Artificial Intelligence Driven Smart Hierarchical Control for Micro Grids―A Comprehensive Review

Abstract

The increasing demand for energy combined with depleting conventional energy sources has led to the evolution of distributed generation using renewable energy sources. Integrating these distributed generations with the existing grid is a complicated task, as it risks the stability and synchronisation of the system. Microgrids (MG) have evolved as a concrete solution for integrating these DGs into the existing system with the ability to operate in either grid-connected or islanded modes, thereby improving reliability and increasing grid functionality. However, owing to the intermittent nature of renewable energy sources, managing the energy balance and its coordination with the grid is a strenuous task. The hierarchical control structure paves the way for managing the dynamic performance of MGs, including economic aspects. However, this structure lacks the ability to provide effective solutions because of the increased complexity and system dynamics. The incorporation of artificial intelligence techniques for the control of MG has been gaining attention for the past decade to enhance its functionality and operation. Therefore, this paper presents a critical review of various artificial intelligence (AI) techniques that have been implemented for the hierarchical control of MGs and their significance, along with the basic control strategy.

1. Introduction

As conventional energy sources have almost been depleted, renewable sources have been receiving greater attention over the past few decades. In addition to this problem, the excessive emission of carbon dioxide is another reason why the current generation of mankind in leading countries is turning towards strengthening their renewable energy policies [1]. These renewable systems are initially interconnected to the existing power system network as distributed generators (DG) [2] to meet excess demand during emergencies. Soon, the operators realised that the individual installation of these DGs could not solve the problem which led to the concept of a microgrid. A microgrid (MG) by itself is a mini power system network that has one or more distribution generation systems connected to it that can operate as semi-autonomous power systems disconnected from grid at the time of exigencies or operate in co-ordination with the grid at the point of common coupling during normal operating conditions [3]. MGs are expected to meet the established standards during the connection/disconnection process at the common point, as indicated in the IEEE standard series 1547. This can be accomplished by implementing various controllers that ensure the seamless implementation of the process.

MGs formed from various distributed renewable energy generation systems (REGS), such as PVs, WECS and HPGS with or without storage, have great potential for solving current complex power system problems. These MGs have various features that distinguish them from existing conventional power systems and hence demand different control strategies

The major problem faced here is the mismatch in the frequency and voltage of various REGS, especially during the islanded mode of operation. Power electronic interfaces, such as VSIs, are employed to synchronise the frequency with that of the grid and hence regulate the voltage and power output of the system. However, they introduce various other new control issues; hence, it is necessary to develop a power electronic interface that can support many micro-generators operating in these two modes. Hence, the control of these power electronic interfaces plays a vital role in the proper functioning of the MG system in either the grid-connected or islanded mode [4]. Several control strategies have been proposed in the literature for the conventional control of these components, including conventional droop control, virtual impedance control and coordinated control. However, the increased penetration of these REGS, along with the active participation of customers, multiplied the complexity of the control system and hence cannot be handled by numerical calculations [5]. With the advent of computational intelligence (CI) techniques such as Fuzzy, Artificial Neural Networks and Deep Learning, which adapt to these uncertainties, the adjustment of the crucial parameters of these interfaces, which have multiple control layers, has been made easy. Incorporating these techniques in the control and operation of MGs has shown promising improvements in the speed, accuracy and effectiveness of the system. Most researchers have implemented these intelligent techniques in the areas of microgrid control, energy management, stability and protection. However, there is still a gap in their application to areas of microgrid control until a physical system or a laboratory test-bed is developed and brought into existence.

Hence, this study presents an intensive review of the application of these CI techniques in the control of microgrids, thereby paving the way for researchers to explore the gap to determine the right technique for their selected control logic. This paper is organised as follows: Section 2 presents an overview of microgrids and their control strategies, Section 3 presents an overview of AI techniques and Section 4 presents the application of these techniques to the primary, secondary and tertiary control of MGs.

2. Overview of Microgrid and Its Control Strategies

A microgrid is a small-scale power system with one or more DG units that are less than 100 MW with or without energy storage elements and power consumers (controllable loads). These MGs can be dc, ac, or hybrid ac/dc, depending on the availability of renewables in that area and the amount of load that must be met during the islanded operation, as shown in Figure 1. It can operate in either the islanded or grid-connected mode, depending on the abnormal or normal situation of the grid. The plug-and-play function is performed at the point of common coupling to enable a continuous power supply, at least for high-priority loads, in the event of grid failure; hence, it is considered a single controllable entity from the grid perspective. The most important feature of this MG is its control strategy, which distinguishes it from conventional distribution grids. This is also required for efficient and stable operation; hence, the control structure should perform the following functions [6]:

• Voltage and frequency regulation during the two different modes of operation.
• Flawless transition of operation between the two different modes of operation.
• Smooth power exchange among the grids (MG and UG).
• Ensuring a reliable power supply for all the connected loads.
• Monitoring the energy requirements and power flow in critical devices for fault identification and protection.
• Maintain power balance along with optimal load sharing.

These required features call for a complex control architecture, which is mostly a hierarchical control strategy that controls the power flow between the DG units and loads.

This control can be implemented in a centralised or decentralised manner.

2.1. Centralised Hierarchical Control (CHC)

The CHC performs its function at three levels to achieve various objectives, such as maintaining power quality, load sharing among DGs, involvement in the energy market and providing ancillary services, in addition to maintaining the voltage and frequency deviation within the prescribed limit [7]. They are

• Grid level—Tertiary control
• Management level—Secondary control
• Field level—Primary control

In this, the primary controller at the field level is associated with the microsources and is provided with limited degrees of intelligence associated with load control capabilities. These primary controllers in the grid-connected mode allow the connection/disconnection of various sources and loads based on the set values communicated to them by the MGCC, while they operate autonomously in the islanded mode. The secondary controller, that is, the MGCC, restores the voltage and frequency deviations produced by the primary controller to facilitate the transition from the grid-forming mode to the grid-following mode. The tertiary controller at the grid level provides information about the active and reactive power requirements of the grid to the secondary controller [8]. The tertiary controller is also responsible for maintaining the stability of the grid with the necessary protection schemes. The complete CHC is indicated in Figure 2.

2.2. Decentralised Control

Decentralised control is mainly agent-based control, where the primary level controllers are given the major responsibility by providing an enhanced degree of intelligence to maximise their autonomy.

These controllers at various DERs can communicate with each other and form an intelligent layer group. Secondary and tertiary controllers are also incorporated with communication protocols connecting them with various DERs at the primary level as well as at the top level. Some intelligent methods that are actually used in decentralised controllers are [9]

• Multi Agent System (MAS)
• Gossip based Algorithm

The most commonly used is the MAS, which also provides optimal voltage regulation. In MAS, each agent has a specific target to satisfy with the help of skills and resources that can communicate with each other and work in coordination with minimal data storage [10]. Although there are several advantages to this MAS approach, such as easy access to information through communication, low expense and plug-and-play capability, it has yet to gain the attention of researchers. Gossip algorithms are developed for calculating the average value of the real and reactive power deviations pertaining to all the DER units, which are then summed to the reference real and reactive power to make zero deviations in voltage and frequency [7].These gossiping algorithms are also developed to improve the supply and demand balance economics by matching the marginal cost functions of DER units

These control strategies are not only important for ensuring a flawless transition between the two modes of operation while maintaining the voltage and frequency deviation, but they are also required for the protection, optimal management and transmission of power. The widely implemented and discussed control strategy is the centralised approach because decision-making with the information collected and the prioritisation of events is easier from a central point. All these requirements call for complex control architecture that have different significances as well as different time scales when compared to the conventional power networks [11] and are implemented to optimally tune the control parameters using any one of the below mentioned controllers: PI/PID controllers, sliding mode controllers, linear quadratic controllers (Regulator/Integrator), Robust or H-infinity controllers, Model Predictive Controllers etc. However, owing to the increased complexity of MG networks, it becomes a very tedious task to model and tune these controllers manually, either heuristically or by tuning algorithms, because of their interdependency, and it is not easy to fix the power sharing of each DG source owing to their capacity constraints and nature dependency. In addition, fixing the PI control parameters is a vital task; hence, researchers are looking for the implementation of intelligent control techniques that utilise various ML algorithms, such as Artificial Neural Networks (ANN) and fuzzy logic (FL), to automatically tune the control system parameters [12]. This not only increases the reliability and stability of the controller but also enhances the power quality and response time. Even hybrid intelligent control techniques comprising FL controllers to tune the parameters and PSO for optimising them are being researched and implemented.

