Comparison of Classical and Artificial Intelligence Algorithms to the Optimization of Photovoltaic Panels Using MPPT

Abstract

This work investigates the application of artificial intelligence techniques for optimizing photovoltaic systems using maximum power point tracking (MPPT) algorithms. Simulation models were developed in MATLAB/Simulink (Version 2024), incorporating conventional and intelligent control strategies such as fuzzy logic, genetic algorithms, neural networks, and Deep Reinforcement Learning. A DC/DC buck converter was designed and tested under various irradiance and temperature profiles, including scenarios with partial shading conditions. The performance of the implemented MPPT algorithms was evaluated using such metrics as Mean Absolute Error (MAE), Integral Absolute Error (IAE), mean squared error (MSE), Integral Squared Error (ISE), efficiency, and convergence time. The results highlight that AI-based methods, particularly neural networks and Deep Q-Network agents, outperform traditional approaches, especially in non-uniform operating conditions. These findings demonstrate the potential of intelligent controllers to enhance the energy harvesting capability of photovoltaic systems.

1. Introduction

Given the increasing global emphasis on sustainable and renewable energy solutions, photovoltaic (PV) energy plays a central role in the transition towards a sustainable energy paradigm. The global climate crisis, combined with the reduced availability of fossil resources, has accelerated the adoption of renewable energy solutions, positioning solar technology as a key pillar in decarbonization strategies worldwide. In this context, PV systems have emerged as a viable and scalable alternative, not only for their ability to generate clean energy but also for their flexibility in applications ranging from small-scale residential installations to large utility-scale power plants [1].

The overall efficiency of PV systems is directly linked to their ability to continuously extract the maximum available power from the solar panels. However, this task is inherently complex due to the dynamic nature of environmental conditions such as irradiance, temperature, and especially the occurrence of partial shading. These non-linearities introduce multiple local maximum in the power–voltage (P–V) characteristics curve, making the tracking of the global maximum power point (GMPP) a significant challenge. As a result, MPPT algorithms have become an essential component in PV system design, aimed at dynamically adjusting the operating point to ensure optimal energy harvesting under all conditions [2].

Over the past decades, MPPT has been the subject of extensive research, with traditional methods such as Perturb and Observe (P&O) and Incremental Conductance (InC) being widely adopted due to their simplicity and low computational cost. Nevertheless, these algorithms are known to suffer from several limitations, particularly under rapidly changing irradiance or partial shading conditions, where they tend to converge to local maxima or induce persistent oscillations around the operating point [3,4].

To overcome these limitations, the integration of intelligent control techniques has gained significant attention. Among these, Artificial Neural Networks (ANNs) and Deep Q-Networks (DQNs), a type of Deep Reinforcement Learning (DRL), have shown remarkable performance in handling the complex, non-linear, and dynamic behavior of PV systems. These AI-based approaches offer faster convergence, improved tracking precision, and greater stability under challenging conditions such as partial shading. By learning from environmental variations and adapting their control strategies in real time, ANN and DQN algorithms significantly enhance the system’s ability to identify and follow the true GMPP, even in highly variable scenarios.

The work by J and SY [5] presents an MPPT strategy using Artificial Neural Networks (ANNs), comparing it against the conventional P&O method. The study revealed that the ANN-based approach provides more accurate tracking and faster response times under dynamic irradiance conditions, outperforming P&O in terms of efficiency and stability.

In a more recent study, Giraldo et al. [6] applied Deep Reinforcement Learning (DRL), specifically a Deep Q-Network (DQN), for MPPT in real PV systems. The work includes both simulation and experimental validation, showing that, while DQN outperformed P&O in simulations, especially under partial shading conditions (PSCs), the real-world results were mixed. Notably, in PSC scenarios, P&O often got stuck in a local MPP, whereas DQN was able to extract up to 63.5% more energy, demonstrating the advantages of autonomous learning strategies. An early contribution to the use of reinforcement learning in MPPT is presented in the study by Kofinas et al. [7]. In their work, a tabular Q-learning algorithm was developed to track the MPP, showing strong convergence stability and reduced computational effort when compared to traditional metaheuristic techniques. Their results demonstrated that the method consistently outperformed the conventional P&O algorithm across multiple scenarios involving variations in irradiance and temperature. This line of research has since been extended by authors such as CSE [8] and Bavarinos et al. [9], who investigated alternative tabular reinforcement learning strategies for MPPT. These studies introduced comparative analyses between Q-learning and SARSA agents, as well as hybrid schemes incorporating fuzzy logic and sliding mode control, further validating the effectiveness of RL-based methods under dynamic environmental conditions.

Similarly, Phan et al. [10] explored DRL methods, including DQN and Deep Deterministic Policy Gradient (DDPG), for MPPT in MATLAB/Simulink simulations. Their work focused on the challenges of partial shading and showed that DQN significantly improved tracking efficiency compared to P&O. Under uniform conditions, DQN achieved a 5.83% efficiency gain, while DDPG yielded a 3.21% improvement, supporting the applicability of DRL in optimizing PV performance.

Remoaldo and Jesus [11] presented a comparative analysis between traditional P&O and a fuzzy logic-enhanced version (FLP&O). Using a system with five series-connected PV panels and a boost converter, the study demonstrated that FLP&O offers quicker convergence to the MPP under rapid irradiance changes, confirming the benefits of fuzzy logic in improving MPPT response time and accuracy.

Furthermore, the comprehensive review by Katche et al. [12] highlighted the limitations of conventional MPPT algorithms like P&O under partial shading. The authors advocated for hybrid and intelligent optimization techniques, which offer improved performance at the expense of higher computational complexity and cost, suggesting that the trade-off is justified in dynamic environments.

In addition to conventional approaches, the scientific community has increasingly explored the application of soft computing and evolutionary algorithms for MPPT control in photovoltaic systems. As highlighted by Rezk et al. [13], advanced control techniques such as fuzzy logic control (FLC) [14] and adaptive neuro-fuzzy inference systems (ANFISs) [15,16] have shown strong capability in managing the non-linear and time-varying behavior of PV systems. In parallel, a wide range of bio-inspired optimization algorithms have emerged as promising alternatives for global optimization. Among these, techniques such as genetic algorithms (GAs) [17], cuckoo search (CS) [18], ant colony optimization (ACO) [19], bee colony algorithm (BCA) [20], bat-inspired optimization (BAT) [21], and the memetic salp swarm algorithm [22] have demonstrated considerable effectiveness. These methods are particularly well-suited to address the challenges posed by partial shading conditions as they enable good exploration of the search space and enhance the system’s ability to reliably detect the GMPP. According to Jiang Jiang et al. [23], both soft computing and evolutionary techniques exhibit enhanced adaptability and are capable of delivering efficient and reliable tracking performance even under highly dynamic and non-linear operating conditions.

