Avant-Garde Solar Plants with Artificial Intelligence and Moonlighting Capabilities as Smart Inverters in a Smart Grid

Abstract

Intelligent inverters have the capability to interact with the grid and supply supplemental services. Solar inverters designed for the future will have the ability to self-govern, self-adapt, self-secure, and self-heal themselves. Based on the available capacity, the ancillary service rendered by a solar inverter is referred to as moonlighting. Inverters that communicate with the grid but are autonomous can switch between the grid forming mode and the grid following control mode as well. Self-adaptive grid-interactive inverters can keep their dynamics stable with the assistance of adaptive controllers. Inverters that interact with the grid are also capable of self-adaptation Grid-interactive inverters may be vulnerable to hacking in situations in which they are forced to rely on their own self-security to determine whether malicious setpoints have been entered. To restate, an inverter can be referred to as a “smart inverter” when it is self-tolerant, self-healing, and provides ancillary services. The use of artificial intelligence in solar plants in addition to moon-lighting capabilities further paves the way for its flexibility in an environment containing a smart grid. This perspective paper presents the present as well as a more futuristic outlook of solar plants that utilize artificial intelligence while moonlighting advanced capabilities as smart inverters to form the core of a smart grid. For the first time, this perspective paper presents all the novel ancillary applications of a smart inverter while employing Artificial intelligence on smart inverters. The paper’s emphasis on the Artificial Intelligence associated with PV inverters further makes them smarter in addition to ancillary services.

1. Introduction

All large-scale solar farms rely on inverters to transform the DC electricity generated by solar panels into the AC power used by the electrical grid. These solar farms rely on inverters to transform the DC electricity produced by the sun into the AC electricity required by the grid during daylight hours. Inverters are essential for large-scale solar farms because they change the DC electricity generated by solar panels into the AC power used by our electrical grid. The inverters at these solar farms convert the direct current (DC) electricity generated by the sun during the day into alternating current (AC) electricity for the grid, however, they can only run during the day. Since so much renewable energy has been added to the grid, solar inverters now serve a dual purpose, operating both during the day and at night. The stability of the renewable energy system may depend on solar inverters working longer shifts. Solar farms, wind farms, and energy storage devices are just a few examples of the rapidly expanding renewable energy sources that are changing the face of the energy sector. When connecting non-constant energy sources (AC or DC) to the grid, solid-state converters are used. One component of these converters is an inverter that connects to the public power grid. However, the term “grid-interactive inverters” is used to refer to these multi-stage converters in the business world and the actual devices that make up the decentralized power grid are called “distributed generators” (DGs). The most important function of DGs is provided by inverters, and the widespread use of these devices has allowed electricity networks to become more adaptable. A “smart” inverter can proactively and objectively make decisions based on information gathered from both its internal and external environments. By offering auxiliary services to power networks, inverters, for example, can be designed to enhance power quality under abnormal conditions. When first installed, inverters are set to a configuration that is grid-friendly. When inverters are set to grid-forming mode, a standalone microgrid can be created in the aftermath of a widespread power outage caused by, for example, a natural disaster.

Another application where inverters are essential is clustering a power grid, also referred to as a network of microgrids or a grid of microgrids. This is another moniker for this arrangement. While there are many benefits to the widespread use of DGs, it is also challenging for electricity grids to maintain their customary levels of security, stability, and dependability. While droop controllers and other distributed control schemes can help with DG power sharing, a supervisory control scheme is still necessary for optimal energy management and economic dispatch. A utility operator and the inverters must be in constant contact for this plan to work. When modules, sensors, and processors are integrated into a smart inverter in addition to data packet connections, the resulting system takes on the characteristics of a cyber-physical system [1,2]. The integrity of safety safeguards may be compromised when inverters are linked to communication or a cyber network [3].

Furthermore, because of the lack of inertia in inverter-based DGs, low-inertia microgrids can be created. This leaves them vulnerable to power outages. Internal switch short circuits are another potential cause of inverter failure that could compromise the system. In addition to good governance, other important qualities of a smart inverter are the ability to fix itself, the ability to change and adapt to its environment, and security. Finally, an inverter having multiple operating modes should be able to flip between them without negatively impacting the device’s normal performance. Smart inverters in a supervisory system should be able to self-regulate via communication with agents or a human supervisor. Information from their local communication or cyber networks is combined with that from their local sensors. A smart inverter could foresee problems and take corrective measures by considering the status of other distributed generators (DGs), smart meters, and forecasted data [4,5]. In the following sections, we will also take a high-level look at some of the more recent advances and present technical issues related to each function. A typical association of a smart inverter in a smart grid environment is shown in Figure 1. This paper presents the perspective of smart inverters with advanced capabilities employing artificial intelligence and as moonlighting grid-interactive smart inverters for ancillary services.

Over the past few years, ref. [6] favorable regulatory rules have been designed to increase the penetration of renewable energy sources, and the number of distributed generation systems that are connected to the grid has increased dramatically. Electricity producers from all over the world are scratching their heads over how to incorporate most effectively solar photovoltaic (PV) generating units into existing power grids despite the fact that utility companies do not know how high volumes of variable renewable energy output will affect grid operation, power quality, and worker and equipment safety. An inverter’s usefulness is crucial in this situation. Formerly, a fault or other out-of-the-ordinary grid condition necessitated the disconnection of the inverter from the grid before it could be reconnected. Newer grid systems, however, need supplementary tools for monitoring voltage and current [7,8,9,10]. It wasn’t until recently that inverters became a practical tool for grid management and control. Electrical power generated by renewable resources and stored in batteries typically takes two forms: direct current (DC) and variable frequency (VFR). All this energy must go through a frequency converter before it can be added to the grid as a reliable alternating current (AC) [11,12,13]. In a microgrid, an inverter is what links the various power generation and consumption nodes together. Thus, it can be used to regulate power, detect issues, and cut power, when necessary, in addition to converting AC-DC or vice versa. Some people refer to inverters as the “brains” of a microgrid because they are responsible for processing and analyzing the inputs and outputs that have the greatest impact on the microgrid’s ability to exert control over its power distribution. They gather information and adjust their infrastructure to work best in a safe, standardized, and simplified setting. Besides its primary role, it also has grid support features such as voltage and frequency regulation and ride-through capabilities. As a result, smart inverters are cutting-edge pieces of equipment that can ease the process of connecting solar PV and other DERs to the electrical grid. This is because it is outfitted with state-of-the-art power electronics, which allow it to simultaneously compensate for DC voltage fluctuations of varying intensities and provide a variety of services. This is effective with PV cells, whose optimum operating voltage changes with environmental temperature and light intensity. Solar panels generate direct current (DC), and inverters (both traditional and smart) change this into alternating current (AC) that can be used indoors. An increase in the number of distributed energy resources (DERs) connected to the grid has necessitated the use of more robust inverters.

Figure 2 presents the block diagram of a simplified version of an inverter system comprising of a DC source, a power processing unit, an output filter, grid synchronization, and a controller function.

It is possible that distributed power sources and automated processes could overwhelm centralized command, management, and information-sharing infrastructures. Multi-agent systems and other forms of distributed control are integral components of complex power systems that are needed to meet challenges such as decomposition and global performance. It is highly desired to have plug-and-play subsystems that are autonomous, adaptive, scalable in functionality, and require little to no system administration in order to lessen the burden of control and communications. This holds true for a large class of control architectures.

(a)

Self-cognizant

Self-awareness is becoming an increasingly necessary component of distributed systems, which makes higher degrees of autonomy dependent on it. For the sake of this discussion, say for instance that a system is self-cognizant if and only if it is cognizant of its own internal state. When used in safety-critical systems, power electronics with built-in self-awareness can greatly improve operational dependability and lifetime prediction, enable fail-safe or maintenance actions, and greatly reduce the likelihood of catastrophic accidents [16].

(b)

Versatility

A good measure of a system’s adaptability is its ability to adapt to new or different conditions quickly and efficiently, as well as to any changes that are made to the system’s component parts. There is potential for fault tolerance, control tolerance, frequency tolerance, and impedance tolerance to be built into the design of intelligent inverters. Self-tuning controllers based on adaptive laws, such as adaptive maximum power point tracking (MPPT) and droop and current control, can achieve Lyapunov stability despite the presence of uncertainty and a broad spectrum of operating conditions. This holds true for any variety of self-tuning controllers.