3. Overview of Computational Intelligence Techniques

Artificial Intelligence (AI) is a vast domain of comprehensive science and technologies that deploys various systems, such as computers, robots and software, that combine to think and act with intelligence, similar to humans, by learning from their past experiences to solve complex problems. Its huge data handling capability, combined with powerful hardware and high-speed computation, has opened up a new era in the field of power systems and power electronics. The AI methods that are mainly used in power systems for computations and control (CI) can be broadly categorised into four groups: Machine Learning (ML), Fuzzy Logic (FL), Expert Systems (ES) and metaheuristic methods (MM), each of which is applied for different applications [13].

ML is mainly data-driven algorithms that learn from the dataset and help make accurate decisions or recognise patterns. These algorithms may be Supervised or Unsupervised, or even reinforced, depending on the availability of datasets. In contrast, FL uses linguistic variables for human reasoning to address uncertainties in decision-making processes. Expert systems (ES) use well-defined logical rules and knowledge to replicate decision-making processes and respond to complex queries [14]. MM mainly focuses on solving optimisation problems using various strategies, such as particle swarm optimisation (PSO) and genetic algorithms (GA), to mention a few. Figure 3 illustrates the classification of AI techniques. Among the four techniques, FL and ES are consistently validated in physical microgrid hardware, whereas ML and MMs remain largely simulation-centric with limited real-world deployment due to various issues like need of large training datasets, stability concerns, high computational burden and complex nonlinear dynamic control scenarios, to mention a few.

ML Techniques find vast applications in the fields of modelling, control and monitoring of complex power system problems. Techniques such as artificial neural networks (ANN), Deep Reinforcement Learning (DRL) and support vector machines (SVM) are currently being applied to solve security and stability issues in power systems and increase its operational efficiency [15] because they can understand and learn the characteristics of dynamically changing loads, network configurations and frequencies, which are the predominant characteristics of existing systems.

These algorithms also help to significantly improve control schemes to solve various control and decision processes [16]. The performance analyses of these AI techniques are compared against each other for their application in microgrid control and are presented in

Table 1. Performance Analysis of AI techniques with respect to MG control.

Aspect		Fuzzy Logic (FLC)	Machine Learning (ML: NN, SVM, RL, ANFIS)	Expert Systems (Rule-Based/Knowledge-Based)	Metaheuristic Methods (GA, PSO, DE, ACO, GWO)
1. Stability		√√√	√	√√√	×
2. Adaptability		√√	√√√	√	√√
3. Handling Nonlinearity		√√	√√√	√	√√√
4. Robustness to Uncertainty and Noise		√√√	Depends	√√	Depends
5. Computational Requirements		√	√√	√	√√√
6. Data Requirements		√	√√√	√	√
7. Interpretability		√√√	√√	√√√	√
8.Real-Time Suitability	Primary Control	√√√	√√	√	√
	Secondary Control	√√√	√√√	√√	√√
	Tertiary Control	√√	√√√	√√	√√√
11. Scalability		√	√√√	√	√√√
12. Fault Detection and Self-Healing		√	√√√	√	√√
13. Deployment Maturity		√√√	√√	√	√√
14. Required Expertise		√√√	√√√	√√√	√√√
15. Application in Microgrid Layers		Primary & Secondary control	Secondary (adaptive) & Tertiary EMS	Supervisory decision-making	Tertiary EMS optimisation & tuning
16. Strengths		Simple, robust, interpretable, fast	Adaptive, predictive, high performance	Clear logic, explainable decisions	Strong global search, multi-objective optimisation
17. Limitations		Not scalable, manual tuning	Needs data computation, validation	Rigid, poor for nonlinear dynamic environments	Not suitable for real-time; stochastic behaviour

√√√—High;√√—Medium; √—Low.

. To summarise, AI techniques act synergistically across hierarchical microgrid control by aligning their intrinsic strengths with the timescales and objectives of each layer. Fast, rule-based and fuzzy controllers stabilise real-time inverter dynamics, while ML and expert reasoning in the secondary layer provide adaptive supervisory control. At the tertiary level, global optimisation and forecasting deliver economically optimal dispatch decisions. Information flows upward as real-time data and downward as optimised set-points, establishing a self-learning, coordinated control ecosystem. This cross-layer synergy enhances stability, resilience, energy efficiency and economic operation, making AI-integrated hierarchical control superior to single-layer or single-method approaches. This study focuses on the review of the application of these AI techniques for controlling the operation of microgrids.

4. Artificial Intelligence Techniques Applied to Microgrid Control

AI techniques such as FL and ANN are extensively applied for tuning control system parameters owing to their capability of learning the mapping between input and output variables even when the dynamics of the system under consideration are partially known [17]. In addition, they can effectively compensate for the performance of the system in the event of a disturbance which is more important from the MG point of view, as it is regularly subjected to various disturbances. Therefore, researchers are now focusing on utilising them to effectively control MG. This paper mainly focuses on the application of these AI techniques for CHC at each level.

4.1. Basics of Primary Control

As mentioned in Section 2, this control action is initiated by the LCs at the terminals of the interfacing converter (either grid or load). Primary control is applied to supply the required load with existing power generation and storage while maintaining voltage and frequency deviations within prescribed limits by controlling local variables, such as voltage, frequency and current injection. This is achieved using appropriate controllers. During islanded operation, microgrids operate in autonomous mode to increase the reliability of the system. Hence, their time level is the fastest compared to other levels and includes islanding detection, active and reactive power sharing and successive changes in control modes. Although there are several ways to implement this primary control as mentioned in Figure 4, conventional droop methods and improved droop methods have been extensively explored and used for MG control in the literature, owing to their simplicity and efficiency. The output of these droop controllers is sent to the inner voltage control and current control loop, whose output generates the PWM pulse for the power electronic interfaces after applying the necessary transformation [18].

Droop controller implements the relationship between f-P and V-Q that can be represented as

$$V_{mag}=V_n-D_{Q}Q$$

(1)

$$\omega ={\omega }_n-D_{P}P$$

(2)

where ω and $V_{mag}$ are the frequency and magnitude of the voltage of the DG source, $P$ and $Q$ are the active and reactive power measured at the DG source, $D_P$ and $D_Q$ are the droop coefficients selected according to the rating of the DG sources, and $V_n$ and ${\omega }_n$ are the references for the primary control set by the secondary control.

$$D_P={\displaystyle \frac{∆\omega }{P_{max}}} and D_Q={\displaystyle \frac{∆V}{Q_{max}}}$$

(3)

where $∆\omega$ and $∆V$ are the maximum allowed deviations for the frequency and magnitude of the voltage, and $P_{max}$ and $Q_{max}$ are the maximum real and reactive powers of the DG sources (optimal design of the islanded MG). Thus, the frequency and voltage magnitude are influenced by the P/Q sharing from each of the DG sources as well as the droop coefficients.

Conventionally, proportional integral (PI) controllers are used for DC MGs because of their zero steady-state error in tracking the reference, and Proportional Resonant (PR) controllers are used for AC MGs, as they are more appropriate for following AC command signals. With the advent of ML techniques, these conventional controllers can either be replaced or cooperated with an intelligent controller for improving the performance of MGs.

4.1.1. Application of Fuzzy Logic for Primary Control of MG

Various studies have utilised various fuzzy rules and controllers to tune either the droop coefficients or the PI control parameters. With the advent of fuzzy self-tuning and self-organising gains of the PID controller in real-time to improve the response of the system [19], FL-based PI controllers have started gaining attention. Then, Li (2008) [20] proposed a self-organised/self-tuned fuzzy controller for frequency control of MG in grid-connected and autonomous modes by controlling the power of the electrolyser. The proposed control system comprises an API controller and an FL self-tuning mechanism for adjusting the KP and KI parameters. Although they proved that the proposed controller minimised the tie-line power fluctuation despite frequency control, there was no mention of the voltage profile of the system which cannot be bypassed. C.N. Papadimitriou (2010) [21] proposed fuzzy logic-based local controllers for the primary control of frequency and local bus voltage following a disturbance in either islanded or grid-connected mode, as fuzzy controllers are flexible and adaptive in nature to nonlinear systems. They used three fuzzy controllers for the combined Fuel Cell and Battery bank system and two fuzzy controllers for the DFIG system, which were designed with the heuristic knowledge of the complete system and thoroughly iterated for fine-tuning of the control parameters. Hiroaki Kakigano (2013) [22] proposed a fuzzy control with gain scheduling control techniques for voltage control of a primary controller under islanded operation of MG with a storage system and also accomplishes energy management as well. They proposed an implemented fuzzy control that is applicable for energy balance control, which reduces losses compared to conventional droop control. He considered electric vehicles and double-layer capacitors as storage systems and experimentally validated his results.

