Control Methods and AI Application for Grid-Connected PV Inverter: A Review

Abstract

Grid-connected PV inverters (GCPI) are key components that enable photovoltaic (PV) power generation to interface with the grid. Their control performance directly influences system stability and grid connection quality. However, as PV penetration increases, conventional controllers encounter difficulties in managing nonlinear dynamics and weak-grid conditions. This paper reviews both conventional and artificial intelligence (AI)-based control methods for GCPI. It compares their performance characteristics, application scenarios, and limitations and summarizes current research progress and remaining challenges. The potential and issues of applying AI to enhance system intelligence are also highlighted. Finally, future development trends are discussed, emphasizing high efficiency, strong adaptability, and intelligent integration in GCPI technologies.

1. Introduction

The accelerating global energy transition toward decarbonization and sustainability has become a strategic priority for many countries, driven by international agreements such as the Paris Accord and the United Nations’ Net-Zero 2050 roadmap [1,2]. According to the IEA, the share of renewables in global electricity generation is projected to reach about 43–46% by 2030, with PV emerging as the fastest-growing renewable technology because of its renewability, cost reduction, and modular deployment advantages [3]. Figure 1 shows the distribution of renewable energy by technology from 2000 to 2030. As the penetration of renewable energy sources increases, a high proportion of power electronic devices has become a defining feature of modern power systems, bringing challenges for grid stability, voltage regulation, and harmonic distortion, especially as PV systems become more decentralized and integrated into weak-grid environments [4,5,6].

To ensure secure and efficient integration, many countries and regions have established standards for inverter performance. Standards such as IEEE 1547, IEEE 519, and IEC 62116 define technical requirements for converter design and grid compliance [7,8,9]. These standards specify steady-state indicators, including limits on total harmonic distortion, power factor, and DC injection, as well as dynamic indicators such as fault ride-through (FRT) capability and grid-support functions for voltage and frequency [10].

GCPIs are essential for maintaining stable power exchange between renewable sources and utility grids, and their control strategies strongly affect system efficiency, reliability, and dynamic performance [11]. Conventional controllers such as PI, PR, Sliding Mode Control (SMC), and Model Predictive Control (MPC) are widely applied in PV systems [12,13,14,15,16,17]. In recent years, AI-based techniques, including fuzzy logic control (FLC), neural networks (NNs), adaptive neuro-fuzzy systems (ANFISs), reinforcement learning (RL), and metaheuristic algorithms (MAs), have gained increasing attention [18,19,20]. While conventional methods are simple and reliable, they lack adaptability to nonlinear and time-varying conditions [21]. With advances in AI and data-driven modeling, intelligent approaches can perform adaptive tuning and optimization without relying on precise mathematical models, improving inverter performance under complex grid conditions [22,23,24].

Many reviews on GCPI control have examined both traditional and intelligent control strategies. Conventional surveys mainly focus on model-based controllers, explaining their mechanisms and improvements under various grid conditions [10,25,26]. However, these studies often evaluate traditional methods independently, without considering how intelligent optimization or hybrid integration could enhance their performance. Reviews on AI-driven control highlight self-learning and model-free adaptability in renewable energy systems, but most concentrate on algorithmic design rather than comparative performance with classical controllers under consistent conditions, including weak grids, harmonic distortion, or hardware constraints [27,28].

A systematic evaluation of AI-optimized traditional control frameworks is still lacking. Applications such as using AI to tune PI/PR parameters, modify MPC weighting factors, or reduce SMC chattering have not been comprehensively reviewed. Although hybrid approaches like FLC–MPC, NN–SMC, and ANFIS–PI have improved transient response and harmonic suppression, their comparative benefits remain insufficiently quantified.

To address this gap, this review develops a unified framework that connects traditional, AI-based, and hybrid control strategies. It compares their structures, performance, and limitations under consistent criteria and summarizes future directions for intelligent, adaptive, and hardware-efficient GCPI control in next-generation smart grids. The paper focuses on single-phase and three-phase inverters under high renewable penetration and low inertia, emphasizing both model-based and AI-based data-driven algorithms that enhance power quality, stability, and real-time adaptability in weak-grid conditions.

Finally, this paper reviews the control methods of GCPI and the applications of AI. The remainder of this paper is organized as follows: Section 2 introduces the literature screening and workflow. Section 3 describes PV grid-connected systems and explains the principles and differences between grid-forming inverters (GFMIs) and grid-following inverters (GFLIs). Section 4 discusses conventional control methods, and Section 5 presents AI-based control strategies. Section 6 provides a comprehensive discussion, and Section 7 concludes the paper.

2. Methodology

The literature sources used in this review were collected from several databases, including Web of Science, IEEE Xplore, and ScienceDirect. In addition, Google Scholar was used for supplementary searches to obtain relevant studies that might not be indexed in the major databases. The detailed search strategy adopted in this study is summarized in

Table 1. Search strategy, including search string, inclusion and exclusion criteria.

Category	Description
Database	Web of Science, IEEE Xplore, and ScienceDirect were used as the main databases. Google Scholar was used for supplementary searches.
Search string	TS = ((“photovoltaic inverter” OR “Grid-connected inverter” OR “PV inverter” OR “power converter”) AND (“PI control” OR “proportional integral control” OR “PR control” OR “proportional resonant control” OR “sliding mode control” OR “model predictive control” OR “predictive control” OR “fuzzy logic control” OR “fuzzy controller” OR “fuzzy system” OR “adaptive neuro-fuzzy inference system” OR “ANFIS” OR “neural network” OR “artificial intelligence” OR “machine learning” OR “reinforcement learning” OR “intelligent control” OR “hybrid control” OR “optimization algorithm”)) AND (“photovoltaic” OR “solar energy”)
Inclusion criteria	Reviewed and accepted publications. Energy- and engineering-related areas. Academic journals, conference papers, and reviews. Publications between 2015 and mid-2025.
Exclusion criteria	Book chapters, technical reports, and articles under review. Publications without full access or DOI. Studies unrelated to inverter control.

. The keywords were selected based on the core concepts of GCPI control, covering both traditional control methods and intelligent control methods. The literature search mainly focused on studies published between 2015 and mid-2025, since significant progress has been made in recent years in areas such as nonlinear control and artificial-intelligence-assisted control techniques. This time range was established to ensure that the review reflects the latest developments in inverter control, particularly the transition from conventional model-driven control to data-driven and hybrid control frameworks.

To quantitatively illustrate this evolutionary trend, Figure 2 presents the publication statistics of major control strategies for PV systems between 2015 and 2025. The publication data presented in Figure 2 were collected from the Web of Science Core Collection [29]. The search used keywords including “photovoltaic inverter,” “model predictive control,” “fuzzy logic control,” “neural network control,” and “reinforcement learning.” The study period was 2015–2025, and duplicate or irrelevant records were removed. The figure clearly demonstrates the transition from conventional control methods—such as PI, PR, and SMC—toward model-based and intelligent approaches like MPC, NNC, and RL. This shift reflects the growing research interest in adaptive, data-driven, and hybrid control frameworks that bridge the gap between reliability and intelligence in future GCPI applications.

Figure 3 illustrates the screening and exclusion process adopted in this review, following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) methodology.

An initial search across the Web of Science, IEEE Xplore, and ScienceDirect databases yielded 1214 records related to GCPI control strategies. These records were first included in a bibliometric analysis to assess the scope of the retrieved literature.

During the screening stage, articles without a Digital Object Identifier (DOI) (n = 80), duplicate records (n = 113), and papers filtered out due to irrelevant or overly broad keywords (n = 85) were excluded. After this step, 936 articles remained for eligibility assessment.

In the eligibility stage, a detailed examination of titles, abstracts, and full texts was performed to ensure relevance to inverter control topics. Studies focusing on unrelated areas—such as general power electronics, robotics, or communication systems—were removed. Based on the applied inclusion criteria, 118 papers met the final eligibility requirements and were included in the in-depth review and classification.