(c)

Autonomy

The ability of a system to determine its own behavior without external control is what is meant by “autonomy” here. It is necessary for any distributed system to provide the means for individuals to independently carry out the responsibilities that are assigned to them. Instead of relying on a traditional hierarchy that is organized from the top down, modern automation is moving toward a structure that is flatter, more decentralized, and is based on CPS. This contrasts with the traditional hierarchy, which is organized from the top down. It is possible that autonomous smart inverters will prove to be an essential component if there is a disruption in communication within a cyber-physical system [17].

(d)

Coordination

Cooperative controllers must be implemented in each smart inverter for the system or sub-system to reach the desired collective goal. However, this must be done while still maintaining the system neighbor’s stability and maintaining the quality of the system. Distributed solutions, on the other hand, call for information to be shared between neighbors and smart inverters that are in proximity. Centralized solutions are the more common type of solution. Self-organization and resistance to dynamic uncertainty are just two of the potential benefits that could result from a decentralized control and decision-making architecture [18].

(e)

Plug-and-Play

A smart inverter can be easily integrated into an existing system thanks to plug-and-play (PnP) technology, which eliminates the need for any additional configuration or set-up. Scalability, interoperability, robustness, and dependability are just some of the benefits of PnP functionality, which can be enjoyed with or without communication. To establish distributed intelligence based on power electronics equipment, the PnP concept was only recently introduced. This contrasts with the widespread use of the PnP concept in computing interfaces and industry. When used in microgrids and smart grids, it is possible to use PnP in power converters with smart inverters and modular hardware structures to control the voltage and frequency at each individual node.

The ability of smart inverters to switch, conduct, and store electromagnetic energy is one of the ways in which power electronics are embodied in these devices. In addition to the internal operations that were just described, the intelligent features will be implemented by utilizing a combination of different types of hardware and control strategies. Market price signals are monitored by three distinct controllers: an autonomous controller, a cooperative controller, and a transactive controller. Figure 3 is a diagram depicting the internal structure of a typical smart inverter.

A smart inverter is not constrained by any one control architecture for the system and functions most effectively in a decentralized setting. Hierarchical layouts allow for easier practical application. A microgrid with many feeders and smart inverters can be represented by the holarchy, a four-level hierarchical structure made of holons, beginning with the microgrid level and proceeding downwards through the feeder level, the smart inverter level, and finally the function of the smart inverter.

2. Artificial Intelligence Based Solar Plants as Smart Inverters

Having the ability to regulate grid voltage and frequency, as well as provide autonomous ancillary services for grid maintenance, is what is meant by “self-governing.” This allows for the construction of microgrids and networked microgrids. Setting up microgrids and networked microgrids, as well as providing support for the grid through the provision of independent ancillary services requires the ability to modify the voltage and frequency of the grid. Other modes of operation for grid-tied inverters offer greater versatility than the conventional method of feeding the grid, which entails injecting all available power into the grid while keeping the power factor at 1.

PV inverters were developed with the goal of sending as much of the solar array’s active power P (kW) as possible to the point of the common connection while keeping the power factor at unity. We did this by supplying as much active power as we could while maintaining a unity power factor. Reactive power Q (kVAR) consumption and generation capabilities of the three-phase inverter have garnered significant attention from utilities and independent power producers in recent years. In recent times, this ability has received a lot of attention. Due to its obvious value, this competency has been receiving a lot of attention as of late. A photovoltaic inverter will operate at currents that are more than 95% below its rated output when converting DC solar electricity to AC. Direct current (DC) and alternating current (AC) are abbreviations for two different types of electrical current. Reactive power can be generated using the inverter’s excess capacity, which has a variety of applications. The most efficient operation of the greatest generating resources, all of which are synchronous machines, occurs at exactly 50 or 60 hertz, making it crucial to keep the frequency as stable as possible within a transmission network. Further, for the generation load to be distributed equitably amongst them, the speed governors on these machines must work in unison and in accordance with the set timetable. Frequency stability requires that the active power produced be exactly equal to the passive power consumed. Since smart inverters are unable to absorb energy, regulating the active power they produce is only effective in reducing the frequency. Large PV projects often use active power curtailment and ramp rates to assist alleviate site-specific problems and boost grid stability. Optimal performance in terms of true power P and VAR component of apparent power- Q flow throughout the grid is likely to need two distinct control schemes. Managing active power is related to managing grid frequency. However, managing reactive power and controlling grid voltage are intertwined.

Reactive power can only be managed at big generation units, but the voltage can be regulated across the transmission and distribution network by injecting and absorbing VARs at different points. Overvoltage is dangerous to equipment and loads. Transmission line losses can be drastically reduced with VAR management, which also improves grid stability. Transmission lines can either supply or absorb reactive electricity, depending on the load and the line’s length. When voltages are very high, the reactive power loss proportion of the total power loss typically predominates over the resistive power loss proportion. A static VAR (volt-ampere-reactive) compensation is one use for the reactive power capacity of a smart PV inverter. This kind of compensator has the ability to either lower or raise the AC voltage further down the line in response to instructions from the supervisory control and data acquisition (SCADA) system, or to do so on its own. The main benefit of this implementation is the low cost of the components used.

Using inverters as a reactive power source necessitates configuring the inverters to automatically reduce their active power when the peak current limit is reached. A sudden failure of all inverter-based distributed generators could cause problems far more severe than a simple drop in voltage due to the increasing prevalence of inverters. To comply with modern regulations, inverters must stay connected to the grid and continue to supply reactive power even if the grid voltage drops. This is required by IEEE Std. 1547-2018. A permanent drop in voltage, however, necessitates cutting power to the inverter. An inverter will trip if the active power injected into the grid is allowed to grow too large in tandem with the increase in reactive power. The active and reactive power available to the inverter at any given time varies depending on the DC bus voltage Vdc, the modulation index m, and the impedance between the inverter and the grid. Furthermore, this strength is proportional to the maximum current limit, which is shown as a circle with a radius of Smax. The second circle is depicted as well; its radius is mVdc Vg/|ZTh|. Two diagrams are shown in Figure 4. The left-hand diagram depicts how a smart PV inverter operates in each of the four available power domains, injecting real and reactive power as needed. The diagram on the right shows how a grid-tied inverter operates in a distribution system’s four different quadrants. The R/X ratio and the Thevenin impedance are used to determine the efficacy of the system in this operation. This means that the smart PV inverter can function as a moonlighting smart inverter in either a constant power factor mode or a variable power factor mode, depending on the capacity that is available after the real power is injected. The smart inverter’s capacity determines whether it operates in real power mode, reactive power mode, or a combination of both. The Thevenin impedance of the system determines this.

Asymmetrical faults in the power grid and an uneven distribution of single-phase loads in the distribution grid are just two of the many potential causes of asymmetrical anomalies. Besides the positive-sequence reactive power support that a smart inverter typically provides, negative-sequence compensation services can be of assistance to the grid in these cases. With the addition of q and d axis current control channels, negative-sequence compensation for the grid currents can be provided. This holds true regardless of whether the reference frame for developing the control scheme is synchronous. These channels send the controller the negative-sequence components of the grid current at zero references, even though the local load may be asymmetrically distributed. The voltage at PCC includes a negative-sequence component that is not eliminated by this modification. Therefore, a small amount of active and reactive power oscillations may persist, resulting in a slight ripple in the inverter’s DC bus voltage. Overvoltage waves can cause severe damage to DC bus capacitors. The frequency of oscillations and the amplitude of waves are reduced when the controller’s setpoints are altered from zero. The active and reactive power fluctuations on the grid can be reduced by adjusting the controller settings, which also stabilizes the grid’s current.

It is the goal of the negative-sequence controllers’ setpoints to keep the currents within safe limits while also minimizing power oscillation. The voltage’s negative-sequence component at PCC is crucial to accomplishing this goal. Whether or not such services can be offered depends on whether the smart inverter incorporates a negative-sequence compensation mechanism and whether an appropriate control target is chosen. Due to the imbalanced nature of the negative sequence correction, some of the inverter references may enter the overmodulation zone when the power is brought up to its proper level. To keep the inverter in the linear modulation zone and take advantage of additional services such as negative sequence and harmonic compensations, a low-frequency common-mode signal is added to the PWM reference signals. Figure 5 and

Table 1. Ancillary grid applications of a Smart PV inverter.

