Advanced Solar Panel Fault Detection Using VGG19 and Jellyfish Optimization

Abstract

Solar energy has become a vital renewable energy source (RES), and photovoltaic (PV) systems play a key role in its utilization. However, the performance of these systems can be compromised by faulty panels. This paper proposes an innovative framework that combines the deep neural network VGG19 with the Jellyfish Optimization Search Algorithm (JFOSA) for efficient fault detection using aerial images. VGG19 excels in automatic feature extraction, while JFOSA optimizes feature selection and significantly improves classification performance. The new framework achieves impressive results, including 98.34% accuracy, 98.71% sensitivity, 98.69% specificity, and 94.03% AUC. These results outperform baseline models and various optimization techniques, including ant colony optimization (ACO), genetic algorithm (GA), and particle swarm optimization (PSO). The system demonstrated superior performance in detecting solar panel defects such as cracks, hot spots, and shadow defects, providing a robust, scalable, and automated solution for PV monitoring. This approach provides an efficient and reliable way to maintain energy efficiency and system reliability in solar energy applications.

1. Introduction

Over the years, incremental enhancements in panel voltages, cell current, and efficiency have been incorporated into the design and fabrication of solar panels [1]. Like any other power generation system, photovoltaic (PV) systems are subject to loss mechanisms, including series resistance loss [2]. Solar energy is regarded as the most abundant renewable energy source (RES) because it originates from the primary source [3]. In areas where solar panels are installed, it is important to appreciate that the sun’s movement requires operation and adjustment with respect to them [4]. Several factors determine the quantity of sun rays reaching the surface of the Earth, such as time of the year, terrain, weather, and geographical positioning [5]. Moreover, the curvature of the Earth affects the angle of incidence, leading to variability in solar intensity, commonly explained by the air mass principle. The use of artificial intelligence (AI) and machine learning (ML) methods in the solar energy sector has proven to be highly beneficial in recent years, especially in monitoring and detecting faults in photovoltaic systems [6,7,8]. This study presents a systematic review of recent research and development in the use of AI and ML for defective solar panel detection, highlighting key results, approaches, challenges, and future potential in this rapidly evolving field. In the past, faults in solar panels were detected through manual observation, visual evaluation, or basic heuristic techniques. Although these methods can identify certain flaws, they often fail to detect more subtle defects or predict impending failures. In contrast, AI- and ML-based techniques are capable of analyzing large datasets from PV systems, identifying internal patterns, and providing early warning of potential defects [9,10,11]. These advanced approaches are more flexible, scalable, and automated, fundamentally transforming the way solar panels are maintained.

High-quality data is a primary consideration when applying AI and ML in PV systems. Many studies have been focused on the acquisition, data cleaning, and analysis of data collected from various sensors in PV systems, such as current and voltage sensors, temperature sensors, and irradiance sensors. Using this data, researchers have developed complex models that can accurately detect and classify different fault types, including degradation, soiling, shading, as well as electrical mismatch [7,12,13]. There has been widespread interest in the use of AI and ML for solar panel fault detection. In this regard, many supervised learning techniques have been implemented, such as support vector machines (SVMs), decision trees (DTs), random forests, and neural networks. Additionally, there is increasing dependence on unsupervised learning, including clustering and anomaly detection techniques [6,14,15]. Although significant progress has been made, challenges remain. These include the need for standardized datasets, real-time fault monitoring, incorporation of domain-specific knowledge into models, and better integration with operational PV systems.

1.1. Research Gap and Motivation

While many AI and ML techniques have been explored for detecting faults in PV systems, several challenges have not been fully addressed:

• Many existing models are designed for static conditions and fail to perform effectively under dynamic environmental factors, such as changing irradiance and temperature.
• The lack of standardized benchmarking datasets hinders the objective comparison of performance across different methods.
• Conventional ML algorithms often struggle with feature selection, leading to suboptimal performance and increased computational cost.

To address these issues, this paper introduces a novel approach that combines VGG19 for feature extraction with the Jellyfish Optimization Search Algorithm (JFOSA) for optimal feature selection. The hybrid model aims to improve fault detection accuracy, reduce computational complexity, and enhance model generalizability across a wide range of operational conditions.