Additionally, recent works have explored data-driven and neural network-based methods that, while applied in different domains, offer valuable insights for adaptive control strategies in PV systems. For example, Lu et al. [24] proposed a transfer learning framework to enhance adaptability in dynamic environments, while Aizenberg and Tovt [25] introduced a multilayer neural network approach for intelligent frequency domain filtering. These contributions reinforce the relevance of learning-based techniques in handling non-linear time-varying conditions.

Lastly, Sharma et al. [26] compared traditional and metaheuristic MPPT techniques. Their findings reinforce that, although conventional algorithms such as P&O are easier to implement, they struggle in PSC scenarios.

Although previous studies have provided valuable insights, the majority of works in the literature still present notable limitations that prevent a comprehensive evaluation of MPPT strategies. Many focus on comparing only two or three algorithms, often within limited and idealized scenarios. Furthermore, performance assessments are frequently restricted to a few basic metrics, such as efficiency or steady-state accuracy, overlooking crucial dynamic aspects like convergence time or tracking stability under complex PSC. In particular, convergence time is an essential indicator of system responsiveness that remains underexplored and rarely quantified rigorously. Additionally, detailed scenario characterization, especially for PSC cases, is often lacking or simplified.

To address these gaps, this work presents a comprehensive comparative analysis of MPPT algorithms, including conventional methods (P&O and InC), control logic (FLC), hybrid (GA-based), and intelligent approaches (ANNs and DQNs). All the methods are evaluated under both uniform and partial shading conditions using a modular MATLAB/Simulink framework. Multiple performance metrics—MAE, IAE, MSE, ISE, efficiency, and convergence time—are considered. Convergence time is calculated using formal signal analysis with graphical validation. Test cases are clearly defined with specific irradiance and temperature settings, ensuring transparency. This study offers deeper insight into the strengths and limitations of MPPT strategies, especially AI-based ones, to support their practical application in PV systems.

The article is organized as follows. Section 2 describes the photovoltaic system modeling and simulation setup. Section 3 outlines the buck converter design. Section 4 defines the test scenarios, including both uniform and partial shading conditions. Section 5 presents the implemented MPPT algorithms, covering conventional, hybrid, and AI-based methods. Section 6 introduces the evaluation metrics, followed by the results and comparative analysis in Section 7. Section 8 discusses the key findings, and Section 9 concludes the work with suggestions for future developments.

2. PV System Description

The implemented PV system models were developed in MATLAB/Simulink to analyze the performance of different MPPT control strategies under realistic conditions. Figure 1b presents the system’s block diagram and its component interactions. Two configurations were considered: (i) a photovoltaic model with a single bypass diode, representing the actual SOLARPOWER XUNZEL 30 W 24 V panel (Figure 1a), and (ii) an extended model with three series-connected cell groups, each equipped with an individual bypass diode to simulate PSC (Figure 2). This latter approach allowed independent adjustment of irradiance and temperature for each cell group, enabling a comprehensive evaluation of MPPT algorithms under non-uniform operating conditions.

Figure 1a shows the Simulink implementation, structured into colored blocks:

• Blue (MPPT Controller): Computes PV voltage, current, and power, implementing the MPPT algorithm to generate the optimal duty cycle for the Pulse Width Modulation (PWM) generator.
• Green (DC–DC Buck Converter): Models the buck converter circuit, which adjusts the output voltage and current to transfer power efficiently to the load based on the MPPT control.
• Orange (Measurements): Monitors key variables such as voltages, currents, power, efficiency, temperature, and irradiance, displaying them for performance analysis.
• Yellow (Performance Metrics): Calculates and exports input/output power, efficiency, and tracking error to the MATLAB workspace for further evaluation.

This modular structure enables a clear and organized simulation environment for testing and comparing MPPT algorithms under different operating conditions.

The photovoltaic panel operation was characterized through detailed electrical calculations based on the manufacturer’s specifications presented in

Table 1. Specifications of the SOLARPOWER-XUNZEL-30W-24V photovoltaic panel [28].

Parameter	Value
Maximum power ($P_{max}$)	29.88 W
Number of cells per module ($N_{cell}$)	72
Open-circuit voltage ($V_{oc}$)	43.20 V
Short-circuit current ($I_{sc}$)	0.92 A
Voltage at maximum power point ($V_{mpp}$)	36.00 V
Current at maximum power point ($I_{mpp}$)	0.83 A
Temperature coefficient of voltage (${\beta }_{Voc}$)	−0.27 %/°C
Temperature coefficient of current (${\alpha }_{Isc}$)	+0.05 %/°C
Temperature coefficient of power (${\alpha }_{Psc}$)	−0.35 %/°C

.

As shown in Figure 3, the red curve on both I–V and P–V plots indicates the MPP of the panel.

The PV parameters, summarized in

Table 2. Calculated PV system operating parameters.

Parameter	Value
Input voltage ($V_{in}$)	36 V
Output voltage ($V_{out}$)	20 V
Input power ($P_{in}$)	29.88 W
Output power ($P_{out}$)	26.89 W
Converter efficiency ($\eta$)	90%
MPP current ($I_{mpp}$)	0.83 A
Output current ($I_{out}$)	1.34 A
Load resistance ($R_o$)	14.88 $\Omega$

, were calculated using the following formulas:

$$P_{in}=V_{mpp}\times I_{mpp}$$

(1)

$$P_{out}=\eta \times P_{in}$$

(2)

$$I_{out}={\displaystyle \frac{P_{out}}{V_{out}}}$$

(3)

$$R_o={\displaystyle \frac{V_{out}^2}{P_{out}}}$$

(4)

3. DC/DC Buck Converter Design

The DC/DC buck converter was carefully designed to interface with the PV panel while maintaining maximum power point tracking. The complete design process is presented below, with the circuit topology shown in Figure 4 [29,30].

Figure 4 illustrates the buck converter configuration, which includes a control switch, a diode, an inductor, and an output capacitor. The converter steps down the PV array voltage to the desired load level (20 V) while maintaining operation at the MPP, as dictated by the MPPT algorithm.

The key parameters of the converter, such as duty cycle (D), inductance ($L_o$), and capacitance ($C_o$), were calculated using classical design equations to ensure continuous conduction mode (CCM) and minimize ripple in both current and voltage. These calculations are based on the following expressions, respectively:

$$D={\displaystyle \frac{V_{out}}{V_{in}}}$$

(5)

$$L_o\ge {\displaystyle \frac{(V_{in}-V_{out})\times D}{f_s\times \Delta I}}$$

(6)

$$C_o\ge {\displaystyle \frac{\Delta I}{8\times f_s\times \Delta V}}$$

(7)

Table 3. Theoretical buck converter parameters used in simulations.

