithm has more advantages in multi-objective optimization, avoiding prematurity and dealing with large-scale optimization problems, especially in the case of changing wind conditions. In the future, combining GA with machine learning in the face of more complex scenarios may further improve the optimization effect.

3.1.2. Application of Particle Swarm Optimization in Wind Farm Layout Optimization

PSO is another random search method that mimics the cooperative group dynamics of birds during foraging. The broad consensus is that it belongs to the class of swarm intelligence algorithms, and that the multi-agent optimization system may use it. PSO was developed by Drs. Eberhart and Kennedy [28] and is more effective than GA at handling variables that are constantly changing. It provides an effective solution in optimizing the layout of WFs. Reference [29] compared PSO with classical binary coded GA, and the PSO is based on penalty function to solve the problem. The objective of optimization is to optimize WFs’ power while maintaining an acceptable separation between each WT and a sufficient amount of WTs. It can be shown from the results that PSO outperforms GA.

Reference [30] used the gbest model of PSO to tackle the problem of minimizing the unit price of wind power generation. The overall amount of power generated and the number of WTs in WF determine this cost. The WF’s installed turbine count, each WT’s power output, and the influence of wind speed on WT power are all taken into account. The PSO approach was utilized by reference [31] to optimize the WT arrangement in WFs. To optimize the amount of energy produced, the positions of each WT within a designated region can be adjusted. Based on the results of the simulation, this study increases the total cost of WTs while also increasing the efficiency of WFs. Reference [32] looked into how PSO might be used for WFLO. Three different wind conditions are taken into account: changing wind direction and speed, steady wind direction and speed, and changing wind condition and speed. The resource is described in all three scenarios using 36 distinct wind directions, each of which is utilized in the evaluation function’s wake modeling and AEP computation. The PSO employed for these three cases resulted in layouts with improved LCOE as compared to previous research using GA.

Reference [33] proposed a mathematical model including wind direction change and wake loss. PSO is used to maximize energy output while minimizing total investment costs and to ultimately effectively reduce LPC. The optimization program is suitable for the optimal arrangement of WTs in WFs and can be extended to various wind conditions and farm capacities.

A number of mathematical formulas for hybrid energy cost are presented in reference [9], taking into account the geometric properties of WTs and the Weibull parameters of wind power. An optimization technique composed of IPSO and an iterative approach is constructed as a composite. While the latter is iteratively optimized to determine the proper position of WTs, the former is utilized to minimize the energy investment associated with the distinctive parameters of a single WT.

This paper conducted a statistical analysis of the articles reviewed in Section 3.1.2 and compiled the relevant authors and optimization characteristics of the articles in

Table 4. Application of PSO in WFLO.

Authors and Literature	Optimization Characteristics
	Wake Model	Data Sources	Optimization Objectives
Chunqiu Wan et al. [29]	Jensen	—	The output power
Vlachos Aristidis et al. [30]	—	—	LCOE
Rasoul Rahmani et al. [31]	Jensen	—	LCOE
Ajit C. Pillai et al. [32]	Larsen	—	LCOE
Peng Hou et al. [33]	Jensen	The reference wind farm is located near FINO3, 80 km west of the Seat Island, Germany	LPC
Longfu Luo et al. [9]	Jensen	Newport nearshore wind park wind site in the USA, Xiangshui intertidal Pilot project offshore wind farm in China and Rønland offshore wind farm in Denmark	COE

.

In summary, PSO is more suitable for complex wind conditions and multi-objective scenarios in WFLO, especially in cost control and dynamic response. Compared with GA, PSO has faster convergence speed and can be dynamically adjusted by inertia weight to avoid premature convergence, and PSO performs more efficiently in optimization cost than GA.

3.1.3. Other Algorithms in Wind Farm Layout Optimization

Similar to GA and PSO, the ant colony optimization algorithm (ACO) is also a type of swarm intelligence optimization algorithm. Reference [34] used ACO and compared it with the evolutionary strategy algorithm in the literature. The results show that the performance of ACO is better than the existing strategy, which maximizes the energy obtained by WFs and makes WFs more suitable in terms of layout.

