Analysis of Wind Turbine Equipment Failure and Intelligent Operation and Maintenance Research

Abstract

Power generation from wind farms is growing rapidly around the world. In the past decade, wind energy has played an important role in contributing to sustainable development. However, wind turbines are extremely susceptible to component damage under complex environments and over long-term operational cycles, which directly affects their maintenance, reliability, and operating costs. It is crucial to realize efficient early warning of wind turbine failure to avoid equipment breakdown, to prolong the service life of wind turbines, and to maximize the revenue and efficiency of wind power projects. For this purpose, wind turbines are used as the research object. Firstly, this paper outlines the main components and failure mechanisms of wind turbines and analyzes the causes of equipment failure. Secondly, a brief analysis of the cost of wind power projects based on equipment failure is presented. Thirdly, the current key technologies for intelligent operation and maintenance (O&M) in the wind power industry are discussed, and the key research on decision support systems, fault diagnosis models, and life-cycle costs is presented. Finally, current challenges and future development directions are summarized.

1. Introduction

Wind energy generation is an important measure to develop a circular economy and to alleviate resource constraints [1]. In contrast to the substantial consumption and environmental pollution brought on by traditional energy sources, wind energy, as a technology for electricity conversion using renewable resources, does not generate pollution during power generation [2]. It is therefore of great importance in terms of energy savings and emission reduction. Over the long term, the economic and environmental benefits of wind energy are superior to those of conventional power, and its energy conservation and emission reduction are outstanding, setting a new trend in the future development of electric power generation and thus establishing a clear trend for the future development of electric power [3]. According to statistics from the Global Wind Energy Council, the operation and maintenance (O&M) costs of wind turbines can be as much as 10% to 20% of the total electricity produced and as much as 20% to 25% of the total electricity produced by offshore wind turbines [4].

Wind energy O&M costs primarily comprise routine maintenance costs, breakdown maintenance costs, spare part purchase costs, insurance premiums, and administrative costs. Offshore wind turbines cost about three times as much to run as their onshore counterparts. The difficulty and cost of maintenance has increased with the advent of large-capacity offshore wind turbines. The cost of operating and maintaining an offshore wind farm accounts for around 40% of the total life-cycle cost of the project. The cost structure of offshore wind energy O&M is shown in Figure 1 [5]. High failure rates and heavy maintenance workloads are the most significant challenges in wind energy O&M. Thus, it is important to ensure efficient and intelligent O&M of wind turbines to reduce costs and to increase the efficacy of wind farms.

This paper summarizes the current research, challenges, and countermeasures of intelligent O&M in the wind energy industry and forecasts its development trend based on a summary of the working principle of wind turbines, fault analysis, and O&M costs. Difficulties in the O&M process are discussed to provide a reference for future research on intelligent O&M in the wind energy industry.

2. Main Components and Faults Analysis of Wind Turbines

Currently, the most widely used type of wind turbine is the horizontal axis wind turbine with three blades. This type of turbine is mainly composed of a hub, blades, an engine room, a diversion hood, a generator, an environmental control system, a pitch system, a wind gauge system, and a yaw system [6]. Wind turbine structures are designed according to the working principles and functions of their components to extract kinetic energy from the air flow and to convert it into electricity. The structure of a three-blade-type horizontal axis wind turbine is shown in Figure 2.

2.1. Analysis of Main Components in Wind Turbine

The main components used in a wind turbine are as follows.

• Blade

In wind turbines, the blades absorb wind energy. They are often made of fiber-reinforced composite materials. In the flow field, the blade generates upward lift force through the pressure difference given by the wind field to the upper and lower surfaces, thus driving the rotation of the wind wheel and converting mechanical energy into electrical energy [8].

• Pitch System

The pitch system is the central component of any large wind turbine control system. It adjusts the angle of the blades, maximizing the wind energy output of the wind turbine. To a large extent, the variable paddle system guarantees the working reliability of the wind turbine [9].

• Gear box

Gearboxes generally consist of gears, bearings, shafts, and other parts. The quality and design of the gears directly affect the reliability, service life, and operating efficiency of the gearbox [10]. Wind turbine gearboxes are usually made of high-strength, low-alloy steel materials, which are treated with precision heat treatment and surface strengthening processes to improve their load-bearing capacity and wear resistance. In general, wind turbine gearboxes are designed for a service life of about 20 years, while requiring low noise, low vibration, and reliable operation [11].

• Yaw system

The yaw system is used to control the vectoring of the wind turbine in the direction of the wind to ensure that the wind turbine is always in the best position. The wind yaw system consists of a wind sensor, a control system, and an actuator [12]. The wind sensor is responsible for detecting the wind direction. The control system compares the wind direction signal with the set value and adjusts the orientation angle of the wind turbine via the actuator so that it always maintains a certain angle from the wind direction, thus maximizing the wind energy available for power generation [13]. The design and operation of the wind turbine yaw system is extremely important, as it has a direct impact on the operating efficiency and power generation of the wind turbine. Therefore, the yaw system requires sensitive response and high accuracy, as well as strong durability, anti-interference capability, and adaptability to different climatic environments [14].

Wind turbines work by converting wind energy into mechanical energy, which is then converted into electrical energy through a gearbox. As the wind turbine rotates, it drives the main shaft, which turns the generator through the gearbox and transmits the electrical energy to the grid [15]. However, due to complex mechanical structures and harsh environments, wear in wind turbine components such as the blades, the hubs, the gearboxes, and the pitch systems is inevitable and wind turbines are at risk of failure at any time, leading to a significant increase in their operating and maintenance costs. The purpose of a fault analysis is to determine the cause of the fault; to clearly understand the nature, location, and reason for the fault; and to take targeted maintenance actions to ensure normal operation of the wind turbine while reducing cost and maintaining efficiency [16]. Therefore, a timely fault analysis is important for wind turbines.

2.2. Analysis of Common Faults in Wind Turbines

In the past, in order to meet the requirements of wind energy generation, wind farms were usually built at sea or in remote mountainous areas and thus operated in extremely harsh environments [17]. As an important part of a wind farm, a wind turbine is bound to suffer from the harsh environment. Unpredictable loads due to dust, humidity, temperature, air pressure, and wind gusts can subject the main load-bearing components in the unit to severely alternating loads, making the blades and towers highly susceptible to tribological effects such as wear, fatigue, and corrosion, thus leading to unit failure [18]. However, new problems arose as the number of industrial wind turbine installations increased. Some failures are due to damage to the wind turbine and exceeding the design fatigue limit, while others are due to new failure modes as a result of material ageing, manufacturing defects, and changes in rotor size standards. The number of accidents is also increasing as the number of onshore and offshore wind turbines increases [19].

Figure 3 shows the number of cases of damage to wind turbines around the world from 2000 to 2017. It shows that as the number of wind turbines installed around the world has increased, so has the number of accidents. The number of wind turbine accidents averaged 57 per year from 2000 to 2005 and increased to 118 per year from 2006 to 2010. Furthermore, between 2013 and 2017, the total number of accidents per year rose even higher, to 167 [20].

Common types of failure in wind turbines include blade failure, gearbox failure, pitch system failure, and yaw system failure. The common fault characteristics and causes are summarized as follows.

• Failure of blade

The cost of manufacturing wind turbine blades is approximately 15% to 20% of the total cost of the turbine [21]. Usually, the blade as a whole is exposed in the field, and working conditions can be harsh. These conditions include high altitudes; atmospheric radiation; sand and dust; lightning; heavy rain; freezing rain or ice, and snow; salt corrosion; and typhoons. Such variable conditions can result in blade failure events; thus, blade condition monitoring, early warning systems, and effective O&M analyses are increasingly important [22].

The common failure modes of blades are (1) fracture and cracking failure; (2) damage caused by lightning strikes; and (3) local surface abrasion, cracks, etc. Blade failure is often accompanied by defects such as folding, poor curing, and unequal layering [23]. The causes of blade failure are analyzed below.