3. PV Grid-Connected Inverter System

3.1. Photovoltaic System

A PV system converts DC electricity generated by solar modules into AC power that meets grid requirements through power electronic converters and feeds it directly into the public grid. Figure 4 shows several possible system topologies. Converters can be classified as single-stage or multi-stage according to the number of conversion stages. The two-stage topology is the most common configuration, consisting of a front-end DC–DC converter that performs maximum power point tracking (MPPT) and DC voltage regulation, and a rear-end inverter that carries out DC–AC conversion [30]. In contrast, the single-stage topology eliminates the DC–DC stage, allowing the inverter to perform both MPPT and grid-connection control simultaneously [31]. This design offers a compact structure, low cost, and high efficiency, making it suitable when the PV array voltage matches the inverter DC bus requirements. However, because MPPT and grid-connection control share the same power loop, sudden changes in solar irradiance or grid conditions may cause DC-link voltage coupling and degrade power quality [32,33]. To mitigate these issues, researchers have proposed improved modulation strategies, adaptive control techniques, and multi-objective optimization methods to balance MPPT performance with grid integration quality [34,35,36].

3.2. Grid-Connected Inverter

GCPI can be divided into two main types based on their grid operation modes: GFLI and GFMI.

The primary function of GFLI is to synchronize with the grid in magnitude, phase, and frequency and to inject active and reactive power through current control. Because these inverters depend on the grid as a reference, they cannot independently establish voltage or frequency. Therefore, they usually include a Phase-Locked Loop (PLL) to ensure synchronization [37]. Their control strategy employs closed-loop current regulation and power adjustment with protection mechanisms to handle grid disturbances. Owing to their mature design, simple control, and good grid adaptability, GFLIs are widely used in large-scale PV systems. However, their performance deteriorates under weak grid or islanded conditions, motivating the development of GFMI and intelligent control strategies.

Unlike GFLI, GFMI can independently establish voltage and frequency, functioning as a virtual voltage source. It can perform black-start restoration after grid collapse and emulate generator characteristics to provide inertia and voltage support, improving system stability and resilience [38,39]. However, GFMI requires more complex control strategies due to its dual capability of autonomous operation and seamless transition to grid-following mode when an external reference is available. This flexibility allows coordination with the grid or other inverters while addressing challenges such as grid fluctuations and load variations [40,41]. Figure 5 compares the features of GFMI and GFLI.

4. Conventional Control

In a PV system, inverter control must achieve two primary objectives: efficient energy transfer and grid stability. Conventional control methods have been widely adopted for many years due to their theoretical maturity and high reliability. This chapter presents an overview of several representative control strategies.

4.1. PID/PR Control

Proportional–Integral–Derivative (PID) control is one of the most classical and widely applied feedback strategies in power electronics. Because of its simple structure, robustness, and mature industrial implementation, it remains a fundamental technique for current regulation in GCPI [42,43]. Figure 6 illustrates a typical PID controller. In practical PV applications, the derivative term is usually omitted to avoid amplification of switching noise [44], leading to the widely used PI controller:

$$G_{PI}(s)=K_p+{\displaystyle \frac{K_i}{s}}$$

(1)

where $K_p$ and $K_i$ are proportional and integral gains. The PI controller offers easy tuning and fast response. Operating in the synchronous dq frame, it regulates active and reactive currents independently and eliminates steady-state error [45]. Under normal grid conditions, a well-tuned PI controller achieves total harmonic distortion below 3% with a settling time of 10–20 ms [46].

However, its fixed parameters limit performance under weak-grid conditions, where voltage or frequency variations may affect current tracking. To improve control accuracy, enhancements such as anti-windup compensation, decoupling control, and adaptive tuning using metaheuristic optimization have been proposed [47,48].

To better handle AC quantities, the PR controller was developed. Its transfer function is

$$G_{\mathrm{PR}}\left(s\right)=K_p+{\displaystyle \frac{K_{r}s}{s^2+{\omega }_0^2}}$$

(2)

where $K_p$ and $K_r$ are proportional and resonant gains, and ${\omega }_0$ is the grid angular frequency (50 or 60 Hz). As shown in Figure 7, the resonant term provides high gain at the fundamental frequency, ensuring zero steady-state error for AC signals without requiring Park transformation [49]. This feature makes PR control well suited for single-phase and stationary-frame systems. A multi-resonant PR structure can further suppress low-order harmonics such as the 3rd, 5th, and 7th, improving current quality [50].

In practice, grid-frequency deviation may reduce PR performance, so the resonant frequency is often synchronized with the PLL output (${\omega }_{PLL}$) to maintain accuracy. However, PLL noise and bandwidth can broaden the resonance peak, requiring a compromise between fast response and stability.

Overall, PI and PR controllers are effective in strong-grid conditions but less suitable for weak or dynamic environments. Their fixed-gain structures limit adaptability to nonlinearities and parameter variations, motivating the use of more advanced nonlinear control techniques such as SMC and MPC, discussed in the next section.

4.2. Sliding Mode Control

SMC is a nonlinear control strategy based on state feedback, well suited for variable-structure systems with switching elements such as inverters. Its core principle is to design a sliding surface in the state space and drive the system trajectory toward it through a discontinuous control law. Once the trajectory reaches the surface, the system becomes insensitive to parameter variations, providing fast transient response and stable operation [51,52].

Figure 8 illustrates the overall structure of an SMC controller. In PV systems, SMC has been effectively applied at different control levels. It is used for current regulation in both single-phase and three-phase inverters under weak-grid and unbalanced conditions [53,54]. The controller generates a discontinuous switching signal based on the error between reference and measured currents. SMC-based voltage control has also been implemented in GFMI to provide virtual inertia and maintain stable operation during grid faults [55]. Moreover, SMC has been used in outer-loop designs for MPPT and DC-link voltage regulation, improving system dynamics and stability [56].

However, classical SMC suffers from chattering, which arises from high-frequency switching around the sliding surface [57]. Chattering can increase power losses, electromagnetic interference, and device stress. Reference [58] proposed a continuous super-twisting algorithm that replaces the discontinuous sign function with a smooth control term, effectively reducing chattering while maintaining control accuracy. Reference [59] introduced an adaptive boundary-layer adjustment strategy that dynamically modifies its thickness according to disturbance intensity, achieving a balance between precision and switching effort. Reference [60] developed a second-order SMC with adaptive gains integrated with a high-gain observer to enhance disturbance rejection while minimizing chattering.

Recent research also emphasizes hybrid approaches that combine SMC with other control methods, such as MPC or FLC, to exploit complementary advantages. MPC–SMC hybrids improve dynamic response and reduce switching losses, while FLC–SMC systems enhance adaptability and smoothness [61,62].

In summary, SMC provides a robust and computationally efficient framework for GCPI control under parameter uncertainties and grid disturbances. Although chattering remains its main limitation, modern variants and hybrid architectures have substantially improved dynamic stability and control precision, making SMC a strong candidate for next-generation renewable-energy inverters.

4.3. Model Predictive Control

MPC is an advanced strategy that uses a mathematical model of the system to predict future responses and compute optimal control actions by minimizing a cost function at each sampling instant [63]. Figure 9 shows the typical structure of an MPC controller, which includes a prediction model, rolling optimization, feedback correction, and a reference trajectory generator.

Two main variants are widely used: Finite Control Set MPC (FCS-MPC) and Continuous Control Set MPC (CCS-MPC). FCS-MPC directly selects the optimal switching vector from a finite set, achieving fast transient response and direct modulation without PWM [64,65]. Reference [66] implemented FCS-MPC in a three-level Neutral-Point-Clamped (NPC) inverter by constructing predictive models for current and DC-link midpoint voltage and defining a multi-objective cost function to ensure accurate current tracking and voltage balancing. However, the method is limited by the finite number of switching states [67].

To improve model accuracy, Reference [68] integrated a Kalman filter with an online correction mechanism, enhancing prediction precision and system reliability. Reference [69] proposed a power control scheme to mitigate voltage fluctuations in distribution networks under high renewable penetration, though it increased computational complexity. To further reduce computation, Reference [70] decomposed control objectives into multiple cost functions and adopted a cascaded evaluation approach, eliminating manual weighting-factor tuning and significantly reducing computational burden.