Component	Calculated Value	Simulation Purpose
Duty cycle (D)	0.555(5)	Fundamental conversion ratio for ideal voltage step-down
Inductor ($L_o$)	1658.71 µH	Minimum value to maintain CCM at $\Delta I=0.268$ A
Capacitor ($C_o$)	1675 µF	Exact value to achieve $\Delta V=1$ mV ripple
Switching freq. ($f_s$)	20 kHz	Balance between switching losses and dynamics
Current ripple ($\Delta I$)	0.268 A	20% of $I_{out}$ for stable MPPT operation
Voltage ripple ($\Delta V$)	1 mV	0.005% of $V_{out}$ for algorithm precision

summarizes the theoretical values obtained and used in the simulations.

4. Simulation Scenarios

To evaluate the performance of the implemented MPPT algorithms under realistic operating conditions, two sets of simulation scenarios were defined: one with a single bypass diode and another with three bypass diodes to simulate PSC.

Table 4. PV panel scenarios with single bypass diode.

Case	Scenario Description	Fixed Parameters
1	Uniform irradiance (1000 W/m2)	T = 25 °C
2	Variable irradiance (G)	T = 25 °C
3	Variable temperature (T)	G = 1000 W/m2
4	Simultaneous G and T variation	—

summarizes the four scenarios for the single bypass diode configuration. These include uniform irradiance, varying irradiance, varying temperature, and simultaneous variation of both parameters, enabling assessment under static and dynamic environmental conditions. The corresponding temperature and irradiance profiles for each scenario are illustrated in Figure 5, which shows the environmental input conditions applied to the PV model in these simulations.

Table 5. PSC with three bypass diodes.

Case	PSC Configuration	Fixed Parameters
1	G = [900, 1000, 200] W/m2	T = 25 °C
2	G = [200, 400, 300] W/m2	T = 25 °C
3	G = [900, 100, 150] W/m2	T = 25 °C
4	G = [300, 200, 100] W/m2	T = 25 °C

presents the PSC scenarios using three series-connected cell groups, each with its own bypass diode. These configurations simulate different shading patterns by assigning independent irradiance levels to each cell group while maintaining the ambient temperature at 25 °C.

Analysis of I–V and P–V Curves Under PSC Conditions

To understand the formation of multiple local maxima and the GMPP under partial shading, it is essential to analyze the I–V and P–V characteristic curves for each PSC scenario illustrated in Figure 6.

These characteristic curves validate the importance of testing algorithms under non-uniform conditions to assess their capability to detect the true GMPP efficiently.

5. MPPT Algorithms

Table 6. Classification of MPPT algorithms implemented in this study.

Class	Subclass	Acronym
Conventional Algorithms	Perturb and Observe	P&O
	Incremental Conductance	InC
Control Logic-Based Algorithm	Fuzzy Logic Controller	FLC
Artificial Intelligence Algorithms	Artificial Neural Network	ANN
	Deep Reinforcement Learning	DQN
Hybrid Algorithms	InC combined with P&O	InC + P&O
	GA combined with InC	GA + InC
	GA combined with P&O	GA + P&O

presents the classification of the MPPT algorithms implemented in this study. Conventional techniques such as Perturb and Observe (P&O) and Incremental Conductance (InC) were included as baseline methods due to their simplicity and widespread use in PV applications.

Additionally, a control logic-based approach using the fuzzy logic controller (FLC) was implemented for its ability to handle system non-linearities through rule-based decision-making. An optimization-based algorithm, the genetic algorithm (GA), was also tested for its global search capability in identifying the maximum power point, particularly under partial shading conditions.

Hybrid algorithms were explored to combine the advantages of different methods, including InC combined with P&O (InC+P&O), GA combined with InC (GA+InC), and GA combined with P&O (GA+P&O), aiming to enhance convergence speed and tracking accuracy in dynamic environments.

However, particular emphasis in this work was placed on artificial intelligence (AI)-based algorithms due to their superior adaptability and learning capabilities. Specifically, the Artificial Neural Network (ANN) was employed for its powerful pattern recognition and non-linear mapping ability, enabling accurate prediction of the optimal operating voltage under varying irradiance and temperature conditions. Furthermore, the Deep Q-Network (DQN) algorithm, which integrates reinforcement learning with deep neural networks, was implemented to achieve autonomous learning and optimal decision-making without requiring an explicit PV system model, demonstrating good performance even under complex partial shading scenarios.

Overall, the integration of AI-based algorithms, namely ANN and DQN, formed a key focus of this study, aiming to improve MPPT accuracy, adaptability, and robustness beyond traditional approaches.

5.1. Artificial Neural Network (ANN)

The Artificial Neural Network (ANN) approach leverages its ability to model non-linear systems and recognize patterns to predict the maximum power point (MPP) under varying environmental conditions. The ANN used in this study was trained with irradiance and temperature as input features, while the output was the predicted maximum power point voltage ($V_{mp}$). This prediction served as the reference for a PI controller, which adjusted the duty cycle of the buck converter to drive the PV panel towards the MPP.

The main characteristics of the implemented ANN are summarized in

Table 7. Architecture and training parameters of the implemented ANN.

Component	Details	Remarks
Input Layer	2 neurons	Irradiance and temperature
Hidden Layers	1st: 20 neurons (tansig) 2nd: 10 neurons (tansig)	Hyperbolic tangent sigmoid activation function
Output Layer	2 neurons (purelin)	$V_{mp}$ and $P_{mp}$ with linear activation
Training Algorithm	Levenberg-Marquardt (trainlm)	Implemented using fitnet in MATLAB
Data Division	75% training 15% validation 10% testing	Ensures generalization and performance evaluation

.

5.1.1. Dataset

To train the Artificial Neural Network (ANN), a comprehensive dataset was generated under realistic photovoltaic (PV) operating conditions, including dynamic variations in irradiance and temperature over a period of 1000 s. The simulation was carried out using a discrete step size of $T_s=0.001\mathrm{s}$, resulting in a total of one million samples per variable:

$$N={\displaystyle \frac{t_{\mathrm{total}}}{T_s}}={\displaystyle \frac{1000}{0.001}}=\mathrm{1,000,000}$$

(8)

The time vector was defined as

$$t=(0:N-1)·T_s$$

(9)

To ensure representative behavior under Standard Test Conditions (STCs), three time windows (at 200 s, 600 s, and 850 s) were defined with fixed irradiance of 1000 W/m² and temperature of 25 °C. Outside these intervals, both irradiance and temperature followed sinusoidal functions with added Gaussian noise to emulate natural environmental fluctuations:

$$G\left(t\right)=G_{\mathrm{mean}}+A_G·\sin \left({\displaystyle \frac{2\pi t}{T_G}}\right)+\mathcal{N}(0,{\sigma }_G^2)$$

(10)

$$T\left(t\right)=T_{\mathrm{mean}}+A_T·\sin \left({\displaystyle \frac{2\pi t}{T_T}}\right)+\mathcal{N}(0,{\sigma }_T^2)$$

(11)

where

• $G_{\mathrm{mean}}=500\mathrm{W}/{\mathrm{m}}^2$, $A_G=350\mathrm{W}/{\mathrm{m}}^2$, $T_G=300\mathrm{s}$, ${\sigma }_G=50\mathrm{W}/{\mathrm{m}}^2$;
• $T_{\mathrm{mean}}=30°\mathrm{C}$, $A_T=15°\mathrm{C}$, $T_T=600\mathrm{s}$, ${\sigma }_T=1.5°\mathrm{C}$.