A similar work for controlling the DC grid voltage using an FL based PID controller was done by R.K. Chauhan (2015) [23]. The authors used real-time data from a project in the USA with only PV systems and proved that the steady-state error of voltage was reduced to less than 0.3%, and the peak time, peak overshoot and settling time decreased. This justifies the fact that the FL-PID reduces the voltage fluctuations in the MG, making it more stable. In contrast, Eldessouky (2016) [24] implemented fuzzy model reference learning control (FMRLC) with PID into the control of a static VAR compensator for voltage regulation of an MG in islanded mode. The learning mechanism of the FLC compensates for load variations by updating the controller parameter and hence maintains the system parameters as is required for longer operating conditions. The author also tested the same logic and MHSA technique for the adaptive tuning of the parameters of the PI controller based on the load—frequency deviation in an islanded MG with load—generation mismatch.

A two-stage strategic control was proposed by Fattahi (2016) [25], where the first stage incorporates a fuzzy controller to find the adaptive droop co-efficient, while the second stage incorporates another fuzzy controller to fix the Q sharing errors by controlling the small real power injections. The results were validated using both time and frequency domain analyses. Chin Hsing. Cheng (2016) [26] proposed a simple FLC for each of the DGs connected to the MG, each having one input as the deviation of that specific converter power and the other input in common, which is the deviation in the load power. The output is the control signal for the duty cycle of the converter to regulate the voltage in the MG, whereas optimal power dispatch is performed using dynamic programming to maintain the power flow balance. The entire system was experimentally validated for different load conditions.

Jayapriya M (2016) [27] proposes a FLC that stabilises the frequency during islanded operation by managing the charging and discharging of batteries as per the target power and the battery’s state of charge (SOC). The rules are implemented using Python with the help of the FL library in Raspberry Pi, thereby providing a cost-effective solution, fast fuzzification and quicker logical decisions. Shivam (2017) [28] designed a fuzzy interface system to reduce DC bus voltage degradation by adjusting the droop resistance using fuzzy logic for source converters in a DC MG. Mohammed Hassan Khaboon (2017) [29] tuned the PI parameters automatically by a general type—II fuzzy-based PI with online measurements, where the parameters of controllers are calculated using a heuristic algorithm named Modified Harmony Search Algorithm (MHSA). These three components are considered in the optimisation process: usage of harmony memory, pitch adjustment and randomization and proved that through this rolling optimisation process and adaptive technique, the proposed controller could be able to coordinate the Electric Vehicle’s and DGs output. Hamed Moaziami (2017) [30] combines optimal location and sizing of DG units, considering fuel cost, voltage stability index and variation in voltage as objective functions to minimise/optimise the droop parameters using Fuzzy approach.

Ainur Rofiq Nansur (2018) [31] proposes an FLC for maintaining a constant Voltage at the DC bus when PV coupled with MPPT is used as a DG. With the FLC, the voltage error is maintained at less than 3%. Yang Mi (2019) [32] proposed droop control with optimised droop coefficients by the Takaji-Sujeno Fuzzy method so that the load changes on the DC MG are regulated in a better way. Feng N (2019) [33] proposed an FLC for the virtual capacitor droop control of HESS in a DC microgrid, where the droop coefficients are varied in accordance with the line impedance and load variation. Here, the system realises the full advantage of HESS by using a decentralised controller rather than a centralised controller. Concurrent-learning-based DRL can also replace the conventional droop control with a droop-free control algorithm, which can vary the droop resistance while maintaining the trade-off between the variation in current and voltage [34]. Adapting nonlinear control for real-time modification of droop coefficients is one of the best solutions for overcoming the uneven distribution of reactive power from micro-sources. Therefore, Ding (2020) [35] proposed an ANFIS-based neuro-fuzzy controller to compensate for these droop coefficients by considering various inputs, such as load disturbances, active power and reactive power. The results prove that this adaptive method balances the distribution of reactive power, even when there is a mismatch in the line impedance. The same problems were solved by a vector-controlled feed-forward droop controller by Yadav (2021) [36], where a fuzzy logic control system tunes the loop coefficients aiming at a power balance between parallel inverter operations based on the loads.

In contrast, Guo (2022) [37] designed a fuzzy controller to correct the droop coefficients for a bipolar (multimode) DC microgrid to improve the unbalanced voltage. Javeria Noor (2024) [38] also implemented a fuzzy logic-based control strategy using Takagi-Sugeno’s approach for real-time power distribution in a bipolar DC microgrid. He incorporated a SEPIC-Cuk converter for voltage balancing of bipolar transmission lines. The application of Fuzzy logic in primary control of microgrid is summarized in

Table 2. Summary of Fuzzy Logic applications in primary control of microgrid.

Ref.	Intelligent Technique Used	System Considered	Software/Hardware	Control Application	Standalone/Grid-Connected
[20]	FLC/PID	PV/Wind/FC electrolyzer/MT	Matlab/Simulink	Power balance/Frequency	Grid-connected/Standalone
[21]	FLC	FC/battery/Wind/	Matlab/Simulink	Voltage and Frequency	Grid-connected/Standalone
[22]	FLC	EDLC	Lab based Hardware	Voltage control/Droop control	Grid-connected
[23]	FL/PID	PV	Matlab/m-file	Voltage Control	Grid-connected
[24]	FMRLC	Wind/SG	Matlab/Simulink	Voltage Control	Standalone mode
[25]	FLC	PV	Matlab/Simulink	Voltage and Frequency	Grid-connected
[26]	FLC & DP	PV/WG/FC/ESS	Lab based Hardware	Voltage and Power control	Standalone mode
[27]	FLC	WG/DESS	Matlab/Simulink/Hardware	Power/Frequency control	Standalone mode
[28]	FIS	DC source	Matlab/Simulink	Voltage and power	Standalone mode
[29]	FLC and PID	Wind/EV/Diesel Gen	Matlab/Simulink	Voltage	Standalone mode
[30]	Fuzzy approach	DGs	Matlab/Simulink	Optimal point of location	Grid-connected
[31]	FLC	PV	Matlab/Simulink	MPPT/Voltage control	Standalone mode
[32]	Fuzzy sliding mode	DGs	Matlab-PSIM	Droop control/Voltage	Standalone mode
[33]	FLC	PV/Multiple HESS	PSCAD/ EMTDC	Voltage control	Standalone mode
[34]	FLC	PV/Wind/DEG/ESS	Matlab/Simulink	Voltage control	Grid-connected
[35]	ANFIS	DGs	Matlab/Simulink	Voltage control	Standalone mode
[36]	FLC	DGs	Matlab/Simulink	Frequency and Power sharing	Standalone mode
[37]	FLC	PV/WT/SC/Battery	Matlab/Simulink	Droop Control/Voltage control	Grid-connected
[38]	FLA	PV/Battery	Matlab/Simulink	Voltage control	Standalone mode

.

4.1.2. Application of Artificial Neural Network for Primary Control of MG

While some researchers have dealt with FLC, many have also worked on ANN-based controllers for MG control and stability enhancement. They are mainly used to improve the performance of the maximum power point tracker in the case of PV and WES and can also be used to predict and analyse the SOC of various types of batteries that are used as storage elements in the MG. Eddahech 2011 [39] developed an FFNN-based model for the voltage control of Li-Po batteries. Refs. [40,41,42,43,44] discuss the application of NN models for designing and improving the performance characteristics of various types of batteries that form part of an MG.

In addition, AI is used to improve stability and control to ensure an efficient and reliable supply to consumers. MDNN was designed by Sabahi in 2007 [45] to minimise frequency derivation in the tie-line power, while Baek in 2008 [46] designed NN for non-linear parameter optimisation of first-order derivatives in the power system. An Inverter Control based on ANN was proposed by Singh in 2014 [47] for an MG based on PV and Fuel cells. The authors claimed to have achieved a better THD with ANN control compared to conventional control methods. An ANN-based controller trained by app dynamic programming was introduced by Weizen Dong (2018) [48] for a DC microgrid. Droop control was implemented for each DG for power sharing, and its reference voltage output was used to train the ANN to maintain the voltage within the limit from the primary control point of view. Although he claimed that there was a rapid change in the converter current to meet the new load with less dip in voltage, the system under study only had a DC source, and the condition of the system when connected to the grid was not specified. T.Vignesh (2016) [49] proposes a feed forward Neural Network for proper sharing of real and reactive power as a part of primary droop control for a two DG system. The study shows that despite proper sharing of power; the proposed controller also controls the voltage and frequency of the MG under all operating conditions. For the same system, he also compared the settling time, peak overshoot and rise time with respect to the PID controller and proved that the FLC was successful in improving these parameters, in addition to controlling V and F. Habibi (2019) [50] employed ANN to tune the coefficients of a generalised droop control-based DR, which regulates the voltage profile and maintains a stable frequency using flexible loads.

