AI-Driven Control Strategies for FACTS Devices in Power Quality Management: A Comprehensive Review

Abstract

Current power systems are facing noticeable power quality (PQ) performance deterioration, which has been attributed to nonlinear loads, distributed generation, and extensive renewable energy infiltration (REI). These conditions cause voltage sags, harmonic distortion, flicker, and disadvantageous power factors. The traditional PI/PID-based scheme of control, when applied to Flexible AC Transmission Systems (FACTSs), demonstrates low adaptability and low anticipatory functions, which are required to operate a grid in real-time and dynamic conditions. Artificial Intelligence (AI) opens proactive, reactive, or adaptive and self-optimizing control schemes, which reformulate FACTS to thoughtful, data-intensive power-system objects. This literature review systematically studies the convergence of AI and FACTS technology, with an emphasis on how AI can improve voltage stability, harmonic control, flicker control, and reactive power control in the grid formation of various types of grids. A new classification is proposed for the identification of AI methodologies, including deep learning, reinforcement learning, fuzzy logic, and graph neural networks, according to specific FQ goals and FACTS device categories. This study quantitatively compares AI-enhanced and traditional controllers and uses key performance indicators such as response time, total harmonic distortion (THD), precision of voltage regulation, and reactive power compensation effectiveness. In addition, the analysis discusses the main implementation obstacles, such as data shortages, computational time, readability, and regulatory limitations, and suggests mitigation measures for these issues. The conclusion outlines a clear future research direction towards physics-informed neural networks, federated learning, which facilitates decentralized control, digital twins, which facilitate real-time validation, and multi-agent reinforcement learning, which facilitates coordinated operation. Through the current research synthesis, this study provides researchers, engineers, and system planners with actionable information to create a next-generation AI-FACTS framework that can support resilient and high-quality power delivery.

1. Introduction

1.1. Power Quality Challenges in Modern Grids

In modern power grids, maintaining power quality (PQ) is challenging owing to the increased infiltration of nonlinear loads of various forms, distributed energy resources (DERs), and renewable energy sources. The efficient planning of electric power systems is essential to meet both the current and future energy demands. In this context, reinforcement learning (RL) has emerged as a promising tool for control problems modeled as Markov decision processes (MDPs). Recently, its application has been extended to the planning and operation of power systems. This study provides a systematic review of advances in the application of RL and deep reinforcement learning (DRL) in this field. The problems are classified into two main categories: Operation planning including optimal power flow (OPF), economic dispatch (ED), and unit commitment (UC) and expansion planning, focusing on transmission network expansion planning (TNEP) and distribution network expansion planning (DNEP). The theoretical foundations of RL and DRL are explored, followed by a detailed analysis of their implementation in each planning area. This includes the identification of learning algorithms, function approximators, action policies, agent types, performance metrics, reward functions, and pertinent case studies. Our review reveals that RL and DRL algorithms outperform conventional methods, especially in terms of efficiency in computational time. These results highlight the transformative potential of RL and DRL in addressing complex challenges within power systems. These factors cause voltage sags, flickers, harmonics, and poor power factors, thus negatively affecting industrial, commercial, and residential infrastructures [1,2,3,4,5,6]. Traditional proportional-integral (PI) or proportional-integral-derivative (PID) controllers often fail to adapt well to such dynamic conditions. This failure is especially severe in an industrial environment, such as semiconductor manufacturing, where a PQ lapse can cause losses of up to USD 90,000 to even USD 1 million hourly. The use of renewable energy sources can also cause additional complexities, such as harmonics generated by inverters and the introduction of stochastic changes in production. Modern microgrids are more susceptible to such disruptions, which may cause disruption of sensitive electronic components and interfere with the production processes. IEEE Standard 1159-2019 [7,8] categorizes power quality phenomena according to their magnitude, duration, and spectral characteristics, and establishes measurement protocols and acceptance thresholds for diverse operating environments. The installation of renewable energy sources creates further complications in the regulation of power quality, such as inverter-generated harmonics, fluctuations in resource production settings, and unaligned phases in distribution networks. Modern microgrids are particularly vulnerable to these disturbances and can experience fast transients (or even rapidly moving clouds) that disrupt sensitive electronic devices and manufacturing processes [9]. These challenges necessitate an intelligent adaptive control paradigm capable of operating under uncertainty, ensuring a real-time response, and maintaining consistent PQ in distributed systems. Traditional proportional integral (PI)/PID-based controllers are inadequate in these environments. This encourages the combination of power-electronic-based modern devices with Artificial Intelligence (AI), which has the potential to predict, self-optimize, and scale control strategies for power quality enhancement [10].

1.2. FACTS Devices, Principles and Capabilities

FACTS is a decisive element in the development of the quality of power and stability of systems by dynamically regulating the values of line parameters, such as voltage, impedance, and phase angle [11,12]. The power electronic controllers underneath it can quickly inject or accept reactive power, which is used to stabilize the voltage levels, increase the transmission capacity, and reduce transient disturbances [9,13]. Its main benefits include controlled load flow, reduced generation costs, enhanced stability, and minimized reactive power losses [12,14]. FACTS devices can be classified as shunt compensators (e.g., SVC and STATCOM), series compensators (e.g., TCSC), or hybrid devices (e.g., UPFC and IPFC), each designed to perform a unique role, such as voltage regulation, harmonic mitigation, and power flow management [15,16,17,18,19,20,21,22]. However, traditional FACTS control methods demonstrate limitations in flexibility, which supports the need for more innovative and data-targeted control paradigms in modern grids typified by high influxes of distributed energy resources [23,24,25,26,27]. As illustrated in Figure 1, power quality disturbances affect the stability of the grid and equipment operation.

Figure 1. Overview of Flexible AC Transmission Systems (FACTS).

1.3. Limitation of Conventional FACTS Control

Traditional FACTS control schemes, the models of which are usually based on PI regulators, rely on statistical parameters resulting from small-signal analyses and tuning through trial and error. These methods are satisfactory in the operating conditions of steady-state conditions, but display extreme deficiencies in modern power systems [11]. First, they are not predictable, particularly within the framework of integrating renewable energy, which leads respondents to not implement control policies. Second, the restrained flexibility requires manual retuning when the user changes the network settings and load profiles. Third, the approaches face challenges in multi-objective optimization when the power quality goals are incompatible with other operational goals [2,12]. These limitations highlight the need for intelligent controllers that not only react to but also allow the prevention of disturbances in advance, which artificial intelligence methods are beginning to provide [28,29].

1.4. Artificial Intelligence Controllers

FACTS control is being reshaped by artificial intelligence (AI) to allow real-time optimization with adaptive and predictive capabilities. Neural networks, fuzzy logic, and evolutionary algorithms can improve the performance of FACTS more than traditional methods can [15,30,31]. Unlike conventional methods, AI can use historical data to learn control and be optimized heuristically, thus proving to be especially effective in handling the issue of power quality under intricate grid conditions [11,12,31]. Recent studies have shown that AI can be utilized to enhance the audit of power quality, such as the control of harmonic distortion and the forecasting of voltage sag [32,33,34]. Such abilities can be particularly useful in microgrids and systems that typically use renewable resources, where dynamic variations and inverted power can occur [2,12,14]. Deep learning, fuzzy, and reinforcement learning have shown promising results; however, they have not been systematically analyzed to comprehend how various AI tools can be applied to tackle specific power quality concerns and the use of FACTS applications, particularly in light of practical implementation challenges such as explainability, real-time operation, and data requirements.

1.5. Research Gaps and Paper Contributions

This review determines the essential gaps in the research on AI integration with Flexible AC Transmission Systems (FACTS) in four main ways.

• Systematic Classification of AI Solutions to PQ-Focused FACTS Control: This study develops a new taxonomy based on which AI methods, including Long Short-Term Memory networks (LSTM), Convolutional Neural Networks (CNN), Graph Neural Networks (GNN), fuzzy logic, and reinforcement learning, are classified in respect to their ability to be used with a particular device and power quality targets. Taxonomy is classified in terms of control functionality, which can be voltage sag mitigation, harmonic suppression, and flicker reduction, consequently offering a guide on the relevant choice of AI models for each application.
• Quantitative analysis of AI-enhanced versus traditional Controllers: The study conducts a strict performance analysis of the AI-enhanced FACTSs in comparison to the traditional PI/PID controllers. Various power quality tests were analyzed, including the accuracy of voltage regulation, total harmonic distortion (THD) reduction, flicker reduction, response time, and effectiveness of reactive power compensation.
• Assessment of practical implementation issues: The authors evaluated the practical limitations of deployment, which comprise computational latency, model integrity, interpretability, and data quality, in addition to integration with the existing grid infrastructure. Special attention is paid to the differences in simulation results and hardware-in-the-loop (HIL) or field-level models.
• Roadmap to Future Research and Future Trends: Future promising research opportunities are nominated, including the use of digital twins to validate AI-FACTS, federated learning to create distributed intelligence, physics-based neural networks (PINNs) to create constraint-based control, and multi-agent reinforcement learning (MARL) to achieve coordinated microgrid control.

Together, these works address the gap between theoretical AI-FACTS studies and actual implementation and provide a new framework for future development of smart grids, in the presentation of applications of AI in smart grids. The remainder of this paper is structured as follows: Section 2 provides background information on the basic types of power quality disturbances and their effects on the performance of modern power grids. Section 3 provides an overview of the proposed advanced AI techniques, such as deep learning, reinforcement learning, and hybrid techniques, which are applied to power system control. Section 4 presents the manner in which such AI techniques have been combined with FACTS devices to enhance power quality parameters such as voltage stability, harmonic filtering, and flicker limitation. Section 5 describes the implementation issues, such as computational limitations, sensor accuracy, and interoperability. Section 6 concludes the paper, summarizes the findings, and lays out future ways in which intelligent and AI-assisted FACTSs can be achieved in a smart grid environment.

2. Overview of Power Quality Issues and FACTS Solutions

The definition of power quality (PQ) disturbances can incorporate a range of deviations in voltage, current, or frequency, with the capacity to impair the normal operation of equipment. In the current IEEE 1159-2019 standard [7]. PQ issues are categorized into several classes based on their severity, duration, and frequency of occurrence (FO). Such instabilities, whether transient or steady-state, are becoming increasingly significant problems in modern grids, especially those with large amounts of nonlinear loads and low renewable energy penetrations. In order to be able to plan effective control strategies, all these phenomena must be classified and their effect on grid reliability and operational efficiency understood.

2.1. Classification of Power Quality Disturbances and Cost Loss

PQ is defined as the set of electrical characteristics that allows electrical equipment to function properly without any significant loss of performance or life expectancy [12]. Various power quality disturbances are discussed in the following sections.

These disturbances have significant economic and operational impacts on industries worldwide. In industrial sections, equipment lifetime is reduced, and power losses are increased owing to problems that can affect the grid power quality. Interruption costs are related to faults in equipment repair, damaged products during the production stage, spoiled raw materials, protection systems, energy backup systems, and additional staff working hours. In some sectors, the loss can be staggering; per-minute interruption costs in the chemical sector reach approximately $135,000, in the food industry about $2500, and in the mechanical sector around $75,000 [35,36,37]. From a high-level perspective, power supply variations and voltage disturbances cost approximately $119 billion per year for industrial facilities in the US, according to an Electric Power Research Institute (EPRI) report. In the meantime, 25 EU states suffer the equivalent of $160 billion in financial losses per year due to improper ID PQ [38,39]. This classification is illustrated in Figure 2.

Figure 2. Power quality disturbances.

2.1.1. Voltage Magnitude Variations

Voltage Sags (Dips)

The disturbance caused by a reduction in the root mean square (RMS) voltage at 0.1 to 0.9 per unit over 0.5 cycles up to 1 min is called a voltage sag. As shown in industry data, the percentage of power quality disturbances experienced in industrial premises is about 70 per cent due to voltage sag [11,40,41]. Sags are generally categorized based on the level (strength) and duration, with severity increasing with increasing levels [40]. The major sources of such agitation are system faults (line-to-line, line-to-ground, etc.), large motor starting (in-rush current), transformer energization, and heavy load switching [42]. The mathematical characterization of this phenomenon is as follows:

$${\mathrm{V}}_{\mathrm{s}\mathrm{a}\mathrm{g}}= {\displaystyle \frac{{\mathrm{V}}_{\mathrm{d}\mathrm{u}\mathrm{r}\mathrm{i}\mathrm{n}\mathrm{g}}}{{\mathrm{V}}_{\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}\mathrm{a}\mathrm{l}}}} \times 100\%$$

(1)

where ${\mathrm{V}}_{\mathrm{s}\mathrm{a}\mathrm{g}}$ is the sag voltage (between 0 and 1), ${\mathrm{V}}_{\mathrm{d}\mathrm{u}\mathrm{r}\mathrm{i}\mathrm{n}\mathrm{g}}$ is the actual voltage level measured during the sag event, and ${\mathrm{V}}_{\mathrm{n}\mathrm{o}\mathrm{m}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}}$ is the standard operating voltage of the electrical system.

Voltage Swells

It is a transient rise in RMS voltage between 1.1 and 1.8 per unit occurring over a duration of 0.5 cycles to 1 min [37]. The formula associated with this fault is identical to that of a voltage sag. They are less common than sags, but may be equally dangerous and comparable [35,40]. The usual causes are line-to-ground faults (no faulty phases), load rejection, capacitor bank energization, and incorrect tap-up/down settings of transformers.

Long-Duration Voltage Variations

A long-lasting voltage abnormality (less than 0.9 per unit or greater than 1.1 per unit) that takes more than one minute is considered a long-duration voltage variation. This is because of load variations, switching operations, and poor system voltage regulation rather than fault conditions.

2.1.2. Waveform Distortion

Harmonics

Periodic waveforms in which sinusoidal components have frequencies that are integer multiples of the fundamental frequency are referred to as harmonics. With the growth of nonlinear loads and the infiltration of renewable energy, harmonic distortion has become a ubiquitous challenge in contemporary power-grid systems. IEEE 519-2014 [43,44] recognized the limits of harmonic voltage and current distortion [40,45] at the PCC, with somewhat higher values and different conditions. The major causes of these disturbances are switch-mode power supplies, variable-frequency drives, LED lighting, arc furnaces, and battery chargers [40,42,46,47,48]. The measurement of this disturbance is followed by the following formula, which requires the Total Harmonic Distortion (THD):

$$\mathrm{T}\mathrm{H}\mathrm{D}= {\displaystyle \frac{\sqrt{\sum _{\mathrm{h}=2}^{\infty }{\mathrm{V}}_{\mathrm{h}}^2}}{{\mathrm{V}}_1}}\times 100\%$$

(2)

where ${\mathrm{V}}_{\mathrm{h}}$ is the RMS voltage of harmonic order h and ${\mathrm{V}}_1$ is the fundamental voltage. Another quantified method is for the Individual Harmonic Distortion (IHD), which is used as follows:

$${\mathrm{I}\mathrm{H}\mathrm{D}}_{\mathrm{h}}={\displaystyle \frac{{\mathrm{V}}_{\mathrm{h}}}{{\mathrm{V}}_1}}\times 100\%$$

(3)

Beyond classical harmonics, recent studies highlight supraharmonics and common-mode EMI in the 9–150 kHz band of modern converter-rich systems, which can propagate through motor drives and distribution networks [40,43].