Ancillary Grid Applications of Moonlighting Smart Inverters	Description
Active Power filtering	Active power filtering is a function that can be performed by intelligent inverters equipped with the appropriate controller for reducing the harmonics.
Flicker	The frequency and voltage fluctuations could be reduced.
Regulation of Voltage/Frequency	Smart inverters can perform V/f regulation in the system.
Ride through capability	During such an event, smart inverters could give VAR support and maintain system connectivity by executing the ride-through operation.
Line losses	Smart inverters were able to inject and absorb VARs in the system after the release of IEEE 1547.8 and UL 1741 standards. Line losses may be drastically cut down as a result of this which may result in cost savings.
Regulation of voltage power factor correction	A good way to avoid utility fines is to use smart inverters with specialized controllers to perform voltage regulation in the system and power factor correction of local loads.
Virtual detuning	Virtual detuning could be used by intelligent inverters with specialized controllers to lessen the effects of the network harmonic resonance phenomenon. The harmonics could also be decreased with this step.
Mitigation of Temporary OverVoltage (TOV) phenomenon	Smart inverters can be used in the healthy phases of Single Line to Ground Fault and Double Line to Ground Fault conditions to effectively mitigate TOV.
Anti-island detection	Capable of analyzing transient faults using the given technique.
Reverse power flow	The voltage spike generated by DERs’ inward power flow can be mitigated by voltage regulation performed by smart inverters. It also helped the widespread adoption of distributed energy resources (DERs) such as wind and solar power as well as EVs (PEVs).
Power generation	Real electricity generation is feasible. Other than that, reactive power generation/absorption for supplementary services could be accomplished with the inverter’s unused capacity.
Power system restoration	Smart inverters can help restore the power system by maintaining constant VAR levels while also providing real power for black starts and cranking power.
Enhancement in power transfer capability	Installing smart inverters at the line’s midpoint enables them to successfully execute shunt correction in the manner of a STATCOM, hence improving the line’s power transmission capacity. This increased capacity within thermal constraints would allow for the integration of additional DERs into the system. There are many financial gains to be had here as well, and no new electricity transmission lines are required.
Subsynchronous resonance (SSR)	Subsynchronous resonance (SSR) can be reduced with the help of smart inverters with specialized controls. Smart inverters may be able to replace the need for a STATCOM altogether.

graphically depict the supplementary grid applications of a moonlighting smart inverter. The long-term goal is to reduce the need for costly grid-strengthening measures in light of the increasing prevalence of smart inverters. Smart inverters have more of an impact than conventional inverters because of their ability to detect islanding and convert electricity to a higher frequency without missing the maximum power point. This increases the possibility for improvement brought about by smart inverters. The power grid can support more renewable energy sources with the use of smart inverters and the existing infrastructure. Some of the power quality issues that have worsened as the grid has gotten smarter include resonance and harmonic distortion, flicker, stability pertaining to the voltage and frequency, and overall reliability of the system. Power can go in both directions when DERs are broadly implemented into a smart grid. In contrast, power could only travel in one direction through the old infrastructure. This could lead to a gradual increase in voltage at the DERs’ shared connection point which is at the Point of common coupling (PCC). A smart inverter operating in VAR mode can regulate voltage spikes by injecting and absorbing VARs, keeping them within acceptable parameters.

The widespread use of Distributed Energy resources (DERs) in smart grid settings will be greatly facilitated by this. This has the potential to improve the protection systems significantly by preventing the circuit breakers from triggering due to reverse power flow. Many problems associated with power quality, such as voltage flickering, harmonic distortion, and resonance, have their origins in PV interconnections. Active power filters (APFs), such as those found in contemporary smart inverters, can reduce, or eliminate some of these harmonics. In addition, with the aid of a well-designed controller architecture, a smart inverter can perform the function of a virtual detuner, decreasing the likelihood of harmonic resonance in the network. As far as IEEE 519 is concerned, there is no such thing as separate considerations for network resonance and harmonic distortion. It seems like a situation where an intelligent inverter would be very helpful. Uneven load distribution occurs even on a three-phase network. The load balancing function can also be carried out by a smart inverter. Many groups have formed “smart inverter working groups” (SIWGs) with the goal of creating legally binding standards for the most cutting-edge inverter technologies, which are crucial to the stable and secure functioning of the grid. Two such organizations are the Electric Power Research Institute (EPRI) and the California Public Utilities Commission. Two examples of the many organizations that have joined forces to form similar committees are presented here. Additional recommendations for the care and operation of plug-in hybrid electric vehicles have been issued by the Society of Automotive Engineers. These suggestions dealt with a variety of charging methods and the dependability of the power grid.

The ability of an inverter to maintain a grid connection despite transient terminal voltages that differ significantly from the system’s nominal voltage is known as “fault ride-through” (FRT). We refer to a fault as having a “ride through” when it is present and reactive power assistance is still required. Defects are the most likely source of significant voltage changes in a power system. Transient voltage increases, as opposed to terminal voltage falls, may be the result of fault events in power systems exhibiting characteristics. Because they both result in a sizable current flowing between the phases or to the ground, short circuits and lightning strikes are two frequent reasons for voltage drops. The appliance will be disconnected from the power grid if the inverter malfunctions until the problem is resolved. If more inverters fail due to the outage, the current situation might worsen. If the inverters’ voltage excursions are very brief and the voltage returns to normal within a predetermined period of time after the excursions are complete, L/HVRT will allow the inverters to remain connected. The inverter will automatically turn off if a problem continues past the L/HVRT limit. By providing the required reactive power supply, a smart inverter that has been updated to the IEEE 1547.8 and UL 1741 Standard can operate in VAR mode. There is no need to unplug the inverter as a result.

In the event of an asymmetrical fault, such as an SLGF (single line to a ground fault) or LLGF (double line to a ground fault), the magnitude of voltage in each of the defective phases falls to zero. The reason for this is that asymmetrical faults are always located at the ground level. In the healthy phase of a distribution system, the transient voltage rises caused by asymmetrical failures can exceed the 5% safety margin for such spikes. The technical term for this event is “transient overvoltage” (TOV). Smart inverters, which employ specialized controllers, can be used to dampen the TOV phenomena and so reduce the associated risk. The local operating loads’ power factor can be corrected by the intelligent inverter, which also regulates the voltage across all three phases. Power factors for inductive loads, such as induction motors, are typically between 0.60 and 0.75. In order to keep loads running with a power factor greater than 0.9 and without incurring any fees from the utility company, these power factor correction systems frequently rely on capacitor banks as a backup. Nonetheless, harmonic distortion and resonance may result from the capacitor banks’ interactions with the rest of the system. Because of their one-of-a-kind controller, intelligent inverters can effectively alter the power factor, serving as a substitute for a capacitor. This is achieved without triggering any unwanted resonance or harmonics. All auxiliary and augmented services can be performed while the smart inverters’ capacity is underutilized. Smart grid performance may improve if dormant inverters are brought back online.

The proposed control scheme, which includes the algorithm for centralized reactive power dispatch and decentralized Volt/VAR control of a smart inverter, is depicted in a flowchart in Figure 6. An initial X/R value is determined at the PV connection point using the distribution feeder’s default configuration. This value might be changed if the distribution feeder needs to be reconfigured due to a change in the open-loop point assignment, a feeder section outage, or a switch in the capacitor bank. PV inverters can reach the necessary reactive power setpoint by using the instantaneous AC voltage value and the Volt/VAR droop control in accordance with the proposed K (X/R) function. A master/slave Modbus register mapping for the series-connected inverters can be used to manage the reactive power output from each PV inverter independently. The distribution management system’s centralized reactive power dispatch plan is shown in the diagram on the right. The DMS Volt/VAR optimization application coordinates the operations of other legacy voltage regulation devices such as OLTC and shunt capacitors in order to determine the PV reactive power compensation setpoint Qref. The proposed VVC controller modifies the Modbus registers to maintain the reactive power output from all PV inverters at a constant value after receiving the reactive power setpoint Qref via the SCADA IEC 61850 standard protocol.