Liu [51] proposed a DRL algorithm that regulates the voltage and supports current sharing as well as load sharing. Refs. [52,53,54,55] prove that the inability of the PI controller to regulate the inverter output has been overcome by dynamic programming based on NN. FFNN can also be used to improve droop control performance by replacing virtual impedance control [56]. The effectiveness of ANN ANN-based adaptive PI controller was investigated in ref. [57] by reducing the THD, improving power sharing and maintaining IEEE-519 standards. Ref [58] replaced the conventional fixed-gain PID controller with a sliding mode droop control scheme supported by ANN to deal with the high rate of change in frequency owing to the integration of PVs and batteries, which considerably reduced the inertia of the system.

Instead of optimising the internal parameters of ANN using various heuristic approaches, Ref. [59] trains the ANN with optimised dq-axis voltage references obtained by offline training using PSO to compensate for the effects of load variations and transmission line variations. One step further, Akpolat [60] introduced a model predictive control trained by ANN to replace the conventional PI controller trained by ANN for attaining quicker damping. Zaman [61] proposed a primary controller to adjust the gains for disturbance compensation using a PI controller trained by the generalised Hebb’s learning law. An ANN-based PR regulator was proposed by Derban et al. in 2024 [62] for regulating the output voltage with minimum deviations to improve the performance of the VSI. In addition, they incorporated an additional droop control for power sharing among the DGs. The application of AI techniques in primary control of MG is summarised in the

Table 3. Summary of ANN applications in primary control of Microgrid.

References	Intelligent Technique Used	System Considered	Software/Hardware	Control Application	Standalone/Grid-Connected
[39]	ANN	Li-Po Battery	Matlab/Simulink and Hardware	Voltage control	Standalone
[40]	BPNN	Battery	Matlab/Simulink and Hardware	SOC Voltage control	Standalone
[41]	NN	Lead acid Battery	Matlab/Simulink	SOC Voltage Estimation	Standalone
[42]	ANN	Ni-MH battery	Matlab Simulink and Hardware	SOC Voltage Estimation	Standalone
[43]	ANN	Li-Battery	Matlab/Simulink and Hardware	Estimation of charge discharge behaviour	Standalone
[44]	FFNN	SOFC	Matlab/Simulink	Performance Prediction	Standalone
[45]	CNN/MDNN	Two area interconnected systems	Matlab/Simulink	Frequency control	Grid- connected
[46]	FFNN	Two area interconnected systems	Matlab/Simulink	Frequency control	Grid- connected
[47]	ANN	PV/FC	Matlab/Simulink	Power quality/DVR control	Standalone
[48]	ANN	DGs	Matlab/Simulink	Power sharing and voltage control	Standalone
[49]	ANN	PV/BESS/SOFC	Matlab/Simulink	Voltage and frequency control	Standalone
[50]	ANN	DGs	Matlab/Simulink	Frequency control	Standalone
[51]	DRL	DGs/ESSs	PSCAD/EMTDC	Voltage control	Standalone
[52]	HDP	DGs	Matlab/Simulink	Virtual inertia	Grid-connected
[53]	Dual DHP NN	DGs	Matlab/Simulink	P and Q sharing	Grid-connected
[54]	HDP	DGs	Matlab/Simulink and Hardware	P and Q sharing	Grid-connected
[55]	DRL	DGs	Matlab/Simulink	Voltage control	Standalone
[56]	FFNN	DGs/BESS	PHIL	Voltage and Frequency control	Grid-connected
[57]	ANN	DGs/BESS	Matlab/Simulink	Voltage control	Grid-connected
[58]	ANN	PV	Matlab/Simulink	Frequency control	Standalone
[59]	NN	PV/Wind/FC	Matlab/Simulink	Power sharing	Standalone
[60]	ANN MPC	PV/Battery	Matlab/Simulink and Hardware	Voltage and Current control	Grid-connected
[61]	ANN	PV/BESS	Matlab/Simulink and Hardware	Voltage control	Grid-connected
[62]	ANN	DGs	Opal RT	Voltage control	Grid-connected

.

4.2. Basics of Secondary Control

Primary control is implemented to achieve a balance between the required voltage regulation and power sharing at the expense of each other. The primary control leaves a steady-state error, which is fixed by the secondary control. Various secondary control methods are shown in Figure 5. This control level is used to fix the problem of voltage deviation by sending the error signal to all DG sources to restore their output voltage.

So, for the DC grid, the voltage equation now becomes

$$v_o^{\ast }=v_{ref}-R_D{i}_0+\delta v_o$$

(4)

$v_o^{\ast }$: Desired or reference output voltage after compensation.

$v_{ref}$: Base reference voltage.

$R_D$: Droop resistance or a virtual resistance used for load sharing or stability.

$i_0$: Output current.

$\delta v_o$: A small-signal voltage term or disturbance compensation term.

For AC MG, now the equation becomes

$${\mathrm{V}}_{\mathrm{m}\mathrm{a}\mathrm{g}}={\mathrm{V}}_{\mathrm{n}}-{\mathrm{D}}_{\mathrm{Q}}\mathrm{Q}+\mathsf{\delta }\mathrm{E}$$

(5)

$$\mathsf{\omega }={\mathsf{\omega }}_{\mathrm{n}}-{\mathrm{D}}_{\mathrm{P}}\mathrm{P}+\mathsf{\delta }\mathsf{\omega }$$

(6)

${\mathrm{V}}_{\mathrm{m}\mathrm{a}\mathrm{g}}$: Magnitude of the regulated voltage in the system.

${\mathrm{V}}_{\mathrm{n}}$: The nominal or reference voltage.

${\mathrm{D}}_{\mathrm{Q}}$: Coefficient related to the sensitivity of voltage magnitude to changes in reactive power.

$\mathrm{Q}$: Reactive power responsible for regulating the voltage in AC power systems.

$\mathsf{\delta }\mathrm{E}$: Changes in voltage due to dynamic conditions or measurement errors.

$$\mathsf{\omega }={\mathsf{\omega }}_{\mathrm{n}}-{\mathrm{D}}_{\mathrm{P}}\mathrm{P}+\mathsf{\delta }\mathsf{\omega }$$

(7)

$\mathsf{\omega }$: Angular frequency of the grid.

${\mathsf{\omega }}_{\mathrm{n}}$: The nominal frequency or desired frequency.

${\mathrm{D}}_{\mathrm{P}}$: Coefficient relating the sensitivity of frequency to changes in the active power.

$\mathrm{P}$: The active Power.

$\mathsf{\delta }\mathsf{\omega }$: Dynamic deviations or errors from the nominal value.

The voltage and frequency in each MG are measured, and the difference between them and the reference voltage and frequency is calculated as errors that are to be processed by the compensators and sent to all the MGs for restoration.

The output frequency, as well as the voltage, is regulated using these error signals, which are given by

$$\mathsf{\delta }\mathrm{E}={\mathrm{K}}_{\mathrm{P}\mathrm{E}}\left({\mathrm{E}}_{\mathrm{M}\mathrm{G}}^{\ast }-{\mathrm{E}}_{\mathrm{M}\mathrm{G}}\right)+{\mathrm{K}}_{\mathrm{i}\mathrm{E}}\int \left({\mathrm{E}}_{\mathrm{M}\mathrm{G}}^{\ast }-{\mathrm{E}}_{\mathrm{M}\mathrm{G}}\right)\mathrm{d}\mathrm{t}$$

(8)

$$\mathsf{\delta }\mathsf{\omega }={\mathrm{K}}_{\mathrm{P}\mathsf{\omega }}\left({\mathsf{\omega }}_{\mathrm{M}\mathrm{G}}^{\ast }-{\mathsf{\omega }}_{\mathrm{M}\mathrm{G}}\right)+{\mathrm{K}}_{\mathrm{i}\mathsf{\omega }}\int \left({\mathsf{\omega }}_{\mathrm{M}\mathrm{G}}^{\ast }-{\mathsf{\omega }}_{\mathrm{M}\mathrm{G}}\right)\mathrm{d}\mathrm{t}$$

(9)

Here ${\mathrm{K}}_{\mathrm{P}\mathsf{\omega }}$, ${\mathrm{K}}_{\mathrm{i}\mathsf{\omega }}$, ${\mathrm{K}}_{\mathrm{P}\mathrm{E}}$ and ${\mathrm{K}}_{\mathrm{i}\mathrm{E}}$ are the PI control parameters for secondary control to be set to limit the δω and δE to be within deviation limits [63].