Interharmonics

Frequency components that are not integer multiples of the fundamental frequency have become increasingly common in modern power systems owing to the large-scale application of cycloconverters, arcing-type devices, induction motors operating at variable loads, and power electronic converters that utilize asynchronous switching. These Interharmonics may have a variety of negative effects, such as light flicker, thermal loading of capacitor banks, noise that interferes with control and protection signals, and interruptions in ripple control systems [49].

Notching

This phenomenon represents periodic voltage disturbances caused by the normal operation of power electronic devices during phase communication [50].

Noise

Noise is the voltage or current with spectral components below 200 kHz that are superimposed on the power waveform. Practically, this noise is typically divided into two categories: common mode noise, which propagates between the conductors and ground, and transverse mode noise, which influences the conductors of the line [9].

2.1.3. Voltage Fluctuations and Flicker

Voltage fluctuations refer to random changes in voltage, typically between 0.9 and 1.1. These changes in voltage often occur in the frequency range of 0.5–30 Hz, which can cause visible flicker [38,39]. The IEC 61000-3-7 and IEEE 1453-2015 [44] standards establish the planning and compatibility levels for flicker-severity assessments. Flicker severity was quantified using

• Short-term flicker severity (P_st): Measured over 10 min
• Long-term flicker severity (P_lt): Measured over a 2 h period

$${\mathrm{P}}_{\mathrm{s}\mathrm{t}}=\sqrt{\sum _{\mathrm{i}=1}^5{\mathrm{k}}_{\mathrm{i}}.{\mathrm{P}}_{\mathrm{i}}}$$

(4)

$${\mathrm{P}}_{\mathrm{l}\mathrm{t}}=\sqrt[3]{{\displaystyle \frac{1}{12}}\sum _{\mathrm{i}=1}^{12}{\mathrm{P}}_{\mathrm{s}\mathrm{t},\mathrm{i}}^3}$$

(5)

where ${\mathrm{k}}_{\mathrm{i}}$ is the weighting coefficient and ${\mathrm{P}}_{\mathrm{i}}$ is the percentile level of the flicker sensation.

Primary sources of voltage fluctuations include

• Arc furnaces
• Welding equipment
• Rolling mills
• Reciprocating pumps and compressors
• Wind turbines
• Large motor starting (repetitive)

2.1.4. Power System Imbalance

Voltage Unbalance

In the normal scenario of a three-phase voltage system, the magnitudes of the waves should be the same, with a 120° difference in their phases. However, an unbalanced voltage can occur because of abnormal grid conditions. Typical industry standards limit voltage unbalance to 2–3% [51]. This disturbance can be quantified using the following equation:

$$\mathrm{V}\mathrm{o}\mathrm{l}\mathrm{t}\mathrm{a}\mathrm{g}\mathrm{e} \mathrm{U}\mathrm{n}\mathrm{b}\mathrm{a}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e} \mathrm{F}\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{o}\mathrm{r} \left(\mathrm{V}\mathrm{U}\mathrm{F}\right)={\displaystyle \frac{\left|{\mathrm{V}}_2\right|}{\left|{\mathrm{V}}_1\right|}}\times 100\%$$

(6)

where V₁ and V₂ are the positive and negative sequence voltage components, respectively.

This phenomenon arises from several structural and operational factors in the power system. One key contributor is the uneven distribution of single-phase loads over the three phases, which can lead to current imbalance and voltage deviation. Moreover, open delta transformer configurations can often cause asymmetry in voltage magnitudes and related phase angles. Blown fuses in capacitors can also cause reactive power compensation disruption. In addition, asymmetric transformer impedances resulting from aging or manufacturing tolerances introduce unequal voltage drops under the load. Finally, non-transparent transmission lines contribute to mutual coupling, which distorts the problem throughout the entire network.

Current Unbalance

Similarly to voltage imbalance, current imbalance is a critical indicator of power-quality degradation. This term is often quantified as the Current Unbalance Factor (CUF), which is defined as [51]

$$\mathrm{C}\mathrm{U}\mathrm{F}={\displaystyle \frac{\left|{\mathrm{C}}_2\right|}{\left|{\mathrm{C}}_1\right|}}\times 100\%$$

(7)

where C₁ and C₂ are the positive and negative sequences of the current components, respectively.

This metric provides a percentage-based evaluation of asymmetry extension in the current waveform. A higher CUF indicates a greater deviation from the normal operation of the three-phase system, which can adversely affect various equipment, such as motors, transformers, and power electronic converters. Current and voltage imbalances can occur independently of each other. Even when the system voltage appears balanced, other factors such as unbalanced loading or nonlinear devices can cause issues with the CUF.

2.1.5. Power Frequency Variation

Power frequency deviations from the nominal 50 or 60 Hz beyond ±0.1 Hz indicate PQ issues. These arise from imbalances between generation and load [52]. Frequency stability encompasses steady-state variations and dynamic deviations caused by generator failures or load changes. Metrics such as the Rate of Change in Frequency (ROCOF), frequency nadir, and recovery time are used to assess performance and guide frequency control.

2.1.6. Transient Disturbances

Impulsive transients are sudden voltage or current changes with unidirectional polarity caused by lightning strikes, inductive switching, and electrostatics [53]. They have rapid rise times (1–10 µs), short durations (<50 µs), and high magnitudes (up to kV). Oscillatory transients affect both polarities and are categorized by frequency range: low (<5 kHz), medium (5–500 kHz), and high (500 kHz–5 MHz), which are related to switching and lightning events.

2.1.7. Power Factor and Reactive Power Issues

A low power factor has a high negative impact on both power quality and system efficiency [54]. These issues are caused by various factors, such as the current increment for the same active power transfer, high voltage drops across the lines, reduction in equipment capacity utilization, and increased system losses From a system-level perspective, improving PQ and efficiency in conversion chains complements broader sustainability goals identified by comparative life-cycle assessments of renewable electricity systems. For both sinusoidal (power factor, PF) and nonsinusoidal (displacement power factor, DPF) conditions, this occurrence can be quantified using the following equations:

$$\mathrm{P}\mathrm{F}= {\displaystyle \frac{\mathrm{P}}{\mathrm{S}}}={\displaystyle \frac{\mathrm{P}}{\sqrt{{\mathrm{P}}^2+ {\mathrm{Q}}^2}}}$$

(8)

$$\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e} \mathrm{P}\mathrm{F}={\displaystyle \frac{\mathrm{P}}{\mathrm{S}}}={\displaystyle \frac{\mathrm{P}}{{\mathrm{V}}_{\mathrm{r}\mathrm{m}\mathrm{s}}.{\mathrm{I}}_{\mathrm{r}\mathrm{m}\mathrm{s}}}}=\mathrm{D}\mathrm{P}\mathrm{F}.\sqrt{1+{\mathrm{T}\mathrm{H}\mathrm{D}}_{\mathrm{I}}^2 }$$

(9)

$${\mathrm{T}\mathrm{H}\mathrm{D}}_{\mathrm{I}}={\displaystyle \frac{\sqrt{{\mathrm{I}}_2^2+{\mathrm{I}}_3^2+{\mathrm{I}}_4^2+\dots }}{{\mathrm{I}}_1^2}}$$

(10)

where PF is the power factor; true PF is the actual power of the system considering both the displacement and distortion components; P is the active power; Q is the reactive power; S is the apparent power; Vrms is the effective voltage; Irms is the effective current; DPF is the displacement power factor, which is related to the phase angle between voltage and current; and THD_I is a measure of the amount of harmonic distortion present in the current waveform. Table 1 summarizes the characteristics, sources, and economic impacts of major power quality disturbances. Representative waveforms for these disturbances are shown in Figure 3, which contrasts the normal and distorted profiles.

Table 1. Power quality disturbances.

Disturbance Type	Description	Characteristics	Primary Causes	Impacts
Voltage Sags [37]	Temporary drop in voltage (0.1–0.9 pu)	0.5 cycles to 1 min, defined by RMS	Faults, motor starts, transformer	Equipment shutdown
Voltage Swells [35]	Temporary rise in voltage (1.1–1.8 pu)	0.5 cycles to 1 min, same formula as sag	SLG faults, load rejection, capacitor energization	Overheating, insulation stress
Long-Duration Variations	Sustained voltage <0.9 or >1.1 pu for >1 min	Sustained, steady-state voltage deviation	Load variation, poor regulation	Appliance damage, flicker
Interruptions [28,29]	Loss of voltage (<0.1 pu)	Momentary, temporary, sustained	Faults, protection trips	Process downtime, data loss
Harmonics [49]	Integer multiples of fundamental frequency	THD formula used is voltage/current distortion	Non-linear loads, VFDs, chargers	Equipment overheating, waveform distortion
Interharmonics [49]	Non-integer frequency components	0–50/60 Hz, affects stability	Cycloconverters, arcing, converters	Flicker, signal interference
Notching [12]	Periodic voltage dips due to commutation	Notch depth, area, count	Power electronics	Waveform distortion
Noise [55]	High frequency (<200 kHz) superimposed	Common/transverse mode	Switching, EM interference	Communication errors
Voltage Fluctuations/Flicker [35,44]	Random voltage change (0.9–1.1 pu)	Pst (10 min), Plt (2 hr)	Arc furnaces, motors	Light flicker, stress
Voltage Unbalance [51]	Unequal voltage magnitude/angle	VUF = |V2|/|V1|	Asymmetrical loads, open delta	Motor torque pulsation
Current Unbalance [51]	Unequal phase currents	CUF = |C2|/|C1|	Even in balanced voltages	Excess heating, vibration
Frequency Deviations [7]	Deviation from nominal (50/60 Hz)	Normal: ±0.1 Hz, Disturbance: ±0.5 Hz	Load-generation mismatch	Instability, control errors
Oscillatory/Impulsive Transients [53]	Sudden voltage/current shifts	Rise < μs, duration < μs	Lightning, switching	Insulation damage
Power Factor/Reactive Power [54]	Low PF affects efficiency	PF = P/S, True PF includes THDI	Reactive loads, harmonics	Losses, poor voltage regulation

Figure 3. Representative waveforms of major power quality disturbances.

2.2. FACTS Devices, Classifications and Operating Principles

FACTS devices represent a sophisticated family of power electronic-based controllers, and their design is primarily intended to enhance the controllability, stability, and power transfer capabilities of AC transmission systems. This chapter focuses on various aspects of FACTS devices and their impact on increasing the functionality of the grid by mitigating disturbances [56].

$$\mathrm{P}={\displaystyle \frac{\left|{\mathrm{V}}_{\mathrm{s}}\right|\left|{\mathrm{V}}_{\mathrm{r}}\right|}{\mathrm{X}}}\mathrm{S}\mathrm{i}\mathrm{n}({\mathrm{\delta }}_{\mathrm{s}}-{\mathrm{\delta }}_{\mathrm{r}})$$

(11)

$${\mathrm{Q}}_{\mathrm{s}}={\displaystyle \frac{{\left|{\mathrm{V}}_{\mathrm{s}}\right|}^2}{\mathrm{X}}} {\displaystyle \frac{\left|{\mathrm{V}}_{\mathrm{s}}\right|\left|{\mathrm{V}}_{\mathrm{r}}\right|}{\mathrm{X}}}\mathrm{C}\mathrm{o}\mathrm{s}({\mathrm{\delta }}_{\mathrm{s}}-{\mathrm{\delta }}_{\mathrm{r}})$$

(12)

where P is the active power transmission, Q_s is the reactive power at the sending end, ∣V_s∣ and ∣V_r∣ are the voltage magnitudes at the sending and receiving ends, respectively, δ_s and δ_r are the voltage phase angles, and X is the line reactance.

FACTS devices provide dynamic control over the three main parameters.

• Line impedance (X)
• Voltage magnitude (∣V∣)
• Phase angle (δ)

This dynamic control is achieved by inserting reactive elements (such as inductors and capacitors) or direct voltage and current injection using power electronic converters that typically operate at high switching frequencies [42].

2.2.1. FACTS Device Classifications

Series-Connected Controllers

These FACTS devices are connected in series with transmission lines through coupling transformers to control the impedance or voltage sources [11]. With this configuration, direct control of the power flow (i.e., relatively smaller ratings compared to shunt devices) is possible according to the following relationship:

$$\mathrm{P}= {\displaystyle \frac{\left|{\mathrm{V}}_{\mathrm{s}}\right|\left|{\mathrm{V}}_{\mathrm{r}}\right|}{{\mathrm{X}}_{\mathrm{e}\mathrm{f}\mathrm{f}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}}}\mathrm{S}\mathrm{i}\mathrm{n}({\mathrm{\delta }}_{\mathrm{s}}-{\mathrm{\delta }}_{\mathrm{r}})$$

(13)

where X_effective is the effective line reactance after series compensation. Thyristor-controlled series reactors (TCSR) and thyristor-switched series reactors (TSSR) allow continuous adjustment of series inductive reactance using antiparallel thyristors, thereby providing dynamic power flow tuning [11]. A detailed comparative summary of FACTS categories, operating principles, and control objectives is provided in Appendix A.

2.2.2. Shunt-Connected FACTS Controllers

Shunt controllers [37] are connected in parallel with the system at specific points to have the most impact on the system (typical locations are at load buses or midpoints of long transmission lines), functioning as controllable current sources or variable impedances. It includes:

• Static Var Compensator (SVC): Comprised of thyristor-controlled reactors/capacitors, it provides fast voltage regulation and supports system stability under heavy or fluctuating loads [11].
• STATCOM (Static Synchronous Compensator): A voltage-source converter (VSC)-based shunt device offering rapid dynamic reactive compensation, harmonic filtering, flicker mitigation, and rapid support during faults [11]. The mathematical modeling details for these devices, including control equations and converter relationships, are summarized in Appendix A.1.

Practical PV-coupled deployments show that reduced-switch D-STATCOM topologies can maintain voltage and mitigate PQ issues while lowering switching burden, using modified SRF control in grid-tied operation [31].

2.2.3. Combined Series-Shunt Controllers

Combined devices integrate both series and shunt controls for multimodal power optimization.

• Unified Power Flow Controller (UPFC): Simultaneously manages voltage magnitude, impedance, and phase angle using dual VSCs sharing a common DC link; UPFCs represent the most flexible FACTS technology for power flow control and quality enhancement [11].
• Interline Power Flow Controller (IPFC): Coordinates the power flow across multiple transmission lines using several VSCs connected to a shared DC link, optimizing grid utilization, and compensating multiple lines collectively [11].