In [35], the layout optimization problem of multi-hub-height WTs is studied to minimize the power output cost using the three-dimensional greedy approach. Compared with GA, the optimization results using the greedy algorithm have lower computational cost. The multi-hub-height WT layout, particularly for WFs with difficult terrain, can greatly improve the overall output power and minimize the unit output power cost when compared to the WT layout with the same hub height.

The lightning search algorithm (LSA) is an HA based on the mechanism of lightning. Reference [36] proposed a multi-objective LSA, which better optimizes the layout of WFs and reduces the annual energy cost, wake loss and the area of the entire WF.

CRO is a meta-HA inspired by the reproduction and survival process of corals. The distribution of OWF was designed using CRO by reference [37], with the ultimate optimization aim being to maximize AEP. The performance of the CRO algorithm outperforms that of the evolutionary, DEA, and HA, according to the results. The drawback of CRO is that its optimization accuracy is poor. As a result, reference [38] proposes an improved the coral reef optimization algorithm with substrate layer (CRO-SL) evolutionary meta-HA and the coral reef optimization algorithm to the WT layout problem. By introducing various geometric and algebraic processes, the algorithm speeds up the computation of the wake effect in the objective function. Considering the finite state of the wind rise, the convex cone description of the wake, and the sparse matrix processing of the problem, the execution time of the objective function is notably decreased. This article combines simulation and real data to discuss the performance of CRO-SL in different WT design problems. The optimal layout was discovered by contrasting the results with the alternative approach to the problem, taking into consideration all pertinent variables.

SAA uses the principle of temperature drop, combined with certain probability jump characteristics, to find the global optimal solution of the objective function in a larger search space. SAA was employed in [39] to ascertain the best WT arrangement for a sizable OWF. Under both predictable and more complicated scenarios, the algorithm converges to the ideal configuration, increasing WF’s AEP by 1%.

In order to solve the layout optimization problem of irregularly arranged OWFs, reference [40] analyzed the efficacy of GA and PSO. The comprehensive WF model evaluates the leveling cost of energy and incorporates accurate models of wind behavior, wake, and cost. It should be mentioned that these two optimization techniques do not perform as well as first anticipated when it comes to locating the global optimal solution. It is advised to employ PSA to improve each solution produced by PSO or GA in order to achieve the greatest outcomes.

Reference [41] proposed an extended pattern search (EPS)–multi-agent system (MAS) optimization method. The optimization algorithm uses the profit target, and the overall search determines the position of each WT. At the same time, the optimal hub height and rotor diameter of each WT can be determined. Considering two wind conditions, (1) constant unidirectional wind and (2) three wind speeds and changing wind directions, the EPS algorithm is very suitable for complex layout problems, especially WF layout optimization problems, and its performance is better than that of similar GAs.

Reference [42] proposed a hybrid method of balance optimization and pattern search (EO-PS), which optimizes the layout of WFs, improves efficiency and reduces installation cost under different wind conditions and multiple optimization objectives.

In recent years, the optimization methods of wind farm layout based on machine learning and intelligent optimization algorithm have made remarkable progress. The following three articles put forward innovative solutions from different dimensions. In reference [43], aiming at the balance between power grid security and economy, an ANN is combined with optimal power flow calculation with security constraints. Under the premise of considering line faults and multiple uncertainties, a hybrid optimization framework is constructed using the interior point method and Monte Carlo simulation, which realizes the rapid location decision of minimizing operating cost and system stability. Reference [44] establishes a high-precision wake prediction model through machine learning and cooperates with GA to optimize the spatial configuration of wind turbines, which effectively reduces the aerodynamic interference between units and improves the wind energy capture efficiency of wind farms. Reference [45] proposes a data-driven framework of multi-algorithm fusion, which uses a support vector machine to realize the spatio-temporal prediction of wind energy resources, combines PSO to optimize the global layout, and introduces decision tree to realize the dynamic scheme optimization under multiple environmental parameters, forming an intelligent system covering the whole chain of prediction, optimization and decision. These studies promote the deep collaboration between machine learning and traditional optimization algorithms and provide a new technical route for the layout optimization design of wind farms.

This paper conducted a statistical analysis of the articles reviewed in Section 3.1.3 and compiled the relevant authors and optimization characteristics of the articles in

Table 5. Other algorithms in WFLO.