The implementation of AI is used to analyze the difficulties associated with power system design, control, monitoring, forecasting, and security using a variety of techniques. The five AI techniques that are most frequently mentioned in relation to power systems are optimization, data exploration, classification, regression, and clustering. In intelligent PV plants, the best dimensions for photovoltaic (PV) arrays and energy storage systems (ESS) are determined using linear programming, and model predictive control is used to ensure optimal system performance. Furthermore, by addressing issues with optimal power flow (OPF), optimization-based methods are used to manage power system performance. Due to their superior capacity to model complex problems at low cost, optimization-based methods are also being used in the analysis of power system reliability. Network topology optimization (NTO) and dynamic thermal rating (DTR) technologies help to increase transmission assets and improve the reliability of the power system. For the best reliability planning, we then use stochastic dual dynamic programming and Monte Carlo techniques. To recognize and reduce the impact of uncertainties and interferences on system dependability, a reliable optimization model for generation and transmission is developed. Due to easier access to operational data from the power system, artificial intelligence (AI) implementation has expanded significantly and improved in accuracy. The gathered information is then used to train artificial intelligence (AI) models that can quickly identify systemic issues and abnormalities. It is also possible to determine the mission profile for the operation of PV systems the following day by using irradiance forecasting with a long short-term memory network.

A mission profile should theoretically be able to control environmental variables (such as irradiance, temperature, and humidity), energy estimation, annual power generation, and other visible results. In order to predict the operating conditions and average more accurately, minimum, and peak power outputs of PV systems, designers can consult this dataset. It should be easier to spot instances of malicious data injection into the command and control of power systems thanks to a data-driven approach to security that has been developed by researchers. The study’s findings suggest this process can be executed in an online environment using a variety of reinforcement learning techniques. With proper data organization and categorization, it is much simpler to determine what state a power system is in and what values are being measured. This is because processing massive amounts of data rapidly is crucial to the functioning of the power system. By keeping tabs on the power system in real time, its various operating stages can be categorized, and any disruptions can be pinpointed. In order to recognize and classify voltage variations and outages, a professional system analysis is also carried out within the power system. Monitoring the condition of PV modules is done in real-time using a trained database. In order to achieve this, a database of PV panel failure states is compiled, which is later used to guide an ongoing assessment. To obtain the normalized peak amplitude and phase at a sampled instant, a Fourier linear combiner is employed. Based on the collected data, diagnoses are made using fuzzy systems. Regression techniques were repurposed by analysts to be used in power system forecasting, demand-side management, and power flow analysis in order to make the most of the gathered data.

Particle swarm optimization, genetic algorithms, and artificial neural networks are used to forecast PV power. The Gaussian regression technique is used in this case to determine how the input parameters affect the output energy. To enhance power quality, a gradient descent least squares regression neural network technique is also developed. By reducing noise, keeping harmonics to a minimum, and compensating for the DC offset, this technique has the potential to improve power during both normal and abnormal grid operations. In addition to the methods already mentioned, regression-based methods can also be used to perform the power flow analysis. Additionally, the data collected from the different power system operating states is modified using clustering techniques to achieve effective modeling of the system with enhanced performance and operation. This is carried out in order to achieve the goal of successful system modeling. Scaling the various virtual power plants to determine the required power output is how the K-clustering technique is put into practice. In this instance, the distributed dynamic clustering algorithm aids in the heterogeneous deployment of ESS throughout the power system. A multi-cluster optimization algorithm is used to determine the ideal size of the ESS for PV generation while taking the uncertainty of the power system into account. such as how a hierarchical spectral clustering approach is used to examine the connections within a power grid.

The four categories of optimization, classification, regression, and data structure comprise the core components of artificial intelligence. Between AI and power electronic applications, these features act as a functional layer. The Smart inverter’s (power electronics domain) functional layer of artificial intelligence could be categorized as follows:

• Optimization: Finding the best option from a range of alternatives, taking into account constraints, equality requirements, and inequalities, to maximize or minimize objective functions is referred to as this. To accomplish this, choose the option that maximizes or minimizes the objective function. For instance, optimization can be used to investigate the ideal set of parameters that maximizes or minimizes design goals while maintaining design constraints.
• Classification: This is the process of affixing a label to a piece of input information or data that indicates which of the k discrete classes it belongs to. In particular, the detection of anomalies and the diagnosis of faults in maintenance are examples of typical classification tasks that are performed in order to determine fault labels using condition monitoring information.
• Regression: By figuring out the relationship between the input variables and the target variables, regression aims to predict the value of one or more continuous target variables given the input variables. This is done by figuring out how the two sets of variables relate to one another. For example, it may be simpler to implement an intelligent controller if there is a regression model between the electrical input signals and the output control variables. The values of the control variables can be predicted using this model as well.

Data structure exploration: Data compression is the process of projecting high-dimensional data down to low-dimensional data in order to decrease the number of features. This includes data clustering, which is the process of identifying groups of data within a dataset that are similar to one another; density estimation; and data compression. For instance, the category of “data structure exploration” can be found in the maintenance section of the phrase “degradation state clustering”. For instance, 12.4% of all maintenance tasks, 78% of control tasks, and 9.8% of design tasks all make use of AI. Optimization constitutes 33.3% of all functions, classification 6.6%, regression 58.4%, and data structure exploration 1.7%. It demonstrates that the majority of artificial intelligence (AI) work in power electronics can be boiled down to regression and optimization. Expert systems, fuzzy logic, metaheuristics, and machine learning are some of the broad classifications that can be applied to the various AI approaches. A total of 0.9%, 21.3%, 32.0%, and 45.8% are the percentage of AI applications. This data points to machine learning as being the primary area of artificial intelligence in power electronics. It should be noted that the investigation is thorough but not complete. In this context, we restrict our attention to those areas of AI that have found widespread use in power electronics and with a special focus on smart inverters. The crisp value is modified in a fuzzy space to finish the nonlinear mapping between the input and output with carefully considered design principles.

2.1. Expert System

Expert systems have advanced significantly since their inception as one of the first AI methods to be used in industry. Expert systems are essentially databases that organize specialized knowledge into catalogs of Boolean logic, which can be used to model the IF-THEN logic rules of the human brain. It is an intelligent database-based system that simulates the process of inference in order to respond to why and how queries. The information, statements, and facts in the IF database come from either real or imagined computer simulations. It has the capacity to obtain regular updates. The expert system’s reliance on system principles and rules—which are specific to the system in question and thus not universal—is the cause of this. It only works in areas with established, well-defined rules that were developed by experts. Additionally, other cutting-edge AI techniques (such as fuzzy logic and machine learning) can now take the place of expert systems, offering significantly improved inference and approximation capabilities. This is made possible by the quick evolution of computational platforms.

2.2. Fuzzy Logic

Fuzzification, rule inference, knowledge base, and defuzzification are typically the four main parts of a fuzzy logic method. First, membership functions such as triangular, trapezoidal, Gaussian, bell-shaped, and singleton are fuzzified for the input linguistic variables. Second, the inference module combines the signals in line with IF-THEN fuzzy rules stored in the knowledge base and derived from expert experience. The third step involves defuzzifying the output signal. The fuzzy rule has an Antecedent example. Z IS POSITIVE IF Y IS ZERO AND X IS MEDIUM. The degree of fulfillment for both the antecedent and the consequent is determined by the membership functions. Fuzzy inference schemes fall into the Mamdani type and Takagi-Sugeno-Kang-type (TSKtype) categories. In the Mamdani-type fuzzy inference scheme, the membership functions of the antecedent and the consequent are shape-based functions, such as triangles. The membership function of the TSK-type fuzzy inference scheme’s antecedent component is the same as the Mamdani-type, whereas the consequent component’s membership function is singleton at various constant values. For the same task, the Mamdani-type scheme typically needs more fuzzy sets than the TSK-type scheme. The TSK-type scheme’s membership function can be functionally typed as either linear or constant, which is more accurate and effective in nonlinear approximation than the Mamdani-type’s fuzzy terms. Fuzzy logic can be used to handle prior knowledge and expert experience before being combined as a hybrid method with other AI techniques, according to this point of view.

2.3. Metaheuristic Methods

Once the optimization task for a particular application has been defined, the optimal solution can be found using either a deterministic programming method or a non-deterministic programming method, such as the metaheuristic method. The gradient and Hes-sian matrices are difficult to use for most optimization tasks in power electronics because they are complex and difficult to calculate using deterministic programming methods. Metaheuristic methods serve as a general, end-to-end tool that is efficient and scalable for a variety of optimization tasks while requiring less specialized knowledge.