2.2.4. Operating Principles and PQ Enhancement

FACTS devices enhance real-time voltage profiles and stability of the system by requiring active control of the magnitude of voltages, impedance, and phase angle, minimizing transmission losses, and facilitating power transfer up to thermal constraints [11], alleviating the problem of power quality disturbances, including voltage sags, flickers, and harmonics, and offering adaptive load flow management to overcome the challenges of integrating renewable energy.

To make its description more complete and thorough, Table 2 summarizes, characterizes, and classifies each type of FACTS device according to Table 2 and provides a schematic diagram as a visualization to better understand them. An example is the STATCOMs, which offers voltage support as well as harmonic/flicker reduction and is therefore crucial in renewable and load-dominated grids [11]. As detailed in Table 3, various FACTS devices differ broadly in terms of their response time and harmonic filtering capacity.

Table 2. FACTS characteristics.

FACTS Device	Configuration Type	Operating Principle	Characteristics	Applications	Response Time	Limitations
TCSC [11,57,58]	Series (via coupling transformer)	Variable capacitive reactance using a fixed capacity or a thyristor-controlled reactor	Impedance varies by firing angle; operates in bypassed, blocked, inductive/capacitive vernier modes	Power flow control, oscillation damping, SSR mitigation, fault current limitation	1–2 cycles	Resonance risks, complex control, harmonic generation
SSSC [59,60]	Series (via coupling transformer)	VSC injects controllable voltage in series with line	Emulates variable reactance or phase angle shift	Power flow control, voltage stability, transient stability improvement	<1 cycle	High cost, advanced control, protection complexity
DVR [61,62]	Series (with load)	VSC with energy storage injects compensating voltage	Offers in-phase, pre-sag, and minimum energy compensation	Voltage sag/swell mitigation (up to 50%), harmonic filtering	10–25 ms	Limited by storage size, less effective for deep sags
SVC [63]	Shunt (at buses or midpoints)	Thyristor-switched/controlled reactors/capacitors (TCR/TSC)	Reactive power control via firing angle; slope-based voltage regulation	Voltage regulation (±3%), flicker mitigation, unbalance compensation	2–3 cycles	Voltage-dependent performance, harmonics, limited at low voltages
STATCOM [64,65]	Shunt (at critical buses)	VSC generates reactive power through voltage magnitude control	Voltage regulation, power factor correction, reactive reserve support	Harmonic filtering (up to 13th), flicker & unbalance compensation	<1 cycle	High cost, complex control, cooling needed
D-STATCOM [66]	Shunt (distribution systems)	STATCOM variant for distribution (1–10 MVA)	Higher frequency switching (2–5 kHz), enhanced voltage support	Flicker mitigation (e.g., arc furnaces), harmonic filtering (up to 25th)	<1 cycle	Lower rating vs. transmission-grade STATCOM
UPFC [67,68]	Combined Series–Shunt (common DC link)	Integrated STATCOM + SSSC for full power flow control	Controls impedance, voltage magnitude, and phase angle simultaneously	Real-time control of active/reactive power, damping, load balancing	<1 cycle	Highest cost and control complexity
UPQC [27,69]	Combined Series–Shunt (distribution-level)	Combines series & shunt active filters for voltage/current issues	Series filter mitigates voltage issues; shunt filter mitigates current issues	Voltage sag/swell mitigation (up to 50%), harmonic suppression (up to 25th), unbalance and flicker reduction	<1 cycle	Expensive, bulky for small enclosures, high complexity
IPFC [70,71]	Multiple Series Controllers (shared DC link)	Multiple SSSCs controlling flows across lines	Shares active power between lines, provides reactive compensation	Power redistribution, reactive support, system loss reduction	<1 cycle	Complex coordination, high cost, limited practical deployment

Table 3. FACTS connection types and their construction features.

Parameter	Shunt (Thyristor)	Shunt (VSC)	Series (Thyristor)	Series (VSC)	Combined (VSC)
Connection Type	Shunt	Shunt	Series	Series	Combined
Base Technology	Thyristor	VSC	Thyristor	VSC	VSC
Response Time	2–3 cycles	<1 cycle	1–2 cycles	<1 cycle	<1 cycle
Control Range	Limited by V2	Independent of V	Limited range	Independence of I	Comprehensive
Maximum Compensating Current	Proportional to V	Rated value	N/A	N/A	Rated value
Maximum Compensating Voltage	N/A	N/A	Limited to MOV	Rated value	Rated value
Reactive Power Capability	Power V2	Independent of V	N/A	N/A	Comprehensive
Active Power Control	No	With ESS only	Indirect	With ESS only	Direct
Overload Capability	30% for 1 s	1.5–2x for 1 s	30% for 1 s	1.5–2x for 1 s	1.5–2x for 1 s
Harmonic Generation	2–3% THD	<1% THD	2–3% THD	<1% THD	<1% THD
Harmonic Filtering Capability	Limited	Up to 25th	None	None	Up to 25th
Maximum Practical Rating	800 MVAr	400 MVAr	350 MVAr	200 MVAr	300 MVAr
Response to System Faults	Poor at low V	Excellent	Limited	Good	Excellent
Space Requirements	0.08–0.12 m2/MVAr	0.03–0.05 m2/MVAr	0.06–0.08 m2/MVAr	0.05–0.07 m2/MVAr	0.08–0.12 m2/MVAr
Losses at Rated Output	0.7–1.0%	1.2–1.8%	0.7–1.0%	1.4–2.0%	1.6–2.5%
Life Expectancy	30–40 years	25–30 years	30–40 years	25–30 years	25–30 years
Control Loop Bandwidth	5–10 Hz	100–300 Hz	3–5 Hz	100–300 Hz	100–300 Hz
Relative Cost ($/kVAr)	30–50	50–70	40–60	70–90	90–120
Acoustic Noise	60–80 dB	65–75 dB	55–75 dB	65–75 dB	65–80 dB
Cooling Requirements	Air/Oil	Air/Liquid	Air/Oil	Air/Liquid	Air/Liquid
Control Complexity	Medium	High	Medium	High	Very High
Operating Temperature Range	−40 to +50 °C	−35 to +45 °C	−40 to +50 °C	−35 to +45 °C	−35 to +45 °C
Altitude Derating (above 1000 m)	1%/100 m	1%/100 m	1%/100 m	1%/100 m	1%/100 m

3. Fundamental of Artificial Intelligence in Power Systems

Although FACTS devices can help diminish the impact of numerous PQ disruptions owing to real-time reactive power and voltage control, their capability is sometimes limited by the static control logic. The response times to various PQ phenomena are often rapid and unpredictable, particularly in renewable-rich systems, necessitating adaptive responses. This makes the field ripe for AI to improve the detection, classification, and control of these disturbances, so FACTS devices can become proactive rather than reactive.

AI [31] is now revolutionizing power system operations, control, and analysis, thereby improving grid resilience and reliability. The given discussion presents the key AI technologies that can be used in the power domain and further elaborates on their implementations with reference to Harmonic Distortion mitigation in power-grid applications.

3.1. Core AI Technologies for Power System Applications

AI comprises computational methods that mimic human decision-making. Several branches of AI have proven to be highly effective in power system applications.

Machine Learning (ML) [28,72] is the heart of AI-driven power system solutions. It aids the system decision-making process to enhance its performance on data rather than only on pre-defined rules. ML algorithms analyze historical operational data for pattern recognition and relationship analysis [73,74,75,76]. The learning process is as follows:

$$\overbrace{\mathrm{y}}=\mathrm{f}\left(\mathrm{x};\mathrm{\theta }\right)$$

(14)

Here, f maps the inputs x (e.g., system measurements) to the outputs (e.g., load forecasts or equipment status) with parameter optimization through iterative training. It is a key subset of ML with wide-ranging applications in power systems. An artificial neuron [77,78] computes

$$\mathrm{y}=\mathrm{\phi }(\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{i}}{\mathrm{x}}_{\mathrm{i}}+\mathrm{b})$$

(15)

where x_i is the inputs, w is the associate weight to each input, b is the bias, and $\mathrm{\phi }$ is the activation function.

In modern Deep Learning applications, it is useful to recognize two main architectures (feedforward networks to address a static prediction problem and recurrent networks to process temporal sequences) to address such intricate power problems. Another extension, hierarchical features, provides a more accurate and effective solution by combining information at multiple levels.

Such an approach to uncertainty management suggests the use of degrees of truth, rather than binary logic. Fuzzy Logic [79] instead of using the conventional grid classification assigns values of membership, thus enabling a more flexible classification of operating conditions and appropriate distribution of control responsibility. The reduction in instability that accompanied the resulting increase in operational efficiency was significant.

$$\mathrm{A}=\{(\mathrm{x},{\mathrm{\mu }}_{\mathrm{A}}\left(\mathrm{x}\right))|\mathrm{x}\in \mathrm{X},{\mathrm{\mu }}_{\mathrm{A}}\left(\mathrm{x}\right)\in \left[0.1\right]\}$$

(16)

where A represents the fuzzy test, x represents an element belonging to the universal set X, X is the domain where elements x are drawn from, and ${\mathrm{\mu }}_{\mathrm{A}}\left(\mathrm{x}\right))|$ is the membership function.

Nature-inspired algorithms and swarm-type intelligence are subsets of optimization techniques [80] that have been implemented to optimize power system performance. These techniques follow a repetitive process to improve the population of candidate solutions using selection and variation mechanisms to achieve optimal results. They are particularly effective when implemented for optimization problems with nonconvex, discontinuous, and mixed-integer search spaces, where traditional methods tend to provide suboptimal results.

$${\mathrm{P}}_{\mathrm{t}+1}=\mathrm{s}\mathrm{e}\mathrm{l}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\left(\mathrm{v}\mathrm{a}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\left({\mathrm{P}}_{\mathrm{t}}\right)\right)$$

(17)

3.2. AI Applications Across the Power System Value Chain

Artificial intelligence technologies will be used in the entire lifecycle, both in the creation and use stages. An overview of the AI methods applied for power-quality enhancement is presented in Table 4.

• Deep Learning Renewable Forecasting: Deep learning-based models combine historical weather data, measured irradiance, and power generation measurements to predict wind and solar generation for better grid and dispatch planning [81].
• Predictive Maintenance: AI models based on convolutional neural networks (CNNs) and recurrent neural networks (RNNs) use sensor compositions, such as vibration, temperature, and acoustic sensor readings, to identify early signs of equipment and component deterioration, thus facilitating cost-efficient maintenance plans [82].
• Dynamic Line Rating (DLR): Models of neural networks combine ambient and operating variables to forecast real-time transmission limits; therefore, the utilization of the grid is optimized, and grid congestion is avoidable [83].
• Distribution Operations: Graph neural networks (GNNs) can model the network topology, spatial and temporal dependencies in load prediction, outage detection, and non-technical loss detection [84,85].
• Microgrid and Adaptive Control: Reinforcement learning algorithms find control policies to use energy management in uncertain conditions, thus maximizing the utilization of a grid and reducing costs [86].

Table 4. AI applications in the power industry.

AI Application Area	AI Techniques Used	Key Benefits
Power Generation—Forecasting	Deep Learning (incorporating G: solar irradiance, T: temperature, C: cloud coverage)	Improved forecast accuracy, reduced reserve requirements, lower integration costs
Power Generation—Predictive Maintenance	Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs)	Early malfunction detection, reduced breakdowns, minimized unnecessary servicing
Transmission System—Dynamic Line Rating	AI models analyzing ambient temperature, wind speed, solar irradiation	Increased transmission capacity, reduced grid congestion, enhanced renewable integration
Distribution Management—Load Forecasting and Loss Detection	Graph Neural Networks (GNNs)	Improved forecasting, better network reconfiguration, non-technical loss reduction
Microgrid Management—Optimal Control Policies	Reinforcement Learning (RL)	Optimized control policies, cost efficiency, reliability under uncertainty

Table 4. AI applications in the power industry.

AI Application Area	AI Techniques Used	Key Benefits
Power Generation—Forecasting	Deep Learning (incorporating G: solar irradiance, T: temperature, C: cloud coverage)	Improved forecast accuracy, reduced reserve requirements, lower integration costs
Power Generation—Predictive Maintenance	Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs)	Early malfunction detection, reduced breakdowns, minimized unnecessary servicing
Transmission System—Dynamic Line Rating	AI models analyzing ambient temperature, wind speed, solar irradiation	Increased transmission capacity, reduced grid congestion, enhanced renewable integration
Distribution Management—Load Forecasting and Loss Detection	Graph Neural Networks (GNNs)	Improved forecasting, better network reconfiguration, non-technical loss reduction
Microgrid Management—Optimal Control Policies	Reinforcement Learning (RL)	Optimized control policies, cost efficiency, reliability under uncertainty

3.3. Advantages of AI in Power System Application

AI-driven approaches offer the following benefits.

• It has superior forecasting accuracy and reliability compared with traditional statistical methods.
• Adaptive modeling and control strategies continuously evolve as system patterns change, thereby reducing manual recalibration and operational overhead [81].
• Proactive fault detection, predictive disturbance mitigation, and dynamic optimization of conflicting objectives (quality, cost, and emissions) [87].
• Real-time optimization via gradient-based update rules:

$${\mathrm{\theta }}_{\mathrm{t}+1}={\mathrm{\theta }}_{\mathrm{t}}+\mathrm{\alpha }{\nabla }_{\mathrm{\theta }}\mathrm{L}({\mathrm{\theta }}_{\mathrm{t}})$$

(18)

where ${\mathrm{\theta }}_{\mathrm{t}}$ is the model parameters at time t, ${\nabla }_{\mathrm{\theta }}\mathrm{L}({\mathrm{\theta }}_{\mathrm{t}})$ is the gradient of the loss function, $\mathrm{\alpha }$ is the learning rate, and ${\mathrm{\theta }}_{\mathrm{t}+1}$ is the updated parameters for the next iteration.

3.4. Implementation Considerations and Challenges

The implementation of AI in power systems is affected by numerous obstacles that have been successfully addressed in previous studies. For uniformity, the specific PQ indices used for evaluating AI-based controllers are defined belowand in Appendix A.2.

• Data Quality and Integration: AI models require high-resolution, contextually relevant, and complete datasets. As mentioned in reference [88], the mitigation measures include advanced infrastructure for measurements, systematic data purification, and synthetic data generation.
• Model Interpretability: AI algorithms (many of which are opaque) compromise regulation and trust between operators. Explainable AI (XAI) techniques, including attention mechanisms, feature-importance analyses, and symbolic regression, are solutions to this issue and contribute to the understandability of the decision-making process.
• Computational Complexity and Latency: It centralizes AI solutions, high levels of latency and reliability issues can be expected. According to reference [89], edge AI platforms are used to compute on and around substations and control nodes, which is a significant improvement in the responsiveness and robustness of the system [90,91].