These profiles simulate realistic irradiance oscillations every 5 min and slower temperature drifts every 10 min.

The PV system was simulated (Figure 7) using a resistive load calculated to match the maximum power point (MPP) at STC using the manufacturer specifications:

$$R={\displaystyle \frac{V_{mp}}{I_{mp}}}={\displaystyle \frac{36}{0.83}}\approx 43.37\Omega$$

(12)

The temporal evolution of the temperature and irradiance profiles is illustrated in Figure 8.

Table 8. Statistical summary of environmental variables.

Parameter	Temperature (°C)	Irradiance (W/m²)
Number of Samples	1,000,000	1,000,000
Mean	31.09	583.44
Standard Deviation	10.11	272.92
Minimum	15.00	1.00
Maximum	45.00	1000.00

summarizes the main statistics of the temperature and irradiance profiles.

This large and diverse dataset enabled the ANN to learn the non-linear relationship between environmental inputs and the optimal MPP voltage, ensuring good generalization across varying PV conditions.

5.1.2. Training Results of ANN

Figure 9 shows the training results of the ANN. The performance plot (Figure 9a) highlights a best validation performance of 0.94349 at epoch 266, indicating effective convergence. The regression plot (Figure 9b) reveals a strong correlation between predicted and actual values, with a regression coefficient of $R=0.99575$ for the test dataset, confirming the model’s ability to generalize. A dataset containing one million samples was used for each variable (irradiance, temperature, $V_{mp}$, and $P_{mp}$), ensuring a broad representation of different operating conditions and enhancing the generalization capability of the ANN.

The trained network is subsequently integrated as a Simulink block within the MPPT control system, as illustrated in Figure 10.

To evaluate the generalization performance of the network, 10 random test samples (10% of the dataset) were selected from the testing dataset.

Table 9. Real and predicted values of $V_{mp}$ and $P_{mp}$, with respective absolute and relative errors.

Test Pt.	Irrad.	Temp.	${\mathit{V}}_{\mathbf{mp}}$	Pred.	Abs. Err.	Rel. Err.	${\mathit{P}}_{\mathbf{mp}}$	Pred.	Abs. Err.	Rel. Err.
	(W/m2)	(°C)	(V)	(V)	(V)	(%)	(W)	(W)	(W)	(%)
1	477.21	31.58	18.20	18.56	0.36	2.00	7.64	7.94	0.30	3.94
2	163.71	23.19	6.52	6.44	0.08	1.25	0.98	0.96	0.02	2.13
3	597.23	43.73	22.79	23.14	0.35	1.54	11.97	12.34	0.37	3.06
4	433.10	42.56	16.71	16.99	0.28	1.70	6.44	6.65	0.22	3.35
5	868.99	18.45	32.66	32.73	0.06	0.19	24.60	24.70	0.10	0.41
6	531.06	32.95	20.60	20.58	0.02	0.07	9.78	9.78	0.001	0.01
7	797.51	17.32	30.52	30.01	0.50	1.65	21.47	20.76	0.71	3.29
8	856.85	23.42	32.94	32.47	0.47	1.43	25.01	24.32	0.70	2.78
9	883.10	43.10	33.58	33.15	0.43	1.27	25.99	25.42	0.58	2.22
10	294.77	44.72	11.80	11.66	0.13	1.13	3.21	3.14	0.07	2.22

combines the real and predicted values for $V_{mp}$ and $P_{mp}$, along with absolute and relative errors.

It is worth noting that the performance saturation can be primarily attributed to the size of the dataset used for training. With approximately one million samples covering variables such as irradiance, temperature, voltage, and power, the dataset is sufficiently large to ensure that the network effectively learns the mapping between input and output variables. As a result, increasing or decreasing the number of neurons in the hidden layers did not lead to significant improvements in prediction error since the network already achieves good generalization.

5.1.3. ANN Training Under PSC

To enhance the model’s performance and adaptability under PSC, the neural network was retrained using updated datasets that included PSC scenarios. The network architecture remained identical, with two hidden layers comprising 20 and 10 neurons, respectively, while the input layer was expanded to six neurons to accommodate additional input features. Notably, the training algorithm and hyperparameters were kept unchanged to ensure evaluation consistency with the baseline model.

The retrained network was integrated as a Simulink block within the MPPT control system, as illustrated in Figure 11.

5.2. Deep Q-Network (DQN)

Figure 12 presents a simplified block diagram of the Deep Q-Network (DQN) algorithm used for MPPT control in PV systems. It illustrates the key components involved in the agent–environment interaction, including state observation, action selection, Q-value estimation, and policy updating.

The agent receives the system state, comprising the PV voltage, current, duty cycle, and its variation, and evaluates possible actions using a Q-network (a deep neural network). The agent selects the next action using an $\epsilon$-greedy policy, balancing exploration (random actions) and exploitation (choosing the action with the highest Q-value). This policy is formally defined as follows [10,31]:

$$a_t=\left\{\begin{array}{cc}\arg \underset{a\in A}{\max }Q(s_t,a|\theta )\hfill & \mathrm{with}\mathrm{probability}1-\epsilon \hfill \\ \mathrm{random}(a_t\in A)\hfill & \mathrm{with}\mathrm{probability}\epsilon \hfill \end{array}\right.$$

where

• A is the set of all available actions;
• $\epsilon \in [0,1]$ is the exploration rate.

This approach allows the agent to continuously explore new strategies while also exploiting the best-known actions, facilitating learning in complex and high-dimensional environments. The selected action is applied to the PV system by modifying the duty cycle of the buck converter, which changes the operating point of the panel. The environment then moves to a new state and returns a reward that reflects the change in power output resulting from the agent’s action.

In DQN, the traditional Q-table is replaced by a deep neural network (Q-network) to approximate the action-value function $Q(s,a|\theta )$, where $\theta$ are the network weights. Two networks are used: the predict Q-network (with weights $\theta$) and the target Q-network (with weights ${\theta }^{\prime }$) [10]. The loss function minimized during training is the mean squared error (MSE) between the predicted and target Q-values described in Equation (13) [10,32,33]:

$$L\left(\theta \right)={\mathbb{E}}_{s,a}\left[{\left(Q_{\mathrm{target}}-Q_{\mathrm{predict}}\right)}^2\right]$$

(13)

The target Q-value is computed as

$$Q_{\mathrm{target}}=\underset{\mathrm{reward}}{\underbrace{r}}+\underset{\begin{array}{c}\mathrm{discount}\\ \mathrm{factor}\end{array}}{\underbrace{\gamma }}\underset{a}{max}Q(s_{t+1},a_{t+1}|{\theta }^{\prime })$$

(14)

And the predicted Q-value from the online network is

$$Q_{\mathrm{predict}}=Q(s_t,a_t|\theta )$$

(15)

The critic network evaluates the Q-values and calculates the loss, which is used to update the weights of the predict network through backpropagation. Over time, this allows the agent to improve its decision-making, leading to convergence to an optimal policy that maximizes long-term power extraction from the PV system.