(a) 

Failure of fracture and cracking

• Design defects

The design of some blades leads to safety issues, including the blade root and blade middle section area being too small, the section shape not meeting strength and stiffness requirements, and the blade’s actual operating load exceeding the design of the predicted limit. Figure 4a shows design defects causing blade fracture. In addition, the redundancy amount of each component of the wind turbine, and the spacing and mass of the blades are often insufficient in the design. For example, the wind energy capture rate will be reduced if the blade spacing is too large. Due to disturbance in the airflow between the blades, there will be air leakage and the wind energy cannot be fully utilized. Additionally, there will be an increase in the volatility of the turbine, and then, the stability will become worse. If the distance between the blades is too small, it will cause the blades to interfere with each other and produce their mutual influence. This will not only reduce the efficiency of using wind energy but also cause the turbine to become overloaded and unsafe. Moreover, a blade spacing that is too small will also increase the amount of vibration in the unit, accelerate fatigue failure of the unit, and affect the service life of the unit, leading to problems during operation [24,25].

• Leaf material quality does not meet requirements

Manufacturers often use unqualified gel-coat, resin, or fiber materials, which have poor homogeneity and are prone to local soft ribs, leading to sudden leaf failure [26].

• Change in production technology without authorization

At present, high-quality composite blades are mostly manufactured via PRIM (polyamine reaction injection molding), RTM (resin transfer molding), winding, prepreg, or hot-pressing processes. However, some manufacturers use hand-paste processes to reduce costs and to gain market advantages. This phenomenon leads to uneven glue content, poor fiber penetration, and incomplete curing of the blade during production, resulting in problems such as blade deformation and breakage [27].

• Inexperienced manufacturers

Some manufacturers have insufficient experience, unclear control protocols, and simple methods. Faced with hundreds of main and auxiliary materials, tools, tooling processes, molds, and equipment involved in blade manufacturing, quality and production control is often lacking and inspections are limited to the surface, which cannot guarantee blade quality [28].

(b) 

Damage caused by lightning strikes

With the majority of wind energy plants being installed in the open countryside and in mountainous regions, in addition to the longer operating time of plants and the deeper soiling of blade surfaces, wind energy plants are increasingly affected by lightning strikes [29]. Figure 4b shows a blade damaged by a lightning strike. Failure due to lightning strikes on a blade is common at the lightning catcher and other parts of the blade [30].

• Cracks in the lightning conductor and blades

The cementation between the lightning initiator and the blade is not strong, and rain or moisture absorption can occur, changing the path of mine induction [31]. When lightning strikes the blade, the huge energy released by the lightning leads to a sharp increase in the temperature of the water-soaked material in the blade structure; then, the decomposed gas expands at high temperature, and the pressure rises, resulting in damage [32].

• Fouled blade surface

Dust deposition corrodes the blade surface, accelerating blade surface weathering, burrowing, and elastic weakening and resulting in micro-cracks. These cracks lead to further corrosion from dust, and mixtures formed by electrostatic dust accelerate blade corrosion again [33].

(c) 

Local surface abrasion and cracking

Local surface abrasions occur mainly on the tip-windward side, the mid-windward side, the leading edge, and the trailing edge of a blade, as shown in Figure 4c [34].

• Abrasion by wind sand and water vapor

A rotating blade will produce friction and collision with particles (dust and water vapor) in the air [35]. In most cases, the tip velocity of the blade exceeds 70 m/s. At this speed, the particles in the air will cause wear and tear on the leading edge of the blade and vortex abrasion on the trailing edge. Though the leaves worn by gravel first appear to be depressed, this depression is actually due to tiny sand holes, and the ageing of materials is greatly accelerated following the formation of sand holes [36]. Water is stored in these sand holes, and humidity in the depressions increases; thus, this greatly increases the possibility of the blade being damaged by lightning [37].

• High-speed wind, shear wind, bad weather, and fatigue life

Although blades stop rotating when wind speeds exceed the limit, high-speed winds, shear winds, or high gusts can cause the blade load to exceed the design load, resulting in blade damage. Severe weather such as heavy rain, snow, and hail can also cause leaf damage. With the extension of the service cycle of the blade, the anti-fatigue ability of the blade is weakened, which will also lead to blade damage [38].

Figure 4. Forms of blade failure. (a) Blade fracture accident caused by design flaw [24]; (b) lightning-damaged blade failure [30]; (c) locally abraded blade [34].

Figure 4. Forms of blade failure. (a) Blade fracture accident caused by design flaw [24]; (b) lightning-damaged blade failure [30]; (c) locally abraded blade [34].

• Failure of gearbox

The gearbox contains a large number of gears, and failure of the gears accounts for about 60% of total gearbox failures. Failure of the gears is mainly concentrated in the teeth, including tooth surface corrosion, tooth surface wear, tooth surface bonding, and broken teeth [39].

(a) 

Tooth surface corrosion

As shown in Figure 5a,b, tooth surface corrosion mainly includes rust and electric corrosion. The most common chemical rust is tooth surface rust, which is mainly due to the excessive water content of lubricating oils or long-term high humidity from air inside the gear box, which mostly occurs in areas with abundant summer rain [40]. Corrosion occurs when the gear surfaces are in close contact and there is little movement between the gears (generally, an amplitude of movement less than 120 μm). Under such conditions, the film of the lubricating oil between the gear teeth cracks due to compression, which leads to direct contact between the metals and results in adhesion of the micro-bulges onto the gear surface. Under these relatively small movements, there can be tearing of the gear contact surface by fusion welding adhesion as well as the production of a fine reddish-brown iron oxide powder similar in color to coffee beans. As a result, the gear develops abrasive wear. During this process, close contact between debris or worn edges of the gear will further prevent lubricating oil from reaching the surface of the gear. Corrosion results in deterioration over time and often occurs in relatively stationary gear drives.

(b) 

Tooth surface wear

Tooth surface wear can be divided into micro-pitting, macro-pitting, tooth surface spalling, and fretting wear according to different causes and levels of severity [41].

Micro-pitting, also known as gray spot, is a gray damage phenomenon appearing on the tooth surface of the gear during long-term engagement, as shown in Figure 6a [42,43]. The causes of wear are relatively complex. It is generally believed that due to problems in machining accuracy during gear handling, there are tiny bumps on the surface of the teeth, which results in the oil film not covering 100% of the gear surface or the thickness of the cover being insufficient. During meshing of these bumps, the teeth suffer from friction and shear stress, fatigue, and local overheating, resulting in micro-pitting.

Macro-pitting is the formation of small potholes when the gear contact surface is split by fatigue. Not only relative rolling but also relative sliding occurs during gear meshing. Due to material defects or excessive short-term loads caused by the impact force of the gear, the shear stress generated by relative sliding of the tooth surface will exceed the fatigue limit of the material for a short time, and fatigue cracks will occur on the tooth surface [44]. As the operation continues, the crack expands, and eventually, metal particles begin to flake off at the crack boundary; this phenomenon is called macroscopic pitting [45]. Figure 6b shows the gear damage caused by macroscopic pitting on the tooth surface [46]. If the fatigue crack continues to expand, eventually there will be massive spallation of the surface.

Fretting wear, also known as a “black line”, is a composite type of wear generated between metal surfaces that are pressed together, as shown in Figure 6c [47]. The reason for this wear is that the gearbox stops running for an extended period of time, and the tooth surface above the lubricating oil level does not receive lubrication. Small vibrations bring about constant reciprocal meshing between the gears, and alternating contact stresses appear; a black line will appear on the tooth surface of the gear bite. With time, pitting and even cracks will gradually appear from fretting wear due to high temperatures and oil shortages; subsequently, a tooth fracture will occur [48].

Figure 6. Forms of partial gearbox failure. (a) Micro-pitting of tooth surface [43]; (b) macroscopic pitting of tooth surface; (c) gear fretting wear [46].

Figure 6. Forms of partial gearbox failure. (a) Micro-pitting of tooth surface [43]; (b) macroscopic pitting of tooth surface; (c) gear fretting wear [46].

(c) 

Tooth surface gluing

Tooth surface gluing is the direct contact between two tooth surfaces due to rupture of the oil film under pressure, which results in adhesion and tearing. Depending on the mechanism of production, there are two types of gluing: hot gluing and cold gluing [49]. In the event of gear wear, vibration, or overload, the worn gear surface will generate high frictional heat. At this time, if the lubricant temperature is too high, the oil film fails and the tooth surfaces come into direct contact, resulting in thermoplastic deformation. Even if the lubrication is restored and the deformation produced by the tooth surface is memorized, the tooth surface will continue to undergo thermoplastic modulus deformation, which is also known as the gear hot glue phenomenon [50]. Cold gluing is caused by the low sliding speed of the gear. The meshing part thus does not easily form an oil film, resulting in a local pressure increase. The two contact surfaces of the oil film are punctured, resulting in adhesion tear [51].