In contrast, CCS-MPC employs continuous control signals with PWM modulation, producing lower harmonic distortion and fixed-frequency operation but requiring greater computational effort [71,72]. Reference [73] proposed a CCS-MPC method that achieves high-precision current control for GCPI by searching for discrete offsets of the modulation voltage at a fixed switching frequency.

In summary, MPC provides precise control and excellent transient performance but demands high computational resources. Its evolution toward data-driven and intelligent predictive frameworks positions it as a promising bridge between classical control and AI-enhanced strategies.

4.4. Comparative Analysis and Technical Limitations of Classic Control Methods

Classic control strategies—including PI/PR, SMC, and MPC—have been extensively applied in GCPI due to their mature theoretical foundations and proven industrial reliability. However, as PV systems operate increasingly under fluctuating irradiance, weak grids, and distributed configurations, the performance boundaries of these classical algorithms have become more evident. A detailed comparative analysis is therefore essential to understand their respective strengths, limitations, and the motivations for transitioning toward intelligent control frameworks.

• Robustness, Adaptability, and Model Dependence

From a robustness standpoint, SMC and MPC outperform PI/PR under weak-grid and high-disturbance conditions. SMC’s variable-structure mechanism maintains control invariance despite system uncertainties, while MPC’s predictive capability compensates for transient mismatches. However, both still depend on accurate measurement and parameter estimation. MPC suffers performance degradation when model errors accumulate or delay compensation is inaccurate; SMC, though less model-dependent, requires careful gain tuning to balance robustness and chattering suppression. PI/PR, though robust under nominal operation, lacks adaptability once grid impedance or frequency deviates from nominal values.

2.

Computational and Implementation Constraints

A major differentiator among the three lies in real-time computational demand. PI/PR control involves static linear operations suitable for DSP implementation with minimal delay. SMC introduces moderate computational effort due to discontinuous switching logic, but its simple structure maintains real-time feasibility. MPC, however, requires iterative optimization and matrix operations at every sampling cycle, making its implementation computationally intensive. This constraint limits MPC’s use in large-scale GPCI unless combined with simplified cost functions, parallel computation, or FPGA acceleration.

3.

System-Level Performance and Trade-Offs

The trade-offs among robustness, optimality, and simplicity define the practical application boundaries of each algorithm.

• PI/PR: High steady-state precision, low cost, but poor dynamic adaptability.
• SMC: Excellent disturbance rejection, good dynamic response, yet chattering-induced energy losses.
• MPC: Optimal transient and harmonic performance, but with high computational burden and model sensitivity.

These characteristics are summarized as a triangular performance relationship:

No single classical controller can simultaneously maximize robustness, adaptability, and simplicity; enhancing one typically compromises another.

4.

Collective Technical Limitations

Despite their distinct mechanisms, conventional control methods share three fundamental technical limitations:

• Limited adaptability: Most require offline tuning and cannot autonomously adjust to real-time grid or load changes.
• High sensitivity to modeling and measurement errors: Even advanced controllers degrade under model mismatch, noise, or communication delay.
• Scalability and computational limits: As PV systems evolve toward multi-inverter, high-bandwidth networks, the computational overhead of model-based or discontinuous control laws becomes a critical bottleneck.

To provide a clearer comparison of the advantages and limitations of conventional control methods in GCPI,

Table 2. Classic Control Methods.

Method	Robustness	Dynamic Response	Model Dependence	Implementation Complexity	Key Advantages	Technical Limitations	Improvements
PI/PR	+	+	Low	Low	Simple, mature, good steady-state accuracy under stable grid.	Fixed gains → poor adaptability under grid disturbances; sensitive to harmonic distortion and frequency variations.	Multi-resonant PR controller for harmonic suppression [50]; PR tuning with zero steady-state error [49].
SMC	+++	+++	Medium	Medium	Strong robustness, fast dynamic response.	Chattering problem increases switching losses; performance degrades under noise and sampling delay.	Super-twisting/second-order SMC to reduce chattering [58]; adaptive boundary-layer and gain tuning [59]; observer-assisted disturbance rejection [60].
MPC	++	+++	High	High	Excellent dynamic performance; handles multivariable constraints.	High computational cost; strong dependence on accurate model parameters.	Kalman filter prediction [68]; cascaded objective evaluation [70]; fixed-frequency MPC for low THD [73].

“+++” → Excellent; “++” → Good; “+” → Fair.

summarizes their key characteristics and commonly adopted improvements.

In conclusion, classical control algorithms form the foundation of GCPI technology, each demonstrating specific strengths under certain operating conditions. However, their limited adaptability, accuracy, and scalability constrain further progress in modern power systems. These limitations have driven the shift toward AI-based control approaches to meet the demands of future renewable energy applications.

5. AI-Based Control

In recent years, AI methods have become an increasingly significant research focus in the control of GCPI. Techniques such as FL, NN, ANFIS, RL, and MA have been widely studied to enhance dynamic performance, adaptability, and power quality. Under complex grid conditions and multi-objective constraints, these intelligent control methods offer greater flexibility and optimization potential than conventional strategies. Figure 10 shows the differences between the classical method and the artificial intelligence method. Notably, AI is typically used alongside traditional control rather than as a full replacement, enabling complementary advantages and improved system performance. The following section introduces typical AI control methods and their applications.

5.1. Fuzzy Logic Control

FLC has been widely applied in GCPI due to its strong nonlinear adaptability. Unlike traditional control methods, FLC does not rely on precise system models. Instead, it uses rules derived from expert knowledge or data to dynamically adjust control signals, thereby achieving smoother current regulation and higher power quality. This capability makes FLC particularly effective under challenging operating conditions such as rapidly changing solar irradiance, grid impedance variations, and nonlinear load disturbances.

Figure 11 shows the basic structure of an FLC, which includes four main components: fuzzification, knowledge base, fuzzy inference, and defuzzification. The system measures grid voltage and inverter current, computes the instantaneous error and its rate of change, and determines control actions through the fuzzy rule base. The controller output modulates inverter switching signals to maintain synchronization and current quality. FLC is often used directly as an independent controller for inverter current and voltage regulation, making it one of the simplest intelligent control methods. Under nonlinear or dynamic disturbances, it offers better dynamic response, stronger anti-interference capability, and improved robustness compared with conventional PI controllers [74].

To overcome the drawbacks of conventional control—such as slow response, large steady-state oscillations, and high harmonics caused by PLL dependence—reference [75] introduced a PLL-free FLC direct power control strategy. FLC can also be integrated with other methods to form hybrid control structures that enhance overall performance. Reference [76] proposed a hierarchical H∞–FLC hybrid control scheme that significantly improved voltage quality in high-PV-penetration distribution networks.

Beyond replacing traditional controllers, FLC can optimize control systems. When combined with PI control, it can adaptively adjust proportional and integral gains, maintaining fast response under varying operating conditions [77]. Reference [78] presented a Disturbance-Observer-Based Fuzzy SMC (DOB-FSMC) method, where FLC dynamically tunes the upper bound of the sliding-mode switching gain. This approach adaptively suppresses chattering, reduces control signal fluctuation, and preserves robustness. Figure 12 illustrates the control structure. References [79,80] combined FLC with MPC, where fuzzy rules adaptively adjust cost-function weights in real time to achieve optimized performance.

In GFMI, FLC can emulate synchronous-machine behavior by adapting virtual inertia and damping based on frequency deviation and the rate of change in frequency (RoCoF). This enhances frequency recovery and voltage stability during islanding or weak-grid conditions. Fuzzy tuning of virtual impedance further stabilizes power sharing and suppresses circulating currents in multi-inverter systems [81]. These applications meet the growing need for dynamic grid support—such as black start, LVRT/FRT, and frequency/voltage regulation—in low-inertia networks. Reference [82] employs a fuzzy controller to adaptively adjust virtual impedance, compensating for reactive power errors caused by line impedance mismatches and maintaining voltage levels through a voltage compensation link. Reference [83] employs fuzzy rules to adaptively adjust the PWM modulation depth and vector switching rate, thereby improving current quality and dynamic stability. FLC offers a computationally efficient and flexible solution for GCPI control, especially under parameter uncertainty, grid impedance variation, and nonlinear operating conditions. In hardware implementations, the real-time performance of FLC depends mainly on the number of rules and membership functions. Compact rule bases implemented on embedded platforms such as DSPs or low-cost microcontrollers have achieved stable three-phase GCPI operation, demonstrating feasibility for medium-frequency control loops [84].