Being a model-free method, the DQN agent adapts autonomously to varying conditions, including non-uniform irradiance and PSC, demonstrating strong performance in identifying the GMPP without requiring prior knowledge of the system dynamics.

Additionally, Figure 13 shows the detailed Simulink implementation of the DQN-based MPPT control strategy developed in this study. This model integrates the trained agent within the PV system environment to perform real-time duty cycle adjustments based on learned policies.

5.2.1. Agent Architecture and Reward Function

The Deep Q-Network (DQN) agent used in this work was designed to operate based on three core elements: the state space, the action space, and a multi-objective reward function. These are defined as follows:

• State Space (S).

The agent observes the system state at each timestep through the following vector:

$$S=\left\{V_{pv},I_{pv},D,\Delta D\right\}$$

where $V_{pv}$ and $I_{pv}$ are the PV panel voltage and current, D is the duty cycle of the DC/DC converter, and $\Delta D$ is the perturbation applied to D.

• Action Space (A).

The agent can select from a discrete set of actions that modify the duty cycle:

$$A=\{-0.03,-0.01,-0.005,-0.001,0,0.001,0.005,0.01,0.03\}$$

These values allow the agent to fine-tune the converter operation in small or large steps, or maintain the current setting.

• Reward Function (R).

The total reward at each timestep is composed of seven distinct components:

$$R=r_1+r_2+r_3+r_4+r_5+r_6+r_7$$

The components are:

• $r_1$: Relative Power (base reward)

  $$r_1={\displaystyle \frac{P_{t+1}}{P_{MPP,STC}}}$$
  This ensures higher rewards when operating near the GMPP.
• $r_2$: Reward/Penalty for Power Change

  $$r_2=\left\{\begin{array}{cc}{\left({\displaystyle \frac{P_{t+1}}{P_{MPP,STC}}}\right)}^2,\hfill & \Delta P>-{\delta }_1\hfill \\ -{\left({\displaystyle \frac{P_{t+1}}{P_{MPP,STC}}}\right)}^2,\hfill & \Delta P<-{\delta }_1\hfill \end{array}\right.\mathrm{with}{\delta }_1=1$$
• $r_3$: Duty Cycle Limit Penalty

  $$r_3=\left\{\begin{array}{cc}0,\hfill & 0\le D\le 1\hfill \\ -1,\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$
• $r_4$: Proportional Penalty for Power Drop

  $$r_4=\left\{\begin{array}{cc}\Delta P·\left({\displaystyle \frac{5}{P_{MPP,STC}}}\right),\hfill & \Delta P<-2\hfill \\ 0,\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$
• $r_5$: Significant Power Gain Reward

  $$r_5=\left\{\begin{array}{cc}{\displaystyle \frac{P_{t+1}}{P_{MPP,STC}}},\hfill & \Delta P>3\hfill \\ 0,\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$
• $r_6$: Current Stability Reward

  $$r_6=\left\{\begin{array}{cc}+0.5,\hfill & |\Delta I|<0.001\hfill \\ -1.5·|\Delta I|,\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$
• $r_7$: Exponential Reward for Power Stability

  $$r_7=\left\{\begin{array}{cc}\alpha ·\exp \left(-{\displaystyle \frac{|\Delta P|}{\beta }}\right),\hfill & |\Delta P|<{\delta }_2\hfill \\ 0,\hfill & \mathrm{otherwise}\hfill \end{array}\right.\mathrm{with}\alpha =1.5,\beta =0.05,{\delta }_2=0.005$$

This multi-component reward function was carefully designed to guide the agent towards fast convergence, stable operation, and global maximum power tracking while penalizing unnecessary oscillations or control saturation.

5.2.2. Simulation Setup

The training of the Deep Q-Network (DQN) agent was conducted in the MATLAB/Simulink environment using the Reinforcement Learning Toolbox [34]. A stochastic training strategy was adopted, where the irradiance and temperature applied to each of the three PV panel segments were randomly varied at the beginning of each episode. This approach ensured generalization of the learned policy across a wide range of environmental conditions, including STC and PSC.

Table 10. DQN agent configuration parameters.

Parameter	Value
Simulation time per episode	0.5 s
Agent sample time	0.01 s
Maximum number of episodes	300
Steps per episode	50
Discount factor	0.9
Mini-batch size	64
Experience buffer length	$1\times {10}^6$
Initial exploration rate ($\epsilon$)	1.0
Minimum exploration rate	0.001
Exploration decay rate	0.0001
Double DQN	Enabled
Critic optimizer	Adam
Critic learning rate	0.0001
Policy for saving agent	Episode reward ≥ 150
Stopping criterion	Average reward ≥ 10,000
Parallel training	Disabled

summarizes the main parameters used to configure the agent and the training process.

5.2.3. Training Results of DQN

The training process of the DQN agent is depicted in Figure 14, where the light-blue curve represents the episode reward, the dark-blue line indicates the moving average of the reward, and the yellow curve corresponds to the Episode Q0 metric. The Episode Q0 provides an estimation of the expected future reward based on the current policy and is a useful indicator of learning progress.

It is evident that the training converged around episode 100 as the average reward stabilized and episode rewards consistently exceeded the defined performance threshold. The final episode reward and Q0 values confirmed the agent’s ability to maximize the power extracted from the PV system.

A summary of the training performance is presented in

Table 11. Summary of the training performance for the DQN agent.

Metric	Value
Mean reward per episode	28.73
Final episode reward	11.74
Episode Q0	3.89
Average steps per episode	50
Total agent steps	15,000
Total training time	2 h 3 min

, with hardware specified in

Table 12. Specifications of the hardware used for simulations and training of ANN and DQN agents.

Component	Specification
PC Model	HP ProDesk 600 G4 Mini PC
Processor	Intel® CoreTM i7-7700T CPU @ 2.90 GHz
RAM	16 GB DDR4 2400 MHz
Storage	512 GB SSD
Operating System	Windows 10 Pro 64-bit
Integrated GPU	Intel® HD Graphics 630
Parallel computing	Disabled
GPU acceleration	Disabled

.

5.3. Benchmark Algorithms

To ensure a fair and consistent comparison across all MPPT algorithms evaluated in this study, a common benchmarking strategy was adopted.