1.2. Related Work

PV installations may encounter issues that reduce their lifespan and impact performance. These faults, affecting various PV components, can stem from multiple causes. The study [16] investigated the causes of these faults and their consequences. Depending on the fault’s nature, these instances lead to reduced efficiency, reliability, and the potential for component damage. Ground, arc, and line-to-line faults elevate the risk of fire. Common faults include partial shading, temperature variations, and line-to-line faults, leading to power losses of up to 31.67% [17]. Swift detection and resolution of these faults are essential to ensure optimal and reliable PV system performance. However, existing safety mechanisms like differential relaying, overcurrent relaying, and distance relaying do not cover PV fault detection. Critical components such as the encapsulant and junction box are prone to failures, with delamination being a significant risk factor [18]. Data classification algorithms are often inadequate or unsuitable for identifying PV system issues. These algorithms must be trained to recognize fault attributes and differentiate between healthy and faulty PV conditions. The initial phase involves training, followed by testing to evaluate the efficiency of the categorization methods. Recent studies have proposed machine learning techniques for proactive maintenance, enhancing reliability, and minimizing downtime [19]. Based on their time-related features, anomalies in a PV array can be classified as either emerging anomalies or persistent anomalies. It is crucial to swiftly identify and rectify such anomalies to ensure uninterrupted service of a PV facility and prevent severe malfunctions. Ground and line-to-line anomalies are common in PV arrays and are often associated with arcs.

ANNs and manifold learning combined with multilayer perceptron networks have shown promising results in identifying faults, achieving accuracy improvements over traditional methods [20,21,22]. Multiclass Support Vector Machine algorithms have been effective in classifying various defects, demonstrating high efficiency in fault detection [23], yet the research did not contrast detection approaches to ascertain their effectiveness. Faults in PV arrays such as open circuits and short circuits have been identified using voltage and current thresholds [24]. However, fault detection methods based on voltage and current thresholds can falter under changing irradiance conditions, leading to inaccuracies [22,25]. Nevertheless, this method becomes inaccurate when PV arrays are subjected to continuously changing irradiance. Moreover, the presence of safeguard diodes complicates the identification of faults that develop over time, necessitating more sophisticated detection models [26].

1.3. Contribution

This study introduces a new framework for detecting defects in solar panels through the VGG19-DL model with JFOSA for effective feature selection. The main contributions include:

• A comprehensive end-to-end approach for automating defect detection using aerial imagery.
• The use of JFOSA to optimize feature selection, leading to improved classification performance.
• Significant improvements in accuracy, precision, and computational efficiency over traditional methods.
• Extensive empirical validation using publicly available datasets, with comparisons to existing optimization algorithms.

1.4. Organization

Section 1 provides the general introduction, literature review, and research motivation. Section 2 introduces the proposed framework in detail. Section 3 describes the materials and methods, especially the VGG19 architecture, the JFOSA algorithm, and the datasets used. Section 4 presents the quantitative, comparative results, and the findings are discussed. Finally, Section 5 offers conclusions and suggestions for future work.

2. Proposed Framework

This research presents an innovative framework for detecting solar panel defects in aerial images. The framework combines two powerful methods:

• Deep Feature Extraction with VGG19: First, the VGG19 convolutional neural network architecture is used to extract rich and deep features from aerial images of solar panels. These features provide comprehensive information about the patterns and details in the images.
• Optimal Feature Selection with JFOSA: In the next step, the JFOSA approach is used to select the most effective subset of the extracted features. This helps reduce the dimensionality of the data and increase the efficiency of the model.
• Classification with Neural Network: Finally, the selected features are fed into an artificial neural network (ANN) to distinguish defective solar panels from healthy ones.

The results show that this proposed framework achieves higher accuracy and efficiency in classification compared to traditional deep learning methods and baseline models.

3. Materials and Methods

3.1. Dataset

The compiled dataset includes a total of 2624 grayscale solar cells, each sized at 300 × 300 pixels, depicting healthy and impaired cells at varying degrees of loss. The images were obtained from 44 solar modules and display two types of damage, intrinsic and extrinsic, which are known to affect the performance of solar modules.