When the rule base grows, computation and memory demands increase, which can hinder high-frequency operation [85]. Robust fuzzy schemes improve tolerance to parameter uncertainty and grid disturbances but still require careful implementation to maintain real-time stability [86]. Moreover, fixed-point arithmetic and quantization in embedded systems may slightly reduce numerical precision, especially during defuzzification. Hence, FLCs are most effective for low- to medium-frequency control loops or as supervisory components in hybrid architectures. Current research trends focus on ANFIS and AI-assisted hybrid FLC, aiming to develop self-tuning intelligent control for next-generation renewable energy systems.

5.2. Neural Network Control

NN is the computational model that mimics the structure and information-processing mechanisms of biological neurons. As shown in Figure 13, it typically consists of three layers: an input layer, one or more hidden layers, and an output layer. NN exhibit strong self-learning and nonlinear mapping capabilities. Through iterative training, they can adjust their weights to approximate complex input–output relationships. With minimal modeling requirements, high adaptability, and strong generalization ability, NNC has become a major research direction in intelligent control in recent years.

Depending on their role, NNs can serve different functions in control systems. They may directly generate control signals (direct controllers) or act as compensators and parameter regulators for conventional controllers (indirect controllers).

During training, the NN learns the nonlinear relationship between input variables (e.g., error, current, voltage, power) and the desired control output. After training, it can generate real-time control signals based on system states, thereby bypassing traditional control architectures and achieving end-to-end control. The use of NNC in switch-mode inverters has been demonstrated as a feasible alternative, as shown in Figure 14.

Compared with conventional control methods, NNC offers a key advantage: even when system parameters vary, the dynamic characteristics remain stable within a certain range. Reference [87] employed a genetic algorithm to generate training data for an ANN, achieving robust inverter control. Experimental results showed that the ANN controller maintained a constant fundamental output voltage amplitude despite DC bus voltage fluctuations. Reference [88] proposed a vector control strategy based on a Recurrent Neural Network (RNN), which achieved better current-tracking performance and robustness than traditional PI and PR controllers. Moreover, it did not require high sampling or switching frequencies, effectively reducing energy consumption. Reference [89] used data generated by SMC to train an NN controller, achieving control performance comparable to SMC while simplifying the control structure. To mitigate the computational burden of MPC, References [90,91] used MPC to generate offline training data for ANN training, allowing the ANN to emulate MPC behavior. This enabled direct generation of optimal switching signals without online optimization, significantly reducing computational effort while maintaining MPC-level performance. Reference [92] employed two independent neural networks to emulate PI and MPC controllers, where the RNN–MPC surrogate achieved a 2.4 ms dynamic response—matching MPC performance with a 42% reduction in computation time.

Figure 15 illustrates the general procedure for training an ANN to emulate an MPC. In this framework, measured system parameters—such as current error, voltage error, and DC-link voltage deviation—are used as inputs. The ANN learns to predict the optimal switching vectors and weighting factors that would otherwise be determined by the computationally intensive MPC optimization process.

The process involves three main steps:

• Data collection—Generate training data using a controller or optimization algorithm to reflect system inputs and corresponding desired outputs.
• Offline training—Train the ANN with the collected data to learn the mapping between inputs and outputs.
• Online testing—Deploy the trained ANN in a real system and evaluate its performance under actual operating conditions.

Due to factors such as component aging and grid fluctuations in PV systems, controller parameters must be continuously updated to adapt to changing conditions. To address real-time adaptability, researchers have integrated NN controllers with traditional control structures. Reference [93] used an NN to predict the weighting factor λ in FCS-MPC, modeling the relationship between target switching frequency, current peak, and optimal weight as a nonlinear mapping. This allows real-time adjustment in each control cycle, maintaining the average switching frequency and current-tracking performance. The method, based on inverse mapping modeling, can be extended to optimize additional control objectives or weighting parameters. The ANN reduced switching frequency variation by 48%, maintaining approximately 3.2 kHz ± 5%, compared with ±25% in conventional MPC. Reference [94] proposed a PI controller based on a Modular Neural Network (MNN), composed of multiple MNN modules, each designed for specific operating conditions (e.g., different irradiation levels or fault types). Their outputs are combined through weighted summation to generate final PI parameters, enabling real-time dynamic adjustment. Fault recovery time was reduced by 37%, enhancing transient stability during grid-voltage dips.

Another approach involves using NN as an inverse system model for feedforward compensation, effectively linearizing nonlinear inverter dynamics. Reference [95] proposed an SMC scheme integrated with a Fuzzy Neural Network (FNN), which estimated system uncertainties and adaptively updated control gains, reducing control-input chattering amplitude by 65%. Reference [96] introduced a low-switching-frequency MPC method for modular inverters that employed an NN predictor for modeling error compensation and online weighting adjustment. This method reduced switching frequency by 49% (from 3.5 kHz to 1.8 kHz), decreased switching losses by 43%, improved THD (from 3.2% to 1.4%), and reduced dynamic overshoot (from 8.6% to 3.1%).

Real-time implementations of ANN controllers have been reported on DSP and FPGA platforms for GCPI [97,98]. Experiments show that ANN-based voltage and current controllers can produce stable sinusoidal waveforms with low harmonic distortion, confirming their suitability for medium-frequency switching [97]. However, achieving high accuracy requires many neurons and weights, which increases computational complexity and memory use [98].

Controller performance depends on the quality of training data; poor datasets can cause weak generalization and instability. Neural networks also lack interpretability, limiting their use in safety-critical systems. Physics-informed neural network (PINN) helps address this issue by embedding known physical laws, such as Kirchhoff’s or power balance equations, into the learning process.

Even so, online training remains difficult on embedded hardware with limited resources, and quantization or fixed-point arithmetic can reduce numerical precision. ANN-based controllers are therefore best suited for supervisory optimization, adaptive tuning, or medium-frequency inverter loops with sampling periods of several tens of microseconds.

5.3. ANFIS Control

The ANFIS combines the learning capability of neural networks with the reasoning mechanism of fuzzy logic to create a hybrid controller that inherits the advantages of both paradigms. ANFIS was introduced to address the limitations of standalone FLC and NN methods—specifically, the lack of learning ability in fuzzy control and the interpretability problem in neural networks.

As illustrated in Figure 16, a typical ANFIS architecture consists of five layers:

• Input layer: maps input variables to membership degrees.
• Rule layer: computes the firing strength of each rule.
• Normalization layer: normalizes firing strengths to obtain weighting coefficients.
• Consequent layer: produces rule-specific outputs.
• Output layer: aggregates all rule outputs via weighted summation to generate the final system output.

To address transient and voltage stability issues that may arise in PV power plants during grid faults, Reference [99] proposed an ANFIS-based control strategy capable of adjusting inverter active and reactive power outputs in real time according to grid voltage and power variations, thereby maintaining system stability during disturbances. Reference [100] developed an ANFIS controller to independently regulate the active and reactive power of the inverter under normal grid-connected conditions. Experimental results demonstrated superior dynamic response compared with conventional PI control. Reference [101] proposed an ANFIS–PID control scheme, in which ANFIS dynamically adjusts PID gains in response to voltage deviations and system variations.

To mitigate instability caused by insufficient system inertia during grid connection, Reference [102] introduced a hybrid control strategy integrating ANFIS with a Virtual Synchronous Generator (VSG). In this method, measured and estimated PV system parameters are used as inputs to the ANFIS block to adaptively tune VSG parameters in real time. The corresponding control structure is shown in Figure 17.

Here, ω and dω/dt denote the grid angular frequency and its RoCoF, respectively. The ANFIS controller uses these quantities to estimate the optimal modulation factor m, dynamically adjusting the virtual inertia and damping coefficients within the VSG model to ensure frequency and voltage stability under transient conditions. During LVRT tests, the fault-clearing time was reduced from 0.35 s to 0.10 s, while the minimum voltage improved from 0.79 p.u. to 0.82 p.u., demonstrating improved voltage recovery and transient stability.