Authors and Literature	Optimization Characteristics
	Wake Model	Optimization Algorithm	Data Sources	Optimization Objective
Yunus Eroglu [34]	Jensen	ACO	—	The output power
K. Chen et al. [35]	Jensen	the three-dimensional greedy algorithm	—	COE
Sinvaldo Rodrigues Moreno et al. [36]	Jensen	MO-LSA	—	COE
S. Salcedo-Sanz et al. [37]	—	CRO	Real offshore wind farm data for the Baltic Sea	AEP
J. Perez-Aracil, D et al. [38]	Jensen	CRO-SL	Simulated data and real data of a wind farm in Spain	The output power
Rajai Aghabi Rivas et al. [39]	Jensen	SAA	Horns Rev offshore wind farm.	AEP
Javier Serrano Gonzalez et al. [40]	—	PSA	Horns Rev 3 offshore wind farm	LCOE
Bryony DuPont et al. [41]	Jensen	EPS-MAS	—	COE
Rizk M et al. [42]	Jensen	EO-PS	The reality of Egypt’s Suez Bay–the Red Sea	COE
Shahzad, U [43]	—	ANN	—	LPC
Yang, K et al. [44]	—	SVM, GA	—	AEP
Kun Yang et al. [45]	—	SVM, PSO, DR	—	AEP, LPC

.

In the field of WFLO, traditional optimization algorithms exhibit distinct characteristics but are inherently limited. GA excels in multi-objective collaborative optimization, yet it is prone to local optima and high computational costs. PSO demonstrates rapid convergence but relies heavily on parameter tuning and lacks dynamic adaptability. Hybrid algorithms (e.g., MOGA, IGA) enhance global search capabilities through multi-strategy integration, but their implementation complexity remains a significant challenge. While SAA effectively avoids local optima, its practical application is constrained by slow convergence rates. These limitations indicate that single traditional algorithms struggle to comprehensively meet the requirements of high precision, efficiency, and robustness in WFLO. In the future, the deep integration of machine learning and traditional optimization algorithms will drive the evolution of wind farm layout design from a “static experience-driven paradigm” to a “dynamic data-physics co-driven paradigm”, enabling more adaptive and intelligent solutions for complex real-world scenarios.

3.2. Application of Artificial Intelligence in Wind Turbine Component Optimization

AI-based WTCO includes studies on the geometric structure design of WTs’ blade R and airfoil section, the design of hub height, and the design of the WT drive train. Firstly, the optimization of the geometric characteristics of the wind blade section enables the WT to obtain higher energy collection efficiency. Secondly, high-hub-height WTs are exposed to more wind energy. The wind turbulence factor at high altitude is small, and the wind fluctuation is small. However, when the hub height increases, the WT’s manufacturing costs rise as well, making building and shipment more challenging. Therefore, it is necessary to select the appropriate hub height under the premise of considering AEP and PC.

The goal of the WT drive-train optimization design is to reduce costs and boost generator efficiency by optimizing the internal generator, gearbox, and power converter system. PM wind generators are the most efficient and energy-producing type of typical WTs, and they do not require an external power source for magnetic field excitation. PM wind generators are more promising since they have cheaper costs and improved dependability due to the absence of mechanical components like slip rings. The generator-related articles reviewed in this paper all involve the optimization of PM wind generators. This will be broken down into two sections: drive-train optimization and external structure optimization of the WT.

3.2.1. The Optimization of the External Structure of Wind Turbines

Reference [46] used three distinct wind fields to apply nested GA. The output power of WFs employing WTs with varied hub heights is higher for the same total number of WTs when compared to WFs using WTs with the same hub height. Several cost models are considered in the investigation. The first is the simplified cost model defined in reference [11], while reference [47] establishes the comprehensive model. The findings indicate that raising the WT hub height can enhance WFs’ unit power.

The geometric properties of the MOGA blade section for HAWT were used in reference [46] to determine the Pareto optimal solution set. The performance of the airfoil cross section in the energy conversion process is evaluated by a two-dimensional incompressible unsteady computational fluid dynamics solver and a second law analysis. The least amount of energy lost, the greatest efficiency, and the best stability are ultimately attained. The data set that the solver obtained is utilized to train the ANN to represent the goal function in the procedure.