(d) 

Broken gear teeth

Broken gear teeth include fatigue-broken teeth and overload-broken teeth, as shown in Figure 7a,b [52]. This is failure caused by alternating loads exceeding the fatigue limit of materials [53]. Fatigue fractures can be divided into three zones: the fatigue crack source zone, the crack extension zone, and the transient fracture zone. Their development process can be roughly divided into root bending stress, initial fatigue crack, continuous crack propagation, and tooth fracture [54]. An overload fracture occurs when the load on the tooth surface is greater than the maximum load. The macroscopic characteristics of the fracture are the same as that for a tensile fracture in the material, with a pronounced crinkling phenomenon. The development process of overload-broken teeth can be roughly divided into local tooth fracture caused by short-term overload impact, local fracture of gear with smaller tooth width, and fracture of helical gear or spur gear with larger tooth width [55].

• Failure of pitch system

There are three main forms of failure in the pitch system.

(a) 

Pitch shaft cabinet battery under voltage

If there is a failure of the pitch shaft cabinet battery, it is possible that one of the batteries in the fan blade battery cabinet has been damaged due to prolonged use or that the battery charging circuit is faulty, in which case the battery power provided by the battery cabinet is below the critical lower limit, and the paddle cannot change pitch [56,57].

(b) 

Abnormal hub detection data

If the hub detection data are abnormal, the wind turbine’s pitch communication system will fail. The wind turbine nacelle and hub communicate via a bus of electrical slip rings. When wind turbines are operated for long periods of time, problems in communication due to foreign objects or slip ring failures may occur [58,59,60].

(c) 

Paddle angle deviation exceeds limit

The rotation angle of the paddles can exceed the lower or upper limit of the preset paddle angle due to rotation overruns during operation or shutdown of the wind turbine [61,62].

• Failure of yaw system

The yaw system is an important components of a wind turbine. Its function is to track the change in wind direction and to drive the nacelle to rotate around the central axis of the tower, so that the wind turbine is always in the windward position [63]. In the course of wind energy O&M, faults in the yaw system constantly appear, especially in the mechanical parts, which are characterized by frequent maintenance difficulties and various types of faults [64]. Common failures in wind turbine yaw systems are as follows.

(a) 

Abnormal noise

• Yaw drive motor

Yaw drive motor issues include the following: the motor rotor adopts a sealed bearing; the bearing lubricating oil deteriorates due to age [65,66]; the raceway, roller, and bearing cage of the bearing are damaged; the position of the brake grinding disc at the rear of the motor is not correct; the grinding disc is worn; and the cooling blade is broken or the blade is rubbing the rear end cover of the motor [67].

• Yaw transmission

Yaw transmission issues include the following: a lack of lubricating oil or deterioration of the lubricating oil in the planetary gear train of yaw transmission; excessive content of mechanical impurities in the lubricating oil; transmission bearing damage; a lack of lubricating grease in the transmission; and yaw gear occlusion [68].

• Yaw big gear ring

Gear ring issues include the following: the lubrication grease on the tooth surface of the gear ring can be infiltrated by sand; the positioning pin or fixing bolt assembling the rack of the big gear ring is loose; and the big gear ring is broken [69,70].

• Yaw gear pair fit clearance

Each blade yaw mechanism is provided with two to four yaw drivers [71]. If the fit clearance between the drive gear and the yaw gear with large ring teeth is uneven, noise from the gear bite clearance is likely to occur during yaw [72].

• Yaw bearing

There are two types of yaw bearings: sliding bearings and rolling bearings [73]. Due to the low speed of the yaw bearings and the lack of lubricating grease in the rolling bearings, the bearings are corroded by the current, causing scorching of the bearing rollers [74]. The sliders in the sliding bearings are made of Teflon material, and some are lined with friction-reducing material between the sliders and the slewing disc; noise can come from slider failure [75].

• Yaw brake

Cracks in the friction discs, faulty adjustment of the caliper body compensation device, and uneven damping torque values in the brake are caused by excessive damping pressure in the yaw brake or by the presence of sand and mechanical impurities on the friction surface [76].

(b) 

Yaw drive machinery is stuck

First, yaw motor bearings are sealed bearings, and staff are unable to regularly replenish the lubricant; this results in fatigue damage to the bearings due to high and low temperature differences [77,78,79]. Second, the yaw drive is mounted vertically, with the yaw drive bearing at the lower end. The bearing carries a large inertial force from the large gear ring, and mechanical magazines within the drive tend to enter the bearing gap [80]. Third, the yaw drive gear is susceptible to over-torque damage at the moment of start or stop of the yaw system, especially during gusts in periods of high wind [81]. Fourth, the sliding yaw bearing liner can crack. This failure is related to the quality of the liner material and the yaw caliper balancing torque [82].

(c) 

Inaccurate yaw positioning

Inaccurate yaw positioning can be caused by brake failure of the yaw motor, insufficient yaw braking pressure, and low yaw braking torque values [83,84]. Broken teeth of the yaw gear ring and yaw drive device are shown in Figure 8a,b [85,86].

Often, a failure has multiple causes and is accompanied by multiple failure modes. Figure 9 summarizes the failure modes of key wind turbine components. Due to the harsh operating conditions of wind turbines, the constant changes in environment, temperature, and load make it extremely likely for individual wind turbine components to fail;

Table 1. The main causes of failure modes for each component.

Failure Mode of Each Component	Main Causes
Fracture and cracking	Design defects; poor quality; change in process; lack of production experience
Lightning strikes	Cracks in components; fouled blade surface
Local surface abrasion and crack	Abrasion by wind, sand, and water vapor; extreme weather
Tooth surface corrosion	Excessive water content of lubricating oil; wet working environment
Tooth surface wear	Machining accuracy; material defect; short-term overload; inactivity for an extended period
Tooth surface gluing	High-speed or low-speed heavy load
Broken teeth	Fatigue; overload
Pitch shaft cabinet battery under voltage	Shaft cabinet battery under voltage; charging loop fault
Abnormal hub detection data	Communication failure
Paddle angle deviation exceeds limit	Rotational overrun
Abnormal noise	Deterioration of driving motor oil; transmission bearing damage; ring tooth broken; uneven ring clearance; lack of grease in yaw system; brake damping pressure exceeds limit
Yaw drive machinery stuck	Planetary gear train failure
Inaccurate yaw positioning	Brake failure of yaw motor

lists the main causes of failure modes for each component. If researchers are able to analyze these failure modes and their causes, combined with a signal processing analysis and determination of component operating conditions, it will be easier to identify specific component failure types, making fault diagnosis more efficient and further enabling a reduction in costs and an increase in efficiency.

3. Cost Analysis of Intelligent Operation and Maintenance for Wind Energy

The main purpose of a wind power intelligent O&M cost analysis is to determine the cost structure and scale of this O&M, to evaluate its feasibility, and to provide important guidance and reference for its implementation. First, by analyzing the cost structure and scale of wind power intelligent O&M, the specific costs and budgets for each phase can be understood; this is conducive to achieving more effective cost control and budget management. Second, a cost analysis can evaluate the feasibility and economic benefits of wind power intelligent O&M and can clarify whether there is a commercial value and market prospect for investment. In addition, a cost analysis can provide specific guidance and reference for the implementation of intelligent O&M, including equipment selection, personnel training, management processes, and other aspects. At the same time, based on the analysis results, a more scientific and reasonable implementation plan can be developed, improving the efficiency and success rate of intelligent O&M implementation. Without a cost analysis in the early stage of implementation, it is impossible to determine a sustainable business model.

3.1. Analysis of Wind Power Intelligent Operation and Maintenance Cost Components

Currently, intelligent O&M for wind turbines can be broken down into online monitoring and diagnosis and offline maintenance [87]. The online monitoring and diagnosis system mainly includes three functions: intelligent detection, intelligent maintenance, and advanced application [29]. The intelligent detection function will comprehensively monitor the performance parameters and power parameters of online and offline blade equipment and upload the monitoring data to the wind farm and headquarters [88]. The intelligent maintenance function classifies the data signals collected from all dimensions of the device blade to detect major faults or hidden dangers during device operation in a timely manner. Advanced application functions include multi-angle and overall data analyses and intelligent judgments on the problems arising during equipment detection and maintenance and reasonably estimate the possible problems based on large-scale data mining [89].