However, the major limitations of ANFIS lie in its computational burden and memory requirements. As the number of inputs and fuzzy sets increases, the number of rules grows exponentially (rule explosion problem), leading to significant increases in training time and hardware cost. To address these challenges, most practical designs rely on offline training or rule-reduction techniques before deployment. Furthermore, fixed-point arithmetic can introduce quantization errors that slightly affect accuracy.

5.4. Reinforcement Learning Control

RL is an interaction-driven approach to intelligent decision-making. Unlike supervised learning, which relies on labeled datasets, RL enables an agent to iteratively refine its decision-making policy through continuous interaction with the environment, guided by reward signals, with the objective of maximizing long-term cumulative rewards. Theoretically, this process can be modeled as a Markov Decision Process (MDP) and is proven to converge toward an optimal policy under the principle of optimality. The inherent exploration–feedback–update mechanism makes RL particularly suitable for control and optimization problems in systems that are difficult to model precisely or exhibit highly dynamic and nonlinear operating conditions.

RL controllers continuously accumulate experience through interaction with physical systems or their simulation models, progressively learning control strategies that maximize predefined performance objectives. Reference [103] proposed a DC control method based on the Deep Deterministic Policy Gradient (DDPG) algorithm. Leveraging DDPG’s capability to manage continuous action spaces, the method employs an actor network to directly generate d–q axis voltage reference signals, effectively replacing conventional PI and MPC controllers to achieve precise current regulation. To address the overestimation problem of Q-values inherent in DDPG, Reference [104] adopted the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, which enhances both training stability and generalization performance. Hardware-in-the-loop (HIL) experiments demonstrated that the TD3-based controller achieved a THD of 2.93%, compared with 3.94% for PI and 4.69% for MPC controllers, while maintaining robust current tracking under ±20% filter parameter deviations.

Figure 18 shows the adopted DRL control framework diagram. The adopted DRL framework primarily uses the d–q axis currents, voltages, and DC-link voltage as system inputs. Based on the state vector, the controller evaluates the operating condition and outputs optimal voltage commands, which are processed by the PWM module to generate switching signals for a three-level NPC inverter, producing sinusoidal output currents. During operation, the controller dynamically adjusts the weighting factors through a reward mechanism to minimize current error, DC-link voltage deviation, and harmonic distortion.

The measured current, voltage, irradiance, and grid-impedance signals are provided as inputs to the observation space, where the actor–critic agent generates optimal d–q axis voltage control actions. The reward function is designed based on current-tracking precision and dynamic stability, guiding policy updates to enable the inverter to maintain optimal operation under varying irradiance and grid conditions.

Reference [105] applied a Proximal Policy Optimization (PPO)-based RL algorithm to achieve model-free and adaptive controller design. Reference [106] introduced an off-policy integral reinforcement learning method, which reformulates the grid-connected power quality optimization problem as a zero-sum game within a voltage–current dual-loop framework. By solving the Hamilton–Jacobi–Bellman (HJB) equation, an optimal control law was derived to mitigate power quality degradation under unknown or time-varying system dynamics.

Traditional control methods are often insufficient to meet the dynamic operational demands of modern PV systems. Recent research has therefore shifted toward hybrid control frameworks that integrate RL with classical control methodologies. In Reference [107], RL was employed to design an adaptive weighting factor mechanism for MPC, enabling model-free control through imitation learning and resolving issues related to manual weight tuning and model dependency. Reference [108] combined the TD3 algorithm with SMC by superimposing an RL-generated correction signal on the reference signal, improving both steady-state accuracy and transient response under large disturbances and parameter variations. Reference [109] presented a dual Deep Q-Network (DDQN)-based MPC weighting optimization method that dynamically adjusts weighting factors to reduce grid-current THD and enhance capacitor voltage regulation across diverse operating conditions, outperforming both fixed-weight and other adaptive weighting strategies.

In practice, most RL-based inverter controllers are trained offline and then validated in HIL or PHIL before deployment on embedded targets, which helps ensure safety during transients and grid events [106,110]. Real-time execution on low-cost controllers is feasible when using pre-trained or lightweight policies; a DSP-class implementation without external compute demonstrates that timing and memory constraints can be met with careful design [111]. Hybrid designs that pair RL with established schemes—such as VSG and MPC—reduce computation and improve interpretability while maintaining robust grid support under disturbance [110,112]. Recent studies on grid-forming control further indicate that DRL can coordinate synchronization and frequency regulation, but reward design and training-environment fidelity remain critical to stability and convergence [110,113]. Overall, RL controllers are promising for supervisory adaptation and medium-frequency inverter loops; fully online learning at microsecond sampling is still challenging on embedded hardware.

5.5. Metaheuristic Algorithm Optimization

MAs are general optimization methods that use stochastic search and iterative updating to approximate global optima. They assess candidate solutions through fitness functions without relying on specific models, giving them wide applicability. Unlike other AI-based controllers, MAs are mainly designed for optimization and are often used as auxiliary optimizers. In PV systems, they are applied to tune controller gains, choose switching sequences, and adjust cost-function weights in MPC or RL frameworks. PSO and GA are the most widely used, showing strong global search ability and efficient convergence in complex nonlinear problems.

5.5.1. Particle Swarm Optimization

PSO is a global optimization algorithm based on swarm intelligence, inspired by the collective behavior of birds and fish. It simulates the cooperative movement of multiple “particles” within a multidimensional search space. Each particle continuously updates its position and velocity according to its individual best experience and the global best position found by the swarm, gradually converging toward the optimal solution. The algorithm is well-known for its simple structure, few tuning parameters, fast convergence, and strong numerical robustness.

In [114], PSO was used to optimize the membership functions of a fuzzy logic controller to minimize the mean square error of the output voltage and avoid the manual trial-and-error tuning process. The optimized controller achieved fast convergence and stable sinusoidal voltage waveforms under resistive, inductive, and nonlinear loads (MSE ≈ 0.005). Reference [115] applied PSO for automatic tuning of PI controller parameters. The optimized controller reduced voltage error and achieved a low THD of about 3.9%, ensuring system stability under different load conditions and inter-phase faults.

5.5.2. Genetic Algorithm

GA is a population-based global optimization method that simulates the processes of natural selection and genetic evolution. Through selection, crossover, and mutation operations, GA evolves a population of candidate solutions toward optimality. Compared with PSO, GA offers greater flexibility and adaptability, making it particularly suitable for problems with high-dimensional search spaces, multiple variables, or multi-objective optimization requirements.

In [116], GA was utilized to optimize the PI controller parameters, while [117] used GA to fine-tune the initial weights of a Backpropagation Neural Network (BPNN), improving settling time and reducing overshoot in BP–PID control. Reference [118] further applied GA for multi-objective parameter optimization of a fuzzy SMC, which resulted in lower output error and reduced switching losses. However, GA is computationally intensive, which limits its direct real-time applicability. Future research is expected to focus on developing faster hybrid or adaptive GA variants to enable real-time and online optimization in power electronic control systems.

While metaheuristic algorithms provide strong global optimization capability, their iterative and stochastic nature introduces challenges for real-time implementation in GCPI controllers.

Most studies report offline optimization, where algorithms such as PSO or GA run on desktop environments to precompute optimal controller parameters, later embedded into DSP or FPGA firmware. This approach avoids excessive on-board computation but limits adaptability during runtime.

In practical hardware setups, real-time implementation requires significant computational resources. Standard PSO or GA algorithms often require hundreds to thousands of fitness evaluations in each optimization cycle, which makes real-time execution difficult for inverter control loops operating within 10–50 µs. Implementations on FPGA or GPU platforms can exploit parallel computation to significantly reduce execution time. However, such hardware acceleration increases system complexity and cost, and the use of fixed-point arithmetic may introduce quantization errors that affect accuracy.

Moreover, convergence reliability depends on algorithm parameter tuning (population size, inertia weight, mutation rate, etc.), which may vary under different operating conditions.

Hybrid strategies—where offline metaheuristic search provides coarse tuning and online lightweight refinement handles dynamic changes—are emerging as promising compromises between performance and real-time feasibility.