To ensure consistency, the images were all scaled to size and adjusted to the same angle of view, with all lens aberrations removed prior to cropping out the solar cells. Each image has a defect probability value ranging from 0 to 1, indicating the proportion of a defect in the image. The type of solar module (mono vs. polysilicon) from which each cell image was taken is also included in the provided dataset. For the simulation we used the Matlab 2024a version.

The dataset is freely accessible and downloadable from GitHub at “https://github.com/zae-bayern/elpv-dataset” (accessed on 5 March 2024), along with the illustrative examples given in Figure 1.

According to Figure 1, specialists examined the original images, categorized each solar cell based on its operational condition, and assigned a label denoting the defect probability. These cells were divided into four different types with detected defect probability levels of 0.0, 0.33, 0.66, or 1.0. Figure 1a illustrates a defect-free solar cell with a probability of 0.0, while Figure 1b shows a cell with a minor defect probability of 0.3. A probability value of 0.66, as indicated in Figure 1c, represents a moderately defective cell, whereas Figure 1d displays a fully defective cell with a value of 1.0. This defect probability value reflects the extent of damage each cell has sustained within the PV unit.

Table 1. Dataset summary.

Condition	Defect Probability	Number of Samples
Defective	% 33	295
	% 66	106
	% 100	715
Non-Defective	% 0	1508

classifies the data into two classes, defective and non-defective, along with the provided defect rate probabilities and sample sizes for each state.

The dataset illustrated in Table 1 has been categorized into two general classes: defective and non-defective. It was further reorganized into categories representing specific fault types, enabling targeted training and validation. The non-defective class consists of samples with defect probabilities of 0.0 and 0.33, while the defective class comprises samples with probabilities of 0.66 and 1.0. The probabilities presented in Table 1 are derived from the softmax activation function applied to the final classification layer, indicating the model’s confidence in each prediction.

3.2. VGG19 Architecture

The VGG19 architecture is a deep convolutional neural network (CNN) that consists of 19 layers (16 convolutional and 3 fully connected layers) used for feature extraction and image classification. Below is the mathematical expression representing each key component of the VGG19 architecture, layer by layer.

The input layer takes an image of size $224\times 224\times 3$ (height, width, and color channels for RGB). The convolutional layers use filters to detect features. The convolution operation can be mathematically represented as:

$$Z^{\left[l\right]}=W^{\left[l\right]} \ast A^{[l-1]}+b^{\left[l\right]}$$

(1)

where $Z^{\left[l\right]}$ is the output of the l-th layer, $W^{\left[l\right]}$ is the filter (weight) applied in the l-th layer, $A^{[l-1]}$ is the activation from the previous layer, and $b^{\left[l\right]}$ is the bias term.

The VGG19 modeling technique comprises 16 convolutional layers divided into sets, each followed by a max-pooling layer [26]. Each convolutional layer uses a 3 × 3 filter with stride 1 and padding 1, thus maintaining the spatial dimensions of the input. Convolution is immediately followed by a ReLU activation function, as shown in Equation (2) [27]:

$$A^{\left[l\right]}=\mathrm{R}\mathrm{e}\mathrm{L}\mathrm{U}\left(Z^{\left[l\right]}\right)=max\left(0,Z^{\left[l\right]}\right)$$

(2)

Max pooling layers are then employed to downsample feature maps. Each max-pooling operation uses a $2\times 2$ filter with stride 2, which reduces the spatial dimensions of the feature maps. This occurs after each convolutional block. The downsampling process is represented by Equation (3):

$$A^{\left[l\right]}=max\left(A^{[l-1]}\right)$$

(3)

where the output $A^{\left[l\right]}$ is the maximum value from each 2 × 2 window of the input $A^{[l-1]}$.

After the last max-pooling layer (Equation (4)), the output is flattened into a 1D vector of size NFLAT and then passed through three fully connected layers. Each fully connected layer computes as follows:

$$Z^{\left[l\right]}=W^{\left[l\right]}{A}^{[l-1]}+b^{\left[l\right]}$$

(4)

This is followed by the ReLU activation, as given in Equation (5):

$$A^{\left[l\right]}=\mathrm{R}\mathrm{e}\mathrm{L}\mathrm{U}\left(Z^{\left[l\right]}\right)$$

(5)

In the last fully connected layer, class probabilities are predicted using the softmax activation function, as given in Equation (6):

$${\widehat{y}}_i={\displaystyle \frac{e^{Z_i^{\left[L\right]}}}{\sum _j e^{Z_j^{\left[L\right]}}}}$$

(6)

where ‘${\widehat{y}}_i$’ refers to class i, for which the probability estimates are being made, $Z_i^{\left[L\right]}$ represents the output for the j-th class form the final layer, and $L$ indicates the final output layer of the neural network structure. The architecture of VGG19 is illustrated in Figure 2.