To ensure that equipment and systems work stably and safely during normal operation of a wind energy project, wind farm staff also need to regularly maintain and inspect wind turbines; these costs are known as the maintenance costs [90]. For onshore wind farms, maintenance costs for wind turbines include maintenance personnel costs, material costs, and technical transformation costs [91]. Maintenance costs vary depending on the type of equipment used on a wind farm. Maintenance costs refer to the total cost of manual inspection and repair in order to restore the wind turbine to normal operation after failure, including labor, machinery, and spare parts costs [92]. For different types of wind turbines, the recovery rate and utilization rate of equipment failure are different, as are the maintenance costs [93]. The O&M of offshore wind farms differs significantly from that of onshore wind farms, and these points of difference clearly affect the cost of offshore wind O&M [94].

The cost of intelligent O&M for wind power installations consists of seven components, as shown in

Table 2. Components of the intelligent O&M costs for wind power systems.

Cost Type	Cost Content
Hardware costs	The cost of hardware equipment required by wind power intelligent O&M software, such as servers and data storage devices [95]
Software costs	Copyright fee and usage fee for wind power intelligent O&M software [96]
Labor costs	Labor maintenance costs, including technical support, system updates, etc., as well as labor costs for maintaining wind power equipment [97]
Communication costs	Communication costs related to intelligent O&M [98]
Monitoring equipment costs	The cost of monitoring and testing equipment, such as sensors and measuring equipment [99]
Operating costs	Includes costs for manpower, materials, and funds required for daily operation management, supervision, maintenance, and troubleshooting; predictive maintenance and scheduled maintenance; coordination and organizational management [100]
Training costs	The cost of training personnel to use intelligent wind power O&M systems, including the cost of training facilities and labor costs [101]

.

3.2. Analysis of Wind Power Intelligent Operation and Maintenance Cost Control

As the scale of wind power development continues to expand, traditional maintenance methods are becoming obsolete. First, the wind farm station service environment is harsh and largely unattended, and manual inspection is difficult and costly. Second, the unit is scattered and complex, with variable operating conditions and low reliability, resulting in maintenance difficulties. Third, the high cost of repair and replacement of large components accounts for a high proportion of the infrastructure and power generation costs. Fourth, regular maintenance is not timely, and after-the-fact maintenance affects the overall efficiency of power generation [102,103,104]. Compared with traditional maintenance technologies, wind power intelligent O&M technology uses big data, artificial intelligence, and other technologies to manage systems more efficiently and to ensure their safety and stability through automatic monitoring and preventive measures. In addition, this O&M technology can address issues faster and more efficiently, thus reducing downtime and losses, lowering maintenance and repair costs, and improving the sustainability of wind power generation systems [105,106].

However, with the large-scale development in the past decade, wind power generation technology has matured in various aspects, such as model design, micro-siting, construction and installation, and operation and control; as a result, problems have gradually increased, and the cost of wind power intelligent O&M has also increased [107]. Wind power intelligent O&M requires significant technical investment, including software and hardware equipment and network communications infrastructure. Such devices must be constantly updated and maintained, and they require professional technical staff to manage, operate, and maintain equipment [108]. Intelligent O&M requires the collection and processing of data from numerous sensors to detect equipment failures and to provide early warning. These data require long-term storage, analysis, and processing, which necessitates appropriate storage equipment and data analysis software [109]. Again, intelligent O&M requires professional technicians for such maintenance and management, and personnel must have advanced technical skills and professionalism. Therefore, companies need to invest appropriate human resources and funds for recruitment, training, and management [110]. Finally, intelligent O&M requires continuous equipment updates and maintenance to ensure equipment performance and reliability. These updates and maintenance require a large amount of capital and human resources [111].

As shown in Figure 10, between 2018 and 2022, the number of wind turbines installed in China increased from 62,000 to 117,000, while the cost of wind power intelligent O&M increased from USD 14.7 billion to USD 20.1 billion. It can be predicted that the cost of wind power intelligent O&M will continue to rise in the next five years, suggesting an urgent need to control the cost of wind power intelligent O&M [112].

As the cost of wind power intelligent O&M increases year by year, more and more scholars in the field are conducting in-depth research on wind power intelligent O&M cost control. Through collation and rational analysis, the following four categories of methodological systems are presented.

• Decision support system

The decision support system can help power companies prevent equipment failures and damage by providing real-time equipment monitoring and maintenance information, thereby reducing the number of repairs and the time needed to complete them and thus lowering maintenance costs [113]. It can also integrate advanced intelligent optimization algorithms, such as genetic algorithms and fuzzy cluster analysis, to optimize the processes and parameters of wind power intelligent O&M, thereby reducing operating costs [114]. By analyzing and visualizing the large amount of data generated during wind power intelligent O&M, trends and anomalies can be identified to better understand the operational status and conditions of the equipment, thereby improving operational efficiency and accuracy and reducing maintenance costs [115]. By comprehensively controlling these aspects, the decision support system can effectively reduce the cost of wind power intelligent O&M.

• Fault warning and health management

Fault warning and health management can detect the fault information of wind turbines early on, leading to timely repair or replacement and thus improving the reliability and stability of the equipment [116]. It can also accurately determine the timeline and scope of equipment to be repaired and replaced, avoiding manual inspection and maintenance time and thus reducing O&M costs. Through the analysis and prediction of equipment operation data to develop a more accurate maintenance plan, timely monitoring and adjustment of equipment operation status can take place, reducing unnecessary maintenance work and improving maintenance efficiency and quality [117].

• Life-Cycle Assessment

Life-cycle assessment is a method used to evaluate and, in turn, control the cost of wind power intelligent O&M. First, in the planning stage, factors such as the use of wind power, intelligent O&M, implementation factors, and environmental impact must be considered to ensure the rationality and feasibility of the system. This assessment also collects and analyzes a large amount of data from wind turbines to evaluate their operational efficiency and economy over their life cycle [118]. Data collection and analysis can be performed using tools such as sensors, monitoring equipment, and data analysis software. Second, by evaluating the actual operation of the wind power intelligent O&M system, the performance in terms of efficiency, cost, and sustainability over its life cycle can be evaluated, summarized, and analyzed [119]. Third, by continuously improving the system, its efficiency and economy can be improved, thus reducing costs. Fourth, by decommissioning the system, the life cycle of wind power intelligent O&M can be completed, and further maintenance and repair can be avoided [120]. The control of wind power intelligent O&M costs through life-cycle assessment can effectively improve the efficiency and economy of the system.

• Digital twin system

The digital twin system can quickly detect equipment faults, digitally determine the cause of faults, and provide repair solutions and data support, which can reduce the time and cost of manual inspection and equipment repair and can improve the automated O&M of wind farms [121]. In addition, it can warn of wind farm failures, enable early intervention of maintenance teams, and provide optimization recommendations through simulation solutions to help wind farms improve power generation efficiency by reducing troubleshooting time and costs, thereby balancing costs and benefits [122].

In summary, intelligent O&M cost control is key to improving the economic benefits of wind power operation. Through scientific and reasonable O&M cost control, wind farms can improve their profits and increase their economic benefits. Conducting wind power intelligent O&M cost control can also ensure that staff pay more attention to the operation status of wind power plants, finding and correcting faults in a timely manner, thus improving the operation efficiency of wind power plants. With the rapid development of the wind power industry, the competition between enterprises is becoming more and more intense. Scientific and reasonable control of wind power intelligent O&M costs can reduce the cost of enterprise products and improve the competitiveness of enterprises.

4. Research on Key Technologies for Intelligent Operation and Maintenance of Wind Turbines

The intelligent O&M of wind turbines is based on the integration of cloud computing and big data to build a new energy intelligent platform for full data sharing and data value deep mining. An online O&M solution based on a wind turbine condition assessment model can improve station O&M capabilities and operational decision making [123,124,125]. It mainly focuses on preventive maintenance and optimization of spare part storage and supply strategies. In order to improve the safety and reliability of wind turbines and reduce O&M costs, it is essential to study the key technologies [126]. The development of intelligent wind turbine O&M can be classified into several stages, as follows. The first stage is based on decision support systems. The second stage is centered on the fault diagnosis model. The third stage involves the use of life-cycle assessment. In this section, a summary of existing and potential key technologies is presented.