In summary, AI-based methods demonstrate strong adaptability in controlling GCPI, providing improved dynamic performance under complex conditions compared to conventional methods. When integrated with conventional strategies, these methods enhance steady-state accuracy and system robustness.

Table 3. Comparison of AI Methods.

Method	Advantages	Disadvantages	Typical Applications
FLC	Does not require an accurate mathematical model. Provides strong nonlinear adaptability and robustness under grid disturbances. Enables adaptive parameter tuning (e.g., PI/PR gains, PLL, reactive power).	Rule-based and membership designs rely on expert experience. Performance depends on rule granularity and scaling factors. Limited scalability for high-dimensional systems.	Adaptive PI/PR tuning [77,86]; PLL-free fuzzy direct power control [75]; fuzzy–MPC hybrid weight tuning [79,80].
NNC	Strong nonlinear mapping and self-learning capability. Can emulate MPC or SMC behaviors with reduced computation time. Enables online adaptive tuning of control parameters.	Requires large, high-quality training datasets. May face overfitting and poor generalization under unseen conditions. High computational demand for real-time inference.	ANN-based inverter voltage control [87]; RNN vector control [88]; MPC-data-driven ANN [90,91].
ANFIS	Combines fuzzy interpretability and NN learning ability. Automatically generates rules and membership functions. Excellent adaptability to nonlinear and time-varying PV systems.	Requires careful training and parameter initialization. Sensitive to data quality and convergence issues. Moderate computational complexity for real-time use.	ANFIS for P/Q regulation [99,100]; ANFIS–VSG dynamic inertia tuning [102]; ANFIS–PID adaptive gain control [101].
RL	Model-free and data-driven optimization. Learns optimal control policy through continuous interaction. Achieves high adaptability in dynamic and uncertain grids.	Long training process and high data requirements. Risk of instability during exploration. Difficult interpretability and certification.	DDPG-based dq-control [103]; TD3 control for current tracking [104]; RL-tuned MPC weighting [107,109];
PSO/GA	Effective for global optimization and multi-objective problems. Can tune controller parameters, switching frequency, or filter coefficients. Simple to implement, robust to local nonlinearities.	Offline optimization dominates—limited real-time capability. Convergence speed and repeatability depend on parameter settings. May get trapped in local optima.	PSO-optimized FLC membership [114]; PSO-tuned PI controller [115]; GA-based fuzzy SMC optimization [118].

compares their advantages, limitations, and applications.

Table 4. Comparison of GCPI Control.

Control Method	THD (%)	Response Time (ms)	Voltage Stability	Computational Complexity	Examples of Industrial Applications
PI/PR	2–4	5–10	Moderate	Low (O(1))	Commonly adopted in commercial GCPI for stable grid conditions
SMC	1.5–3	3–6	High	Medium (O(n))	Frequently used in research prototypes and laboratory-scale PV systems
MPC	1–2	2–5	High	High (O(n2))	Demonstrated in academic studies and experimental validation platforms
FLC	1.5–2.5	3–7	High	Medium–High	Widely evaluated in laboratory and pilot-scale hybrid PV systems
NNC	1–2	2–4	High	High	Tested in AI-assisted control simulations and experimental prototypes
ANFIS	1–2	2–5	Very High	High	Reported in the literature for hybrid GCPI simulations and testbeds
RL	0.8–1.5	2–3	Very High	Very High	Applied in recent experimental research on intelligent inverter control
PSO/GA	1–2	4–8	High	High–Very High	Utilized in optimization-based control design studies and simulation environments

The data on THD, settling time, and computational complexity are summarized from multiple studies in the literature. These values represent typical ranges reported for classical and intelligent control methods, including PI/PR, sliding-mode, model predictive, and fuzzy or neural-network-based control.

compares these methods. From a quantitative standpoint, traditional methods like PI/PR and SMC maintain moderate THD and fast response with low computational overhead, making them suitable for stable grids. In contrast, AI-based methods exhibit superior adaptability and dynamic response under weak-grid or nonlinear conditions but require greater computational resources

5.6. AI-Based Parameter Optimization

In addition to optimizing primary control loops and controller gains, AI techniques are widely used to optimize operational parameters that strongly affect the performance and reliability of GCPI. Parameters such as switching frequency, filter component values, MPPT settings under partial shading, and power-quality indices determine conversion efficiency, current distortion, and dynamic grid compliance.

5.6.1. Switching-Frequency Optimization

Switching frequency critically influences system efficiency, power quality, EMI, and thermal behavior. A higher frequency improves dynamic response and reduces harmonic distortion but increases switching losses and device temperature. Conversely, a lower frequency reduces losses yet degrades power quality and control precision. The objective is to balance efficiency, harmonic suppression, and thermal stress.

Conventional FCS-MPC often yields variable switching frequency due to discrete decisions each sampling period, which produces unpredictable harmonic spectra and complicates filter design. To address this, several fixed-switching-frequency MPC strategies have been proposed. In [119], an optimal-switching-sequence MPC for LC-filtered inverters maintains a constant commutation rate and reduces current ripple. In [120], a modulated MPC integrates PWM into the predictive framework, achieving a fixed frequency with fast transient response. In [121], a double-voltage-vector MPC for multilevel inverters stabilizes the switching frequency and minimizes THD.

AI and MA methods can dynamically tune the switching frequency according to grid impedance, temperature, and irradiance variation. A common formulation is the multi-objective cost function:

$$J=w_1\cdot \mathrm{THD} \left(f_s\right)+w_2\cdot P_{\mathrm{loss}}\left(f_s\right)+w_3\cdot \mathsf{\Delta }{i}_L\left(f_s\right)$$

(3)

where $f_s$ is the switching frequency, $THD\left(f_s\right)$ is the total harmonic distortion, $P_{\mathrm{l}\mathrm{o}\mathrm{s}\mathrm{s}}\left(f_s\right)$ denotes switching/conduction losses, and $\mathsf{\Delta }{i}_L\left(f_s\right)$ is the inductor-current ripple. The weights $w_1,w_2,w_3$ are adaptively tuned by AI to optimize performance. In [122], a two-step efficiency optimization combines PSO with frequency adjustment to minimize inverter losses under varying loads.

5.6.2. Filter Parameter and Damping Tuning

Filter parameters and damping elements are crucial in GCPI. Proper LCL tuning suppresses switching harmonics, improves current quality, and ensures stability; poor design can induce resonance peaks, excessive THD, or instability—especially when grid impedance varies. Thus, adaptive filter tuning and damping optimization are essential.

Traditional designs rely on analytical equations or small-signal models to set inductances, capacitances, and damping resistors. While adequate at nominal conditions, they do not adapt to grid fluctuations or inverter aging. Recent studies employ optimization and AI to automatically tune filter parameters, achieving lower THD, higher efficiency, and stable dynamics.

In [115], a PSO-based multi-objective optimization finds optimal LCL parameters for GCPI, simultaneously minimizing THD and power loss. In [123], a hybrid PSO–GA minimizes THD, reduces resonance frequency, and improves efficiency and stability under weak-grid conditions. Unlike analytical methods, the AI-assisted approach jointly tunes inductances, capacitance, and damping resistance to balance harmonic suppression, dynamic response, and efficiency.

5.6.3. MPPT Under Partial Shading and Nonuniform Conditions

Under partial-shading conditions, the PV array’s P–V curve presents multiple local maxima, causing classical P&O and INC methods to get trapped in local optima. Global-search metaheuristics—such as PSO, ant-colony optimization, and artificial-bee colony—together with AI methods like fuzzy logic and neural networks, reliably track the global maximum power point (GMPP) across diverse shading patterns [124,125].