In reference [48], the blade element momentum simulation code was used to estimate the performance of VAWT, and MOGA was used to optimize WT structure. The optimization results in increased Cp by 6.5% compared with the baseline. The optimized configuration of exergy AEP increased the yield by 8.1% compared with the baseline. In [49], IGA was used to optimize the blade shape of the free-flow Savonius WT. The Cp is set as the objective function. Compared with the traditional semi-circular-blade Savonius turbine, the Cp of the improved Savonius turbine with GA-optimized blades has been significantly improved, with an increase of 33%.

Reference [50] combines the advantages of the GA and inverse design method and embeds the inverse design algorithm into GA to directly specify the required aerodynamic characteristics so as to determine the corresponding blade geometry, determine the optimal blade spacing and blade chord and twist distribution, and maximize the annual power generation.

A fully automatic method for optimizing the cross section of VAWT airfoils was introduced and shown in reference [51]. The goal is to increase torque while maintaining the solidity, tip speed ratio, and blade profile that are standard WT design restrictions. The parallel DEA is used in conjunction with the modular design and simulation methodology to create the best possible blade design that maximizes the WT’s efficiency.

PSO and the finite element method (FEM) are coupled in reference [52] to optimize the design of the HAWT blade structure with the intention of lowering blade weight. The overall performance of the blade can be enhanced by reducing its weight. Lighter blades can reduce the load of WT structure and reduce the cost.

Reference [53] introduced the heuristic search algorithm and the covariance matrix adaptive evolution strategy (CMAES) into the double multi-stream tube model (DMST). To satisfy the demands of the input wind speed range, a distinct fitness function modifies the geometric variables of the wind wheel. The performance of the φ-shaped Darrieus WT is optimized. Compared with the baseline, the Cp at the optimal wind speed is increased by 12.5% and reaches 0.3735.

Since deeper-water and higher-power WTs are the current trends in offshore wind power plant development, it is highly desirable to lower jacket costs. It is projected that the number of offshore wind energy converter jacket substructures will rise in the far future. The jacket substructure design for an NREL 5 MW turbine is optimized using a meta-heuristic PSO method in reference [54], which lowers the offshore WT cost but still requires improvement in the optimization time.

Reference [55] focuses on the aeroacoustic multi-objective trade-off problem, utilizing a deep neural network (DNN) to establish a nonlinear mapping relationship between blade geometric parameters and performance metrics (e.g., power coefficient and noise level). By employing inverse optimization to generate blade configurations with low noise emission and high power output, the study overcomes the limitations of traditional single-objective optimization approaches, achieving a breakthrough in balancing aerodynamic efficiency and acoustic performance.

A approach based on an ANN and EA was proposed in [56]. An ANN comprising rotor and stator cascades was trained with an established data to determine the geometrical parameters of individual WTs and groups of WTs based on various design assumptions. The goal is to maximize the WT’s overall efficiency.

In reference [57], the advanced aeroelastic simulator was used to design the bending shape of WT blades by using an ANN under turbulent inflow conditions. The author uses the vortex particle method, in which the WT blades are represented by the lift line theory, and the WT structural dynamic is modeled using a multi-body method based on finite elements. An NN and gradient-based optimizer can give better curved blades in complex scenarios. Compared with the straight blade design, the blade design given in the NN increases the pre-bending and back-swept parameters and produces about 1% more power on average, and the average thrust on the rotor is slightly increased by 0.02%.

Reference [58] optimized the airfoil design using GABP-ANN. The Bessel polynomial reduces the airfoil curve to eight pairs of coordinates. Next, using 1446 arrays as the training set and 50 data sets as the test set, the NN is trained to predict the lift coefficient and the maximum lift-to-drag ratio of the airfoil. This technique significantly reduces the optimization time and offers a novel approach to airfoil optimization.

In reference [59], a virtual cloning method based on an ANN was employed to develop an ANN-based virtual cloning model. The time required for this proposed model to yield results is significantly reduced, being only one-fifth that of the existing simulation method, which combines ANN with GA. This approach is aimed at determining and optimizing the performance of WTs, focusing on minimizing time, cost, and effort.