4.1. Decision Support Systems

Many factors affect the choice of maintenance policy, such as failure probability and weather conditions. It is unreasonable to expect O&M staff to take all these factors into account when making decisions [127]. Hence, the emergence of decision support systems, which can be used to provide decision makers with multiple and cost-effective options. The decision support system (DSS) is an intelligence-based human–computer system that supports decision-making activities based on the principles of management science, operational research, simulation, and behavioral science and uses computer technology, simulation, and information technology as a means to address semi-structured decision-making problems. The system is able to provide the decision maker with the necessary data and background to help clarify the objectives of the decision, to carry out problem identification, to build or modify decision models, and to consider different alternatives [128]. Various options can be evaluated and analyzed, compared, and judged through human–computer interaction functions, providing the necessary support for correct decision making [129,130].

Stamatescu et al. [131] described the development of a decision support system (DSS) for low-voltage networks incorporating renewable energy sources (photovoltaic panels and wind turbines), with the aim of achieving energy balance in pilot micro-grids, reducing network energy consumption, and lowering operating costs. However, during operation, the method only considers the qualitative benefit of the energy provided by the micro-grid subsystem, rather than the exact value provided by the government, which may lead to small fluctuations in the experimental results. Rinaldi et al. [132] proposed a corrective maintenance decision model, which can restore components to a good state and restore reliability values to the initial state. Figure 11 shows the workflow diagram of the model. Aiming at minimizing maintenance costs, Nguyen et al. [133] proposed a single maintenance plan and a group maintenance plan for offshore wind energy systems. The proposed method takes into account system reliability, weather conditions, maintenance time, generation losses during maintenance, location of offshore wind systems, and other parameters neglected in existing studies, which improves the accuracy of the method. Although the method reduces the cost of operating and maintaining offshore wind energy, it is only used in some parts of Taiwan and has not been tested in other regions.

Effective O&M aims to maximize economic efficiency. While tariffs are commonly used to calculate the profit or cost of electricity for wind farms, when they are used as evaluation metrics in decision support models, they are mainly used to calculate the production loss due to wind turbine downtime. Dinwoodie et al. [134] mentioned two methods for calculating production losses. The basic approach is to use the product of the rated power of the wind turbine as the capacity factor and downtime as the influencing factor. Another more precise approach is to use a time series of wind speeds to create a function and to determine the loss from this function curve. However, the literature does not take into account the cost of employing staff and the replacement of parts for wind turbines. Ju et al. [135] proposed a price-based demand–response virtual power plant and constructed a risk-averse model based on a conditional value-at-risk approach and robust optimization theory, considering the operational risks of wind power plants and photovoltaic power generation. However, experimental data show that the system is not suitable for simple scheduling scenarios where decision makers must fully account for uncertainty in the factors and require improved prediction accuracy. Qin et al. [136] proposed a new framework for wind farm resilience modeling to simulate how different types of disturbances can lead to damage and failure of wind farm systems and subsystems, as well as the associated direct and indirect consequences, to support optimization of asset integrity management decisions. Relevant performance metrics for wind farms, such as expected values for downtime and adequate inventory of essential spare parts, are easily quantified within this framework. However, in terms of maintenance regime selection, performing multiple condition substitutions to target different failure modes can achieve expected production performance above current values. Benalcazar et al. [137] provided a framework capable of identifying the optimal component size and determining the optimal generation schedule for an autonomous hybrid energy system, based on the formulation implementation of an optimization model with a linear programming approach and taking into account the meteorological and economic parameters of the experimental area. The techno-economic feasibility of hybrid energy system projects in remote areas of Ecuador was addressed. The framework reduced investment costs more effectively than Qin’s cost results. Although the difference between construction and production costs is not significant, Benalcazar’s model is clearly superior in terms of maintenance and recovery costs at a later stage. The comparison is shown in Figure 12. However, the model does not consider other techno-economic factors, such as payment and tariff methods, taxes, distribution costs, and energy saving methods, and thus has significant limitations.

All decision support models should be able to integrate and evaluate different O&M scenarios. Most decision support models operate in a complex environment, and operators using the models must already have an in-depth understanding of how to use them and be familiar with the reasons that affect their scheduling. There is currently no operational model that optimizes maintenance policies without input from operators or researchers. Therefore, there is an urgent need for a decision support model that can be operated without the need to have experience in wind energy O&M. Over the past decade, wind turbine decision support models have become a mature research field and are gradually being applied to manage and operate the wind power industry. These models not only contribute to the sustainable development of the wind power industry but also help wind power plant managers and operators to optimize the performance and efficiency of wind turbines. Verhelst et al. [138] developed corrosion measurement data visualization software that identifies standards for effective structural corrosion analysis as a scalable, SCADA-compatible, secure, and network-accessible tool, showing that SCADA-compatible visualization software tools are possible and can significantly reduce the technical and experiential requirements for O&M service personnel. However, since it is difficult to quantify the corrosion process theoretically, both the quantity and quality of data are extremely important, which also requires relevant organizations to raise risk awareness and collect adequate data in a timely manner. Petrichenko et al. [139] established a high-resolution methodology for measuring economic performance through a proposed index under different development scenarios of renewable energy diffusion and different network configurations by comparing the profitability of two business models: individual distributed consumption and energy communities; the latter proved to be a more profitable framework. However, the initial investment in energy communities is huge and requires governments and investors to improve their decision-making capabilities. Dinçer et al. [140] developed a multi-criteria decision-making approach to evaluate energy technology investment options based on a multi-stage weighting ratio analysis to determine the relative importance of factors and designed five evaluation criteria: maintenance frequency, ease of installation, environmental adaptability, transmission technology, and cost efficiency. The results show that cost efficiency is the most important factor in the effectiveness of energy investments and that investment decision makers should focus on this element. However, there are no actual data to confirm the results obtained using this method, so their applicability is questionable.

The maintenance and operation of wind turbines is costly and uncertain, and decision support systems can assist O&M personnel in identifying, classifying, and predicting failures or damage to wind turbines [141]. In the current technical research on decision support systems, there are several problems.

• Data acquisition

Decision-support-system-based wind power intelligent O&M technology relies on the accuracy and timeliness of equipment data, while current data collection and transmission equipment has a relatively high failure rate and low data quality. Thus, more reliable data collection and transmission equipment are needed to improve the accuracy and timeliness of data.

• Data processing

Data derived from different devices, formats, and processing methods can vary, leading to quality, temporal, and processing efficiency issues. In addition, data processing requires sufficient computing power and storage capacity. Therefore, new technical means for data processing and integration need to be explored. Among them, artificial intelligence technology is currently the most promising means.

• Intelligent diagnosis

Machine learning is a common intelligent diagnostic technique used today. The accuracy and reliability of this technique depends on the model. Therefore, models with high accuracy need to be developed. Some new models, such as deep learning models and decision tree models, are possible solutions to this problem.

• Maintenance forecasting

Maintenance prediction is an important element of decision-support-system-based technology. Existing maintenance prediction methods are limited by historical data and do not make good use of advanced algorithms and data from new energy monitoring devices. These methods also do not use big data and have very limited accuracy in fault prediction. The use of smarter algorithms, combined with big data means, can achieve more convincing results.

In response to the above problems, several improvement methods can be proposed, as follows. First, optimize data collection and transmission equipment to improve data accuracy and timeliness. Second, apply artificial intelligence technology and data mining technology to improve the quality and efficiency of data processing. Third, explore new algorithm models, such as deep learning models and decision tree models, to improve the accuracy of intelligent diagnosis. Fourth, use big data means to combine intelligent algorithm models and machine learning methods for maintenance prediction to improve the prediction effect. Meanwhile, multiple data sources can be combined to improve the accuracy of fault prediction [142]. In conclusion, the research on wind power intelligent O&M technology based on decision support systems needs to be improved to solve the existing problems and to provide better technical support for efficient and stable development of the wind power industry.