In summary, droop control strategies are often used in primary control to maintain the generation-level balance that makes the MG stable. However, it introduces a voltage and frequency deviation that is dependent on the load. This problem can be solved using PI controllers. However, conventional PI or droop-based secondary controls often respond slowly to the voltage and frequency deviations introduced by the primary control. In addition, inaccurate control can lead to poor power sharing or unstable voltage/frequency profiles. In addition, secondary control typically requires centralised or distributed communication among DERs.

Therefore, communication delays, losses or cybersecurity issues can desynchronise DERs and compromise system reliability. Improper communication or coordination among control units can result in the desynchronisation of inverter-based sources, frequency/voltage instabilities and poor power quality. They also struggle to handle fault-induced dynamics without advanced fault detection and control schemes [12]. To tackle these issues, intelligent control techniques such as fuzzy logic, artificial neural networks, reinforcement learning and model predictive control (MPC) have been proposed. These can offer:

• Faster, adaptive and decentralised control.
• Reduced communication needs.
• Fault tolerance and resilience.
• Better real-time decision-making.

The following section deals with the literature review of intelligent secondary control of microgrids in detail.

4.2.1. Application of Fuzzy Logic for Secondary Control of MG

FLCs are very effective for controlling systems that are usually complex, nonlinear or poorly defined; hence, they are suitable for applications where conventional control methods, such as PID controllers, are not effective. They enhance the PID’s performance by compensating for the areas in which the gains do not perform well. The combination of PI control and a fuzzy control strategy combine the advantages of both and makes them suitable for complex applications, such as the control of power system networks.

Hettiarachchi [64] reviewed the application of FL for the control of an AC-DC MG system in a multi-agent-based approach. He presented a summary of papers that used FL for DC bus bar control, frequency control and AC voltage control. This was helpful for many researchers in consolidating the application of FLC for MG control. Bevrani H [65] proposed an intelligent PSO-based Fuzzy PI control for mitigating the variations in frequency and demonstrated its effectiveness by comparing it with the pure fuzzy PI method. However, the frequency deviations are still high from the standard value. Therefore, Annamraju [66] introduced a two-stage adaptive FL-based PI controller and optimised the membership function by adding GWO along with PSO with a reduced number of variables. The controller was robust and efficient. A dynamic secondary LFC based on a sliding mode approach was proposed in ref. [67], where an artificial gorilla troop optimisation technique was used for tuning the coefficient, and its effectiveness was validated on the IEEE 14-bus system by comparing it with PSO and GWO. African Vulture optimisation (AVOA) was used in ref. [68] to optimise the control variables and obtain fixed gains in the cased PI-FL controller in a simulation, as well as tested in HIL Opal RT and a rapid prototype for its effectiveness. Prusty [69] also justified the result of the Opal RT for a fractional-order type 2 Fuzzy PID controller for frequency control with the impact of storage devices.

While Alizera [70] proposed the ANFIS technique to adaptively optimise the PI controller coefficients to compensate for the deviations in voltage and frequency for different study cases, Elian [71] introduced a fuzzy secondary controller with reduced communication links with local measurements to maintain the deviation in voltage and frequency within an acceptable tolerance. Even when subjected to an unintentional landing, the controller maintained the voltage and frequency without using an islanding detection system. Furqan [72] developed a new mechanism for frequency and voltage stabilisation with battery storage instead of diesel. The proposed controller maintained the deviations and controlled the charging of the battery storage. Refs. [73,74] also used the battery of electric vehicles as an alternative to BESS from an economic perspective. In contrast, Ref. [75] proposed an FLC with separately defined frequency and voltage references necessary to guarantee minimum deviation at each instant of time.

Refs. [76,77] also proposed an FLC with a self-tuning fractional order for load frequency control of a microgrid to minimise the violation in the amplitude of the frequency. On the other hand, Ref. [78] proved that the FL controller is better than the PID and ANN-based controllers in terms of efficiency and precision to control them under different loads and solar irradiation conditions. While all these studies focused on Standalone systems, Neves [79] proposed a multitask fuzzy secondary controller that changes from grid-tied to Standalone control in the event of isolation.

Table 4. Summary of Fuzzy Logic applications in Secondary control of Microgrid.

References	Intelligent Technique Used	System Considered	Software/Hardware	Control Application	Standalone/Grid-Connected
[65]	FL-based PI controller	PV/MT/CHP/FC/BESS/Flywheel	Matlab/Simulink	Secondary Frequency control	Grid-connected
[66]	FL-based PI controller	PV/Wind/DEG/FC/BESS	Matlab/Simulink	Secondary Frequency control	Standalone
[67]	FLSMC	Wind/Diesel Gen.	Matlab/Simulink	Secondary Frequency control	Standalone
[68]	FLC	DGs	Matlab/Simulink/Opal-RT	Voltage control	Standalone
[69]	Fuzzy-PID	DGs/EV	Matlab/Simulink/Opal-RT	Secondary Frequency control	Grid-connected
[70]	ANFIS	PV/WT/MT/FC	Matlab/Simulink	Secondary Voltage and Frequency control	Grid-connected
[71]	Fuzzy Sec controller	DGs	PSCAD Simulation	Frequency control	Standalone/Grid-connected
[72]	FLC	Diesel gen/PV/WT/DEG/BESS/EV	Matlab/Simulink	Secondary Voltage and Frequency control	Standalone
[73]	FLC	Multiple MG with EV	Matlab/Simulink	Frequency control	Grid-connected
[74]	Fuzzy PI	PV/DEG/Wind/BSS	Matlab/Simulink	Frequency control	Standalone
[75]	FLC	DGs	Simulation *	Secondary Voltage and Frequency control	Standalone
[76,77]	FO-FLC	PV/WT/FC/ESS	Matlab/Simulink	Load Frequency control	Standalone MG
[78]	FLC	PV/WT/Diesel Gen/ESS	Matlab Simulink	Voltage control	Standalone MG
[79]	FLC	DGs	d-SPACE and Hardware	Secondary Voltage and Frequency control	Standalone/Grid-connected

* Simulation tool not mentioned.

consolidates the FL application for the secondary control of MG.

4.2.2. Application of ANN for Secondary Control of MG

ANN-based tuning for regulating the voltage and frequency of microgrids has been practiced for a long time [80]. Rohit Trivedi et al. [12] recently reviewed how these AI techniques have been implemented in microgrid control. They discussed the use of ANN for deviation control, as well as its application to mitigate communication delay, fault restoration and protection. However, Ref. [81] implemented secondary layer voltage control without a communication channel for an islanded MG using a distributed ML technique. The authors in refs. [82,83,84] used ANN for the control of V and f deviations, with differences in how the parameters are tuned; ref. [82] used a multilayer perceptron for the selection of parameters based on various aspects of the controller, while ref. [83] used GA for parameter initialisation and ANN for online tuning. Adaptive dynamic programming (DP) is used to manage current sharing and regulate DC voltage in refs. [85,86]. Refs. [87,88,89,90,91,92] used DRL to regulate voltage and current sharing and improve the efficiency of the MG.

Various DRL algorithms have been used for secondary voltage and frequency control of MG by adjusting battery power, offline and online training, gain scheduling, etc. [93,94,95,96,97,98,99]. Ali M. Jasim [100] proposed ANN-based online tuning of PI controllers for intelligent V and f variation mitigation as well as ANN-based reactive power controllers for secondary control. They proposed PR-based primary control for accurately sharing power among the DGs. He also proposed GA-optimised ANN-based controllers to minimise V and f fluctuations and share the active and reactive power under primary control [101]. Recently, ref. [102] employed a GA-based Pi controller with fine-grained online tuning to mitigate the V and f deviations and validated their results on the ThingSpeak platform with real-time data. Similar work was performed by Alshalawi [103], who proposed an adaptive controller with a GA optimiser and ANN for real and reactive power control in a microgrid. The combination of GA with ANN for tuning PI controller parameters can improve the stability of the system by mitigating the fluctuations in frequency, as justified by Dashtdar [104]. The PSO-based ANN technique is also used to tune the PID controller in an islanded MG with an EV incorporated [105]. Their results showed that the system dynamic characteristics and stability improved with the use of an ANN controller. Ref. [106] used HBA (Honey badger algorithm) to train the gain values of a PID controller at the first stage, and ANN is trained in the second state to match them with the tie-line power. The trained controller exhibited a better dynamic response and stability with a minimum deviation in frequency. In contrast, Ref. [107] used a BFOA tuned-PID controller for reducing the frequency deviations in the MG system powered by various DGs.