Table 13. Benchmarking configuration for MPPT algorithms.

Algorithm	Step Size $\Delta$	Initial Duty $\left({\mathit{D}}_{\mathbf{init}}\right)$	Bounds ${\mathit{D}}_{\mathbf{min}}/{\mathit{D}}_{\mathbf{max}}$	Training Required	Dynamic Tuning	Main Inputs
P&O	0.000125	0.5555	0 / 1	No	Fixed $\Delta$	$V_{pv},I_{pv}$
InC	0.000125	0.5555	0 / 1	No	Fixed $\Delta$	$V_{pv},I_{pv}$
FLC	Not applicable	Not defined	0 / 1	No	Implicit (fuzzy rules)	$V_{pv},I_{pv}$
InC + P&O	0.000125	0.5555	0 / 1	No	Threshold switching ($\left|dV\right|>5\times {10}^{-4}$ V)	$V_{pv},I_{pv}$
InC + GA	[0.00005, 0.0002]	0.5555	0 / 1	No	GA-optimized $\Delta$	$V_{pv},I_{pv}$
P&O + GA	[0.00005, 0.0002]	0.5555	0 / 1	No	GA-optimized $\Delta$	$V_{pv},I_{pv}$

summarizes the key implementation parameters defined for each method, including initial duty cycle, perturbation step size, input variables, and control structure. This standardized setup ensures that observed performance differences are driven by algorithmic behavior rather than configuration disparities.

Additionally, for the hybrid algorithms GA-P&O and GA-InC, a small exponential moving average (EMA) filter was applied to the input measurements of panel voltage and current ($V_{pv}$ and $I_{pv}$) in order to mitigate the effect of high-frequency measurement noise. The filtered signal is computed using Equation (16) [35,36]:

$$x_f\left(k\right)=\alpha ·x\left(k\right)+(1-\alpha )·x_f(k-1),$$

(16)

where x represents the measured signals $V_{pv}$ and $I_{pv}$, and $\alpha =0.01$ is the smoothing coefficient. This filtering strategy enhances the stability of the GA-based controllers by providing cleaner input data during operation.

6. Evaluation Metrics

To quantitatively assess the performance of the implemented MPPT algorithms under various operating conditions, six key metrics were employed. These metrics evaluate accuracy, stability, and dynamic response characteristics, providing comprehensive insights into each algorithm’s capabilities.

6.1. Error Metrics

• Mean Absolute Error (MAE): Measures the average absolute deviation between the PV panel’s input power ($P_{in}$) and converter output power ($P_{out}$):

  $$\mathrm{MAE}={\displaystyle \frac{1}{N}}\sum _{t=1}^N|P_{in}\left(t\right)-P_{out}\left(t\right)|$$

  (17)

  where N is the number of samples. Lower MAE values indicate better steady-state accuracy.
• Integral Absolute Error (IAE): Evaluates the cumulative tracking error over time, penalizing persistent deviations:

  $$\mathrm{IAE}={\int }_0^T|P_{in}\left(t\right)-P_{out}\left(t\right)|dt$$

  (18)
  This metric is particularly sensitive to prolonged errors during transient conditions.
• Mean Squared Error (MSE): Emphasizes larger errors through quadratic penalization:

  $$\mathrm{MSE}={\displaystyle \frac{1}{N}}\sum _{t=1}^N{[P_{in}\left(t\right)-P_{out}\left(t\right)]}^2$$

  (19)
• Integral Squared Error (ISE): Combines temporal accumulation with quadratic error weighting:

  $$\mathrm{ISE}={\int }_0^T{[P_{in}\left(t\right)-P_{out}\left(t\right)]}^{2}dt$$

  (20)
  ISE is especially relevant for assessing performance under partial shading conditions where large power fluctuations occur.

6.2. Efficiency and Dynamic Response

• Conversion Efficiency ($\eta$): Quantifies the energy harvesting capability:

  $$\eta =\left({\displaystyle \frac{\overline{P_{out}}}{\overline{P_{in}}}}\right)\times 100\%$$

  (21)

  where $\overline{P}$ denotes average power over the simulation period.
• Convergence Time ($t_{conv}$): Determined when the system reaches and maintains steady-state operation, calculated using MATLAB’s lsiminfo function. The threshold (2%) was set to ensure stable power delivery, with validation through moving average analysis of the error signal.

7. Results

Before presenting the performance results of the implemented MPPT algorithms, it is important to outline the methodology used to evaluate their behavior under both uniform and non-uniform operating conditions. The simulation scenarios were carefully designed to replicate realistic environmental profiles, including variable irradiance and temperature, as well as complex PSC. Key performance indicators such as tracking efficiency, convergence time, output power stability, and dynamic adaptability were used as evaluation metrics. This section presents a detailed comparison of conventional methods (P&O and InC), control-logic strategies (FLC), metaheuristic approaches (GA), and artificial intelligence algorithms (ANN and DQN), highlighting their relative strengths and limitations across various operating scenarios.

All the simulations, as well as the training of the ANN and DQN agents, were executed using the hardware configuration detailed in Table 12.

7.1. Scenario 1: Constant Irradiance and Temperature

As shown in

Table 14. Performance metrics for Scenario 1 (uniform irradiance at 1000 W/m²).

Algorithm	MAE	IAE	MSE	ISE	Avg. Eff. (%)	Conv. Time (s)
P&O	2.2740	23.6469	8.4456	96.7815	93.6013	0.2029
InC	2.2681	23.6453	8.4224	96.7687	93.6261	0.2029
Fuzzy	1.3028	13.5664	3.4857	45.5630	95.3713	0.2385
InC + P&O	2.2755	23.6486	8.4512	96.7889	93.5962	0.2030
InC + GA	1.3144	13.6370	2.7629	33.8812	95.3102	0.3783
P&O + GA	1.3124	13.6443	2.7561	33.8601	95.3165	0.3787
ANN	1.2931	13.4424	3.3865	45.5965	95.4995	0.1763
DQN	1.3396	14.2425	2.7681	43.0119	95.2981	0.2118

, the ANN achieved the best performance, presenting the lowest MAE and IAE values, combined with the shortest convergence time (0.1763 s). Although DQN recorded a slightly lower MSE, it required more time to converge. Conventional algorithms such as P&O and InC demonstrated higher error metrics and lower average efficiencies, confirming their limitations under stable conditions.

To better visualize the dynamic behavior of the AI-based algorithms, Figure 15 shows the convergence performance of the ANN and DQN during the first 0.5 s of operation. The ANN agent demonstrates faster stabilization and lower initial error compared to the DQN agent.

Additionally, Figure 16 presents the error evolution of all the tested MPPT algorithms up to 0.5 s. It is evident that conventional approaches such as P&O and InC exhibit larger and more persistent oscillations. In contrast, intelligent controllers, particularly ANN and DQN, provide smoother convergence with reduced transient errors.