4.2. Fault Diagnosis Models

Early detection and diagnosis of wind turbine failures is essential for the application of possible maintenance and control strategies to avoid catastrophic events [143]. Nevertheless, turbines’ complex and large-scale structures, as well as their operation in remote locations with harsh environmental conditions and highly variable random loads, make the occurrence of faults inevitable. Therefore, the use of algorithms that can continuously monitor and diagnose potential faults and mitigate their effects before they become failures can significantly improve wind turbine fault mitigation [144].

In the study of wind energy bearing fault diagnosis models, Wang et al. [145] proposed the instantaneous frequency selection strategy guided by the l-kurtosis fluctuation spectrum (LFS) and the weight coefficient determination strategy guided by the cuckoo search algorithm (CSA) through the simulation of fault signals. It was verified that the optimized adaptive chirp mode decomposition (OACMD) method has outstanding advantages in defect detection in wind turbine bearings. However, this method is less efficient and cannot detect multiple sources of damage in wind turbine bearings under fluctuating speed conditions. Guo et al. [146] used the KS method to study the temperature–power dispersion distribution of gears of similar units and convolutional neural networks to construct a prediction model for bearing over temperature, which can detect abnormalities of the turbine in advance and reflect the fluctuation of turbine parameters well; however, for short-term fault warning, this approach needs to be remodeled based on historical data. Zhao et al. [147] combined Gaussian kernel support vector machines with polynomial kernel support vector machines and proposed a fault diagnosis method based on stochastic subspace identification (SSI) and multicore support vector machines (MSVMs) to accurately identify wind turbine bearing faults. However, this method only tests for a single fault signal and requires more in-depth analysis of the causes of multiple faults occurring simultaneously. Gu et al. [148] proposed two models based on different principles: an adaptive SR algorithm model based on quantum particle swarm optimization (QPSO) and a wind energy bearing fault diagnosis model based on frequency information exchange (FIE) combined with frequency conversion. However, these methods have a large data bias in the frequency exchange process, and the experimental results are not accurate. Liu et al. [149] validated a principle of testing for bearing temperature anomalies based on space–time fusion decisions. The principle uses the analytic hierarchy process (AHP) entropy method to model the temperature rise in a bearing based on historical inspection data uses the temperature signal to analyze the temperature characteristics of the bearing at different time and space distributions. After comparing the actual temperature rise with the predicted temperature rise, a diagnosis of the abnormalities in the bearing can be achieved. This can reduce the false positive rate of diagnoses of abnormal temperature rises in the bearing and has a strong early warning capability.

In the area of fault diagnosis models for wind energy gearboxes, Xu et al. [150] proposed an improved hybrid attention module using a multi-scale convolutional neural network as a feature learning model, in which a multi-label classifier outputting single or multiple labels was used to identify single or compound faults. The method was verified to have higher accuracy and stability than other transmission compound fault diagnosis models for two transmission data sets; the diagnosis process is shown in Figure 13. However, the method cannot be applied to early fault diagnosis. Mao et al. [151] effectively identified structural and nonstructural defects by combining fused domain adaptive convolutional neural networks (FDACNNs) with vibration signals and introducing adversarial networks to achieve state identification of structural and nonstructural defects in the unlabeled target domain. Chen et al. [152] proposed an intelligent fault diagnosis model for wind energy gear boxes, combining CNN and discrete wavelet transform. The deep training forward pass rule is used to train a CNN with a deep hierarchical structure, constructed by alternating convolutional layers and subsampling layer by layer to transform the input low-level features into high-level features. The parameters of the CNN are fine-tuned through a back-propagation process to establish a mapping between the feature space and the fault space. However, the system parameters and structural choices of the proposed model are not discussed in detail, leading to insufficient generality. Zhao et al. [153] developed a variant of the deep residual network, namely, a dynamic weighted wavelet coefficient, to improve model diagnostic performance. Compared with CNN- and DRN-based methods, the developed method improves the average training accuracy by 11.43% and 10.60% and by 3.74% and 3.87%, respectively, in terms of test accuracy. However, this method is only suitable for general data-driven fault diagnosis with small changes and cannot accurately diagnose faults with large data fluctuations. Zhang et al. [154] proposed a hybrid improved residual network-based wind turbine gearbox fault diagnosis method in conjunction with the above research. The method implemented a wavelet transform on the initial signal, highlighting the characteristics of the wavelet coefficients. The proposed method was then validated with a simulated data set from a drive train diagnostic simulator (DDS) and actual measurement data from a wind farm. Compared with the literature [152,153,154], the accuracy of the experimental results obtained using the HA-ResNet method under the DDS simulated data was significantly higher than that of other methods; the mean comparison of the experimental results is shown in

Table 3. Comparison of HA-ResNet results with other methods under DDS simulated data [152,153,154].

Method	Raw Signals		Wavelet Coefficients
	CNN	WDCNN	CNN	WDCNN	ResNet	HA-ResNet
Accuracy (%)	87.13 ± 2.57	94.13 ± 1.39	95.43 ± 1.08	93.14 ± 0.61	96.50 ± 1.26	98.79 ± 0.34

. However, the use of raw data for fault diagnosis in the time domain is not recommended due to the time-shifted nature of raw vibration signals.

In terms of overall fault diagnosis models for wind turbines, most scholars use the SCADA data set as a modeling reference. Liu et al. [155] developed an ambient temperature regression prediction model for wind turbine components using the limiting gradient boosting algorithm, which can be used to calculate the residuals between the predicted and actual values. The exponential weighted moving average (EWMA) control chart principle was used to monitor the variation trend of the residual difference. However, wind turbines are typically nonlinear multi-coupled systems, where the relationships between different faults are complex and the coupling is strong, so that the entire fault of a wind turbine cannot be accurately predicted. Watson et al. [156] used existing Supervisory Control and Data Acquisition (SCADA) data collected for a brand new wind turbine to model operational behavior and to construct a Markov space to be used as a reference space. The fault diagnosis of a wind turbine was performed by comparing the wind turbine behavior predicted by the trained model with the reference space and analyzing the distribution and correlation of the wind turbine SCADA data. However, the fault model was combined with the acoustic signal, which reduced the success rate of fault detection. Leahy et al. [157] established an accurate training database by automatically identifying historical wind turbine alarm sequences and using rule sets to infer wind turbine downtime and the reasons behind it and proved that failures could be predicted in advance. However, due to flaws in the machine learning experimental setup, the experimental results over-relied on the reliability of the data set. Marti-Puig et al. [158] proposed a method of using a single hidden-layer feed-forward neural network (SLFN) to establish a wind turbine health status monitoring and evaluation model, including using the remaining variables of the system or subsystem to predict the target variable, calculating the error deviation from the actual target variable, and finally comparing the high error value with the alarm event selection of the system. The performance of the model was validated.

Table 4. Comparison of selected signals from the Leahy et al. and the Marti-Puig et al. [157,158].

Signals of Leahy	Description	Signals of Marti-Puig	Description
az	Yaw system faults	Power	Generator output power
ba	Backup battery system faults	Wind speed	Average wind speed in 30 s
bk	Braking system faults	Rotor speed	Generator rotor speed
fc	Frequency Converter faults	Winding temperature	Generator winding temperature
gb	Gearbox faults	Drive-end bearing temperature	Generator drive-end bearing temperature
gn	Generator faults	Non-drive-end bearing temperature	Generator non-drive-end bearing temperature
mi	Miscellaneous	Nacelle temperature	Temperature inside the nacelle
pt	Pitch system faults
to	Tower faults

compares the differences between the model signals selected by Leahy et al. and Marti-Puig et al. Although the latter selected fewer signals than the former, these signals show the operating environment of the wind turbine and can adequately represent the operating parameters during operation, thus greatly improving the credibility of condition monitoring. Though it is robust to mislabeled data that may be introduced during the training phase, fault locations cannot be accurately predicted.

Fault diagnosis and fault-tolerant control schemes for wind turbines are critical to their reliability, availability, and cost-effectiveness. However, there remain problems in the current research. For example, the existing fault diagnosis models are mostly based on experience or rules and cannot accommodate adaptive learning and or accurately diagnose newly appearing faults. In addition, there are many kinds of wind power equipment, and various types of faults continually appear, which requires a large amount of data support and complex algorithms to build fault diagnosis models. The fault diagnosis process must involve the comprehensive analysis of many factors, including the structure and operation status of the wind turbine itself, the wind resource situation, and the qualifications of the O&M personnel.