The development of metaheuristic methods frequently draws inspiration from biological evolution. Examples include genetic algorithms that use the natural selection process and ant colony optimization algorithms (ACOs), which mimic ants to find the most effective route to food. The exploration of the ideal solution is sparked by the iterative procedure. The two types of metaheuristic methods are population-based methods and trajectory-based methods. Examples of the former are the tabu search method and the simulated annealing method. For the trajectory-based methods, each exploration stage only contains one candidate solution, which then transforms into another solution in accordance with a set of rules. The standard and efficacy of the rule play a major role in determining the effectiveness of the method. Because of this, for non-convex optimization problems, the final solution is frequently a local rather than a global one, and the trajectory-based methods typically take a while to converge. The population-based methods produce a large number of random candidate solutions.

These candidate solutions are changed at each iterative exploration (for example, crossover in the genetic algorithm) or added and replaced with new candidate solutions in order to improve the quality of the population at the current generation. In order to get closer to the ideal solution, the population’s suitability is consequently gradually increased. They are especially useful for challenging optimization tasks because they are faster at convergent convergence and global searching than trajectory-based methods. The computational burden is greater for population-based approaches. This challenge needs to be considered for online application scenarios where effectiveness and speed are essential. The metaheuristic methods qualitatively differ from one another in terms of several important characteristics, including implementation, simplicity, global convergence, convergence speed, and parallel capability.

Population-based methods are used to solve most optimization problems in power electronics because of their notable benefits. They are developed using a variety of biological inspirations. In addition to the widely used metaheuristic techniques, several other recently developed approaches, such as biogeography-based optimization, crow search algorithm, grey wolf optimization, fire-fly optimization algorithm, bee algorithm, colonial competitive algorithm, teaching-learning-based optimization, etc., have also been used on a small scale. It is significant to remember that deciding which approach is best depends on the application. Particle swarm optimization and genetic algorithms are the two metaheuristic methods that are most frequently used in power electronics that use smart inverters. These techniques serve as the foundation and models for evolutionary algorithms and swarm intelligence algorithms, respectively, on which different variants are built. Although there is no guarantee that metaheuristic methods will lead to a global optimum, most practical applications will find the solution to be generally satisfactory and acceptable.

2.4. Machine Learning

Machine learning is an approach to programming that seeks to automatically learn new skills by gleaning knowledge from existing data or interacting with a system to see what happens. In power electronics, it is broken down into the three subfields of supervised learning, unsupervised learning, and reinforcement learning.

(1)

Supervised learning

With the aid of a training dataset made up of input-and-output pairs, supervised learning aims to establish the implicit mapping and functional relationships between inputs and outputs. This is especially useful in power electronics applications where it can be challenging to define a system model. Activities such as regression analysis and classification are frequently included in supervised learning. The output of the input and output pairs in the training dataset deals with a limited number of discrete categories when used for classification. Fault diagnosis for a multilevel inverter, where the discrete fault label is determined using the input fault information, is a typical illustration of a classification task. The input and output pairs of a regression task produce one or more continuous variables. Regression is well-suited for predicting how long an IGBT will last before needing to be replaced because the result, in this case, the residual useful lifetime, is a continuous variable. The model can be used to analyze data that differs from the set used for training after it has been trained.

A model’s capacity to handle novel data, such as that present in the testing dataset, is referred to as generalization. Given that the training dataset typically only contains a small number of potential input-and-output pairs, generalization on new inputs is one of the most important performance factors for supervised learning techniques. The three main categories of supervised learning techniques are connectionism-based methods (such as the neural network method), probabilistic graphical methods, and memory-based methods. For neural network methods, the data from the training dataset is used to refine and transfer the connection weights and architecture of the network. There has been a lot of research done to improve neural network performance. These improvements offer two advantages for the power electronics sector. Allowing the neural network’s uncertainty capability to deal with the noisy signal is the first step in strengthening the method. By incorporating fuzzy logic into the neural network, the fuzzy neural network, and its variations (such as the adaptive neuro-fuzzy inference system (ANFIS)) make this possible. Neural networks need to improve their dynamic performance in order to handle time-series dataset cases such as intelligent controllers and remain useful in life predictions. Transient performance can be enhanced over a traditional neural network’s completely decentralized structure by allowing weights to be shared between layers and cells. With a convolutional structure (1-D convolutional neural network, time-delayed neural network), weight-sharing can be implemented on a shallow scale. With a recurrent unit in a recurrent neural network, it can be implemented on a full and deep scale. Recurrent unit implementations typically have better modeling abilities than ones that use a convolutional structure. The probabilistic graphical methods infer information from the data by representing input and output pairs graphically. The graphical representation implicitly depicts the conditional dependence relationship between the decision variables. A Bayesian framework is used to state the fundamental relationship of the model, from which probabilistic deductions can be made. This indicates that the model is significantly simpler to comprehend than those created using neural network techniques. Additionally, the probabilistic graphical model performs better in the presence of ambiguity and sparse data. One typical probabilistic graphical technique is the Bayesian network.

When training is complete, the training dataset is thrown away for neural networks and graphical methods. As opposed to kernel methods, which throw away the training dataset after the training phase is over, support vector machines (SVMs) retain the training dataset and use it in the testing phase, and the knowledge they impart is gleaned through the selection of key data points (called support vectors) or a subset of the training dataset. Gaussian processes, a common kernel method, have been applied to the problem of predicting how long an IGBT will continue to function after its useful life has ended. Traditional kernel methods (such as Gaussian processes) are computationally expensive because the entire training dataset is used in the testing phase. Support vector machines (SVMs) and relevance vector machines (RVMs) are two sparse solutions proposed to reduce the computational overhead; in these models, Bayesian methods are used to refine the parameter estimation. The sparse solution outperforms traditional kernel methods because it uses a smaller subset of the training dataset in the testing phase. In comparison to neural network approaches, training datasets for kernel methods tend to be smaller in size. For this reason, cases with a limited dataset benefit more from the use of kernel methods. The kernel methods require more memory than the neural network methods because the training dataset is required in the testing stage. Training dataset usage slows down testing results as well. Online applications (such as control applications) where execution time is crucial may want to give it some thought.

(2)

Unsupervised learning

Unsupervised learning lacks any output data for the learning target during the learning process, in contrast to supervised learning, where the dataset is made up of input-and-output pairs. Data clustering and data compression are two categories that unsupervised learning tasks typically fall under when applied to power electronics applications. In order to cluster data, it investigates the regularities that are present in the smeared dataset and separates the dataset into a number of different groups or clusters based on the similarities that exist between them. In this way, the data characteristics within a cluster are comparable to one another, whereas the data characteristics within other clusters are different from one another. The identification of the discrete health state from the continuous degradation data in the context of power electronic converter condition monitoring is a good example of a typical application of data clustering. Data compression aims to reduce the number of features in a dataset by eliminating extraneous data that is present in the dataset. The dataset’s integrity can be maintained while obtaining a reduced representation of the dataset using principal component analysis (PCA), for example, with a significantly smaller number of features. The number of features in the representation can be decreased to achieve this. These unsupervised learning algorithms typically have the purpose of preprocessing data before sending it on to the subsequent analytics. Even though this step is optional, it can help decrease the amount of computer work required and increase the accuracy of the analytics.

(3)

Reinforcement learning (RL)

Because reinforcement learning (RL) does not require a data set to learn from, it is an alternative to conventional learning techniques such as supervised and unsupervised learning. Rather, it is essentially an optimization or dynamic programming problem that looks for a suitable action plan that maximizes reward for a particular task. This outcome-focused methodology comes from continuously improving one’s interactions with simulation models or systems. It will eventually have accumulated sufficient data to learn a strategy that will produce the best results in relation to the goal you have established. Theoretically, Markov decision-making can be compared to real-time learning (RL). A Q-table in the form of an action selection policy is produced by RL training in order to maximize the total expected rewards over time. The Q-table, a useful policy matrix, details the best course of action considering the pertinent condition variables. It is crucial to remember that RL relies on the interactions between systems rather than on pre-existing datasets. As a result, it is preferred in circumstances where the system has little information or where developing its model is challenging. This is because supervised learning, a flexible tool, is used in most machine learning-related applications in power electronic systems. Figure 7 displays the artificial intelligence methods related to the power electronics field that could be used for smart inverters. Figure 8 depicts the function layer connected to the AI methodologies used in power electronics, and Figure 9 shows the mapping of the function layer of AI used in power electronics design, control, and maintenance that is relevant to smart inverters.