7.2. Scenario 2: Variable Irradiance and Constant Temperature

According to

Table 15. Performance metrics for Scenario 2 (variable irradiance conditions).

Algorithm	MAE	IAE	MSE	ISE	Avg. Eff. (%)	Conv. Time (s)
P&O	2.0134	20.7350	7.1603	81.7637	92.4991	0.2008
InC	2.0333	20.7282	7.2322	81.7327	92.3735	0.2008
Fuzzy	1.2820	13.3091	3.5243	44.9620	94.2180	0.2370
InC + P&O	1.9957	20.7619	7.0857	81.8770	92.6808	0.2010
InC + GA	1.3113	13.5763	3.2909	38.6606	94.0737	0.3749
P&O + GA	1.3100	13.5586	3.2617	38.3420	94.0772	0.3892
ANN	0.9426	9.9433	2.7069	38.8351	95.2038	0.1837
DQN	1.3796	14.1881	3.5495	42.8869	93.7760	0.2316

, the ANN clearly outperformed the remaining methods, with the lowest absolute and integral errors, as well as the highest efficiency (95.20%). The DQN maintained solid performance but showed higher error variance. Hybrid methods like InC + GA and P&O + GA improved upon their base algorithms, although they still lagged behind the AI-based models in accuracy and responsiveness.

7.3. Scenario 3: Variable Temperature and Constant Irradiance

Table 16. Performance metrics for Scenario 3 (variable temperature conditions).

Algorithm	MAE	IAE	MSE	ISE	Avg. Eff. (%)	Conv. Time (s)
P&O	2.2901	23.5039	8.4903	95.9163	93.5644	0.2026
InC	2.3020	23.4960	8.5452	95.8738	93.4919	0.2026
Fuzzy	1.2914	13.4304	3.4588	45.2618	95.4024	0.2389
InC + P&O	2.2737	23.5271	8.4134	96.0351	93.6665	0.2027
InC + GA	1.3036	13.5125	2.9062	35.2575	95.3384	0.3754
P&O + GA	1.3040	13.5121	2.9491	35.6423	95.3367	0.3759
ANN	1.3586	14.1022	3.5430	47.3818	95.1503	0.1773
DQN	1.3499	14.2418	2.9612	43.0156	95.1882	0.2017

indicates that the ANN and DQN remained among the top-performing algorithms, with efficiencies above 95%. The ANN showed better consistency and robustness to temperature variations, while the DQN delivered competitive results. The fuzzy and hybrid approaches also improved tracking performance, whereas the conventional algorithms exhibited greater instability and error accumulation.

7.4. Scenario 4: Variable Irradiance and Temperature

In

Table 17. Performance metrics for Scenario 4 (combined variable irradiance and temperature conditions).

Algorithm	MAE	IAE	MSE	ISE	Avg. Eff. (%)	Conv. Time (s)
P&O	1.9080	19.5159	6.3287	71.3833	92.3418	0.1937
InC	1.8832	19.5411	6.2348	71.4879	92.5729	0.1937
Fuzzy	1.2264	12.6995	3.0669	38.7689	94.0541	0.2303
InC + P&O	1.9232	19.5131	6.3794	71.3694	92.2290	0.1937
InC + GA	1.2610	13.0040	2.8422	33.0738	93.8670	0.3717
P&O + GA	1.2561	12.9948	2.8414	33.1874	93.8860	0.3747
ANN	0.8227	8.8244	2.1729	34.3348	94.8967	0.1746
DQN	1.3351	13.8548	2.8987	38.8920	93.2006	0.2272

, the ANN once again provided the highest efficiency (94.90%) and fastest convergence time, demonstrating excellent adaptability to combined environmental variations. The DQN also delivered a strong performance, although with slightly higher error metrics. The fuzzy and GA-based hybrids offered moderate improvements over the conventional methods, which struggled to maintain accuracy in dynamic conditions.

7.5. Scenario PSC 1

As presented in

Table 18. Performance metrics for Scenario PSC 1 (partial shading conditions).

Algorithm	MAE	IAE	MSE	ISE	Avg. Eff. (%)	Conv. Time (s)
P&O	1.2470	12.7990	2.4616	27.0702	94.3909	0.2261
InC	1.2474	12.8019	2.4625	27.0762	94.3890	0.2254
Fuzzy	0.6244	6.4607	0.8937	11.1054	96.3035	0.1891
InC + P&O	1.2473	12.8012	2.4623	27.0758	94.3897	0.2259
InC + GA	0.6362	6.5798	0.7684	9.5713	96.2199	0.4076
P&O + GA	0.6361	6.5806	0.7636	9.5218	96.2209	0.4116
ANN	0.7540	7.8210	0.9840	13.8583	94.7047	0.1582
DQN	0.9174	9.1121	1.5140	16.2580	93.0104	0.1973

, fuzzy logic and hybrid methods like InC + GA and P&O + GA reached the highest efficiency values (above 96%), with significantly reduced MAE and MSE. The ANN performed well with the fastest convergence time (0.1582 s), while the DQN exhibited slightly higher error values under this moderate shading pattern. The traditional methods showed limited adaptability to PSC conditions.

7.6. Scenario PSC 2

Table 19. Performance metrics for Scenario PSC 2 (severe partial shading conditions).

Algorithm	MAE	IAE	MSE	ISE	Avg. Eff. (%)	Conv. Time (s)
P&O	1.0843	11.5210	1.9253	22.4451	83.5914	0.4561
InC	1.1220	11.5138	2.0071	22.4282	82.9441	0.4557
Fuzzy	1.0036	10.4573	1.3136	15.5130	84.0660	0.9092
InC + P&O	1.0929	11.5257	1.9405	22.4593	83.4341	0.4559
InC + GA	0.6162	6.3335	0.4283	4.7474	89.3252	0.5825
P&O + GA	0.6160	6.3309	0.4308	4.7676	89.3285	0.5809
ANN	0.5115	5.2931	0.3810	4.9916	92.7658	0.1816
DQN	0.2014	2.0017	0.0645	0.6281	93.9154	0.2866

highlights the DQN as the most effective algorithm in this severe shading scenario. It recorded the lowest error metrics across all the categories and achieved the highest average efficiency (93.92%). The ANN followed closely, with strong results and fast convergence. The metaheuristic hybrids and fuzzy logic showed improvements but could not match the precision of the AI-based models.

7.7. Scenario PSC 3

As shown in

Table 20. Performance metrics for Scenario PSC 3 (dynamic partial shading conditions).