Therefore, more machine learning algorithms should be used to learn historical data to improve the accuracy and adaptability of the model, to carry out large-scale and multi-dimensional data collection, to effectively analyze and process data to explore potential correlations in order to improve the prediction accuracy of the model, and to adopt open data to improve the sharing and accessibility of data and open data channels for various application scenarios.

4.3. Life-Cycle Costs

The life cycle of a wind energy project can be subdivided into economic life cycle, construction life cycle, and natural life cycle. The economic service life is defined as the service life with the lowest average cost. The design life is the period of time that a wind farm is designed to be used effectively, without loss of use. Natural life is the actual time a wind farm is in use under natural operation [159]. Life-cycle analysis methods are proposed and widely used in the feasibility analysis of wind energy projects to reflect the details of the entire life-cycle process of a product. Researchers have analyzed the economics of wind energy and provided a series of literature reviews. These reviews provide a positive reference for cost modelling, revenue analysis, and structural optimization at various stages of wind energy development and implementation [160].

With regard to the life-cycle components of wind energy projects, the main components are the planning and design stage, the construction stage, the O&M stage, and the closure stage. Figure 14 shows the workflow for the planning and design phase of a wind energy project. The main tasks of the planning and design phase include wind resource assessment, site selection and land planning, investment and financing, engineering feasibility studies, and environmental impact assessments [161,162]. By 2020, the global average unit capacity of new wind turbines had increased to more than 2.5 MW onshore, and the annual aggregate unit capacity of new wind turbines had increased to more than 7 MW offshore [163,164]. Once the preliminary design is completed, researchers carry out a feasibility analysis of the project, which includes an analysis of the need for the project to be built, a physical feasibility analysis, and an economic analysis. Finally, project and pre-production activities are carried out with regard to ecological restoration, project assurance, base reserves, and pre-production. Major tasks in the construction phase are to organize and plan construction, to construct civil bases, to erect pylons, to install electrical and control systems, and to simulate operation and commissioning [165]. The process of operating and maintaining a wind farm throughout its life cycle is referred to as the O&M phase. Rationalizing production, managing equipment, carrying out O&M, and ensuring the generation and stable operation of the wind farm are the main tasks in this phase [166]. For wind farms, the O&M phase is a long cycle with a large number of uncertainties, and the costs corresponding to this phase include the costs of O&M and the costs of land lease [167]. Currently, wind farm operations and maintenance are carried out jointly by wind turbine manufacturers and operators. The closure phase may include preparation for closure, disassembly of equipment, cleaning and recycling of the site, disposal of recyclable waste, and approval of the works. Most life-cycle analyses only consider full decommissioning, including complete dismantling and site clearance. Figure 15 shows the decomposition structure of the wind farm decommissioning work [168].

With regard to the methods of economic evaluation of wind energy and the selection of evaluation indicators, Mahmud et al. [169] developed a life-cycle inventory considering all inputs and outputs to assess and compare the environmental impacts of solar PV and thermal systems on 16 impact indicators and performed uncertainty and sensitivity analyses for both frameworks. Therefore, the methodology can be analogized in the life-cycle assessment of wind power systems to determine the level of environmental impact of wind power projects and to evaluate the performance of different wind turbine models. Yildiz et al. [170] improved the maintenance and end-of-life steps in the life-cycle assessment of barge-based floating wind turbines by adding recycling and mechanical incineration processes to the life-cycle process to reduce their environmental impact. However, this study failed to account for some calculable costs and showed low accuracy. Zhou et al. [171] used AHP to assess the life-cycle risk of distributed wind plants, providing a perspective on future research directions for distributed wind. However, the AHP is too simplistic for the current evaluation method, and another, more accurate method may be used to assess the risk impact. To facilitate comparison with other studies, Kouloumpis et al. [172] conducted a life-cycle assessment of small vertical axis wind turbines using a range of assumed capacity factors. They also used the Conditional Maximum Likelihood impact assessment methodology to assess 11 environmental impact categories. The results showed that most of the impacts were attributed to support infrastructure. However, since environmental performance is very sensitive to variations in capacity factors, the importance of wind project siting should be emphasized. Typically, NPV (net present value), PVC (present value cost), IRR (internal rate of return), LCOE (levelized cost of electricity), PBP (payback period), ROI (return on investment), ROE (rate of return on common stockholders’ equity), and NAV (net annual value) are widely used in the process of selecting indicators to evaluate wind energy [161]. These indicators reflect the viability of the different directions of project investment and are commonly used for the economic assessment of wind energy projects [173]. PVC and ROI are indicators of absolute value. They can only represent the cash value of a given amount of investment over a given period of time. The difference between investment and generation capacity can be reflected by other relevant indicators, such as LOCE, IRR, PBP, and NPV [174]. PBP increases project risk by allocating the investment over the payback period and does not take into account post-payback cash flows. NPV, PBP, PVC, ROI, and IRR are difficult to use to benchmark different regions, types, and levels of technology, and it is challenging to forecast the cash flows under flexible policies [175]. By comparison, LCOE removes the limitations of various regions, types, and sizes for studying the investment merits of a project, allowing for further analysis of the main factors affecting the cost of electricity generation, as well as accurate horizontal comparisons and choices between diverse projects [176]. The LCOE has become a common economic indicator to assess the overall competitiveness of wind energy projects. In addition, the economic analysis of certain types of projects often uses a more inclusive valuation approach, which combines PBP, IRR, and NPV [177].

In the field of life-cycle cost modelling for wind energy projects, there are many uncertainties in the economic analysis process, leading to a large gap in the results of the analysis and the actual cost of the project. Therefore, scientists have proposed various approaches to uncertainty analysis and prediction algorithms to quantify impacts and to estimate uncertainty factors. Tsvetkova et al. [178] classified the sensitivity analysis output variables, methods, applications, and software, highlighting the problems with this type of analysis. Currently applied computational models are expensive, and an effective sensitivity computational model is urgently needed in the field. Leite et al. [179] suggested that energy prices, production capacity factors, and capital operating costs are the most critical variables affecting economic viability. They used a global sensitivity model to identify the most sensitive parameters affecting the results of economic analysis; however, the model did not establish the corresponding life-cycle model according to the key variables. An et al. [180] designed and discussed a four-level life-cycle evaluation platform for wind turbines based on IoT architecture, developed a model based on IMPACT 2002+, and selected energy recovery time as an important evaluation index for wind turbines; the evaluation model is shown in Figure 16. The platform is oriented toward wind turbines, with good scalability and credibility. However, no reasonable and effective tests have been conducted to obtain valid experimental data. Yuan et al. [181] improved the LCOE model as a life-cycle costing method in three ways. They considered the quantification of external costs, expanded the composition of internal costs, and discounted electricity generation to provide a quantitative reference for electricity production and consumption decisions. However, the method is only applicable to the general domestic environment in China and has not been tested in other countries and thus is not universally applicable [182]. Erfani et al. [183] proposed a risk assessment model for wind energy investment projects, considering expert knowledge and background, based on the risk analysis of improved fuzzy group decision making. The authors established a comprehensive model of risk assessment and cost estimation on this basis, proving that the risk management process has a great impact on investment profitability. However, all cash flow parameters in this model are modeled using the triangle assumption, and the interplay and correlation risk among cash flow parameters is not investigated. Liu et al. [184] developed a new algorithm, called the Cuckoo Search Algorithm with Steepest Descent, based on the improved cuckoo search algorithm and optimizing the weights; it showed higher accuracy and stability in predicting wind speed. However, this method is not sufficiently accurate to predict the upper limit of the wind farm alarm threshold in practical applications. Based on a long short-term memory (LSTM) network and a gray wolf optimizer (GWO) disaggregation technique, Altan et al. [185] developed a new hybrid WSF (Wind Speed Forecasting) model. The model was able to accurately reflect the nonlinear properties of the WSTS (wind speed time series). However, it is more cumbersome and less accurate in its predictions. A new and different hybrid wind speed prediction system based on the combination of two model frameworks, linear and nonlinear, was proposed by Huang et al. [186]. It included the use of linear ARMA (auto regressive and moving average) to catch the underlying linear mode in the data, as well as CS-optimized BPNN (back propagation neural networks) to catch the underlying nonlinear mode in the data. However, other linear and nonlinear models were not considered in the study. Ma et al. [187] combined hierarchical analysis, network analysis, and fuzzy comprehensive evaluation methods to establish a hybrid hierarchical structure model and combined this with a fuzzy comprehensive evaluation method to determine the final risk level. The model fully considered the influence of various disaster factors and was more scientific and accurate than the previous evaluation methods. However, the model used the AHP-fuzzy method, which is too simplistic and less convincing under the depth of current research. Dai et al. [188] proposed a self-attentive neural network (SANN) model based on attention for online learning, which could capture temporal relationships more efficiently with a customized attention module; this online learning algorithm could adapt to weather, operational, and environmental changes to improve prediction accuracy. The SANN model is more cost effective than existing models. A comparison of the results between different models is shown in

Table 5. Test results of SANN model compared with existing models.