In [60], an ANN was linked with GA as a surrogate model after being trained with the data prior to optimization, and the results were compared with a computational fluid dynamics (CFD) simulation. The results show that the airfoil shapes obtained by the two methods are very similar, but the combination of GA and ANN takes less time. Reference [61] innovatively integrates architectural structure design with wind energy capture through a multidisciplinary CFD-ANN-GA coupling approach. By parameterizing shell pavilion designs, the study achieves multi-objective optimization of wind field acceleration effects and structural integrity. The ANN surrogate model significantly reduces computational costs associated with repeated CFD simulations, ultimately yielding a lightweight steel structure solution via genetic algorithms, which enhances wind speed to 14.73 m/s while maintaining a minimal thickness of 1.27 cm.

Reference [62] proposes a multi-objective deep reinforcement learning (MODRL)-based framework for blade geometric optimization. By enhancing the MO-DSPG algorithm with a restricted Boltzmann machine (RBM) to construct an adaptive stochastic agent, the study transforms the collaborative optimization of blade length, airfoil parameters, and twist angles into a stochastic policy search process. Under constraints of maximizing power output, minimizing structural loads, and reducing costs, the framework achieves a 27.48% improvement in computational efficiency while obtaining a blade design with superior comprehensive performance.

This paper conducted a statistical analysis of the articles reviewed in Section 3.2.1 and compiled the relevant authors and optimization characteristics of the articles in

Table 6. The optimization of the external structure of the wind turbine.

Authors and Literature	Optimization Characteristics
	Optimized Parts	Optimization Algorithm	Types of Wind Turbine	Optimization Objectives
Ying Chen et al. [46]	hub height	the nested GA	—	The output power
Seyed Mehdi Mortazavi et al. [47]	blade geometric characteristics	MOGA, ANN	HAWT	Cp
Gabriele Bedon et al. [48]	blade geometric characteristics	MOGA	VAWT	Cp, AEP
C.M. Chan et al. [49]	blade geometric characteristics	IGA	—	Cp
M.S. Selig et al. [50]	blade geometric characteristics	GA and inverse design method	HAWT	AEP
Travis J. Carrigan et al. [51]	blade geometric characteristics	DEA	VAWT	Average torque
Xin Cai et al. [52]	blade geometric characteristics	PSO-FEM	HAWT	Mass of the blade
Yaoran Chen et al. [53]	blade geometric characteristics	CMAES-DMST	φ-shaped Darrieus WT	Cp
Jan Hafele et al. [54]	jacket substructures	PSO	—	The total capital costs of jacket
Zavala J et al. [55]	blade geometric characteristics	DNN	—	The output power
Krzysztof Kosowski et al. [56]	blade geometric characteristics	EA and ANN	—	Efficiency
Matias Sessarego et al. [57]	blade geometric characteristics	ANN	VAWT	Cp
Hao Wen et al. [58]	blade geometric characteristics	GABP	HAWT	The lift and drag coefficients
Abdullah Al Noman [59]	blade geometric characteristics	ANN	Savonius WT	Cp
A.F.P. Ribeiro et al. [60]	blade geometric characteristics	GA-ANN	VAWT	The lift and drag coefficients
Yaxin Li et al. [61]	blade geometric characteristics	GA-ANN	HAWT	Cp
Zheng Wang et al. [62]	blade geometric characteristics	Reinforcement learning	—	The output power

.

In the optimal design of wind turbine components, genetic algorithms can explore the optimal solution from a global perspective, which is especially suitable for dealing with multi-objective problems; particle swarm algorithm is effective in reducing the weight of the blade and optimizing the structure, but it is easily restricted by the local optimal solutions; the differential evolution algorithm is good at coping with the complex design constraints, but it has high requirements for the adjustment of parameters. Neural network and deep learning techniques significantly shorten the design cycle by replacing the traditional simulation computation and, at the same time, realize the synergistic optimization of contradictory objectives such as aerodynamic performance and noise reduction effects. By improving blade shapes, adjusting hub heights, and optimizing material distribution, these methods have resulted in power generation efficiency gains of up to 33%, as well as significant structural cost reductions, and have driven quieter, more stable wind turbine designs. However, these techniques still face technical bottlenecks such as high consumption of computational resources, dependence on historical data accumulation, and so on.