3.5. Regulatory and Organizational Challenges

In various scenarios, non-technical challenges pose greater barriers to AI adoption than technical ones. One major issue is that existing regulations are built around traditional rule-based control systems. AI-driven models that rely on probabilistic decision-making may not be easily accommodated in these frameworks. One approach to address this issue is to update policies and regulatory adjustments to ensure compliance and reliability of the data. Data fragmentation is a significant challenge. Various data-gathering organizations often manage their own isolated data sources. These entities may be reluctant to share critical information, and implementing AI across the entire system could be significantly complex and challenging. The lack of seamless data integration could hinder the full picture needed for optimal decision-making in AI models.

In addition to regulatory and data concerns, organizational resistance can slow AI adoption. Traditional operational frameworks have existed long enough for many power system professionals, who spend years working within them, and AI-driven decision-making system adoptability with new recruits is a challenging area. This requires new training, policy shifts, and changes in mindset. Resistance to change (because of a lack of trust in AI or concerns about job security) can cause an additional layer of difficulty in implementation [92].

3.6. Future Directions and Emerging Trends

With the growing adoption of AI in power system applications, various interceptional trends are expected to significantly influence future development. One of these is the hybrid AI methodology, which combines the strengths of different algorithms to improve the model’s performance and robustness. A notable example of this model is the use of Physics-Informed Neural Networks (PINNs) [93], which embeds physical power system equations directly into the neural network’s loss function. Utilizing this approach could enable the model to learn from both data and known system dynamics, while improving generalization and reducing the dependency on large datasets. The related function in the PINN framework is defined as

$$\mathcal{L}={\mathcal{L}}_{\mathrm{d}\mathrm{a}\mathrm{t}\mathrm{a}}+\mathrm{\lambda }{\mathcal{L}}_{\mathrm{p}\mathrm{h}\mathrm{y}\mathrm{s}\mathrm{i}\mathrm{c}\mathrm{s}}$$

(19)

where ${\mathcal{L}}_{\mathrm{d}\mathrm{a}\mathrm{t}\mathrm{a}}$ represents the error based on the training data, ${\mathcal{L}}_{\mathrm{p}\mathrm{h}\mathrm{y}\mathrm{s}\mathrm{i}\mathrm{c}\mathrm{s}}$ encodes physical constraints derived from power system equations, and $\mathrm{\lambda }$ is a weighting factor that balances the influence of both terms. This combination allows AI models to maintain their alignment with physical laws while capturing complex data-driven patterns.

Another important learning method is federated learning [94], a decentralized training strategy that allows multiple local agents to collaboratively train a global model without any raw data sharing stage. This is useful for power systems, where data privacy and ownership are critical concerns across utilities and regions of interest. This model can be computed using the following formula:

$${\mathrm{\theta }}_{\mathrm{g}\mathrm{l}\mathrm{o}\mathrm{b}\mathrm{a}\mathrm{l}}={\displaystyle \frac{1}{\mathrm{N}}}\sum _{\mathrm{i}=1}^{\mathrm{N}}{\mathrm{\theta }}_{\mathrm{l}\mathrm{o}\mathrm{c}\mathrm{a}\mathrm{l},\mathrm{i}}$$

(20)

where ${\mathrm{\theta }}_{\mathrm{l}\mathrm{o}\mathrm{c}\mathrm{a}\mathrm{l},\mathrm{i}}$ is the model trained on local dataset i, and N is the total number of participating agents. This approach provides users with sufficient assurance regarding data privacy, supports diverse operating conditions, and enables scalable model development without centralizing sensitive data.

In addition, digital twin technology [95] is evolving into a revolutionary concept in which AI-exploiting capabilities engage in real-time interactions with their physical counterparts. Such virtual copies can be used to perform large amounts of scenario testing, help in perfecting systems, and predict faults before they occur without risking real machinery. The mutual communication between physical and virtual systems results in the constant enhancement of both, making digital twins a critical instrument for smart grid advancement.

Finally, quantum computing [96] provides a futuristic method for addressing computational dilemmas in power system analysis. Quantum computing has the potential to solve optimization problems exponentially, which could enable a new level of capabilities in power grid management, planning, and control, particularly in situations in which classical algorithms cannot readily find a solution.

By expanding and developing these methods, AI’s role of AI will expand from supporting isolated system functions to orchestrating comprehensive interactions over different aspects of power systems (generation, distribution, and consumption). Table 5 provides a synthesized overview of state-of-the-art methods, including neural networks, support vector machines, fuzzy logic, deep reinforcement learning, and optimization metaheuristics, and their documented impacts on power quality disturbances, such as harmonics, voltage sags/swells, flicker, and imbalance.

Table 5. AI techniques for power quality improvements.

AI Technique	Application for Power Quality	Advantages	Limitations
Feedforward Neural Networks (FFNNs) [97]	PQ disturbance classification, FACTS optimization, load forecasting	Simple, effective for static patterns, low computation	Limited to time-based patterns, overfitting risk
Long Short-Term Memory (LSTM) [68]	Sag prediction, time-series PQ events, anticipatory FACTS control	Captures temporal dependencies, long-term memory, good for prediction	High complexity, long training, and needs historical data
Convolutional Neural Networks (CNN) [69]	PQ waveform feature extraction, pattern recognition	Good at spatial features, reduces dimensionality, less overfitting	Needs lots of data, computationally heavy
Fuzzy Logic Controllers [70]	FACTS control adjustment, adaptive DVR/D-STATCOM control	Handles imprecision, rule-based, robust	Needs expert rule design, tuning complexity
Particle Swarm Optimization (PSO) [15,98]	Sag detection, harmonic compensation, FACTS placement	Simple, good convergence, parallelizable	Premature convergence risk, sensitive to setup
Genetic Algorithms (GA) [72]	FACTS tuning, harmonic filters, PF correction	Global search, handles constraints	Slow, complex encoding, high computation
Teaching-Learning-Based Optimization (TLBO) [73]	FACTS fine-tuning, controller optimization	No parameter tuning, good convergence, simple	Needs many iterations for large problems
Gray Wolf Optimizer (GWO) [30,99]	FACTS placement, PQ enhancement, controller tuning	Balanced search, fewer parameters	Can stagnate in local optima
SARSA [32]	Adaptive FACTS control, real-time parameter tuning	Stable learning, learn from experience	Slow convergence, exploration parameter sensitive
Deep Q-Networks (DQNs) [100]	High-dimension FACTS control, adaptive PQ systems	Effective in large states, stable with experience replay	Heavy computation, complex, long training

4. AI Applications for Power Quality Enhancement, Monitoring, and Control

The incorporation of artificial intelligence (AI) into power quality (PQ) technologies has evolved the management practices in electrical distribution systems. Different power system issues, such as voltage sags, power factor correction, and real-time monitoring, have seen significant improvements owing to the utilization of AI-driven solutions. This section thoroughly explores this interconnected application, which offers the strength of AI methods to advance power quality management using detailed mathematical frameworks, comparative analysis, and sophisticated optimization strategies.

4.1. Advanced AI Technologies

4.1.1. Neural Network Architecture

Neural networks can be implemented in various PQ applications, where they leverage specialized structures that are suited to different scenarios.

Feedforward Neural Networks (FFNNs) are among the most fundamental architectures in deep learning, consisting of an input layer, one or more hidden layers, and an output layer [101]. In these networks, information travels in a single direction without feedback loops, rendering them suitable for tasks in which temporal dependencies are negligible. The output of each neuron is determined as follows.

$${\mathrm{y}}_{\mathrm{i}}=\mathrm{f}(\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{j}\mathrm{i}}{\mathrm{x}}_{\mathrm{i}}+{\mathrm{b}}_{\mathrm{j})}$$

(21)

where ${\mathrm{x}}_{\mathrm{i}}$ represents the input features, ${\mathrm{w}}_{\mathrm{j}\mathrm{i}}$ denotes the weight connecting input I to neuron j, ${\mathrm{b}}_{\mathrm{j}}$ is the bias term, and $\mathrm{f}$ is the activation function.

Long Short-Term Memory (LSTM) networks are a type of recurrent neural network (RNNs) with structures that specifically address long-term dependency issues. They are particularly useful for forecasting voltage drops and time disruptions in power networks [97]. LSTMs have a long-term cell state that develops over time and is regulated by forget, input, and output gates. The update rule for the cell state is provided by

$${\mathrm{c}}_{\mathrm{t}}={\mathrm{f}}_{\mathrm{t}}\odot {\mathrm{c}}_{\mathrm{t}-1}+{\mathrm{i}}_{\mathrm{t}}\odot {\tilde{\mathrm{c}}}_{\mathrm{t}}$$

(22)

where c_t is the cell state at time, ft is the forget gate’s output, it is the input gate’s output, and ${\tilde{\mathrm{c}}}_{\mathrm{t}}$ is the candidate cell state.

Convolutional Neural Networks (CNNs) are commonly used to extract spatial features from waveform (or image-based) data. CNNs are especially useful for disturbance classification in power quality applications for identifying patterns in voltage and current signals [98]. The convolution operation of each filter is expressed as

$${\mathrm{h}}_{\mathrm{i},\mathrm{j}}^{\mathrm{k}}=\mathrm{f}(\sum _{\mathrm{m}-0}^{\mathrm{M}-1}\sum _{\mathrm{n}=0}^{\mathrm{N}-1}{\mathrm{w}}_{\mathrm{m},\mathrm{n}}^{\mathrm{k}}{\mathrm{x}}_{\mathrm{i}+\mathrm{m},\mathrm{j}+\mathrm{n}}+{\mathrm{b}}^{\mathrm{k}})$$

(23)

where ${\mathrm{h}}_{\mathrm{i},\mathrm{j}}^{\mathrm{k}}$ is the output of the k-th convolutional filter, ${\mathrm{w}}_{\mathrm{m},\mathrm{n}}^{\mathrm{k}}$ represents filter weights, and ${\mathrm{x}}_{\mathrm{i}+\mathrm{m},\mathrm{j}+\mathrm{n}}$ denotes input values within the filter window. The activation function is $\mathrm{f}$ applied element-wise to the resulting sum.

For event recognition, hybrid wavelet–CNN pipelines have achieved accurate decomposition and classification of composite PQ disturbances, supporting near-real-time monitoring [102].

4.1.2. Fuzzy Logic Systems

Fuzzy logic controllers [11,79] are widely used for PQ applications owing to their ability to handle uncertainty and imprecise information. These controllers provide robust decision-making capabilities by mimicking human decision-making behavior and linguistic rules rather than relying solely on mathematical models. The fuzzy control process can be divided into four stages. The first is fuzzification, which involves transforming the numerical values of inputs into fuzzy sets with specific membership functions. Second, the input conditions were interpreted by applying a set of predetermined rules and fuzzy logic operators in the rule evaluation step. The aggregation step then integrates the outputs of each applicable rule into a single fuzzy response. Finally, defuzzification converts the combined fuzzy output to a crisp, implementable control signal that can be used within the system.

In PQ, especially for FACTS devices, fuzzy logic is instrumental in dynamically adjusting the control parameters. This method typically processes inputs, such as errors and changes in errors, to determine an appropriate control action. The mathematical representation of this procedure is as follows.

$${\mathrm{\mu }}_{{\mathrm{A}}_{\mathrm{i}}}(\mathrm{e})\mathrm{\Lambda }{\mathrm{\mu }}_{{\mathrm{B}}_{\mathrm{i}}}(∆\mathrm{e})\to {\mathrm{\mu }}_{{\mathrm{c}}_{\mathrm{i}}(\mathrm{u})}$$

(24)

where ${\mathrm{\mu }}_{\mathrm{A}}$, ${\mathrm{\mu }}_{\mathrm{B}}$, ${\mathrm{\mu }}_{\mathrm{C}}$ are the membership functions representing the linguistic variables for the system inputs and outputs, e is the error, $∆\mathrm{e}$ is the change in error, and $\mathrm{u}$ is the appropriate control action.

These rules help optimize the performance of PQ devices, such as DVRs and DSTATCOMS, by ensuring real-time adaptive control.

4.1.3. Evolutionary and Swarm Intelligence Algorithms

Evolutionary computing and swarm intelligence algorithms have been widely applied to optimize and enhance the efficiency of PQ in control systems. By emulating biological processes and natural behaviors, these methods attempt to solve complex optimization problems efficiently.

PSO [98] is a population-based optimization technique inspired by the behaviors of birds and fish. This method attempts to refine the potential solutions using the particle velocity updates and positions according to the following mathematical equations:

Fuzzy logic controllers provide robust performance in power-quality applications by managing uncertainty and imprecision [103]. The basic fuzzy inference process involves the following:

$${\mathrm{v}}_{\mathrm{i}}^{\mathrm{k}+1}=\mathrm{w}.{\mathrm{v}}_{\mathrm{i}}^{\mathrm{k}}+{\mathrm{c}}_1.{\mathrm{r}}_{\mathrm{a}}.\left({\mathrm{p}\mathrm{b}\mathrm{e}\mathrm{s}\mathrm{t}}_{\mathrm{i}}-{\mathrm{x}}_{\mathrm{i}}^{\mathrm{k}}\right)+{\mathrm{c}}_2.{\mathrm{r}}_2.(\mathrm{g}\mathrm{b}\mathrm{e}\mathrm{s}\mathrm{t}-{\mathrm{x}}_{\mathrm{i}}^{\mathrm{k}})$$

(25)

$${\mathrm{x}}_{\mathrm{i}}^{\mathrm{k}+1}={\mathrm{x}}_{\mathrm{i}}^{\mathrm{k}}+{\mathrm{v}}_{\mathrm{i}}^{\mathrm{k}+1}$$

(26)

where w is the inertia weight, c₁ and c₂ are acceleration coefficients, r₁ and r₂ are random values ensuring stochastic behavior, pbest_i represents the personal best position of each particle, and gbest is the best global position found by the swarm.

This method has been widely used in PQ applications, such as voltage sag detection, harmonic compensation, and optimal placement of FACTS devices.

The genetic Algorithm (GA) provides robust optimization using the principles of natural selection, crossover, and mutation. By evolving candidate solutions to achieve optimal system performance, they can be used in power system quality control for various scenarios. Proportional–integral (PI) tuning in FACTS devices is a common application of this method. The goal of this optimization was to minimize the following fitness function:

$$\mathrm{f}={\int }_0^{\mathrm{T}}\left({\mathrm{w}}_1\left|\mathrm{e}\left(\mathrm{t}\right)\right|+{\mathrm{w}}_2{\mathrm{u}}^2\left(\mathrm{t}\right)\right)\mathrm{d}\mathrm{t}$$

(27)

where e(t) is the error signal between the desired and the actual value, u(t) is the control effort that is applied to the system, w₁ and w₂ are weighting accuracy for implementing the balance among accuracy and control energy.