Adaptive fuzzy-based NN inherited with total sliding mode control (TSMC) for secondary control of MG not only improves the stability of the system, as each DG requires self-information, and of its neighbours only, but is also robust to unpredictable disturbances [108]. IoT-based secondary control of MG is proposed in ref. [109] which maintains the voltage and frequency profile in a stable range with less than 50 ms time delay. With the help of IoT, cloud-based ML provides efficient islanding detection methods without many power quality issues. The duty cycle of the DC-DC converter connected to PV sources is fine-tuned for regulating the power flow outputs in accordance with other sources as well as solar irradiance by an ANN controller [110], which not only improves the transient performance but also ensures satisfactory power delivery to the grid in case of excess generation and power reversal in the event of a deficit.

Kumar et al. [111] forecasted solar and wind power using a Long Short-term Memory Recurrent Neural Network (LSTM-RNN) and injected it into the developed MG small-signal model with auxiliary and secondary sources. When the frequency deviates owing to load disturbance, the controller responds by initiating secondary sources to vary the power shared with the MG. OME-DRL was incorporated for data-driven secondary frequency control in ref. [112] to obtain multi-objective optimality and reduce generation cost. Ref. [113] proposed a nonlinear mode for secondary voltage control under normal operation and a linear mode under large disturbance to maintain stability, and the switching between the modes was performed by proper estimation of the hyperparameters of ANN. While Ho Pham [114] proposed a MIMO-based ANN model for supervisory control of an MG system, Umashankar [115] proposed a MIMO-based ANFIS controller for charging an EV in an MG system powered by a PV and FC. As witnessed from the above, researchers are still exploring the application of AI techniques to improve the control of MG at the secondary level.

Table 5. Summary of ANN applications in Secondary control of Microgrid.

References	Intelligent Technique Used	System Considered	Software/Hardware	Control Application	Standalone/Grid-Connected
[80]	ANN	DGs	Simulation	Voltage and Frequency control	Standalone
[81]	Cluster-based NN	PV/Wind	Matlab Simulink	Voltage control	Standalone
[82]	Adaptive NN	Syn. Generator	Matlab Simulink	Sec. Voltage and Frequency control	Grid-connected
[83]	ANN/GA	DGs	Matlab Simulink	Sec. Voltage and Frequency control	Standalone
[84]	ANN/RLA	Wind Gen./Syn. Generator	Simulation *	Voltage and Current control	Standalone
[85]	NN	DGs	Matlab Simulink	Voltage control	Standalone
[86]	ANN/DDHP	DG	Matlab Simulink	Frequency control	Standalone
[87]	DRLA/DNN	DGs	Matlab Simulink	Voltage and Frequency control	Standalone
[88,89,90,91,92]	ANN/DRLA	DGs	Matlab Simulink	Sec. Voltage and Frequency control	Standalone
[93,94,95,96,97,98,99]	ANN/DRLA	DGs and Battery	Matlab Simulink	Sec. Voltage and Frequency control	Standalone
[100]	ANN	DGs	Matlab Simulink	Voltage and Frequency control	Standalone
[101]	ANN and GA	DGs	Matlab Simulink	Voltage, Frequency control P and Q sharing	Standalone
[102]	ANN and GA	PV/WT/EV/Battery	Matlab Simulink	Voltage and Frequency control	Standalone
[103]	GA-ANN	PV/Battery	Matlab Simulink	P and Q control	Grid-connected
[104]	ANN and GA	VTG/PV/BESS/FESS/FC	Matlab Simulink	Frequency control	Standalone
[105]	PSO-based ANN	DGs/EV	Matlab Simulink	Load frequency control	Standalone
[106]	ANN based Adaptive PID	MGs	Matlab Simulink	frequency control	Standalone
[107]	PID with BFOA	DGs	Matlab Simulink	Load frequency control	Standalone
[108]	AFNN	DGs	Matlab Simulink	Voltage and Frequency control	Standalone
[109]	ANN	DGs	Matlab Simulink	Islanded mode detection	Standalone/Grid-connected
[110]	ANN	PV	Matlab Simulink	Voltage control	Grid-connected
[111]	LSTM-RNN	PV/Wind/FC/BESS/FESS/V2G	Matlab Simulink	Load Frequency control	Standalone/Grid-connected
[112]	DRL-ANN	PV/Wind/Diesel Gen./MT/FC	Simulation *	Frequency control	Standalone
[113]	ANN	PV	Matlab Simulink	Voltage control	Standalone
[114]	Forwarded Adaptive NN	PV	Matlab Simulink	Load power balance	Standalone/Grid-connected
[115]	ANFIS	PV/FC/EV	Matlab Simulink	Voltage and Current control	Standalone

* Simulation tool not mentioned.

provides an overview of the application of ANN for the secondary control of MG.

4.3. Basics of Tertiary Control

This is the last and slowest control level that manages the power flow between the main grid and microgrid, considering optimal and economic concerns. This real and reactive power flow can be managed by adjusting the voltage amplitude and frequency after obtaining their reference values from the below equations

$${\mathsf{\omega }}^{\mathrm{r}\mathrm{e}\mathrm{f}}={\mathrm{K}}_{\mathrm{P}\mathrm{P}}\left({\mathrm{P}}_{\mathrm{G}}^{\mathrm{r}\mathrm{e}\mathrm{f}}-{\mathrm{P}}_{\mathrm{G}}\right)+{\mathrm{K}}_{\mathrm{i}\mathrm{P}}\int \left({\mathrm{P}}_{\mathrm{G}}^{\mathrm{r}\mathrm{e}\mathrm{f}}-{\mathrm{P}}_{\mathrm{G}}\right)\mathrm{d}\mathrm{t}$$

(10)

$${\mathrm{E}}^{\mathrm{r}\mathrm{e}\mathrm{f}}={\mathrm{K}}_{\mathrm{P}\mathrm{Q}}\left({\mathrm{Q}}_{\mathrm{G}}^{\mathrm{r}\mathrm{e}\mathrm{f}}-{\mathrm{Q}}_{\mathrm{G}}\right)+{\mathrm{K}}_{\mathrm{i}\mathrm{Q}}\int \left({\mathrm{Q}}_{\mathrm{G}}^{\mathrm{r}\mathrm{e}\mathrm{f}}-{\mathrm{Q}}_{\mathrm{G}}\right)\mathrm{d}\mathrm{t}$$

(11)

Tertiary control also coordinates to enhance system quality by optimal operation and energy management by considering economic aspects as well as the power flow between the MG and grid or among the MGs [116]. They deal with issues such as (i) the economic dispatch aspect of the trade-off between the MG and energy market at the event of excess generation, (ii) setting the optimal operating point of MG as per the requirements of the main grid, (iii) needs of MG with respect to regulation of voltage and frequency in grid-connected or islanded mode and (iv) active/reactive power management within MG cluster/with grid [117].

The equations governing various operations in tertiary control as given in Figure 6 include

(i)

Economic dispatch

$$\mathrm{Minimise} \mathrm{the} \mathrm{total} \mathrm{generation} \mathrm{cost}: \mathrm{m}\mathrm{i}\mathrm{n}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\mathrm{C}}_{\mathrm{i}}{(\mathrm{P}}_{\mathrm{i}})$$

(12)

Subject to constraints

$$\mathrm{Power} \mathrm{balance} {\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\mathrm{P}}_{\mathrm{i}}={\mathrm{P}}_{\mathrm{L}\mathrm{o}\mathrm{a}\mathrm{d}}+{\mathrm{P}}_{\mathrm{L}\mathrm{o}\mathrm{s}\mathrm{s}}$$

(13)

$$\mathrm{Generator} \mathrm{limits} {\mathrm{P}}_{\mathrm{i}}^{\mathrm{m}\mathrm{i}\mathrm{n}}\le {\mathrm{P}}_{\mathrm{i}}\le {\mathrm{P}}_{\mathrm{i}}^{\mathrm{m}\mathrm{a}\mathrm{x}}$$

(14)

where

${\mathrm{C}}_{\mathrm{i}}{(\mathrm{P}}_{\mathrm{i}})$: cost function of generator i.

${\mathrm{P}}_{\mathrm{i}}$: active power output of generator i.

N: number of distributed generators (DGs).