Using PSO as an example, each particle represents a candidate operating voltage on the P–V curve. The swarm iteratively updates velocity and position to locate the GMPP. The fitness function is the output power:

$$fitness\left(i\right)=P_i\left(V_i\right)=V_i \times {IV}_i$$

(4)

Particle updates follow:

$$v_i^{t+1}=\omega v_i^t+c_1{r}_1\left(p_{best,i}-x_i^t\right)+c_2{r}_2(g_{bset}-x_i^t)$$

(5)

$$x_i^{t+1}=x_i^t+v_i^{t+1}$$

(6)

where $x_i$ and $v_i$ are particle position (PV voltage) and velocity; $\omega$ is inertia weight; $c_1,c_2$ are cognitive/social factors; $r_1,r_2\sim U\left[\mathrm{0,1}\right]$. To balance exploration and exploitation, $\omega$ may be scheduled as

$$\omega ={\omega }_{max}-{\displaystyle \frac{{\omega }_{max}-{\omega }_{min}}{{iter}_{max}}}·iter$$

(7)

An improved PSO-based MPPT in [126] significantly enhanced dynamics under PSC: convergence time dropped from ≈1.8 s to 0.2 s, steady-state oscillation amplitude from ±2.3% to ±0.7%, and total harvested energy increased by ≈7% under complex shading.

5.6.4. Power-Quality and Harmonic Regulation

High power quality is essential for GCPI, especially with nonlinear or time-varying loads. Key indices include THD, power factor (PF), and voltage unbalance. Conventional PI/PR controllers have limited harmonic compensation when grid impedance changes or higher-order harmonics are present, often degrading performance in weak or distorted grids. Adaptive and AI-based control strategies have therefore been developed to regulate harmonics dynamically and maintain grid-code compliance.

FLC and NNC are widely applied due to model-free adaptability and robustness to disturbances. In [75], an FLC-based compensator reduced current THD from 3.8% to 1.8% under nonlinear loads. In [99], an ANFIS controller improved transient and voltage stability, indirectly enhancing power quality.

Overall, AI-driven harmonic regulation enables multi-objective optimization among harmonic suppression, dynamic performance, and energy efficiency. By adaptively tuning compensator parameters, virtual impedance, and current-control loops in real time, AI methods maintain compliance while ensuring stable, high-quality power injection under rapidly changing grid conditions.

AI-based parameter optimization complements primary control by enabling adaptive, multi-objective tuning of critical inverter settings—switching frequency, filter coefficients, MPPT parameters, and harmonic compensators—thereby improving overall performance and reliability of the PV system.

6. Discussion

The preceding analysis shows that both conventional and AI-based control methods have unique advantages and limitations. Conventional controllers offer simplicity, stability, and proven reliability, while AI-based approaches provide adaptability and global optimization that improve performance under nonlinear and time-varying grid conditions. Future PV systems will require intelligent frameworks that integrate these strengths rather than treating them as separate alternatives. Three main directions will guide this evolution: hybrid control strategies, secure and interpretable AI design, and hardware-efficient implementation.

6.1. Hybrid Control Strategies

Hybrid control combines the stability of conventional methods with the adaptability of AI algorithms. This integration improves dynamic response, harmonic suppression, and robustness under changing grid conditions.

For example, combining FLC with SMC can reduce chattering and maintain good dynamic performance. MPC with RL can adjust control parameters automatically to balance current tracking and switching losses. Other hybrid forms, such as FLC–MPC and ANFIS–PI, also show better stability under weak-grid and variable-irradiance conditions.

As AI and computing technologies develop, hybrid control will move toward adaptive structures that merge model-based and data-driven features.

Table 5. Comparative Suitability of Different Control Methods under Various Operating Scenarios.

Method	Strong Grid	Weak Grid	Dynamic Irradiance	Nonlinear Load	Islanded/ Microgrid	Hardware Complexity	Real-Time Feasibility
PI/PR	++++	++	+++	++	++	++++	+++++
SMC	++++	+++	+++	+++	+++	+++	++++
MPC	++++	++++	+++	+++	+++	++	+++
FLC	+++	++++	++++	+++	+++	+++	++++
ANFIS	+++	++++	++++	+++	+++	++	+++
RL	+++	+++++	+++++	+++	++++	+	++

+++++: High suitability; ++++: Suitable; +++: Moderate; ++: Limited; +: Poor feasibility.

summarizes the qualitative performance of major control strategies. Conventional PI/PR and SMC controllers are more suitable for stable grids, while AI-based methods such as FLC, ANFIS, and RL perform better in uncertain and dynamic conditions.

To make the comparison clearer, Figure 19 shows a radar chart that visualizes the trade-offs among control quality, complexity, robustness, and adaptivity. The figure indicates that PI/PR offers good stability and simplicity but weak adaptability, while RL achieves high adaptability with greater computational demand. The radar chart complements Table 5 and gives a clearer view of how each method performs in different aspects.

6.2. Security and Interpretability of AI-Based Control

With the rapid deployment of AI in power electronics, security and interpretability have become critical challenges for practical implementation. Unlike deterministic model-based controllers, AI systems—particularly deep neural networks and reinforcement learning agents—operate as black boxes, making it difficult to predict or verify their behavior under unforeseen grid disturbances, cyber-attacks, or sensor faults. This lack of transparency raises concerns about operational safety and regulatory compliance in smart grids.

To address these challenges, several research directions are emerging:

• Explainable AI (XAI) for control: Integrating interpretability methods such as sensitivity analysis, feature attribution, and surrogate modeling can reveal how input variables (voltage, current, irradiance) influence AI decisions, enhancing understanding and trust in AI controllers.
• Physics-informed and stability-constrained learning: Embedding physical constraints, Lyapunov functions, or energy-based models into AI training ensures that control actions respect system dynamics and stability boundaries.
• Secure learning and anomaly detection: Implementing adversarial training, fault detection, and cyber-resilient architectures can protect inverter control systems from malicious data injection, communication delays, and false measurements.
• Formal verification frameworks: Applying reachability analysis and formal stability proofs can help certify AI controllers for compliance with grid codes and safety standards.

Ensuring explainability and robustness against cyber-physical threats will be fundamental for deploying AI-enabled inverters in mission-critical renewable energy infrastructures.

6.3. Hardware and Real-Time Implementation Requirements

AI-based control methods provide strong adaptability and optimization, but their hardware implementation remains more demanding than traditional approaches. PI, PR, and SMC controllers work well on low-cost DSP or MCU platforms with short sampling cycles, whereas neural networks, ANFIS, and reinforcement learning often require faster processors, higher sampling rates, and larger memory.

Key hardware considerations include:

• Computing capability: Many AI controllers need floating-point DSPs, FPGAs, or SoC processors to meet real-time requirements, especially when switching frequencies exceed 10 kHz.
• Sampling time and latency: Stable control requires microsecond-level computation and actuation. AI models must therefore be lightweight, and inference time should be minimized through model simplification or hardware acceleration.
• Memory and storage: Large neural networks and fuzzy systems may exceed the memory capacity of conventional inverter controllers.
• Hardware acceleration and optimization: Quantization, pruning, and compression can reduce computation time and energy use. New edge-AI chips with dedicated neural processing units may help support AI control in commercial inverters.

Recent studies have demonstrated the feasibility of advanced controllers through simulation, hardware-in-the-loop testing, and laboratory prototypes. Several representative implementations are summarized in

Table 6. Experimental Comparison of GCPI Control Methods.

Year	Control	Key Results	Platform	Reference
2014	FLC	THD: V 2.48%, I 4.64%; Unity PF; Grid disconnect ≈ 2.65 cycles	dSPACE DS1104	[74]
2024	ITTSMC	Current THD 2.1–2.2%; PF 0.998; P ≈ 328 W; Q ≈ 18 VAR	PV + L-filter + Semikron inverter + dSPACE 1104	[127]
2025	Enhanced MPC	THD 3.31%; CMV max 127.48 V; Leakage RMS 0.243 A; ΔVcap 10.09 V; Execution time 25.7 μs	OPAL-RT OP5700 + VC707 FPGA	[128]
2022	ANN-PID	Overshoot 0.54%; THD 2.27%; Response time 0 s; RMSE 0.0466; MAPE 0.0338%; MAE 0.0046	dSPACE 1202 MicrolabBox; IGBT 10 kHz boost; LV25P/LA25NP sensors	[129]

, including MPC-based smart-inverter functions, fuzzy logic control, neural-network-assisted PID regulation, and sliding-mode current control on dSPACE platforms.