3.2.2. The Optimization of the Drive Train of the Wind Turbines

IGA was utilized in reference [63] to maximize the design of seven wind power generation systems with variable speed and constant frequency. The optimization results of WT with various powers are compared, and finally the system cost is reduced and the AEP of the unit cost is improved. While PM generator-based WTs offer significant benefits in terms of efficiency and dependability, they also have drawbacks in terms of volume, expense, and installation complexity. The optimization design model of PM direct-drive wind power generation systems is established by IGA in reference [64], and the performance optimization of a 500 kW PM direct-drive generator is carried out with the minimum cost of the generator active material as the goal, which shows the effectiveness of the optimization design. In [65], the multibrid PM wind generator system was proposed, and the economic range of its gearbox transmission ratio and rated power was studied. The power-generating system’s architecture was optimized using the IGA to reduce costs. In addition, a comparison is made between the multibrid PM wind generator system and the PM direct-drive generator system. The findings indicate that the PM direct-drive generator system is not as cost-effective as the multibrid PM wind generator system concept.

While reference [66] employs a data-driven neural network model for the multi-objective optimization of wind turbines, targeting power maximization, drive-train vibration reduction, and tower vibration minimization, most traditional studies have relied on physical models or focused on single-objective optimization. In contrast, this paper advances beyond those limitations by adopting a non-parametric, nonlinear data-driven approach.

This paper conducted a statistical analysis of the articles reviewed in Section 3.2.2 and compiled the relevant authors and optimization characteristics of the articles in

Table 7. The optimization of the internal structure of the wind turbine.

Authors and Literature	Optimization Characteristics
	Types of Wind Generator	Optimization Objective
Hui Li et al. [63]	PMSG DD, PMSG 1G, PMSG 3G DFIG 3G, DFIG 1G, EESG DD, SCIG 3G	Minimize the cost of generator system
H. Li et al. [64]	PMSG	Minimize the cost of generator system
Hui Li et al. [65]	PMSG	Minimize the cost of generator system
Kusiak, A et al. [66]	—	Power maximization, drive-train vibration minimization, tower vibration minimization

.

To sum up, the traditional algorithm is good at global search, but it faces the bottleneck of low efficiency in high-dimensional solution space and rough dynamic multi-objective trade-off, and machine learning is easily constrained by data quality. In the future, it should focus on hybrid intelligent optimization architecture and combine the advantages of different algorithms to improve the wind turbine design.

3.3. Application of Artificial Intelligence in Wind Farm Electrical Collection System Optimization

WFECSO mainly studies the optimization of the number and location of WF substations, collection system layout design, the selection of electrical system voltage level, etc. Early research [67] focused mostly on the cost of the electrical components’ investments and power loss, and the structure of the collecting system was determined by comparison. Several mathematical formulas for the CCLO problem are presented in [68]. These formulas are computed to minimize the total of the expenses associated with infrastructure and energy loss. It should be mentioned that building costs for OWFs are far higher than those for onshore wind farms, and that transportation and civil construction will become considerably more challenging from land to water. According to statistics, the internal electrical connection system (or grid connection) expenses of onshore wind farms account for 8% of the total investment cost, while the proportion of OWFs is 18% [69]. Therefore, offshore WFECSO is more important.

The goal of WFECSO is mainly to reduce investment costs and maintenance costs. The electrical system’s structure and voltage level have a significant impact on the costs and power loss of WFs. The capacity of electrical devices like transformers, collectors, and cables, as well as the associated power loss, are determined by the voltage level. Reference [70] took into account the system’s cost, power loss, and voltage level during the optimization procedure. The GA-based WF optimization platform is unveiled. The primary parts and significant performance metrics of WF are used as input parameters to optimize the electrical system in terms of production cost and system dependability.

Reference [71] discussed three kinds of cable structures on the three-dimensional seabed. In the case of high cable failure rate and average repair time, the proposed multi-loop structure improves reliability and is the most economical. PSO is used to determine the distribution of substations to minimize the total length of cables. Reference [72] used APSO to optimize the layout and cable distribution of OS. Multiple optimization objectives are proposed, and various wind conditions are considered to provide an accurate power loss model. Compared with APSO-SST-SPLCM, the optimization result of APSO-SST-WDPLCM reduces the capital by 4.38% and the PLC increases by 3.95%.