The mapping of functional layers of AI-to-AI methods must handle power system design, control, monitoring, forecasting, and security issues [19,20,21,22,23,24,25,26]. The design, control, and maintenance could also be referred to as three life phase cycle domains of smart inverter systems (applicable to all power electronic devices) are depicted in Figure 9. Five AI uses can improve power grids. Uses include optimization, data exploration, classification, regression analysis, and grouping. In recent years, research in academia has focused on using AI technology to optimize systems and regulate them. Model predictive control stabilizes smart PV facilities, and linear programming-based optimization establishes the best solar panel-to-battery ratio. More data improves performance, which underpins these two methods. Optimization-based approaches also solve power system optimal power flow (OPF) issues. Due to their enhanced ability to simulate complex issues at lower costs, their influence has expanded to power system reliability studies. Optimization-based method development is key. Network topology optimization (NTO) and dynamic thermal rating have enabled power system dependability and transmission capacity increases (DTR). Monte Carlo and stochastic dual dynamic programming are used to make the most accurate and complete dependability assessments. A reliable generation and transmission optimization model is needed to identify and minimize interferences and uncertainties that affect system dependability. More power system data has improved AI’s accuracy. Figure 10 presents the three life phase cycle domains of smart inverter systems that is also applicable to all power electronic devices. Figure 11 manifests the application of AI-ML on a smart inverter for analysis and further action in the form of ancillary services.

The acquired data teaches AI learning algorithms to easily spot system flaws and outliers. The Bayesian ascending algorithm-based data-driven method can achieve this. A network with long-term and short-term memory can forecast PV system mission profiles. We use approximated irradiance. An ideal mission profile would supervise visual outputs including expected energy output, annual power generation, and other climate-specific information (such as irradiance, temperature, and humidity, amongst others). This information is useful for PV system developers to extract the minimum, average, and maximum power outputs and PV operational parameters. This data set helps explain PV functional features. Data-driven security can also detect malicious data injection into power system control. Virtual reinforcement learning approaches may enable this.

More precise predictions of the power generation supplied into the grid are required due to an increase in photovoltaic (PV) systems connected to the grid in recent years [27,28,29,30,31,32,33,34,35,36,37,38,39,40]. The main cause of the increase is the decline in investment costs, which will be between 10–20% between 2019 and 2022. Other factors that have contributed to the increase include incentives, restrictions on technical requirements for building works, and other directives. Grid-connected photovoltaic (PV) systems will increase power outages and possibly lead to grid instability because of the unpredictable nature of the weather. This growth is anticipated to last for several more years. Because of the liberalization of the electricity markets, which has significantly helped to rebalance supply and demand, spot markets for electricity have also emerged. It is imperative that local communities, major generators, retailers, and big-end consumers all provide accurate forecasts of their production and consumption. To accomplish this, these market participants have heavily relied on a variety of forecasting methodologies.

When the whole system is considered, it is possible to build energy markets and re-serve control grids in such a way that electricity generation and consumption are in sync. Even if PV’s popularity has increased, so has the difficulty system owners and managers confront in keeping up with installations. This is because both solar irradiance and PV outputs are notoriously difficult to anticipate. In addition, the study finds that if a generator or retailer fails to achieve their expected output or demand, they will have to turn to the balancing market and pay very high fees to remedy the imbalance. Since the balancing market is the only option when supply and demand are mismatched. Due to these obstacles, there is widespread consensus that reliable forecasting models are essential to the development of more effective market systems. Preliminary studies showed a wide variety of applications for outcome prediction in the context of managing distributed energy resource networks. However, most efforts have been put toward load forecasting rather than looking at the outcomes of distributed energy resources. Several studies have found that accurate load forecasting can improve system effectiveness. Research that goes beyond using prediction outputs alone is being conducted to make electricity systems more reliable. The results show that the weather, an inherently chaotic system, has a significant role in forecasting. This makes it difficult to foresee outcomes for extended periods of time, such as the next season. In order to better understand the behavior of dispersed energy supply and loads, intelligent methodologies based on statistical and stochastic models were developed.

3. Artificial Intelligence Applied to PV-Based Smart Inverter Systems

3.1. Artificial Intelligence Models for Solar Irradiance Forecasting of PV Interfaced with Smart Inverters

The viability of employing artificial neural networks (ANNs), support vector machines (SVMs), k-nearest neighbors (k-NNs), and deep learning (DLs) to predict solar irradiance. The ANN algorithm fared better than DL, SVM, and kNN when it came to fitting the data. Grid operators, prediction horizons, and solar irradiance forecasting models are all investigated, as are their interactions with one another. The solar perimeter sub-image is obtained radially, and then the irradiance of the sun can be estimated using deep learning and gradient-boosted trees. Successful predictions of solar irradiance have been made using a wide range of DL models. Long short-term memory (LSTM) neural networks are one technique for forecasting future solar irradiance. We were able to fairly distribute our limited resources among a wide range of jobs by employing the tried-and-true method, which improved performance across the board and produced better metric results. In order to forecast solar irradiation, it uses convolutional long short-term memory (LSTM) networks with wavelet decomposition. The wavelet decomposition divides the raw solar irradiance into several subsequences, which aids in network model optimization. This method improves the precision and efficacy of traditional DL-based forecasting techniques. Time series models for forecasting solar irradiance can be made with the aid of DL techniques. In order to increase accuracy, performance, and predictability, the developed models consider univariate data from both single- and multi-location sources. One can build a reliable forecasting system by adjusting the ANN model for the current season.

The newly discovered strategy makes use of Pearson correlation to help input pertinent data into the ANN model. This improves the model’s computational capabilities, allowing for more precise predictions, which is especially helpful when dealing with extreme outliers and dynamic scenarios. Possible strategies for accurately predicting future sun irradiation include investigating the use of Gaussian process regression. Probabilistic renewable energy management systems may now be developed, which will facilitate the functioning of energy trading platforms and provide crucial support to smart grid operators according to the findings of this study. Stochastic power systems, which may be identified by their intrinsic unpredictability, can be developed thanks to the findings of this study. In addition to deep learning and neural networks, other common ways to estimate solar irradiance include multigene genetic programming and machine-learning classifiers such as the multi-layer perceptron neural network and the Naive Bayes method. The advent of deep learning and neural networks paved the way for the development of both techniques [41,42,43,44,45,46,47,48,49,50].

3.2. Artificial Intelligence Model for Controller Action of the Smart Inverters

In order to reduce harmonic distortion, the inverter controller is in charge of managing the power and frequency output on the AC side of the inverter. An inverter control algorithm is used to activate the inverter’s switches. AI improves the responsiveness of inverter controllers to transient errors and the precision of controllers. Standard controllers make use of PI and PR-based algorithms; however, when integrated with AI, these controllers significantly enhance. The PID controller’s accuracy and overall effectiveness as a robust controller benefit greatly from the addition of fuzzy logic inverter control. As the ANFIS-based inverter controller is being detailed, a simulated version using artificial neural networks is being tested. Most inverter controllers using AI technology produce power with a low total harmonic distortion (THD).

Natural disasters, physical attacks, or cyber-attacks could cause the utility electricity that is required for the operation of power systems to go out of service. Isolated microgrids need inverters that can determine if there is a disruption in the utility’s power supply in order to operate properly. A local controller is built into each individual inverter with the responsibility of keeping track of the terminal voltage and reporting any potential changes. It is very likely that grid power has been lost when the measured voltage and frequency are outside of the typical operating range established by IEEE Standard 1547-2018 [51,52,53,54,55,56,57,58,59,60]. Comparing the measured values to the operational range will reveal this. The IEEE advises figuring out how long the voltage has been outside of the typical parameters before attempting to fix the voltage issue. This is a requirement for taking part in any activity. The disconnecting of the connection between the inverter and the power grid is carried out by the static transfer switch. By measuring the voltage buildup at the opposite end of the STS, it is easy to resynchronize a single inverter when it is operating in islanded mode. Resynchronization is a remarkably easy process as a result. Two different PI controllers can be given information about the voltage amplitude and phase differences that have been found across the STS. These controllers’ output signals would be used to modify the voltage of the inverter’s amplitude and phase. High current transients may be avoided, and the system may be able to gradually resynchronize with the grid. The synchronization information might not be available if the STS is placed a long way from the inverter. Because of this, the microgrid requires the presence of a supervisory controller. Through low-bandwidth communication signals, this controller must be able to provide synchronization data. Using systems that completely ignore the grid side of the STS and only rely on measurements of current and voltage taken at the inverter is also conceivable. To reduce the amount of transient overcurrent generated, these systems typically involve a quick detection of the grid phase angle.