Method	MAE (MW)	MAPE (%)	RMSE (MW)	R2	Reference
Persistence	3.53	0.76	5.95	0.991	Lledó. [189]
LSTM	2.83	0.71	4.06	0.995	Li. [190]
CNN-LSTM	3.31	0.72	5.07	0.992	Wu. [191]
AGRU	2.12	0.46	3.71	0.996	Niu. [192]
Bi-LSTM	2.56	0.55	4.03	0.995	Jaseena. [193]
CNN	2.72	0.59	3.99	0.996	Zhang. [194]
MLP	2.23	0.48	3.72	0.996	Khazaei. [195]
TCN	2.26	0.49	3.72	0.996	Li. [196]
SANN	2.22	0.48	3.68	0.996	Dai. [188]

.

Li et al. [197] proposed an optimal planning approach for regional electricity-heat systems with data centers (DC-RIEHS) considering wind power uncertainty and using affine decision rules to hedge the uncertainty of wind power generation into an easy-to-handle single-level mixed integer cone model. However, this study focused only on the granularity modeling of individual data centers and the flexible planning of energy systems in their surrounding areas, without considering detailed energy networks. Alshamrani et al. [198] proposed a transmission expansion planning model that considers the voltage stability assessment of a grid with embedded wind farms under uncertainty conditions, which is unique in that it incorporates the voltage stability margin (VSM) into the planning model. The model is a two-tier model in which the first tier minimizes the total investment and operating cost minus the weighted VSM and the second tier evaluates the optimal expansion plan given by the VSM from the first tier. However, the model is run without the inclusion of transmission line conductors, which can increase the desired loading factor. Maienza et al. [199] collected the latest data and parametric equations from databases and literature to establish a life-cycle cost model for offshore floating wind farms, divided the life-cycle cost into three parts—capital cost, O&M cost, and decommissioning cost—and analyzed the differences between different types of floating platforms in terms of cost components, transportation, and installation. Garcia-Teruel et al. [200] performed a life-cycle assessment of an offshore floating wind farm using an advanced O&M model to quantify the impact of environmental factors and compared the estimated characteristics of the Spar case and the Semi-sub case, as shown in

Table 6. The assumed and estimated characteristics of the two case studies [200].

Characteristic	Spar Case Study [201]	Semi-Sub Case Study [202]
Water depth (m)	95–129	60–80
Distance to shore (km)	25	15
Turbine rating (MW)	6	9.5
Number of turbines	5	5
Turbine model	SWT-6.0-120	v164-9.5
Substructure	Spar	Semi-submersible
Installation port	Peterhead	Dundee
O&M port	Peterhead	Aberdeen
Lifetime (years)	25	25

. The data prove the viability and practicality of the proposed model.

With the continuous development and application of intelligent technologies, research on intelligent O&M technologies for wind power based on life-cycle costs has also seen some progress [203,204]. This section provides an overview of the progress of life-cycle cost-based research.

First, there is a lack of life-cycle cost considerations. Many studies on wind power intelligent O&M technologies do not consider the full life-cycle costs of wind farms, focusing only on a single O&M cost or equipment maintenance cost. This one-sided consideration can lead to excessive reductions in O&M costs, while causing higher losses in later operations and maintenance. Second, there are insufficient data processing and analysis capabilities. Wind power intelligent O&M technology requires the processing and analysis of a large amount of data; however, most wind farms remain weak in this area, and thus, data processing and analysis capabilities need to be further improved. This can lead to a decline in O&M efficiency, which affects the safety and reliability of wind farms. Third, there is a lack of an efficient collaborative O&M mechanism. There is a large amount of O&M work in wind farms; this requires cooperation among multiple O&M teams. However, there is a lack of efficient collaborative O&M mechanisms, which leads to poor cooperation and information asymmetry in the O&M process, which reduces efficiency.

For the above problems, the following countermeasures are suggested.

• Establishing a sound and universal whole life cost model requires in-depth research and discussion for different application scenarios. By quantifying the costs and benefits of wind farms at different stages, a reasonable optimization of the economic benefits of the whole wind farm can be achieved [205].
• It is necessary to strengthen data management and analysis capabilities by adopting efficient data mining and analysis algorithms, as well as more scientific, reasonable, and effective cost estimation methods; this will lead to fast and accurate data processing and will improve cost estimation capabilities.
• It is essential to establish an efficient, collaborative O&M mechanism; to strengthen the training and communication of O&M personnel; and to improve the cooperation between O&M teams, so as to achieve more efficient intelligent O&M for wind power.

5. Summary and Conclusions

In this paper, the common failure modes and causes of wind turbines are discussed in detail, and the failure mechanisms and processes of wind turbine equipment are investigated. This study analyzes current O&M costs in the wind energy industry so that the reader can understand the relationship between wind energy operations and maintenance costs. Finally, we review the current research results on key technologies for intelligent O&M in the wind energy industry, summarize the problems that exist in these research results, and provide an outlook on the future directions of intelligent O&M in the wind energy industry. According to the review of previous studies, a great deal of experimental and theoretical research is still needed in the following areas.

(a) 

Fault analysis and research of wind turbine

• Failure data comes from different platforms and systems, and the differences in data quality lead to certain problems in data stability and reliability. Data quality should be improved by methods such as standardizing defective data and creating a defect database to provide a basis for subsequent data analysis.
• Not only do wind turbines have a large number of installations and complex operational data, but the number of effective analyses for fault samples is small, which results in a lack of real and valid fault analysis data. By establishing a multi-domain fault data platform and providing more fault samples to accelerate the training efficiency of these machine learning algorithms, the accuracy of fault analysis results can be improved.
• The current fault analysis algorithms are based mainly on machine learning algorithms. However, while they have high accuracy, some of the algorithms have low processing efficiency, and it is difficult to address large-scale fault analysis problems. In order to develop more efficient and faster Java application systems for fault analysis decision support, new types of algorithms, such as deep-learning-based algorithms, should be actively practiced and developed.
• Fault diagnosis requires a large amount of domain knowledge and experience. However, in the actual fault analysis process, the analysis of the real situation is sometimes limited by the accumulation of knowledge. A fine-tuned fault knowledge system and modeling method should be established to correlate the wind turbine operational state with the fault mode to achieve accurate and fast fault diagnosis and evaluation.

(b) 

Research on key technologies of wind energy intelligent O&M

• Wind turbines produce a variety of data during operation, and the data are numerous, scattered, and discontinuous; thus, a refined data acquisition system must be established to integrate the data into an analyzable format. A fast and efficient data acquisition system should be established to integrate discrete and scattered data into an analyzable format to provide a basis for wind turbine condition monitoring and prediction.
• Data analysis algorithms and models suitable for wind turbines need to be established, and based on these algorithms and models, intelligent diagnosis and early warning systems need to be developed to diagnose and prevent wind turbine failures in a timely and accurate manner.
• A large amount of data and analysis results must be accurately and reliably transformed into effective decision support to realize automated operation and intelligent decision making. A reliable and practical intelligent decision support system should be established to deeply analyze the meaning behind the data and to provide intelligent decision support.
• The intelligent O&M system undoubtedly faces hidden security problems, such as system attacks and data leakage, which seriously affect the safe and stable operation of wind turbines. Strict measures such as data encryption, system security prevention, and vulnerability repair should be taken to improve the security and confidentiality of the wind power intelligent O&M system.