The impact of OS’s position on the submarine CCL should be considered during the layout optimization design phase in order to optimize cable expense, given the significant correlation that exists between the layout of the electrical system and the OS location. Reference [73] suggested using HGIA as the foundation of an optimization strategy to find the best answer. The model and the approach to solving it are then put to the test in a real OWF close to Shanghai, China. The outcomes demonstrate the suitability and efficacy of the optimization strategy. After locating the substation using a clustering approach, reference [74] optimized the cable path using a navigation grid routing strategy based on Delaunay triangulation. Ultimately, OWF’s substation and inter-array electrical collecting network are efficiently set up, minimizing costs. Reference [75] suggested a platform for OWFES optimization. It primarily takes into account three factors: the quantity and placement of OS, the transmission and collecting systems for cables, and the choice of electrical parts, including voltage level and capacity. The cable arrangement and substation connection method are optimized using APSO-MST. In the end, the cost is 3.01-percent lower than the industrial layout.

In reference [76], IGA is used to optimize the WT collection system of OWF, which reduces the investment amount of WT access to OS. In [77], a clustering-based cable placement algorithm for large-scale WFs is proposed, which improves the reliability of the system and reduces the power loss.

There are many MST algorithms, the most classic of which are Kruskal algorithm and Prim algorithm. The Prim algorithm was utilized by reference [78] to optimize the design of OWF’s electricity-gathering system. The current carrying capacity of the cable was taken into account throughout the cable selection procedure during the optimization in order to meet the needs of the system operation. This technique is used on both regular and irregular WFs to reduce the overall electrical system investment.

Reference [79] proposed a method to optimize the location of cable connections and substations using MST. This method greatly reduces the cost of the cable. Reference [80] introduced the DMST approach to optimize CCL for large-scale OWFs, minimizing the LPC of OWFs. By applying the DMST method to two WFs as examples, cable investment savings of 1.07% and 6.10% were achieved, respectively.

WFECSO is somewhat constrained by WFLO. In order to minimize wake loss, the former often maximizes turbine spacing; nevertheless, this raises the expense of the latter. For the joint optimization of WFLO and WFECSO, reference [81] created a heuristic approach based on the variable neighborhood search method. To balance the two problems, the combination objective function is NPV. This method has been adjusted, and the joint method for combined problems reduces the pre-investment cost by 10%, which is a significant improvement.

This paper conducted a statistical analysis of the articles reviewed in Section 3.3 and compiled the relevant authors and optimization characteristics of the articles in

Table 8. Algorithm for optimal design of electrical collection system in wind power plants.

Authors and Literature	Optimization Characteristics
	Optimized Parts	Optimization Objective
M. Zhao et al. [70]	Voltage level and capacity, topology	COE
Xuan Gong et al. [71]	Substation location and cable length	Cable costs and cable installation costs
Rongsen Jin et al. [72]	Substation location, CCL, cable selection	Cost, PLC
Dong Dong Li et al. [73]	Number and capacity of substations, cable topology	Cost
Pillai AC et al. [74]	CCL	Cost
Peng Hou et al. [75]	Number and location of substations, CCL, voltage level and capacity	Cost
Francisco M et al. [76]	Number, capacity and location of substations, WT data sets and topology of integrated systems	Cost
S. Dutta et al. [77]	CCL	Cost
Peng Hou et al. [78]	Substation location and CCL	Cost
Junxian Li et al. [79]	Substation location and cable length	Cost
Peng Hou et al. [80]	Substation location and CCL	Cable cost
Davide Cazzaro et al. [81]	CCL	NPV

.

In the optimization of wind farm electrical system, GA and its improved version show the advantages of economy and reliability in multi-objective optimization (such as component configuration, voltage adjustment), but there are problems of complex calculation and easily falling into local optima. PSO is efficient in reducing power loss and cable cost by virtue of its global search ability, but it is sensitive to parameters and has weak adaptability to nonlinear problems. The clustering method can optimize the cable layout and substation location, but additional computing resources are needed in complex terrain. The MST algorithm can ensure the economy and feasibility of cable arrangement, but the efficiency is limited in large-scale application. On the whole, these algorithms reduce costs and improve reliability by optimizing design. However, in the face of large-scale complex systems, they are still limited by computational efficiency, parameter dependence and convergence speed. It is necessary to further develop a dynamic hybrid optimization framework to achieve real-time optimization.