3.3. Artificial Intelligence-Based Monitoring System for PV Plants

Precise identification of the various stages and parameters of operation is produced by effective data management and classification. The operation of the electricity grid depends on the ability to analyze large amounts of data quickly and effectively. Continuous real-time monitoring of the power system’s characteristics enables the classification of the system’s various operational phases and interruption detection. The electricity system also undergoes an expert system analysis to detect and categorize voltage fluctuations and outages. It is possible to monitor PV module health by collecting data on PV panel failures under different settings and then evaluating that data in real-time with a trained database.

A Fourier linear combiner is also used to calculate the normalized peak amplitude and phase during the sampling process. Next, fuzzy systems are used to use the data for diagnostic purposes. The collected data can now be used in power system forecasting, demand side management, and power flow analysis more successfully by modifying regression methods. As a result, the information that was gathered became more valuable. A genetic algorithm, particle swarm optimization, and an artificial neural network are used to forecast the power generated by PV systems. Here, we employ the Gaussian regression method to investigate how different input factors affect the output voltage. Additionally, a gradient descent least squares regression-based neural network technique is developed to improve power quality. By reducing noise, lowering harmonics, and adjusting for the DC offset, this technique can increase power during both regular and exceptional grid operations. In addition to the methods described so far, regression-based techniques can also be used to perform the power flow analysis.

In addition, the data collected from the power system’s various operational states is used to refine the clustering methods, leading to a more precise model with enhanced performance and operation. The K-clustering technique is used to determine the necessary amount of output power. With this strategy, several virtual power plants’ capacities are increased. To manage the diverse ESS deployments across the electrical grid, we employ a technique called distributed dynamic clustering. Utilizing a multi-cluster approach, we can maximize PV generation with the smallest possible ESS footprint. This enables us to do so while accounting for the inherent unpredictability of the electricity grid. Furthermore, a hierarchical spectral clustering approach is used to examine the various linkages within the power grid.

3.4. Artificial Intelligence-Based Protection System for Islanding Operation and Fault Ride through Capability

The protection mechanism is designed to cut power to the inverter-based DGs such as PV as soon as an issue is detected. There is a decreasing detectable area and an increasing detectable duration. An anti-islanding defense strategy can be active, passive, or hybrid, depending on how involved the user is in the anomaly detection process. Identifying active islanding requires perturbing the system in issue and paying close attention to the resulting signal. However, when applied to a multi-inverter system, the active technique presents challenges and raises power quality problems.

By closely monitoring the system’s operational metrics, the passive islanding detection method seeks to pinpoint the most likely causes of a problem. Once the threshold has been established, incorrect classification will endanger islanding operations. This led to the development of a third technique, which uses the threshold to find anomalies and then disturbs the system to confirm the discovery. The development of this method was motivated by the drawbacks of the first two approaches. This method can also be referred to as “hybrid islanding detection”. Finding outliers through the combination of the two is effective but time-consuming. By first analyzing the incoming signal to create a database of all possible abnormalities and then instructing the classifier to identify the operating condition by evaluating the real-time signals, artificial intelligence offers a quicker and more accurate approach to the identification of abnormalities than traditional methods. This tactic is justified considering the provided constraint. The accuracy with which problems can be localized is improved by preprocessing the signals and extracting features to expand the data matrix and identification capabilities.

It is not advised to disconnect DGs as soon as a grid anomaly is discovered because doing so could jeopardize grid stability. Due to the grid codes’ support for fault ride-through or low voltage ride-through, the PV system must continue to be connected to the grid and inject a reactive current. This helps with efforts to restore voltage stability. To activate the ride-through function, one can make changes to the inverter’s controller or use an accessory such as a flexible alternating current transmission system (FACTS) device such as a STATCOM [61,62,63,64,65]. LVRT can be made functional with just a small, inexpensive change to the controller. A dual-current controller is required to control the inverter’s reactive power injection and polarity reversal in the event of an error. A droop-based LVRT technique is also applied. In this method, in the event of a dip in DC link voltage, the controller switches from the maximum power point tracking (MPPT) mode to the ride-through mode. The FACTS device and inverter control are introduced with the intention of synchronized control of reactive power injection. Both the injection requirement and the priority distribution have an impact on them.

Artificial intelligence (AI)-based methodologies such as fuzzy logic control (FLC) and computation-based techniques such as particle swarm optimization (PSO) are used to enhance the inverter controller’s LVRT capabilities. FLC-based control reliably achieves LVRT using a vector control plot for the DC link voltage, whereas PSO tends to increase the nonlinear system’s LVRT capacity.

3.5. Artificial Intelligence-Based Approach for Maximum Power Tracking

The inverter must draw the most power from the PV array during DC/DC conversion. The perturb and observation method uses a hill-climbing algorithm to find a PV curve’s maximum value. The system cannot be trusted due to the increase in step size. The optimization and control of mission profiles can happen more quickly thanks to automatic intelligence. The fuzzy logic controller and neural network controller both follow the same maximum operating point when taking the mission profile into account, thanks to the genetic algorithm’s optimization. Transients show harmonics and disturbances in the tracked power output, whereas power consumption shows output power loss caused by a specific MPP algorithm. The research demonstrates that while some MPPT algorithms, such as the P&O and incremental conductance approaches, are easy to construct, they have drawbacks such as slow response times, significant power loss, and output transients. The grid will be cut off from the DGs once a problem has been found and the LVRT has failed to fix it [66,67,68,69,70]. Limiting transient voltage and preventing frequency runaway requires controlling DG disconnects and reconnects. Both modes of operation are controlled by a single control structure, with the outer loop acting as a reference generator for the current loop when it is operating independently and a static control switch facilitating controller switching. With static switch basis control methods, there is a sizable delay in response and transients. Artificial intelligence-enabled methods allow for seamless switching between modes. A transition controller based on fuzzy logic (FL) is used to establish a reference trajectory and smooth the transition, in addition to a model predictive control (MPC)-based transition controller with stable output and simple implementation.

3.6. Artificial Intelligence-Based Approach for the Failure Diagnosis of a PV System

Recent years have shown the reliability and efficacy of AI-based, data-driven, intelligent fault classification approaches for diagnosing failure in grid-connected PV converters. Multilayer H-bridge inverter power switch failures are categorized using a neural network (ANN). With the help of Digital Wavelet Transform (DWT), information such as signal strength, energy, etc. may be gleaned from inverter output voltage data. Then, an ANN is trained using one input layer, one output layer, and a hidden layer. Faults in grid-connected PV systems are categorized using a radial basis function network (RBFN). The time series of data from the inverter’s output is preprocessed using a wavelet approach. The Radial basis function network (RBFN) with a Gaussian kernel is fed data by these features. Using supervised learning, a Probabilistic Neural Network (PNN) can detect problems in diode-clamped multilevel inverters. The Daubechies order 4 (db4) mother wavelet is used in DWT feature mining. After that, we use a PNN with many feedforward layers and no iterations to fine-tune the weights. Intelligent condition monitoring for grid-connected PV systems is achieved using Multilayer Perceptron Neural Network (MLPNN). Data on the voltage and current flowing through a failed inverter switch are used by DWT to calculate characteristics. Dimensionality is reduced via principal component analysis (PCA), and only relevant features are recorded. PV inverters that are connected to the grid rely on fault prediction methods such as rapid clustering and the Gaussian mixture model. Information on the inverter’s current, voltage, and IGBT temperature as they occur in real-time. Defects can be predicted with the help of the Gaussian mixture model, and clusters of comparable data can be organized using the fast-clustering method. In this study, we introduce a modified CNN GAP (global average pooling) technique for inverter switch failure diagnosis. The CNN GAP model receives 1D, first-dimensional time series data directly from the inverter. Multiple convolutions and layer pooling are used to create 2D feature maps for the input layer. The diagnostic outcome is acquired in the output softmax layer after the GAP layer has compressed the output image.