With the optimization of control parameters, GA ensures improvement in transient response, reduces overshoot, and enhances stability in PQ devices. Combining this method with other AI techniques enables intelligent and adaptive power quality management, which significantly improves grid reliability [15].

4.2. AI Techniques for Voltage Sag Prediction and Mitigation

Voltage sag is a crucial issue in power systems that must be addressed. Its prediction and mitigation are vital for power quality management, where AI-driven techniques provide advanced predictive analytics and control mechanisms. Using these methods would enhance grid stability and reliability by enabling proactive responses to voltage disturbances. The representative test systems and simulation platforms referenced in this review are summarized below.

Predictive Analytics for Voltage Sag

In modern voltage sag prediction frameworks, multiple AI modules collaborate to perform specialized tasks, such as feature extraction, classification, and time-series forecasting [32].

For comprehensive voltage signal analysis, wavelet transform techniques are widely utilized [100] for feature extraction and selection; by breaking them down into multi-resolution components, they can effectively capture transient events. The mathematical formula is as follows:

$${\mathrm{W}}_{\mathrm{\psi }}\mathrm{f}\left(\mathrm{a},\mathrm{b}\right)={\displaystyle \frac{1}{\sqrt{\mathrm{a}}}}{\int }_{-\mathrm{\infty }}^{+\mathrm{\infty }}\mathrm{f}\left(\mathrm{t}\right){\mathrm{\psi }}^{\ast }\left({\displaystyle \frac{\mathrm{t}-\mathrm{b}}{\mathrm{a}}}\right)\mathrm{d}\mathrm{t}$$

(28)

where a is the scale parameter for the frequency resolution control, b is the translation parameter shifting the function in time, and $\mathrm{\psi }$ is the mother wavelet function.

Support Vector Machines (SVMs) with Gaussian kernel functions are commonly used for power quality disturbance classification, as they provide high accuracy and robustness [104]. The kernel function is as follows.

$$\mathrm{K}\left({\mathrm{x}}_{\mathrm{i}},{\mathrm{x}}_{\mathrm{j}}\right)=\mathrm{e}\mathrm{x}\mathrm{p}(-\mathrm{\gamma }{\left|\left|{\mathrm{x}}_{\mathrm{i}}-{\mathrm{x}}_{\mathrm{j}}\right|\right|}^2)$$

(29)

The decision function is given by

$$\mathrm{f}\left(\mathrm{x}\right)=\mathrm{s}\mathrm{i}\mathrm{g}\mathrm{n}\left(\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{a}}_{\mathrm{i}}{\mathrm{y}}_{\mathrm{i}}\mathrm{K}({\mathrm{x}}_{\mathrm{i}},\mathrm{x})\right)+\mathrm{b}$$

(30)

where x_i, x_j are feature vectors representing different dataset points, ${\left|\left|{\mathrm{x}}_{\mathrm{i}}-{\mathrm{x}}_{\mathrm{j}}\right|\right|}^2$ is the squared Euclidean distance between the feature vectors, $\mathrm{\gamma }$ is a hyperparameter that controls the influence of a single training, ${\mathrm{a}}_{\mathrm{i}}$ is the LaGrange multiplier, ${\mathrm{y}}_{\mathrm{i}}$ is the class label for each training sample, $\mathrm{K}({\mathrm{x}}_{\mathrm{i}},\mathrm{x})$ is the kernel function output, and $\mathrm{b}$ is the bias term that helps to define the decision boundary in the feature space.

The ability to identify and detect temporal patterns that are reflective in historical voltage measurements is a chief benefit of using Long Short-term Memory (LSTM) networks in the prediction of phenomena involving voltages [105]. Mathematically, the temporal change in the internal state of the cell in the LSTM unit is formalized as follows:

In this formulation:

$${\mathrm{h}}_{\mathrm{t}}=\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}({\mathrm{W}}_{\mathrm{c}}\left[{\mathrm{h}}_{\mathrm{t}-1},{\mathrm{x}}_{\mathrm{t}}\right]+{\mathrm{b}}_{\mathrm{c}})\odot {\mathrm{o}}_{\mathrm{t}}$$

(31)

where h_t represents the latent state of the network at the specific time step t, defining the temporal discovered features; W_c is a matrix containing information on successions between two successive states; x_t is the value of the input data at step t, and here is the voltage signal being considered; and o_t indicates the opening of the output result, which decides the number of contributions of the changed cell condition to the concealer representation.

4.3. Advanced FACTS Devices with AI Control

FACTS devices equipped with AI-based controllers are capable of real-time adaptive solutions for mitigating voltage sags using adjustable reactive power [106]. AI can enhance the precision and efficiency of these controllers, ensuring a rapid response to any grid disturbances of varying severity.

4.3.1. Static Synchronous Compensator (STATCOM) with AI Control

The operation of AI integration to STATCOM [19] can be formulated as

$${\displaystyle \frac{\mathrm{d}{\mathrm{V}}_{\mathrm{P}\mathrm{C}\mathrm{C}}}{\mathrm{d}\mathrm{t}}}={\displaystyle \frac{1}{\mathrm{C}}}({\mathrm{i}}_{\mathrm{L}}-{\mathrm{i}}_{\mathrm{S}\mathrm{T}\mathrm{A}\mathrm{T}\mathrm{C}\mathrm{O}\mathrm{M}})$$

(32)

$${\mathrm{i}}_{\mathrm{S}\mathrm{T}\mathrm{A}\mathrm{T}\mathrm{C}\mathrm{O}\mathrm{M}}={\mathrm{I}}_{\mathrm{m}\mathrm{a}\mathrm{g}}\mathrm{s}\mathrm{i}\mathrm{n}(\mathrm{\omega }\mathrm{t}+\mathrm{\phi })$$

(33)

where V_PCC voltage at the Point of Common Coupling (PCC), C is the capacitance of the DC-link capacitor, i_L is the current load drawn from power system, i_STATCOM is currently injected by the STATCOM to compensate for voltage sags and maintain the grid stability, I_mag is the magnitude of the compensating current, is adjusted by the AI model, w_t is the grid regular frequency term, $\mathrm{\phi }$ is phase angle that is optimized by a neural network controller. At distribution level, ANN-trained D-STATCOM controllers have been shown to enhance voltage regulation, power factor, and harmonic mitigation under time-varying loads [107].

4.3.2. Dynamic Voltage Restorer (DVR) with AI-Enhanced Sliding Mode Control

DVRs utilize advanced Sliding Mode Control (SMC) with AI-optimized procedures to compensate for voltage sags [108,109]. The sliding surfaces for the direct and quadrature-axis error compensation are defined by the following mathematical equations:

$${\mathrm{s}}_{\mathrm{d}}={\mathrm{e}}_{\mathrm{d}}+{\mathrm{k}}_{\mathrm{d}}\int {\mathrm{e}}_{\mathrm{d}}\mathrm{d}\mathrm{t}-{\mathrm{e}}_{\mathrm{d}}(0)$$

(34)

$${\mathrm{s}}_{\mathrm{q}}={\mathrm{e}}_{\mathrm{q}}+{\mathrm{k}}_{\mathrm{q}}\int {\mathrm{e}}_{\mathrm{q}}\mathrm{d}\mathrm{t}-{\mathrm{e}}_{\mathrm{q}}(0)$$

(35)

where s_d and s_q are sliding surface variables in the direct (d) and quadrature (q) reference frames, respectively; ed and e_q are voltage errors in the d-q frame; and k_d and k_q are control parameters tuned by the AI-based optimizer.

In the AI-enhanced novel reaching law, the rapid convergence of the sliding mode control is as follows:

$${\mathrm{s}}_{\mathrm{d}}=-{\mathrm{\epsilon }}_{\mathrm{d}}{\left|{\mathrm{s}}_{\mathrm{d}}\right|}^{{\propto }_{\mathrm{d}}}\mathrm{s}\mathrm{g}\mathrm{n}\left(\left|{\mathrm{s}}_{\mathrm{d}}\right|-1\right)\mathrm{s}\mathrm{g}\mathrm{n}\left({\mathrm{s}}_{\mathrm{d}}\right)-{\mathrm{k}}_{\mathrm{d}}{\mathrm{s}}_{\mathrm{d}}{\left|{\mathrm{s}}_{\mathrm{d}}\right|}^{{\propto }_{\mathrm{d}}}\mathrm{s}\mathrm{g}\mathrm{n}(\left|{\mathrm{s}}_{\mathrm{d}}\right|-1)$$

(36)

where ${\mathrm{\epsilon }}_{\mathrm{d}}$ is the adaptation coefficient controlling the rate of system response, ${\propto }_{\mathrm{d}}$ exponent regulating the nonlinear behavior of the reaching law, $\mathrm{s}\mathrm{g}\mathrm{n}({\mathrm{s}}_{\mathrm{d}})$ is the sing function ensuring the correct switching behavior.

4.3.3. Unified Power Quality Conditioner (UPQC) with AI Algorithm

The UPQC is a multifunctional power quality device that uses both series and parallel compensators for the simultaneous mitigation of voltage sags, harmonics, and power imbalances. AI-based controllers can optimize their performance using a series Active Power Filter by adjusting the reference compensation voltage [19,110]:

$${\mathrm{V}}_{\mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}}^{\ast }={\mathrm{V}}_{\mathrm{r}\mathrm{e}\mathrm{f}}-{\mathrm{V}}_{\mathrm{s}}$$

(37)

where ${\mathrm{V}}_{\mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}}^{\ast }$ is the reference compensation voltage injected by the series APF, ${\mathrm{V}}_{\mathrm{r}\mathrm{e}\mathrm{f}}$ is the desired load voltage after compensation, ${\mathrm{V}}_{\mathrm{s}}$ is the measured distorted source voltage before compensation.

4.4. AI-Driven Dynamic Power Factor Correction and Reactive Power Compensation

AI-based power factor correction [106,111] plays a major role in enhancing power quality with dynamic reactive power compensation. Neural network-driven approaches offer precise real-time adjustments to ensure system efficiency and stability improvement.

4.4.1. Neural Network-Controlled Shunt Active Power Filters

Shunt active power filters (SAPFs) equipped with neural network control are one of the integrations of AI and FACTS. In this system, harmonic currents compensate to improve the power factor and mitigate distortions. The fundamental compensation principle is expressed as

$${\mathrm{i}}_{\mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}}={\mathrm{i}}_{\mathrm{L}}-{\mathrm{i}}_{\mathrm{s}}^{\ast }$$

(38)

where ${\mathrm{i}}_{\mathrm{L}}$ represents the load current ${\mathrm{i}}_{\mathrm{s}}^{\ast }$ is the desired source current.

The neural network can determine the desired source current based on the fundamental power calculations:

$$\mathrm{P}={\mathrm{V}}_{\mathrm{s}}.{\mathrm{I}}_{\mathrm{s}}.\mathrm{c}\mathrm{o}\mathrm{s}\mathrm{\phi }$$

(39)

$$\mathrm{Q}={\mathrm{V}}_{\mathrm{s}}.{\mathrm{I}}_{\mathrm{s}}.\mathrm{s}\mathrm{i}\mathrm{n}\mathrm{\phi }$$

(40)

By continuously adjusting the compensation current, the neural network controls the P and Q, which are the active and reactive powers, respectively, to maintain the unity power factor of the grid.

4.4.2. Adaptive Reactive Power Compensation Algorithms

For advanced reactive power compensation techniques [104,112] that incorporate instantaneous power theory enhanced using neural-network-based optimization, the active and reactive power can be mathematically quantified as follows:

$$\left[\begin{array}{c}\mathrm{p}\\ \mathrm{q}\end{array}\right]=\left[\begin{array}{cc}{\mathrm{v}}_{\mathrm{a}}& {\mathrm{v}}_{\mathrm{\beta }}\\ {\mathrm{v}}_{\mathrm{\beta }}& -{\mathrm{v}}_{\mathrm{a}}\end{array}\right]\left[\begin{array}{c}{\mathrm{i}}_{\mathrm{a}}\\ {\mathrm{i}}_{\mathrm{\beta }}\end{array}\right]$$

(41)

where p represents the instantaneous active power, q represents the instantaneous reactive power, ${\mathrm{v}}_{\mathrm{a}}$ and ${\mathrm{v}}_{\mathrm{\beta }}$ are the transformed voltage components, ${\mathrm{i}}_{\mathrm{a}}$ and ${\mathrm{i}}_{\mathrm{\beta }}$ are the transformed current components.

The neural network plays a vital role in optimizing the extraction of oscillating components (q and q) to enable selective harmonic compensation while preserving the fundamental reactive power component required for unity power factor correction.

The reference current required for compensation is computed as

$$\left[\begin{array}{c}{\mathrm{i}}_{\mathrm{a} \mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}}\\ {\mathrm{i}}_{\mathrm{\beta } \mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}}\end{array}\right]={\displaystyle \frac{1}{{{\mathrm{v}}_{\mathrm{a}}}^2+{{\mathrm{v}}_{\mathrm{\beta }}}^2}}\left[\begin{array}{cc}{\mathrm{v}}_{\mathrm{a}}& {\mathrm{v}}_{\mathrm{\beta }}\\ {\mathrm{v}}_{\mathrm{\beta }}& -{\mathrm{v}}_{\mathrm{a}}\end{array}\right]\left[\begin{array}{c}\mathrm{p}\\ \mathrm{q}\end{array}\right]$$

(42)

Utilizing this approach enables precise reactive power compensation while ensuring a unity power factor and minimizing harmonic distortion.

4.4.3. Distribution Static Compensators with Intelligent Control

A D-STATCOM with the aid of AI-based control strategies could demonstrate enhanced dynamic performance for power factor correction and reactive power compensation. AI-driven approaches can adapt to system conditions in real time, thereby improving response time and stability [107]. The compensating reactive power in the D-STATCOM system is given by

$${\mathrm{Q}}_{\mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}}={\mathrm{V}}_{\mathrm{s}}.{\mathrm{I}}_{\mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}}\mathrm{s}\mathrm{i}\mathrm{n}{\mathrm{\phi }}_{\mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}}$$

(43)

where ${\mathrm{Q}}_{\mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}}$ is the compensating reactive power, ${\mathrm{V}}_{\mathrm{s}}$ is the system voltage, ${\mathrm{I}}_{\mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}}$ is the compensatory current, ${\mathrm{\phi }}_{\mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}}$ is the compensating phase angle.

Based on this equation, the D-STATCOM can rapidly adjust its angle to maintain the power factor close to unity to ensure efficient power delivery and reduced losses.