(ii)

Power Flow Equations (for interconnected microgrids or grid-tied systems)

To ensure optimal power exchange between nodes, i and j, power exchange between the nodes i and j should be

$${\mathrm{P}}_{\mathrm{i}\mathrm{j}}={\mathrm{V}}_{\mathrm{i}}{\mathrm{V}}_{\mathrm{j}}{\mathrm{Y}}_{\mathrm{i}\mathrm{j}}\mathrm{c}\mathrm{o}\mathrm{s}({\mathsf{\theta }}_{\mathrm{i}}-{\mathsf{\theta }}_{\mathrm{j}}-\angle {\mathrm{Y}}_{\mathrm{i}\mathrm{j}})$$

(15)

$${\mathrm{Q}}_{\mathrm{i}\mathrm{j}}={\mathrm{V}}_{\mathrm{i}}{\mathrm{V}}_{\mathrm{j}}{\mathrm{Y}}_{\mathrm{i}\mathrm{j}}\mathrm{s}\mathrm{i}\mathrm{n}({\mathsf{\theta }}_{\mathrm{i}}-{\mathsf{\theta }}_{\mathrm{j}}-\angle {\mathrm{Y}}_{\mathrm{i}\mathrm{j}})$$

(16)

(iii)

Optimal power flow

When the aim of the tertiary control is to optimise power dispatch, the objective function could be

$$\begin{gathered} \min \sum {\mathrm{C}}_{\mathrm{i}}{(\mathrm{P}}_{\mathrm{i}}) \\ {\mathrm{P}}_{\mathrm{i}}{\mathrm{Q}}_{\mathrm{i}} \end{gathered}$$

(17)

Subject to

• Active/reactive power flow constraints.
• Voltage magnitude constraints.
• Line thermal limits.

(iv)

Voltage and Frequency Reference setting

Tertiary control also sets the references for voltage and frequency that primary and secondary controls track as represented by Equations (10) and (11).

As the last control level, the time frame for the action is relatively large; hence, it is a slow control level when compared to the primary and secondary levels discussed earlier, as indicated in Figure 7. Various optimisation algorithms can be implemented to ensure economic operation, and when all the DGs operate at the same marginal cost, optimal economic operation is achieved [118]. Tertiary control may be centralised or decentralised, based on requirements. In the centralised control, the microgrid central controller (MGCC) manages the aforementioned activities, that is, optimises microgrid operation, manages power exchange with the main grid, and provides set points for secondary control. The implementation of these intelligent techniques for the MGCC is the need of the hour, as the time frame available is on the higher side, and hence, more possibilities can be explored to reduce the time.

4.3.1. Application of Fuzzy Logic for Tertiary Control of MG

The fuzzy logic optimisation approach was used to solve the ED in MG as early as 2004 by Pathom [119] for an uncertain deregulated power system by representing these uncertain parameters with fuzzy numbers. They determined the optimal amount of power and reserve. Mahmoud [120] proposed a fuzzy tuning system for a prediction model by FSCM (Fuzzy subtractive clustering method based ANFIS with reduced error for performing the ED in MG. Hui Hou [121] proposed the MSOA multi-objective seeker optimisation algorithm based fuzzy membership function for ED in MG with EVs considering its charging—discharging characteristics. A multi-input single-output (MISO) fuzzy neural network controller for frequency control and to reduce generation cost by managing power distribution for ED is presented in ref. [122]. Lopez [123] developed Fuzzy interface system (FIS) incorporated ED to maintain operational limits of power generation in a MG involving wind and hydropower. In contrast to single-objective optimisation, multi-objective optimisation is proposed for ED in ref. [124], which can result in a set of solutions called Pareto-optimal solutions that are normalised using the fuzzy method. Mukesh Gautam [125] also proposed multi-objective Economic Emission Scheduling (EES) for minimising emission along with minimum generation cost. To obtain the best solution, they used the fuzzified Pareto concavity elimination transformation (PaCcET) and the results indicated that it has a shorter computational time. The multi-objective Bat Algorithm (MOBA) with Fuzzy Set Theory was proposed in ref. [126] for solving the optimal dispatch for optimal energy dispatch and reducing pollutant emissions to obtain a set of solutions at the Pareto optimal front (POF) and FST for identifying the best solution. Ref. [127] also proposed an approach to multi-objective ED considering four objectives, namely environmental impact, COE, distance of supply, load balancing and used fuzzy analytic hierarchy process (Fuzzy AHP) to prioritise them and LP for generating the alternative solution. Ref. [128] introduced a new approach by incorporating ED within the fuzzy logic energy management system (FLEMS) to reduce the cost of generation and increase system efficiency. PSO, GA and ABC optimisation techniques were used on real data, and the results were compared to find the suitable technique for solving the ED problem.

Divya [129] also implemented an FLC-based controller to manage the active power share among the MG and mitigate power quality problems when a nonlinear unbalanced load was applied to the MG. A fuzzy-based coordinate control strategy was proposed by [130] considering the SCO of BESS for the MGMS and proved that this controller is much sensitive to system structure parameter variations compared to the conventional controllers. An FLC-based centralised MG control system for EMS design was implemented in ref. [131] by maintaining the SOC of the BESS, thereby improving the overall grid power profile stability. The authors in ref. [132] proposed two decoupled FLCs for a hierarchical EMS among two energy storage units with hydrogen and electric storage systems in a residential MG. They claimed that this structure provides excellent stability and adaptability compared with PSO and rule-based EMS. A reliable power, irrespective of generation disruption, is guaranteed by [133] using a fuzzy—sparrow search algorithm for MG operation with RES and BESS for EVs. An adaptive fuzzy integrated FOPID controller was proposed in ref. [134] for power sharing among RES and storage that enhances DC bus management and regulates DC bus voltage at reduced battery stress. Nur E Alam [135] designed a FLC for EMS in MG, considering the intermittent nature of RES and varying loads in Malaysia and verified that it offers superior control of SOC thereby preventing the over- and under-charging of batteries. In addition to the above, many researchers have developed fuzzy logic-based EMS for better utilisation of renewables and reduction in costs, as well as for managing power balance either by using BESS [136,137,138,139,140,141,142,143,144] or by controlling the charging of EVs [145,146,147,148,149,150,151], to mention a few. In all these studies, the major differences were between the types of renewables used (PV, Wind, FC, etc.) and the types of FLC they incorporated (conventional, optimal, adaptive, or combined) [152]. The summary of Fuzzy Logic application to tertiary control of microgrid is given in

Table 6. Summary of Fuzzy Logic applications in Tertiary control of Microgrid.

References	Intelligent Technique Used	System Considered	Software/Hardware	Control Application	Standalone/Grid-Connected
[119]	Fuzzy optim. approach	DGs	Simulation *	Dynamic economic dispatch	Grid-connected
[120]	Fuzzy/ANFIS	MG	Simulation *	Economic dispatch	Grid-connected
[121]	Fuzzy-SOA	PV/Diesel Gen./ESS/EV	Simulation *	Multi objective Economic dispatch	Grid-connected
[122]	Fuzzy-NN	DGs	Simulation/HIL	Frequency co- ordination and Economic dispatch	Standalone/Grid-connected
[123]	FLA	Wind//Hydro/Bio Geothermal/Nuclear	Simulation *	Economic dispatch	Grid-connected
[124]	Trapezoidal Fuzzy model/EA	MGs	Simulation *	Economic emission and Load dispatch	Grid-connected
[125]	Fuzzified PaCcET	DERs	Simulation *	Economic emission scheduling (EMS)	Grid-connected
[126]	MODA with FST	PV/WT/FC/MT/ESS	Simulation *	Multi objecting optimal dispatching/EMS	Grid-connected
[127]	Fuzzy-AHP	MGs	Matlab Simulink/Hardware	Multi criteria optimal dispatching	Grid-connected
[128]	FLC ABC techniques	PV/Wind/FC/DEG/BESS	Matlab Simulink	EMS	Standalone
[129]	FLC	PV/Wind/Battery	Matlab Simulink/Hardware prototype	Power flow and power quality improvement	Grid-connected
[130]	FL based control	DGs with BESS	DIgSILENT	Power management	Grid-connected
[131]	FLC	PV/Wind/Diesel Gen./Battery	Simulation *	Economic Operation	Standalone
[132]	FLC	PV/Wind/HESS	Simulation *	EMS	Grid-connected
[133]	Fuzzy SSA	PV/FC/Battery/EV	Matlab Simulink	EMS	Standalone
[134]	Fuzzy-FOPID	PV/Wind/Battery/SC	Matlab Simulink/Opel-RT	Power management	Standalone
[135]	FLC	PV/Battery/	Matlab Simulink	EMS	Grid-connected
[136,137,138,139,140,141,142,143,144]	FLC	Various DGs with BESS	Matlab Simulink	EMS	Grid-connected/Standalone
[145,146,147,148,149,150,151]	FLC	Various DGs with EV	Simulation *	EMS	Grid-connected and Standalone
[152]	FLC ANFIS	DG/EV/BESS	Simulation *	EMS	Grid-connected and Standalone

* Simulation tool not mentioned.

.