4. Conclusions and Prospects

This paper comprehensively reviews the use of AI methods in optimizing wind power system design since 1993, focusing on three key areas: WFLO, WTCO, and WFECSO. The primary optimization objectives are to reduce LCOE, increase Cp, maximize NPV, and raise the ABE of the entire WF. In order to achieve these goals, empirical estimation and numerical calculation methods were used in early work. These methods often need to simplify the problem significantly, which has limitations and affects the optimization of wind power systems. Because the optimization problem of WFs is often a very nonlinear and non-continuous problem, the advantages of using AI methods are obvious. Currently, the most reviewed AI methods are GA, PSO, ANN, etc. GA gradually improves the quality of the solution by screening high-quality individuals, mixing genes and random mutations, but it needs to set an evaluation criterion in advance to guide the direction of evolution. GA methods possess an indirect form of autonomous learning by evolving a population of solutions through operators like selection, crossover, and mutation. GAs adapt and search for optimal or near-optimal solutions without gradient information or prior knowledge about the problem’s nature. However, their learning process is stochastic, potentially requiring more iterations and computational effort to converge, especially in high-dimensional problems. Each particle in PSO adjusts the moving trajectory by referring to its own historical experience and the optimal results of the group and can find a better solution even without mathematical gradient information, but the effect of the algorithm is greatly affected by the parameters set by human beings, Like GAs, it features indirect autonomous learning through iterative solution updates based on the local and global bests. ANNs rely on a large amount of data to repeatedly adjust the parameters of the internal connection and gradually grasp the correlation relationships between the input and the output, and especially when dealing with complex data, they can gradually improve the solution quality and automatically find the law. ANNs learn directly from data and can identify complex, nonlinear patterns without explicit programming, making them highly adaptive to new, unseen data once trained properly. However, their learning relies heavily on the availability and quality of training data and appropriate network architecture. While other rule-based or structured supervised learning models (e.g., support vector machines, decision trees) are capable of learning from data, their autonomous learning ability is typically less adaptive compared to ANNs or evolutionary algorithms, especially for continuously changing or highly dynamic environments.

In the review paper, it can be found that HA (GA, PSO, ACO…) are more often used for optimizing multi-objective functions than ANN. The implementation of these AI techniques has led to significant improvements. For instance, the Cp of WFs can be increased by up to 33%, the NPV by up to 1.02%, and the AEP by approximately 10%, and the overall costs can be reduced by about 10%. Despite these advancements, this review also identifies potential areas for further improvement and exploration in the future:

• Most of the literature only discusses WFLO or WFECSO separately, but these two aspects are often closely linked and should be considered at the same time in order to further improve wind farm design.
• In the literature reviewed in this paper, it is found that GA is used in about 50% of the reviewed papers; GA can optimize a single problem in a single time, but GA may have limitations in solving complex problems. Compared with GA, PSO is easy to implement, has a faster convergence speed, and may obtain an improved LCOE. However, with the continuous expansion of wind farms and the increasing complexity of optimization objectives, the PSO algorithm may need to be further improved to make it converge faster and more accurately. Some other algorithms, such as SAA, CRO, and so on, also have some shortcomings, such as high computational cost, and they are not guaranteed to achieve global optimization, but the integration of some of these methods may also be considered in the algorithm improvements. With the expansion of the scale and complexity of wind power systems, traditional AI models may struggle to cope with massive data and complex nonlinear relationships. In future, advanced AI algorithms such as deep learning and reinforcement learning may be more widely used in wind farm optimization design. These algorithms can better capture the dynamic characteristics of the system, thus providing a more accurate optimization scheme.
• Compared with onshore wind power, offshore wind power has advantages such as stable wind energy resources and no occupation of land resources. However, it also faces huge technical challenges, such as high cost of wind turbine infrastructure, difficulty in laying submarine cables, expensive maintenance cost, etc., so offshore wind farm wind turbine layouts, substations, collection system configuration and capacity selection should be optimized together, in order to further reduce LCOE. In addition, with the rapid development of renewable energy technologies, offshore wind power may effectively be combined with offshore photovoltaic power generation, wave power generation and tidal current power generation to form an intelligent renewable power system which may be optimized with an integrated approach.
• Digital twin technology may be used to build a virtual model of a wind farm and use real-time data for simulation and optimization. Then, AI algorithms will be deeply integrated with digital twin technology to realize real-time optimization and intelligent operation and maintenance of wind farms. This technology can not only improve the operation efficiency of wind farms but also extend the life time of wind turbines.