The Adaptive Neural Fuzzy Inference System (ANFIS) integrates fuzzy logic and Artificial Neural Networks (ANNs) for D-STATCOM improvement control. With this hybrid approach, the adaptability, learning capability, and robustness in handling nonlinear system behavior can be improved. The related equation can be quantified as follows:

$$\mathrm{y}={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{i}}{\mathrm{f}}_{\mathrm{i}}}{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{i}}}}$$

(44)

where y is the ANFIS output, ${\mathrm{w}}_{\mathrm{i}}$ is the rule firing strength, ${\mathrm{f}}_{\mathrm{i}}$ is the output of individual fuzzy rules, and n is the total number of fuzzy rules.

Faster repose and better transient performance compared to traditional controllers can be considered as one of the benefits of this AI-driven controller.

4.5. Real-Time Power Quality Monitoring and Distributed Control

Rather than AI, the integration of other technologies, such as the Internet of Things (IoT) and edge computing, has revolutionized real-time power quality monitoring and control, allowing the system to lead to a more efficient and intelligent decision-making form of operation [100].

4.5.1. AI-Enhanced Disturbance Classification

Multistage AI structures have gained popularity in recent power quality monitoring systems as a means of realizing highly accurate classification of disturbances [32]. Our process starts with signal pre-processing, where we use techniques such as wavelet packet decomposition to extract the most critical features of the voltage waveforms. The next step is the feature selection phase, which uses a metric such as the Information Gain Ratio (IGR) to determine the most pertinent attributes to be used in the classification. Finally, the classification stage uses ensemble learning, in which the outputs of multiple models are integrated to improve the overall accuracy and robustness in detecting different disturbances of power quality.

The mathematical formulation for the IGR in features is as follows:

$$\mathrm{I}\mathrm{G}\mathrm{R}\left(\mathrm{A}\right)={\displaystyle \frac{\mathrm{I}\mathrm{G}(\mathrm{A})}{\mathrm{S}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{t}\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{o}(\mathrm{A})}}$$

(45)

where IG(A) represents the information gain for attribute A and SplitInfo(A) measures the entropy of attribute A.

4.5.2. Disturbance Edge Computing Architecture

The utilization of edge computing, localized AI processing, and the distribution of computational tasks across multiple nodes enables [15]. This technology is important because it runs a lightweight neural network.

$${\mathrm{h}}_{\mathrm{i}}={\mathrm{f}}_{\mathrm{i}}\left({\mathrm{W}}_{\mathrm{i}}{\mathrm{x}}_{\mathrm{i}}+{\mathrm{b}}_{\mathrm{i}}\right)$$

(46)

where ${\mathrm{b}}_{\mathrm{i}}$ is the output at the ith edge node, ${\mathrm{f}}_{\mathrm{i}}$ is the activation function, and ${\mathrm{W}}_{\mathrm{i}}$ with ${\mathrm{x}}_{\mathrm{i}}$ represent the weight matrix and input vector, respectively.

4.5.3. IOT-Integrated Monitoring System

In the framework of IoT-based monitoring [98], a top-down data processing strategy is used to guarantee efficient and scalable system operation. The sensor layer at the bottom measures high-resolution voltage and current waveforms in real time. This data is then sent to the edge processing layer, where the model performs a localized analysis with minimal latency. In addition, the fog computing layer integrates the data of several edge nodes and runs regional-level analytics to support intermediate decision-making. Finally, the cloud computing layer performs system-wide analysis and enables deeper insights and analysis, long-term trends, and global optimization of the entire grid infrastructure.

4.6. Optimization Strategies for Power Quality Systems

4.6.1. Multi-Objective Optimization for FACTS Device Placement

The suitable placement of FACTS devices employs multi-objective optimization to balance conflicting objectives [79]:

$$\min \mathrm{F}\left(\mathrm{x}\right)={|{\mathrm{f}}_1\left(\mathrm{x}\right),{\mathrm{f}}_2\left(\mathrm{x}\right),\dots ,{\mathrm{f}}_{\mathrm{m}}\left(\mathrm{x}\right)|}^{\mathrm{T}}$$

(47)

$${\mathrm{g}}_{\mathrm{i}}\left(\mathrm{x}\right)\le 0,\mathrm{i}=\mathrm{1,2},\dots ,\mathrm{p}$$

$${\mathrm{h}}_{\mathrm{j}}\left(\mathrm{x}\right)=0,\mathrm{j}=\mathrm{1,2},\dots ,\mathrm{q}$$

The common objective functions include:

• f₁(x) is the voltage sag mitigation
• f₂(x) is the installation and operational costs
• f₃(x) is the power system losses

4.6.2. Metaheuristic Algorithms for Controller Parameter Optimization

Metaheuristic optimization techniques fine-tune the controller parameters in PQ systems for enhanced performance and reliability. Metaheuristics are also effective at wide-area damping controller tuning for FACTS; a Grey Wolf Optimizer demonstrated robust channel selection and controller design for oscillation damping. Teaching-Learning-Based Optimization (TLBO) is an approach in which the teacher and student are used for the TLBD algorithm to model the learning process [7]:

$${\mathrm{X}}_{\mathrm{i}}^{\mathrm{n}\mathrm{e}\mathrm{w}}={\mathrm{X}}_{\mathrm{i}}+{\mathrm{r}}_{\mathrm{i}}({\mathrm{X}}_{\mathrm{t}\mathrm{e}\mathrm{a}\mathrm{c}\mathrm{h}\mathrm{e}\mathrm{r}}-{\mathrm{T}}_{\mathrm{F}}{\mathrm{X}}_{\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}})$$

(48)

where ${\mathrm{X}}_{\mathrm{i}}$ represents the ith candidate solution, ${\mathrm{X}}_{\mathrm{t}\mathrm{e}\mathrm{a}\mathrm{c}\mathrm{h}\mathrm{e}\mathrm{r}}$ is the best-performing solution, ${\mathrm{X}}_{\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}}$ is the population mean, ${\mathrm{r}}_{\mathrm{i}}$ is a randomly generated number, and ${\mathrm{T}}_{\mathrm{F}}$ is the teaching factor.

Gray Wolf Optimizer (GWO) algorithm which is inspired by the social hierarchy of gray wolves, the GWO algorithm upgrades its position using:

$$\overrightarrow{\mathrm{X}}\left(\mathrm{t}+1\right)={\overrightarrow{\mathrm{X}}}_{\mathrm{p}}\left(\mathrm{t}\right)-\overrightarrow{\mathrm{A}}.\overrightarrow{\mathrm{D}}$$

(49)

$$\overrightarrow{\mathrm{D}}={|\overrightarrow{\mathrm{C}}.\overrightarrow{\mathrm{X}}}_{\mathrm{p}}\left(\mathrm{t}\right)-\overrightarrow{\mathrm{X}}(\mathrm{t})|$$

(50)

where $\overrightarrow{\mathrm{X}}(\mathrm{t})$ is the current position of a gray wolf in the research space, ${\overrightarrow{\mathrm{X}}}_{\mathrm{p}}\left(\mathrm{t}\right)$ is the position of the prey, or the current best solution, $\overrightarrow{\mathrm{D}}={|\overrightarrow{\mathrm{C}}.\overrightarrow{\mathrm{X}}}_{\mathrm{p}}\left(\mathrm{t}\right)-\overrightarrow{\mathrm{X}}(\mathrm{t})|$ represents the distance vector between the prey and current wolf, and $\overrightarrow{\mathrm{C}},\overrightarrow{\mathrm{A}}$ are the vector coefficient.

4.6.3. Deep Reinforcement Learning for Adaptive Control

Deep Learning (DRL) could enable the adaptive control that allows an agent to learn optimal control strategies through interactions with its environment [87]. This process involves selecting actions based on the states, receiving rewards, and adjusting actions by considering the experience of the agent. The State-Action-Reward-State-Action (SARSA) [113] algorithm updates the Q values each time through the following equation:

$$\mathrm{Q}\left({\mathrm{s}}_{\mathrm{t}},{\mathrm{\alpha }}_{\mathrm{t}}\right)\leftarrow \mathrm{Q}\left({\mathrm{s}}_{\mathrm{t}},{\mathrm{\alpha }}_{\mathrm{t}}\right)+\mathrm{\alpha }[{\mathrm{r}}_{\mathrm{t}}+\mathrm{\gamma }\mathrm{Q}\left({\mathrm{s}}_{\mathrm{t}+1},{\mathrm{\alpha }}_{\mathrm{t}+1}\right)-\mathrm{Q}\left({\mathrm{s}}_{\mathrm{t}},{\mathrm{\alpha }}_{\mathrm{t}}\right)]$$

(51)

where $\mathrm{Q}\left({\mathrm{s}}_{\mathrm{t}},{\mathrm{\alpha }}_{\mathrm{t}}\right)$ is the current estimate of the Q-value for state ${\mathrm{s}}_{\mathrm{t}},\mathrm{\alpha }$ is the learning rate, ${\mathrm{r}}_{\mathrm{t}}$ is the reward received after acting, $\mathrm{\gamma }$ is the discount factor, and ${\mathrm{s}}_{\mathrm{t}+1},{\mathrm{\alpha }}_{\mathrm{t}+1}$ are the next state and action.

The utilization of this algorithm helps the controller adaptively improve and take actions based on real-time outcomes. For more advanced scenarios, Deep Q-Networks (DQN) [99] are used to approximate the Q-values using a deep neural network, particularly when the state or action space has a large number of numbers to be handled with traditional methods:

$$\mathrm{Q}(\mathrm{s},\mathrm{a};\mathrm{\theta })\approx {\mathrm{Q}}^{\ast }(\mathrm{s},\mathrm{a})$$

(52)

where $\mathrm{\theta }$ contains various parameters (e.g., weights and biases) of the neural network and ${\mathrm{Q}}^{\ast }(\mathrm{s},\mathrm{a})$ is the optimal Q-value that network aims to approximate.

This method uses experience (the technique called experience replay), which stores and randomly samples past interactions to improve learning stability and efficiency.

For the model effectiveness evaluation [102], multiple performance indicators were employed across three core functionalities. Some key performance indicators of voltage sag compensation systems are numerous. The response time is the time lag between the detection of a sag and the commencement of the mitigating action. To determine the quality of the waveform after compensation, the Voltage Total Harmonic Distortion (THD) was measured to indicate the capability of the system to provide a less distorted waveform. The voltage recovery profile measures how well the system can recover the voltage to the nominal value after a disturbance occurs. In addition, energy usage is provided during the compensation assessment, which is relevant to systems that use energy-storing elements.

In the power factor correction (PFC) context, performance is determined by the magnitude of power factor correction achieved, with a view to ensuring that the power factor is as close to unity as practical. The current THD measures the quality of the present waveform following compensation, and the reactive power compensation accuracy measures the accuracy with which this system delivers the required reactive power. The transient response characteristics describe how the system reacts to sudden changes in the load or voltage. In addition, monitoring system features are important and are evaluated based on detection accuracy, classification precision, false positive/negative rates, and the computational efficiency of algorithms used to recognize and classify power quality disturbances. A consolidated mapping of AI algorithms to specific FACTS control objectives is presented in Table 5.

4.7. Real-World Applications of AI Technology in Power Quality Management

Neural networks, reinforcement learning, fuzzy logic, and metaheuristic optimization are artificial intelligence technologies that have gone beyond laboratory and theoretical experiments to produce real-world, substantive power system applications. Their performance capabilities in terms of scale-based data analysis, dynamic behavior under scenario changes, and addressing complex control issues make them especially localized to modern electrical grids that must deal with issues arising from the advent of renewable sources and load variability.

Advanced utilities have implemented artificial neural networks (ANNs) for the detection and classification of harmonics in sophisticated monitoring systems. As an example, a utility in Europe has deployed ANN-based predictors that take in all waveforms captured by smart sensors to predict harmonic distortions and, thus, allow a proactive correction of shunt compensators to prevent overheating and failure of equipment [97].

Auto-voltage controls and autonomous microgrids Pilot projects on reinforcement learning (RL) algorithms have been tested to offer autonomous energy management and voltage control. These systems have higher levels of reliability and efficiency than fixed, formal controls, which are made possible by learning the best battery dispatch schedules and inverter settings under changing conditions of renewable generation and load.

To optimize the deployment and tuning of FACTS apparatuses, such as STATCOMs and SVCs, several operators of transmission systems have used Particle Swarm Optimization (PSO) and Genetic Algorithms. Their search control systems are more efficient than manual tuning processes that minimize the loss of power and enhance stability in the voltage profile, and there are examples of their implementation in the North American and Asian market [101].

Fuzzy logic controllers have also been widely used in industrial factories that require high power regulation and reduction in flicker. Fuzzy systems with the ability to cater to sensor errors and nonlinear load behavior provide better compensator operation that is smooth and minimizes maintenance.

High-resolution monitoring and predictive maintenance of the grid with real-time AI and IoT grid sensors allow the incorporation of AI and IoT devices to reduce downtime and prolong asset life. Owing to the use of AI, a digital twin platform mimics real segments of the grid to analyze scenarios that boost operational planning and resilience [96]. Domain-specific applications, such as rail traction systems, report that AI-driven UPQC control improves voltage quality and power factor under highly dynamic loading profiles.

4.8. Comparative Case Studies

This subsection presents brief and to-the-point comparisons of AI-based controllers with conventional PI/PID controllers in typical FACTS applications, providing metrics such as response time, voltage-recovery responses, overall harmonic distortion, and control-point accuracy, evolved in identical operating-point research with an experimental or HIL study [45,46,114].

4.8.1. Flicker Mitigation (STATCOM)

Many experimental and peer-reviewed studies have shown that neuro-fuzzy/adaptive AI-controlled STATCOM attains better var-tracking, reduced overshoot, and additional power-quality indicators compared to suitably tuned PI/PID settings when comparing the performance over identical laboratory conditions, including conditions where the reactive power demand can be unbalanced or time varying. In distribution networks and wind-integrated systems, the coordinated control by STATCOM significantly enhances the voltage regulation and disruptive resilience as measured by quantifiable indices of power quality, which illustrate the out-of-depthless of the learning-based control schemes compared to the conventional baselines when the system is disturbed. These performance improvements become most visible when the device nonlinearities and cross-couplings are dominant, and anti-windup interactions together with saturation and protection boundaries are established in both AI-based and conventional implementations when there are significant events of major transient nature [18,95].

4.8.2. DVR of Voltage-Sag Compensation

According to both the experiment and HIL, the AI-controlled DVRs occupy the reported speeds of sags occurring during the experimental settings along with significantly lower overshoot and injected voltage THD relative to previously noted classical PI/PID representatives under the same disturbance libraries and instrumentation. Proper sag/swell operation and stable converter performance were also verified using well-established hardware testbeds, providing a strong empirical base for comparative claims regarding the recoveryy time, steady-state error, and harmonic distortion under equal operating conditions. Even in the case of safe deployment, which depends on the reliability of fault detection, phase-locked synchronization, and restriction of injected voltage or current with high rigidity constraints to meet thermal and protection criteria, regardless of the control paradigm used [93,94]. Comparative evaluations of DVR control strategies confirm measurable improvements in sag compensation dynamics and steady-state PQ indices across multiple controller designs.