4.3.2. Application of ANN for Tertiary Control of MG

Although many conventional methods are available for tertiary control of MG, ANN plays a major role in streamlining the operation of future grids, as economic scheduling and energy management should be given utmost care for minimising the losses and meeting the loads at the minimum cost. ANN techniques are used mainly to forecast wind velocity, solar irradiation or load, enabling optimal scheduling to be performed accordingly. F. Pilo et al. [153] developed an ANN-based MGCC to predict the optimal set points for the associated DERs operation and hence reduce the overall cost and increase profit. This is accomplished by performing feasibility studies based on various economic evaluation criteria to calculate the operating power capacity of the MG that is most convenient in various market scenarios. This ANN-based MGCC is also used for economic scheduling and optimal participation of MG in a linearised energy market [154,155]. A DRNN-based bidirectional LSTM was proposed [156] for forecasting the hourly PV output and load for optimal scheduling under various scenarios, and it was proven that, in addition to reducing the losses, this approach also improves the system economics. While ref. [157] proposed PSO to predict the optimal number of hidden layer neurons of ANN for optimal scheduling of MG, Ref. [158] employed a nonlinear autoregressive exogenous (NARX) model to train ANN to forecast solar irradiation and wind power for economic dispatch studies. A receding horizon control-based short-term forecasting model with ANN was used for forecasting the primary sources of RES and second-order cone programming for maintaining the optimal operation of a DC MG by Walter Gil-González [159].

A Radial Basis Functional network (RBFN)-based forecast model for estimating wind generation was proposed in ref. [160] along with receding horizon control for an economic dispatch scheme to maximise wind generation and reduce BESS sizing. DNN-based two-stage training was implemented by [161] to overcome the intermittent nature of RES through separate training and proved that this framework needs less data for training. Ref. [162] also proposed a DNN for learning the number of neurons required to solve the ED algorithm, specifically, the λ - iteration algorithm paved the way for real-time control beyond ED.

The performance of the MG can be improved by deploying an intelligent energy management system because the conventional EMS fixes the power references only with the available power generation and load and does not consider the accumulated power at the AC grid and SOC. Joshi [163] conducted a complete literature review on the application of AI techniques for EMS in MG. Ref. [164] proposed a two-step ANN to determine the operation mode and charge/discharge status of an ESS and experimentally proved its effectiveness. However, ref. [78] proposed an EMS based on meeting the load using renewables while monitoring the SOC of BESS using three strategies (PID, ANN and FL) and concluded that FL is more effective in the considered scenario. Ref. [165] also implemented and proved the same, stating that FL is better at keeping the frequency stable, in addition to managing the EMS of the system. Ref. [166] also considered the same data to develop a simple and efficient ANN-based EMS that can simultaneously predict the operating modes of various power converters in an MG. Nonlinear Autoregressive Moving Average Level 2 (NARMA-L2) artificial neural network (ANN)-based EMS was incorporated into an MG consisting of PV/WIND/BESS in ref. [167], and it was claimed that, in addition to EM, the proposed strategy was efficient in DC bus and frequency stabilisation. For demand response control, several DRL techniques were implemented in different realistic scenarios for EMS in ref. [168], while in ref. [169], an EMS aggregator with two-level MPC was employed for decision-making. Metaheuristic algorithms have been used for optimisation in MGs for EMS, as reviewed by [170,171,172,173,174,175,176,177,178], to mention a few others.

Conversely, the current trend towards green transport has paved the way for the use of EVs for EMS in MG. To address the plug-and-play functionality of EV, a PSO-based ANN was used to develop a hierarchical EMS for effective power sharing among multiple ESS, including batteries in ref. [179]. A supervised FF-ANN trained with the Bayesian Regularisation algorithm was used for EMS in an MG with both EV and battery for energy management in ref. [180] and justified a 28% reduction in grid power usage. PSO-tuned ANFIS was proposed by [181] for a similar system under varying power profiles and demonstrated its effectiveness with high revenue, thereby enhancing cost-effectiveness. Ref. [182] developed a modified dragonfly approach for EMS in MG to mitigate the charging effect of HEVs and obtained a 2.5% reduction in the overall operational cost. A combined ANN-AOA (Aquila optimiser algorithm) was proposed by [183] for controlling the dynamics of battery charging/discharging and implementing EMS in a PV/battery/UC system under various load and road conditions. The red-tailed hawk algorithm (RTHA)-optimised ANN was proposed by Richard [184] for optimal energy distribution in an MG consisting of PV, Battery and SC to manage the varying energy demand of EV. He also developed a prototype to confirm the effectiveness of the proposed algorithm, in addition to extending the lifespan of the batteries. Ref. [185] proposed a PSO-based ANN for appropriate energy extraction from MG and an FL-based controller for managing EV charging/discharging and claimed that the system operated at 97% efficiency under various operating conditions. A detailed review of the role of AI techniques for EMS in MG with the advent of EVs was also conducted by Khan [186]. The application of ANN in Tertiary control of microgrids is listed in the

Table 7. Summary of ANN applications in Tertiary control of Microgrid.

References	Intelligent Technique Used	System Considered	Software/Hardware	Control Application	Standalone/Grid Connected
[153,154,155]	ANN	WT/MT/ICE/CHP	Matlab Simulink	Economic and Optimal Generation Scheduling	Grid-connected
[156]	DRNN Bi-LSTM	PV/DEG	Matlab Simulink	Power forecasting and Optimal Gen. Scheduling	Standalone
[157]	ANN based PSO	MGs	Matlab Simulink/Hardware	Optimal energy scheduling	Grid-connected
[158]	NARX trained ANN	PV/Wind	Matlab Simulink	Economic dispatch	Standalone
[159]	ANN	PV/Wind	Matlab Simulink	Economic dispatch and forecasting	Standalone
[160]	Functional NN	Wind/BESS	Simulation *	Forecast and Economic Analysis	Grid-connected
[161,162]	Deep NN	MGs	Matlab Simulink	Economic dispatch	Grid-connected
[164]	ANN	PV/Wind/ESS	Simulation */Hardware	Energy management system	Grid-connected
[165]	ANN/Fuzzy/PID	PV/Wind/ESS	Matlab Simulink	Energy management system	Grid-connected
[166]	ANN	PV/Wind/ESS	Simulation */Hardware	Energy management system	Grid-connected
[167]	NARMA-L2 ANN	PV/Wind/SC/ESS	Matlab Simulink	Energy management system	Standalone
[168]	DRL	DERs	Simulation */Open AI GYM	Energy management system	Grid-connected
[169]	MPC/Q-learning RL model	PV/ESS/EV	Simulation	Power Management and Electricity trading	Grid-connected
[170,171,172,173,174,175,176,177,178]	Metaheuristic Techniques	MGs	Simulation */Hardware	Energy management/Forecasting	Grid-connected/Standalone
[179,185]	PSO based ANN	PV/ESS/EV	Simulation *	Energy extraction/Energy management	Grid-connected
[180]	SFF-ANN with BR Algorithm	PV/EV	Simulation *	Energy management	Grid-connected
[181]	PSO tuned ANFIS ANN	PV/Battery/WT/EV	Matlab Simulink	Energy management	Grid-connected/Standalone
[182]	Modified Dragonfly—ANN	PV/Battery/WT/EV	Matlab Simulink	Optimal Power despatch/Loss minimisation	Grid-connected
[183]	ANN-AOA	PV/Battery/UC/EV	Matlab Simulink	Energy management control	Standalone
[184]	RTHA-ANN	PV/Battery/SC/EV	Simulation */Hardware	Power flow and Energy management	Standalone

* Simulation tool not mentioned.

below.

5. Conclusions

The requirements of MG technologies to meet the growing demand and the advancement in their control strategies have paved the way for the exploration of the application of artificial intelligence techniques. In this regard, this paper provides a clear overview of the recent application of AI techniques for MG control that can confront a paradigm shift. This thorough review highlights the effectiveness of these advanced techniques over conventional methods in the areas of control and energy management, thereby improving the stability of the system. Therefore, despite the complex system model, AI can play a vital role in the flawless integration of DERs into the existing grid and maintain synchronisation. Besides this, it also helps in improving the rise time and settling time while also contributing to a reduction in the THD values of the output. However, most studies have been limited to simulation results and lack experimental validation. Also, in order to defend AI-enabled microgrids against cyber-physical threats, sophisticated anomaly detection, secure communication and robust fallback techniques are also required. Therefore, critical issues pertaining to stability under increasing renewable penetration, communication delays in distributed control and escalating cybersecurity risks must be addressed in future microgrid research, along with finding ways and means for real-time implementation of these AI-based approaches in MG control, knowing its potential benefits in handling these complex systems for a better future.