Algorithm	MAE	IAE	MSE	ISE	Avg. Eff. (%)	Conv. Time (s)
P&O	0.1014	1.0773	0.0627	0.8993	98.5042	0.1958
InC	0.1017	1.0796	0.0628	0.8996	98.5004	0.1958
Fuzzy	0.0977	1.0385	0.0633	0.9023	98.5554	0.1988
InC + P&O	0.0983	1.0369	0.0623	0.8938	98.5481	0.1933
InC + GA	0.0738	0.7924	0.0549	0.8245	98.8231	0.5854
P&O + GA	0.1084	1.1356	0.0598	0.8733	98.3324	0.5760
ANN	0.2525	2.5886	0.0983	1.3707	91.4206	0.4003
DQN	0.2559	2.5900	0.1370	1.5501	96.2822	0.2341

, InC + GA achieved the best performance, with the lowest MAE (0.0738) and IAE (0.7924), confirming its high precision under dynamic shading. Fuzzy and InC + P&O also maintained excellent efficiency (above 98.5%) with minimal errors. In contrast, the ANN recorded the lowest efficiency (91.42%) and significantly higher error metrics. This lower performance is attributed to the neural network’s limited ability to generalize to the specific irradiance profile of PSC 3, which was underrepresented in the training dataset. Enhancing the ANN performance in this scenario would require expanding or adapting the dataset to better reflect such dynamic conditions. Meanwhile, the DQN demonstrated moderate accuracy (96.28%) but exhibited noticeable tracking fluctuations, indicating reduced control stability in this scenario.

7.8. Scenario PSC 4

Based on the results in

Table 21. Performance metrics for Scenario PSC 4 (extreme partial shading conditions).

Algorithm	MAE	IAE	MSE	ISE	Avg. Eff. (%)	Conv. Time (s)
P&O	0.5829	6.1757	0.5657	6.4031	85.5496	0.3757
InC	0.5959	6.1727	0.5790	6.3991	85.1847	0.3751
Fuzzy	0.5532	5.6450	0.3808	4.1320	85.6871	0.5335
InC + P&O	0.6064	6.1711	0.5899	6.3964	84.8987	0.3748
InC + GA	0.1988	2.0226	0.0479	0.5239	91.9530	0.4035
P&O + GA	0.1992	2.0277	0.0515	0.5599	91.9349	0.4025
ANN	0.2053	2.0980	0.0538	0.6348	91.7499	0.1938
DQN	0.2043	2.0352	0.0596	0.5909	91.6215	0.9375

, the ANN provided an excellent trade-off between accuracy and convergence time, achieving 91.75% efficiency with fast response (0.1938 s). The DQN achieved the lowest MSE but at the cost of higher convergence time. Hybrid GA-based methods also performed well, while traditional algorithms like P&O and InC lagged behind with reduced efficiency and greater tracking error.

8. Discussion

The comparative analysis of the MPPT algorithms, summarized in

Table 22. Comparison of MPPT algorithms.

Algorithm	Evaluation Criteria
	Compl.	Cost	Param. Ind.	Training	Conv. Speed	Sensors	Tuning	Stability	PSC
P&O	Low	Medium	Yes	No	Fast	V,I	No	Low	Weak
InC	Medium	Medium	Yes	No	Fast	V,I	No	Low	Weak
FLC	High	Medium	Yes	No	Fast	V,I	No	High	Good
InC + P&O	Medium	Medium	Yes	No	Fast	V,I	No	Low	Weak
InC + GA	High	High	Yes	No	Slow	V,I	No	Medium	V.Good
P&O + GA	High	High	Yes	No	Slow	V,I	No	Medium	V.Good
ANN	High	High	No	Yes	V.Fast	V,I,G,T	Yes	High	Exc.
DQN (RL)	High	High	No	Yes	Fast	V,I,G,T	Yes	High	Exc.

Notes: Abbrev.: Compl. = Complexity, Param. Ind. = Parameter Independence, Conv. Speed = Convergence Speed, Sensors; Stability: Low (oscillations), Medium (small variations), High (stable); PSC: Weak (oscillations), Mod. (moderate), Good, V.Good, Exc. (Excellent); Sensors: V = Voltage, I = Current, G = Irradiance, T = Temperature.

, highlights distinct strengths and limitations across the different approaches.

Conventional methods such as P&O, InC, and InC+P&O offer simplicity and fast convergence but consistently underperform in dynamic conditions, particularly under partial shading (PSC), where they tend to become trapped in local maxima. Their average efficiencies remained below 85% in PSC scenarios, confirming their limited adaptability.

Fuzzy logic control (FLC) improved upon the traditional methods by providing better stability and slightly higher accuracy. While it showed moderate resilience under variable conditions, its tracking capability in PSC scenarios was still outperformed by more advanced techniques.

AI-based methods, namely ANN and DQN, demonstrated superior overall performance across all the scenarios. These algorithms achieved high average efficiencies (above 91%), fast convergence, and excellent results, even under complex PSC conditions. Their ability to learn and generalize from environmental variations enabled them to reliably track the GMPP.

Hybrid methods combining traditional techniques with genetic algorithms, such as InC + GA and P&O + GA, also yielded notable improvements in accuracy and adaptability. However, these gains came at the cost of increased convergence time due to the computational overhead of evolutionary optimization.

Overall, the results reinforce the effectiveness of AI-based and hybrid strategies for MPPT, particularly in challenging and rapidly changing operating environments.

9. Conclusions

This work demonstrates the effectiveness of intelligent control strategies, particularly Artificial Neural Networks (ANNs) and Deep Q-Networks (DQNs), in enhancing the performance of photovoltaic systems through MPPT optimization. The implemented MATLAB/Simulink models, which simulated both uniform and partial shading scenarios, provide a robust framework for comparative analysis. The results show that AI-based methods not only achieved superior tracking efficiency and faster convergence times but also exhibited greater resilience to dynamic changes and adaptability to non-linear environmental conditions. Among all the tested algorithms, the DQN agent consistently outperformed the others in PSC scenarios, confirming its ability to identify the GMPP more effectively. These findings underscore the potential of integrating learning-based controllers into PV systems to improve energy harvesting in real-world applications.

A natural progression of this research involves the integration of a bidirectional DC/DC converter, specifically a buck–boost topology, which would enable both the charging and discharging of an energy storage system. This architecture would facilitate the transition from a passive PV system to a hybrid energy management solution capable of supplying loads autonomously during periods of low solar generation.

Additionally, future work could focus on the experimental validation of the proposed MPPT algorithms under real-world operating conditions. This would involve developing a physical prototype that integrates photovoltaic panels, bidirectional converters, embedded controllers (e.g., microcontrollers), and appropriate sensors for current, voltage, irradiance, and temperature measurement. Implementing the control logic directly in embedded hardware would allow for the assessment of real-time performance, computational constraints, and resilience to disturbances, such as measurement noise or sudden environmental changes.

Another promising research direction includes the development of hybrid renewable energy systems, combining solar and wind energy sources under a unified control framework. The MPPT algorithms could be extended or adapted to coordinate multiple energy sources with complementary profiles, enhancing system reliability and reducing energy intermittency.