4.8.3. New PQ Mitigation as a Combine Approach Using UPCQ

AI-coordinated UPQC controllers, including ANFIS/ANN hybrids, improve simultaneous harmonic rejection and voltage regulation, as well as DC-link stability and angularly couple series-shunt configuration of the microgrid, compared to decoupled PI/PID configurations in microgrid and feeder-level experimental systems. Relative studies systematically catalog the elimination of current and voltage THD and enhanced voltage-deviation qualities during compound disturbances, and workflow simulation-to-hardware validation demonstrates a consistent enhancement of multipurpose performance distributions compared to classical results. The benefits observed depend on the fidelity of the measurement, delay of communication, confidence-cognizant gating, and dangerous fallback protocols to independent loops, which reduce the risk in conditions of poor elaborate telemetry or high-speed topological changes in real-world deployments [89,92]. Table 6 consolidates improvements achieved by AI-controlled FACTS devices under the stated objective functions. As summarized in Figure 4, the comparison highlights AI-driven FACTS devices across case studies and baselines.

Table 6. Improvements achieved by AI-controlled FACTS devices.

FACTS Device (Study)	Metric	Uncompensated	With AI Control	Improvement
UPQC (Rail traction) [20]	Source Current THD	25.16%	1.12% (ANN UPQC)	−95.5% (harmonic cut)
	Load Voltage THD	5.62%	2.07% (ANN UPQC)	−63%
D-STATCOM (Distribution) [88]	Source Current THD	27.24%	10.72% (ANN D-STATCOM)	−60%
	Power Factor (lagging)	0.9	~1.00 (ANN D-STATCOM)	Unity PF achieved
DVR (Lab prototype) [112]	Post-Sag Voltage THD	4.48% (PI DVR)	2.84% (ANN DVR)	−37% vs. PI control

Figure 4. Comparison of AI-driven FACTS devices.

5. Challenges, Future Research Directions, and Implications of AI-FACTS Integration

5.1. AI-FACTS Implication Challenges

Although significant progress has been made, there are considerable technical and practical challenges associated with the combination of artificial intelligence and FACTS devices in power quality management. The dependency on high-quality heterogeneous datasets is an important issue. The lack of sufficient data, noisy data, and prejudiced data undermine model generalization, causing failure under unobserved operating conditions. In addition, the high computational latency and resource demands also limit real-time applications, especially the demands in high-changing grid patterns with response times of less than a millisecond. The second salient challenge is model drift, in which the machine learning models gradually deteriorate in accuracy because of changes in the system, such as changes in the topology of the system, upgrading hardware, or alterations in the load profile. To reduce this drift, periodic retraining, adaptive algorithms, or the adoption of physics-based priors as part of hybrid models are required to maintain stability. In addition, the problem of AI model interpretability continues to interfere with operator confidence and regulatory adherence, highlighting the need to incorporate XAI systems into FACTS decision-making solutions.

5.2. Model Drift and Lifecycle Management

The operational constraints of AI-oriented FACTS controllers result from the mismatch between real-world conditions and the conditions used in training during seasonal load changes, changing mixes of DER, sensor recalibration or aging, and network reconfiguration. All of these individually create a shift in the distribution and affect the controller performance poorly if the conditions are not mitigated. To converter feeders, disturbance patterns and impedances change with topology and dispatch, thus requiring topology-dependent lifecycle reviews instead of offline tuning once [115,116,117].

5.2.1. Monitoring and Detection

Monitors and data should continually test feature-distribution drift by simple statistical procedures, such as the population stability index and Kolmogorov–Smirnov tests, and also through monitoring of residuals or errors between the outcomes of the model and data to detect consistent offered deviation at previously specified levels. Concurrently, Point of quality key performance indicators, such as THD, voltage-recovery time, overshoot of sags or swells, and flicker indices Pst and Plt, need to be plotted on control charts to identify performance degradation when operating in non-stationary mode [27].

5.2.2. Triggers and Policies

The drift tests, residual, and PQ key performance indicators (KPIs) should have quantitative thresholds igniting the alarms and launching the shadow mode of the candidate models on live data and rolling back to the known-good baseline to maintain the safety marginss. Vital data volumes and holdout situations as needed during revalidation must be defined in maintenance windows to ensure that revisions do not cause time constraints and that rare yet significant conditions are included in testing [17].

5.2.3. Failover and Risky Releases

Shadow deployment and A/B testing should be followed by model updates based on a proven baseline, and confidence-conscious gating should be implemented in such a manner that actuation returns to a safe deterministic policy if the confidence or data quality are found to be lower than the preset thresholds. There should be a deterministic contingency of classical controllers, such as PI/PID schemes with anti-windup and hard limits, which should be accessible at all times and guarantee voltage, current, and thermal limits during failover and recovery [17].

5.2.4. Edge-Cloud Orchestration

Algorithms that require a time-varying PQ control loop must run on the edge to achieve sub-cycles to tens to milliseconds latency, and consolidation, retraining, and validation must be performed in the cloud using signed artifacts and batched deployment to edge nodes. This division maintains local control when amputating communications and distributes the compute location with control importance within the distribution systems with FACTS and DERs.

5.2.5. Auditability, Diplomacy and Governance

Both versions of the model must be small but tracked by the provenance of training data, validation measures, and date of deployment, controller activities, and overrides are logged to assist in investigating incidents and regulatory audits. Explanations of the main actuation decisions made by operators are user-friendly and expedite root-cause analysis when the performance is poor in the field.

5.2.6. Retraining Strategy and Validation

Training must be planned proactively based on drift indicators and operation schedules using new data windows and holdout testing that resembles rare but crucial PQ actions in practice. Candidate models should be as effective as the baselines with respect to response time, voltage recovery quality, THD, and flicker, showing performance that meets or exceeds the performance of the baselines given the same disturbance libraries used in previous validation.

5.3. Future Research Directions

To cope with these issues, the further research agenda has several prospective orientations.

• Physics-constrained hybrid AI Hybrid models use data-driven learning models with analytical models to combine resilience to distributional change with reduced dependence on large labeled corpora in the application of PQ control mechanisms in complex grid architectures that use converters in many locations [68,118].
• PINNs incorporate power-system-level differential equations (including constraints on the network, networks, and modeling (and constraining) device dynamics) directly into the objective of the training task with the critical goals of ensuring feasibility and fast convergence [118].
• Federated and privacy-preserving learning models support multi-utility and multi-region model training without necessarily involving the exchange of raw data, preserving confidentiality, and meeting regulatory requirements while capturing the heterogeneity of operational regimes and enhancing the generalizability of AI-FACTS controllers [94].
• Federated architectures use cross-domain aggregation to support heterogeneity in sites, sensor arrays, and DER portfolios; using weighted client updates to reduce non-iid data distributions and concept drift, and this approach helps to increase the stability of globally deployed models that serve to reduce THD mitigation, sag/swell compensation, and properly flicker reduction [94].
• Digital twins and real-time simulations imply that temporally synchronized, high-fidelity, virtual copies of substations, feeders, and FACTS assets are created, such that controllers can be stress-tested in high-fidelity, physically sensitive, in silico, and alternative disturbances can be stress-tested in functions of a rarity database before any controllers can be deployed [12].
• Twin-guided adaptive control is an adaptive control technique that uses the closed-loop feedback of a pair of digital twins to tune artificial intelligence controllers to transiently provide a low-latency response and enhance voltage recovery statistics to mitigate production risk and alleviate response latency and voltage recovery requirements [39].
• Multi-agent reinforcement Learning The (MARL) determines the local juxtaposition goals of specific devices, including STATCOMs, DVRs, UPQCs, and DERs, and rallies the system-wide stability results through objective reinforcement and reward shaping actions in constrained communication among multiple agents [88].
• Safety-conscious grid control reinforcement learning, an explicit grid control problem, is combined with fallback and supervisory layers to ensure that the actions of the agents are within the voltage, current, and thermal parameters, and the responses with respect to the disturbance-rich environment meet the stringent and time-critical PQ feedback requirements [113].
• Decipherable and explicable artificial intelligence architectures provide interpretable attribution measures and post-facto reasoning about FACTS actuation judgments, which implicitly support operator trust, quick incident investigation, and allowance to meet industry norms and regulatory demands for AI-powered grid control.
• Authentic conforming and strong respect are achieved through a combination of formal verification and adversarial belligerence to stochastic noise, sensor flaws, and topology changes to compromise the AI controllers to PQ events and provide what can be confidently described as deterministic performance in the presence of uncertainty.
• Inference and control logic edge computing and distributed intelligence split their logic into local edge nodes, which contributes to sub-cycle to tens-of-milliseconds latency, reduced dependency on backhaul, and persistent provision of services in an intermittently connected scenario [88].
• Hierarchical orchestration is a combination of edge-level execution with regional coordination to offer frequency and voltage backup and make localized PQ remediation conform to the constraints of the bulk system and inter-area stability goals.
• Quantum-prepared optimization focuses on quantum-classical solutions of complex combinatorics optimization of scheduling, protection-constrained reconfigurations, and multi-objective optimal power flow to scale up problems in problem domains where classical algorithms exhibit limitations in scalability.
• Lifecycle and drift management involve keeping track of the distributions of features and residual anomalies and taking action by retraining pipelines with guarded retraining when drift owing to variations in the different seasons, changes in the DER mix, sensor recalibration, and scheduling are observed. Lifecycle and drift management involve monitoring feature distribution changes and residual anomalies and responding by retraining pipelines with rollback safeguards in response to drift detectors, such as seasonal variations, changes in the DER mix, or sensor recalibration [81].
• Open and privacy-protective benchmarking and hardware-in-the-loop verification develop open and privacy-safe benchmarks and testbeds that offer a relative evaluation of AI-FACTS controllers regarding metrics such as THD attenuation, voltage regulation accuracy, flicker mitigation, response latency of controllers, and their robustness at varied operating points [11].
• With a continuously growing collection of associations between methodological families with neural networks, support vendor machines, fuzzy logic, deep reinforcement learning, and meteorology, and with particular categories of the disturbance (harmonics, sags/swells, flicker, unbalance), a growing collection of such associations is integrated into a taxonomy and evidence-curation framework, whereby standard reporting of performance improvements and limitations on functionality are reported [79].

The combination of AI and FACTS equipment indicates that the power quality management policy is turning reactive and adaptive. The application of AI-FACTSs will deliver real-time disturbance detection, automatic compensation, and optimal grid operation in the scope of changing renewable energy infiltration. This development fosters resiliency, efficiency, and sustainability within grids. However, to achieve these benefits, it is necessary to overcome technical challenges, navigate regulatory systems, and establish cross-sectoral cooperation. Robustness, transparency, and security are the main requirements for the mass usage and profitable implementation of AI-enhanced FACTS technologies.

6. Conclusions and Discussion

This manuscript presents a systematic review of the research on the combination of artificial intelligence (AI) and Flexible AC Transmission Systems (FACTS) to improve the quality of power in modern electrical grids. The survey outlines several AI paradigms, such as neural networks, fuzzy logic, reinforcement learning, and metaheuristic optimization, highlighting their respective abilities in detecting disturbances, adaptive control, and operational efficiency [119,120,121].

However, several viable challenges hinder mass adoption. The problems of data quality, model drift, and operational demands do not have rapid computing requirements and require creative solutions, including hybrid modeling, federated learning, and digital twin platforms. In addition, continuous studies should focus on transparency and explanation to create stakeholder and regulatory trust [72,80].

Future studies should focus on the creation of hybrid, adaptive AI systems that integrate the laws of physical systems, enabling privacy-oriented cooperation and working on the network interface to meet real-time needs. These developments will provide greater resilience and flexibility to the power grid, as well as the capacity to absorb renewable energy sources and continue to maintain high power quality [77,79].

Overall, FACTS control with the help of AI can be characterized as a considerable chance of revolutionizing power quality management; however, further work is indispensable in overcoming the obstacles of operation and regulation. Drawing on the available research, additional tasks to be performed in the future encompass scalability, model sustainability, and it is under extreme operational environments. After disturbances to minimize quality degradation

• Continuous Self-optimization: Adapting to changing network conditions without human intervention
• Coordinated Response: Implementing multiple FACTS devices to achieve system-wide objectives

Although significant technical, regulatory, and organizational challenges remain, the trajectory of AI-FACTS integration points toward increasing autonomous power quality management systems, which will form the cornerstone of future smart grid implementations. The ongoing convergence of advanced sensing, communication technologies, and artificial intelligence promises to transform power quality from a reactive maintenance concern to a proactively managed asset, enabling the reliable integration of renewable resources and supporting the increasing electrification of the global economy [119].

Appendix A.1. Static Synchronous Compensator (STATCOM)

The Static Synchronous Compensator (STATCOM) is a shunt-connected, Voltage Source Converter (VSC)-based FACTS device that can be used to regulate reactive power dynamically by adjusting the magnitude of its output voltage relative to the system voltage [46,47]. By operating through a VSC, the STATCOM can absorb and inject reactive power into the grid. The exchanged reactive power between the STATCOM and power system can be calculated using the following equation:

$${\mathrm{Q}}_{\mathrm{S}\mathrm{T}\mathrm{A}\mathrm{T}\mathrm{C}\mathrm{O}\mathrm{M}}={\displaystyle \frac{{\mathrm{V}}_{\mathrm{s}}.({\mathrm{V}}_{\mathrm{s}}-{\mathrm{V}}_{\mathrm{S}\mathrm{T}\mathrm{A}\mathrm{T}\mathrm{C}\mathrm{O}\mathrm{M}}.\mathrm{c}\mathrm{o}\mathrm{s}\mathrm{\delta })}{\mathrm{X}}}$$

$${\overrightarrow{\mathrm{V}}}_{\mathrm{S}\mathrm{T}\mathrm{A}\mathrm{T}\mathrm{C}\mathrm{O}\mathrm{M}}={|\mathrm{V}}_{\mathrm{S}\mathrm{T}\mathrm{A}\mathrm{T}\mathrm{C}\mathrm{O}\mathrm{M}}|.{\mathrm{e}}^{\mathrm{j}{\mathrm{\theta }}_{\mathrm{S}\mathrm{T}\mathrm{A}\mathrm{T}\mathrm{C}\mathrm{O}\mathrm{M}}}$$

where Q_STATCOM is the reactive power exchange, V_s is the system voltage, V_STATCOM is the converter output voltage, X is the coupling reactance, and δ is the phase angle between voltages.