Although these results are promising, large-scale industrial deployment remains limited. Many AI controllers still require more computing power and faster sampling than typical inverter hardware can provide. In addition, the lack of standardized procedures for verifying safety and compliance with grid codes such as IEEE 1547 and IEC 62116 remains a major challenge.

At present, AI-based controllers are mainly applied in academic and small-scale experimental settings. Future progress will rely on advances in hardware acceleration, edge–cloud computing, and unified validation frameworks. Lightweight neural models and improved embedded processors may help close the gap between laboratory demonstrations and commercial GCPI applications.

6.4. Emerging and Future Research Trends

As AI-enabled control continues to advance, research on PV systems is moving toward higher adaptability, interoperability, and industrial scalability. Several promising directions are shaping the next generation of intelligent inverter technologies.

• Federated and Collaborative Learning for Distributed PV Systems

Federated learning offers a decentralized framework for model training across multiple PV units while keeping data local. It enables distributed inverters to learn coordinated control strategies under varying irradiance and grid conditions. Extensions such as transfer learning and meta-learning may further improve adaptability when integrating new PV units or facing unfamiliar operating scenarios.

2.

Cloud–Edge Cooperative Intelligent Control

The growing digitalization of power systems encourages the integration of cloud computing, edge intelligence, and communication networks. In such architectures, the cloud handles large-scale optimization and prediction, while edge devices execute local real-time control. This cooperation supports continuous learning, coordinated operation, and improved resilience. Digital-twin models can further enhance predictive maintenance and fault response.

3.

Hardware Acceleration and Embedded AI Implementation

Hardware limitations remain a major challenge for real-time AI control. FPGA, SoC, and GPU platforms can provide the parallel computing capability required for microsecond-level inference. Future work will focus on lightweight neural models, quantized fuzzy systems, and optimized embedded frameworks that reduce latency and power consumption. Co-design of algorithms and hardware will be essential for practical deployment.

4.

Standardization, Safety, and Explainability

As AI interacts more closely with the grid, safety, transparency, and compliance become critical. Physics-informed neural networks and explainable AI can improve interpretability and support certification under standards such as IEEE 1547 and IEC 62116. Standardized validation procedures that combine hardware-in-the-loop testing, formal verification, and digital-twin simulation will be necessary to ensure reliable operation.

5.

Integration with Energy Storage and Multi-Agent Coordination

Future PV systems will increasingly integrate AI control with energy-storage management, demand response, and grid-forming capabilities. Multi-agent reinforcement learning provides tools for cooperative decision-making among distributed inverters, supporting voltage and frequency stabilization in weak or islanded grids. This approach can enhance performance in hybrid systems combining PV, wind, and storage.

6.

Cybersecurity and Resilience of Intelligent Inverter Networks

The use of communication networks exposes AI-based inverters to cyber-physical risks. Future research should emphasize secure control, anomaly detection, and encrypted learning to protect against attacks and data manipulation. Combining AI-driven detection with decentralized blockchain mechanisms may improve trust and data integrity in networked energy systems.

Looking forward, future research should focus on several key directions to further bridge the gap between theoretical innovation and practical deployment.

First, the integration of AI with edge computing and IoT will enable distributed, real-time, and autonomous control across multiple GCPI, improving scalability and fault tolerance in large power networks.

Second, the application of federated learning can enhance data security and enable collaborative optimization without direct data sharing, promoting privacy-preserving intelligence in smart grids.

Third, adaptive hybrid controllers that combine MPC with deep learning are expected to improve dynamic performance and robustness under complex and uncertain grid conditions.

Finally, hardware implementations on FPGA and SoC platforms will help reduce computational latency and make AI-based control feasible for industrial real-time applications. Together, these advancements will provide a strong foundation for developing intelligent, secure, and efficient PV systems that meet the demands of next-generation smart grids.

In summary, the future of AI-based GCPI control will rely on a comprehensive synergy of distributed learning, cloud–edge collaboration, hardware acceleration, safety assurance, and cyber resilience. These converging technologies will enable intelligent, adaptive, and reliable control frameworks, driving the transformation of renewable-energy systems toward fully autonomous and data-driven smart grids.

6.5. Summary of Key Insights

In summary, the future of GCPI control will rely on the synergistic combination of hybrid control, secure and interpretable AI, and hardware-efficient deployment:

• Hybrid control merges physical modeling and data-driven adaptation, achieving both reliability and intelligence.
• Security and interpretability are prerequisites for trusted, grid-compliant AI control.
• Hardware-aware algorithm design ensures real-time implementation on embedded systems.
• Federated and cloud-based architectures will enable large-scale distributed learning and cooperative control among GCPI.

These aspects shape the readiness of AI technologies for next-generation smart grids and renewable-energy systems. In current industrial applications, most GCPI still use PI or PID control for current and voltage regulation. MPC has also been applied in engineering projects, where field-tested smart-inverter functions improved voltage regulation and increased feeder hosting capacity [130]. AI-assisted functions are now emerging in commercial products. Huawei’s Smart PV Solution applies neural network arc-fault detection with coordinated operation across device, edge, and cloud layers to enhance safety and energy yield [131]. STMicroelectronics has demonstrated a field-tested AFCI solution that uses embedded neural network classification on STM32 microcontrollers to achieve real-time arc-fault detection compliant with UL 1699B [132]. Similar data-driven diagnostic functions are offered by Sungrow’s Isola Cloud platform, which provides cloud analytics and automated IV-curve diagnosis for large PV systems [133]. These developments indicate that AI technologies are moving from theoretical research toward practical deployment.

7. Conclusions

The development of control technology for GCPI is crucial for enabling large-scale integration of renewable energy into power systems. This paper systematically reviews the current progress of inverter control methods and identifies that different techniques exhibit distinct advantages under specific operating conditions. The key findings can be summarized as follows:

• Stable grid environments: Conventional PI/PR and MPC controllers remain the most reliable and cost-effective solutions due to their maturity and simplicity.
• Weak-grid or fluctuating irradiance conditions: AI-based methods, particularly FLC, NN, and ANFIS, offer superior adaptability, improved transient performance, and better harmonic suppression.
• Complex and nonlinear PV system: MPC combined with machine-learning-based weight optimization achieves high accuracy with manageable computational cost.
• Integrated PV–storage and microgrid applications: Hybrid control frameworks (e.g., FLC–SMC, MPC–RL) are the most promising, merging robustness and adaptability for intelligent, autonomous inverter operation.

Despite significant progress, several challenges still hinder large-scale implementation of AI-based controllers. These include high computational overhead, data scarcity, limited interpretability, and hardware constraints. Addressing these issues will require hardware–software co-design, physics-informed and explainable AI, and standardized frameworks for secure and practical deployment.

Looking ahead, AI-based inverter control is expected to evolve from hybrid intelligence toward fully autonomous, grid-interactive operation. Figure 20 presents a potential development roadmap for AI technologies in GCPI over the next decade.

This review has several limitations that should be noted. It mainly relies on literature from WOS, IEEE Xplore, ScienceDirect, and Google Scholar, which may exclude industrial reports, non-English papers, or unpublished research. Most of the analyzed studies are based on simulations or theoretical models rather than large-scale experimental validation. In addition, the reviewed works employ diverse test conditions, hardware configurations, and performance indicators, making direct quantitative comparison across control methods challenging. The evaluation of AI-based controllers further complicates this issue, as their performance depends strongly on the quality of training data, algorithm design, and computational resources, which are not standardized across studies. Consequently, the findings presented in this review should be regarded as indicative trends rather than absolute performance benchmarks. Future research should focus on developing unified experimental frameworks, open-access benchmark datasets, and standardized evaluation metrics to facilitate fair, reproducible, and transparent comparison among different control strategies.

Overall, GCPI control is evolving toward intelligent and adaptive paradigms. As modern power systems demand greater flexibility and autonomy, hybrid control strategies that combine classical and AI-based approaches have become a major research direction. Promoting the transformation of GCPI from “high-efficiency power converters” to “active support nodes within smart grids” will be key to supporting global energy transition and achieving carbon neutrality. However, practical large-scale deployment will still depend on continuous improvements in algorithmic reliability, real-time implementation, standardization, and experimental validation.