3.7. Artificial Intelligence-Based Big Data and Analytics Support for PV System

Using the Internet of Things (IoT), smart devices, and artificial intelligence (AI) data mining inside the Digital Twin (DT) framework enables data-driven product design, manufacturing, and servicing. The DT framework is the result of an IoT platform combining physical system data with historical data from a PV system connected to the grid. Safer transition networks and better data collecting and analysis are made possible by the smart, industrial, and energy internet. Because of this, DT frameworks become more efficient and focused on the needs of their end users, improving AI’s data-handling capabilities. Due to their inherent randomness, raw data are not ideal for use in most PV system applications. That is why it is crucial to use the DT framework to obtain your data in tip-top shape before you try to extract useful features. These characteristics should take PV system dynamics into consideration without compromising the uniqueness of the observed data.

Besides economic dispatch, a supervisory or tertiary controller can exchange synchronization data to integrate microgrids into a grid. A supervisory framework incorporating weather forecast data can also mitigate intermittent renewable energy issues. Wind turbines cease rotating at their wind speed limit. A big cloud passing overhead might also reduce PV array production unexpectedly. Losing a wind or solar farm could cause frequency and voltage swings due to the increased amount of renewable energy. If the weather forecast is known in advance, solar and wind farms can be gradually shut down while the remaining sources, especially high inertia synchronous generators, are gradually ramped up to avoid under-frequency trips.

3.8. Importance of IEC 61850 Standard and Digital Twin towards Smart PV Inverters

Grid-forming and grid-following performance are both enhanced by integrating smart inverters with external data sources. In order to send information to the utility operator, smart inverters use a communication network, which leaves them open to hacking and human error. In order to overcome this issue, a reference system (model) is needed to tell the difference between legitimate power company setpoints and those used by fraudsters. Using a message authentication code (MAC) for encrypted transmission guarantees that a setpoint came straight from the utility and was not tampered with in transit. It is possible to hack the utility computer and transmit setpoints using encrypted tags. Self-protecting inverters can be built according to standard operating parameters and grid regulations. Based on the expected output’s safety, the inverter can determine whether to activate the new setpoints or not. Cyberattacks may prompt more sophisticated countermeasures. Stable, expandable, low-latency, high-range, and sufficient-data-rate communication networks are required for grid-interactive inverters. Connectivity between IEDs, such as smart inverters, is enhanced by IEC 61850 [71,72,73,74,75]. The information models used by GOOSE, MMS, and SMV can all be mapped to IEC 61850. Information Exchange Devices (IEDs) use the generic-object-oriented-substation-events (GOOSE) protocol to publish and subscribe to data, while the manufacturing-message-specification (MMS) protocol is used for client/server communication between IEDs and the utility operator to transfer real-time data and supervisory control data, and the sample-measured-values (SMV) protocol transmits digitized signals from measurement units. Processing and end-to-end data packet transmission delays render centralized control ineffective against fast dynamic phenomena. It is possible for hackers to cause widespread instability by manipulating end-to-end packet delays between smart inverters, sensors, and utilities. Wireless networking can be established entirely via wireless means, or it can be a hybrid of wireless and wired connections, using technologies such as cellular or Wi-Fi. While wired connections are more secure against electromagnetic interference (EMI), they are also less scalable. Although it is the slowest kind of wired communication, power line communication is also the most affordable. Typically, this technology has been used in relay and protection systems. When operating in islanded mode, inverters that rely on power lines to exchange data risk losing this information. Although a mesh wireless network has a higher tolerance for failures without data loss, transmission performance may be lowered due to routing. Inverters can connect to other inverters and smart gadgets in the area via sparse communication, streamlining processes, and allowing for expansion.

When a physical object and its digital counterpart are connected to one another via the internet, they form what is known as “digital twins”, which can share and receive data in real time. This digital duplication can also have real-time conversations with its counterpart. One way to think of a digital twin is as a software copy. Because of their ability to facilitate bidirectional data flow, digital twins simplify the process of merging the digital and physical realms. In addition to improving the accuracy of the twining process, it also aids system operators by allowing for real-time monitoring and control, which boosts the overall performance of the underlying physical system. These advantages emerge from the enhanced functionality of the underlying physical system. VPPs and other organizations can benefit from digitizing EESs in several ways, including the ability to anticipate and prepare for change, increase security, and take part in wholesale energy markets that are abundant in DERs [76,77,78,79]. Additionally, enhanced productivity helps VPPs and other businesses. Blockchain technology, digital twins, and massive amounts of data are used to achieve this purpose. Data is gathered in real-time from the physical asset using Internet of Things (IoT)-connected sensors and communicated to the digital twin in a two-way interaction. Due to its multi-source, multi-scale, noisy, and heterogeneous nature, collected data requires a strong reliance on big data analytics that is inherent in cloud computing-based data processing. In the future, feature extraction and data fusion will be achievable using technology that employs artificial intelligence on large datasets. By incorporating data from both physical and digital sensors, this model can dynamically recognize, forecast, optimize, and regulate any process. What this suggests is that the concept of the digital twin will soon merge with others, such as IoT, big data, AI, and data fusion. If a digital item is to be compared to its physical counterpart, the two must have functionally identical capabilities. One can judge success based on how quickly a digital twin model can be built, how efficiently physical assets are used, and how accurately prognostics and diagnostics can be carried out. Digital twins are being used in more and more contexts, from smart inverters. ISO 23247-1 is a standard for the automation system and integration of the digital twin framework for manufacturing, published by the International Organization for Standardization (ISO) (ISO). Several other normative documents exist, such as ISO 10303, ISO 13399, and OPC Unified Architecture.

It is essential to gather historical and real-time data streams from sources such as weather stations, satellite imagery, consumer behaviors, and volatile electricity prices in order to develop accurate short- or long-term forecasting models for renewable power generation, load pattern recognition, and electricity tariffs in EESs. This is necessary in order to achieve the goal of developing accurate forecasting models for renewable power generation. In addition to this, it is necessary to gather charging sessions from electric vehicle charging stations. The Internet of Things makes it possible for vast information and communication networks to come together to create this high-velocity, real-time heterogeneous data streams. These data streams are produced as a result of these networks. When applied to massive amounts of aggregated raw data, the application of artificial intelligence algorithms, big data analytics, and data fusion methodologies can result in the development of relevant insights and the facilitation of enhanced decision-making. The DT model of Electrical Energy Systems (EESs) that was developed will aid network operators in the following areas: steady-state evaluation; identification of extreme events; monitoring the status of health, and making decisions with confidence in response to rapid changes in the system. Before anything such as this can be performed, the forecasting model needs to be validated first. It is also able to provide advice on the most efficient way to set up EESs, reduce the strain placed on those systems by distributing resources to areas in which they are required the least, and postpone the date on which expensive repairs and upgrades are necessary to be performed. Additionally, independent infrastructures, such as electrical systems, can perform intricate coordination with their neighbors by utilizing DT-DT communication. This is possible because of the independence of these systems.

4. Conclusions

The inverter regulates voltage and maintains grid limits. When utility power fails, grid-adapting resynchronizes and self-heals. Developers added more cyberattack-prevention features. Smart inverters fix internal problems to prevent power outages. Grid-connected inverters are proactive and reactive. Smart inverters benefit from AI-based research projects focusing on system-level issues in the solar PV value chain. ANNs and other task-dependent architectures power today’s AI. ANNs can predict time-series irradiance and power. GAs and population-based optimization are new. A smart inverter meets all the above criteria. AI-powered solar plants can work with smart grids and backup generators. This paper describes AI techniques used in smart inverters. Recent discoveries on the Design, control, and maintenance of smart inverters can be categorized using AI. Expert systems, fuzzy logic, metaheuristics, and machine learning are four types of AI used in power electronics. Relevant AI algorithms are compared in depth, including their use, benefits, and limitations. AI applications optimize, classify, predict, and explore data structures. Difficulties and future research areas are highlighted alongside life cycle examples. This perspective paper presented an overview of the effectiveness of AI-powered solar plants, also known as smart inverters, along with additional functions and moonlighting features that comprise a smart grid.