Through various control modes, the STATCOM can be configured to operate in voltage regulation, reactive power control, power factor control, and var reserve control mode. The advancedd use of FACTS in complex grid environments includes positive-negative sequence control for compensating unbalanced loads, harmonic filtering, and subsynchronous damping via auxiliary modulation. Moreover, flicker mitigation is performed via high-speed reactive current modulation, which is beneficial for fluctuating industrial loads. In addition, the STATCOM offers a superior dynamic response, typically achieving full compensation in a short duration. The advantages of this technology include its compact footprint, minimal harmonic distortion, and the potential for active power exchange.

The Unified Power Flow Controller (UPFC) [49,50] has recently been recognized worldwide as a FACTS device with the broadest range of capabilities, capable of simultaneously controlling the three key parameters that determine power flow on an AC transmission system, namely, the voltage magnitude, phase angle, and line impedance. The UPFC is a structural combination of two VSC: one is a shunt series converter, which is a makeshift STATCOM by a functional characteristic, and the other is a series shunt converter, which is a makeshift Static Synchronous Series Compensator (SSSC) by a functional characteristic. These two converters have a common DC connection incorporating capacitor-based energy storage; therefore, they can exchange active power in both directions. A coordinated control system is used to coordinate the functioning of several control loops in a dynamic response to system conditions. The basis of its operation is to restrain the electrical magnitude of the voltages at the shunt side with the aid of a reactive power source and regulate the power flow at the series side by injecting an additional controllable magnitude and phase of the voltage. Mathematically, the injected voltage of the series converter can be represented as

$${\overrightarrow{\mathrm{V}}}_{\mathrm{i}\mathrm{n}\mathrm{j}}=\mathrm{\gamma }.{\mathrm{V}}_{\mathrm{i}}.{\mathrm{e}}^{\mathrm{j}\mathrm{\theta }}$$

where $\mathrm{\gamma }$ is the injection coefficient (typically between 0.5 and 0.55) and $\mathrm{\theta }$ is the controllable angle.

The UPFC is used to improve system-wide power flow control by injecting the transmission line, and the voltage used changes the apparent impedance and line phase angle. This direct control over the line impedance and phase angle allows the power flow to be controlled between zero and the thermal limit, regardless of the starting point of the system and physical parameters of the line. To do so, the UPFC implements a variety of operational modes: power-flow-control mode (used to simultaneously regulate active and reactive power), line-impedance-emulation mode (that can emulate varying series reactance), phase-angle-regulation mode (designed to control the previous power angle), voltage-regulation mode (used to stabilize the receiving-end voltage), and multifunction mode that combines many goals. With the ability to provide independent simultaneous regulation of active and reactive power, voltage control at the connection point, enhance transient stability using power-flow modulation, and aid in damping oscillations and balancing loads during an asymmetric condition, the UPFC can be regarded as one of the most effective instruments to control and optimize a power system today.

The Unified Power Quality Conditioner (UPQC) design [51,52] is a flexible distribution-level device engineered to handle a wide range of power quality defects simultaneously using series active filters and shunt active filters in the same device and system. Conceptually, it is similar to the Unified Power Flow Controller (UPFC), except that it is specifically designed to improve power quality, as opposed to being designed to regulate power flow. The device combines two converter modules: a shunt converter that only deals with current-related anomalies, that is, harmonics, imbalance, and reactive power demand, and a series converter that deals with voltage-related anomalies, that is, voltage sags, swells, flicker, and harmonic distortion. Common among these converters is that they all use a common DC bus that accommodates a medium-sized energy storage component along with a fast digital control system that allows dynamic coordination of the two subsystems in real time. The related formulas for the UPQC are as follows:

$${\overrightarrow{\mathrm{V}}}_{\mathrm{s}\mathrm{e}}={\overrightarrow{\mathrm{V}}}_{\mathrm{r}\mathrm{e}\mathrm{f}}-{\overrightarrow{\mathrm{V}}}_{\mathrm{s}}$$

$${\overrightarrow{\mathrm{I}}}_{\mathrm{s}\mathrm{h}}={\overrightarrow{\mathrm{I}}}_{\mathrm{L},\mathrm{h}}+{\overrightarrow{\mathrm{I}}}_{\mathrm{L},\mathrm{q}}+{\overrightarrow{\mathrm{I}}}_{\mathrm{L},\mathrm{u}}$$

where ${\overrightarrow{\mathrm{I}}}_{\mathrm{L},\mathrm{h}}$ is the harmonic component, ${\overrightarrow{\mathrm{I}}}_{\mathrm{L},\mathrm{q}}$ is the reactive component, and ${\overrightarrow{\mathrm{I}}}_{\mathrm{L},\mathrm{u}}$ is the unbalanced component of load current.

Most of the latest power-quality control system architectures use instantaneous reactive power theory in a series converter to isolate and cancel a rapid voltage disturbance and a shunt converter, either the synchronous reference frame or p–q theory, to isolate and cancel unwanted current components. A DC-link controller can be used to enable a constant capacitor voltage, whereas a synchronization module tracks the supply voltage phase angle carefully so that it can be used to generate a reference. The UPQC offers comprehensive operational functionality, comprising the correction of sag and swell voltages up to 50% of the nominal voltage, attenuation of harmonic currents to the 25th harmonic with over 90% efficiency, and mitigation of harmonic voltages up to the 13th harmonic with compensation rates greater than 85%. By lowering the adverse sequence current components by approximately 80–90% and 65–80% of the flicker severity, the UPQC is applicable in sensitive locations. The UPQC finds practical use in important manufacturing processes, data centers, telco infrastructure, hospitals, highly renewable penetration microgrids, and industrial parks with highly nonlinear and complex loads [67].

The Interline Power Flow Controller (IPFC) [53,54] is a highly advanced FACTS technique with an extension or generalization of the concept of the SSSC to several transmission corridors. It allows the synchronous control of power flow across more than one transmission path across a meshed network to supervise congestion, power flow balancing, and power destabilization issues within complex transmission designs. The IPFC consists of several series-connected SPFCs (voltage source converters) located in separate lines but sharing a common DC link, through which active power may be exchanged among the converters. This standard energy storage function permits the flexible power transfer between converters [70]. Simultaneously, the IPFC provides reactive compensation at both ends of the line, thereby enhancing network voltage stability. Through this architecture, the IPFC offers several key control capabilities, including the redistribution of active power among interconnected transmission paths, reactive power compensation, minimization of system losses, and enhancement of dynamic stability [71]. The following equation determines the power balance condition of the entire system:

$$\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{P}}_{\mathrm{s}\mathrm{e},\mathrm{i}}=\sum _{\mathrm{i}=1}^{\mathrm{n}}\mathrm{R}\mathrm{e}\left\{{\overrightarrow{\mathrm{V}}}_{\mathrm{s}\mathrm{e},\mathrm{i}}{\overrightarrow{\mathrm{I}}}_{\mathrm{i}}^{\ast }\right\}=0$$

(A1)

where ${\overrightarrow{\mathrm{V}}}_{\mathrm{s}\mathrm{e},\mathrm{i}}$ is the injection voltage in line I and ${\overrightarrow{\mathrm{I}}}_{\mathrm{i}}$ is the current in line i.

The complex interplay of power grids with high penetration of renewable energy sources, emerging market mechanisms, and dynamically relocating customer demand makes adaptive learning especially useful, as it is a technique that keeps the models reliable even at a time of grid stress [93].

ML systems may learn further beyond their traditional, reactive paradigm and proactively predict future situations, which incorporate temporal variations in load and the likelihood of equipment failures. Temporal models [107], including representations of form RNN, can help add past data along with the current input. Such architecture allows applications to anticipate the demand profile, systematic grid stability, and predictive fault events, which facilitates proactive actions that enhance the resilience and efficiency of the system.

$${\mathrm{h}}_{\mathrm{t}}=\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}({\mathrm{W}}_{\mathrm{x}\mathrm{h}}{\mathrm{x}}_{\mathrm{t}}+{\mathrm{W}}_{\mathrm{h}\mathrm{h}}{\mathrm{h}}_{\mathrm{t}-1}+{\mathrm{b}}_{\mathrm{h}})$$

$${\mathrm{y}}_{\mathrm{t}}={\mathrm{W}}_{\mathrm{h}\mathrm{y}}{\mathrm{h}}_{\mathrm{t}}+{\mathrm{b}}_{\mathrm{y}}$$

where ${\mathrm{h}}_{\mathrm{t}}$ is the hidden state at time t, ${\mathrm{x}}_{\mathrm{t}}$ is the input at time t, ${\mathrm{W}}_{\mathrm{x}\mathrm{h}}$ is the input-to-Hidden weight matrix, ${\mathrm{W}}_{\mathrm{h}\mathrm{h}}$ is the hidden-to-hidden weight matrix, ${\mathrm{b}}_{\mathrm{h}}$ is the bias for the hidden state, and Tanh is the activation function.

The use of AI-driven methods in recent years has shed new light on how balancing multiple and potentially competing goals can be achieved, which is a core issue facing the current operations of power systems [72]. Unfortunately, conventional models cannot combine and optimize many of the objectives of an optimization simultaneously, such as minimizing the cost and emissions, improving reliability, and satisfying the complex constraints that define such systems. In comparison, more advanced AI frameworks, especially groundbreaking algorithms and reinforcement learning, can optimize multidimensionally and automatically find solutions with superior quality that can meet and address the complexities of contemporary power grids.

AI techniques have demonstrated significant potential for enhancing various aspects of distribution network operations, including load forecasting, network configuration, and non-technical loss detection. Among these, Graph Neural Networks (GNNs) [80,84] have proven particularly effective because of their ability to capture the inherent topological and spatial dependencies within grids. GNNs can model both local and global relationships across grids. Consequently, GNN-based approaches facilitate topology-aware optimization by incorporating both electrical and physical constraints to improve operational efficiency, fault detection, and energy management across complex, unbalanced, or meshed distribution systems. The following formula demonstrates the aforementioned technique:

$${\mathrm{H}}^{(\mathrm{l}+1)}=\mathrm{\sigma }({\tilde{\mathrm{D}}}^{-0.5}\tilde{\mathrm{A}}{\tilde{\mathrm{D}}}^{-0.5}{\mathrm{H}}^{(\mathrm{l})}{\mathrm{W}}^{(\mathrm{l})}$$

where $\tilde{\mathrm{A}}$ represents the adjacency matrix of the network, ${\mathrm{H}}^{(\mathrm{l})}$ is representing the node features at layer l, $\tilde{\mathrm{D}}$ is the matrix degree, ${\tilde{\mathrm{D}}}^{-0.5}$ is the symmetric normalization, ${\mathrm{W}}^{(\mathrm{l})}$ is the trainable weight matrix, and $\mathrm{\sigma }$ is the activation function.

Reinforcement learning (RL) [86] algorithms have gained traction in microgrid control and the use in energy-optimizing applications, where they may be used to develop adaptive control policies based on iterative interaction with the system. RL controllers increasingly approach an optimal strategy that trades off cost efficiency, reliability of operations, and system restrictions by testing the repercussions of different actions over a gamut of operating conditions. Such controllers have been formidable in minimizing the guesswork that comes with load fluctuation, intermittent renewable generation, and evolving market signals. The actor in the control problem tries to maximize the long-term expected reward by manipulating thinking through the task as a series of correlated good successively staged choices.

$${\mathrm{\pi }}^{\ast }\left(\mathrm{s}\right)={\mathrm{a}\mathrm{r}\mathrm{g}\mathrm{m}\mathrm{a}\mathrm{x}}_{\mathrm{a}}{\mathrm{Q}}^{\ast }(\mathrm{s},\mathrm{a})$$

where ${\mathrm{\pi }}^{\ast }$ represents the optimal policy and $\mathrm{a}$ is the action that maximizes the long-term rewards at given state s.

Despite the significant advantages that AI brings to power systems, its deployment often comes with challenges that must be addressed to ensure its successful implementation.

The quality of AI performance applications in power systems is closely related to the quality resolution and availability of data [88]. Most AI-driven forecasting and control methods rely on historical and real-time datasets to deliver accurate and reliable outcomes. However, in practice, these datasets suffer from limitations such as incompleteness, inaccuracy, and low temporal or spatial resolution. These restrictions can significantly hinder model training, generalization, and deployment in real-world scenarios. To address these issues, several mitigation strategies have been considered prior to the deployment of AI systems. This includes upgrading measurement equipment, robust data preprocessing pipelines, and artificial data generation techniques to ensure the successful integration of AI in modern power system operations.

AI models are often considered a valuable solution because of their high predictive accuracy; however, their black-box nature can cause challenges when applied to critical infrastructure, where transparency and interpretability are essential. To address this issue, developments in Explainable AI (XAI) [122] techniques have been used to make AI decisions more understandable to the human language level to ease the operation of operators and regulators. Methods such as attention mechanisms help identify the number of features that influence the model’s prediction. Feature importance analysis techniques also clarify the impact of each feature on the output for better model behavior evaluation. Moreover, symbolic regression techniques also attempt to transfer complex AI models into interpretable mathematical expressions to bridge the gap between high-performance prediction and human-readable logic.

Computational complexity variation in deep learning methods can also vary significantly in terms of the amount of computation required to train the method and run it in real time. There are several operational challenges of these models relevant to power systems when they are implemented on a centralized cloud-based computing architecture. High communication latency may hinder time-sensitive operations on the grid and its parts, and reliability problems may occur because of both network disturbances and communication breakdowns. To overcome these constraints, Edge AI has become a potential solution, where the power system’s local computational intelligence is divided locally over the system infrastructure. Edge AI [89] can enable low-latency decision-making by processing data at the source (or close to it), that is, substation smart meters or local controllers, without sacrificing system-wide coordination or goals. This distributed setup increases responsiveness and fault tolerance and enables the use of AI in a power grid with limited bandwidth or unstable connections; therefore, it is a viable option for currently designed and developing power grids.

Appendix A.2. Forecasting Uncertainty and Reliability Challenges

Despite the advantageous characteristics of deep learning for renewable forecasting, a major limitation remains in how forecast uncertainty is managed through FACTS control systems. While traditional approaches incorporate clear uncertainty bounds, many AI implementations provide point forecasts without robust quantification of confidence intervals, which could lead to improper control actions during highly uncertain conditions.

One commonly overlooked challenge is model generalization, which ensures that a model works across different types of generation assets. While remarkable accuracy has been demonstrated at specific sites, few studies have rigorously validated their work by considering various geographical regions, turbine technologies, or solar panel configurations, which limits their broader applicability to other regions. This issue becomes more critical during the deployment of FACTS controllers, which rely on these forecasts across heterogeneous power systems.

The main weakness of current AI forecasting methods is their limitation in predicting severe events and rapid ramps that significantly affect power quality. Neural networks are often optimized for average performance, unlike statistical methods that focus on rare or extreme events, potentially missing the event most relevant for FACTS intervention. This limitation undermines the reliability of AI-enhanced FACTSs during critical grid disturbances, while their integration into the grid is most needed.