The VGG19 architecture can be described as follows:

• Input Layer: $224\times 224\times 3$;
• Convolutional + ReLU Layers: 2 layers of 3 × 3 filters, 64 filters each;
• Max Pooling Layer: 2 × 2;
• Convolutional + ReLU Layers: 2 layers of 3 × 3 filters, 128 filters each;
• Max Pooling Layer: 2 × 2;
• Convolutional + ReLU Layers: 4 layers of 3 × 3 filters, 256 filters each;
• Max Pooling Layer: 2 × 2;
• Convolutional + ReLU Layers: 4 layers of 3 × 3 filters, 512 filters each;
• Max Pooling Layer: 2 × 2;
• Convolutional + ReLU Layers: 4 layers of 3 × 3 filters, 512 filters each;
• Max Pooling Layer: 2 × 2;
• Fully Connected Layers: 3 layers (4096, 4096, and number of classes with softmax output).

To extend beyond basic binary classification, the model was enhanced to support multiclass classification, allowing it to detect common PV defects like microcracks, hot spots, shading, and delamination.

3.3. Jellyfish Algorithm

The Jellyfish Optimization Search Algorithm is a bio-inspired metaheuristic introduced in 2021, modeled on the natural movement and foraging behavior of jellyfish [4]. The algorithm relies on two main strategies: (1) drifting with ocean currents (passive motion) and (2) active group movement toward nutrient-rich areas. The efficiency of the algorithm depends on its ability to balance exploration and exploitation within the search space.

$$\left(X^{*}-e_c{X}_i\right){\displaystyle \sum }{\displaystyle \frac{1}{nPop}}=\overrightarrow{trend}$$

(7)

where $X^{*}$ represents the global best position, $e_c$ is the assimilation coefficient, and $X_i$ denotes the position of each jellyfish.

The group movement behavior is defined by:

$$\beta \times rand(0,1)\times \mu =df$$

(8)

where $\beta$ is a control parameter and $\mu$ is the mean position of the population.

Passive movement (random local search) is modeled as:

$$X_i\left(t+1\right)=X_i\left(t\right)+\gamma ·rand(0,1)\times (U_b-L_b)$$

(9)

Active interaction between two jellyfish based on their fitness is governed by:

$$X_i\left(t+1\right)=X_i\left(t\right)+rand·(X_j\left(t\right)-X_i\left(t\right))$$

(10)

To better understand the operation of the JFOSA, the following pseudocode (Algorithm 1) summarizes its operational steps. The algorithm alternates between passive movement with ocean currents and active group movement strategies to balance exploration and exploitation. The method iteratively updates the jellyfish positions based on fitness assessments to find the global optimum.

Algorithm 1: Steps of the Jellyfish Optimization Algorithm (JFOSA).

Start

Input:

▪ Initial population of jellyfish (candidate solutions)

▪ Objective function (fitness evaluation)

▪ Maximum number of iterations (MaxIter)

Output:

▪ Best solution found (optimal feature subset)

2. Procedure

1. Initialize the positions of all jellyfish randomly within the search space.

2. Evaluate the fitness value of each jellyfish based on the objective function.

3. $\mathrm{Determine} \mathrm{the} \mathrm{global} \mathrm{best} \mathrm{position} X^{*}$ among all jellyfish.

4. While (current iteration ≤ MaxIter) do:

4.1. For each jellyfish:

a. $\mathrm{Generate} \mathrm{a} \mathrm{random} \mathrm{number} R\in \left[1,0\right]$

b. $\mathrm{If} 0.5<R$:

$\mathrm{Move} \mathrm{according} \mathrm{to} \mathrm{ocean} \mathrm{current}: X_i \left(t+1\right)=X_i \left(t\right)+rand(0,1)\times \overrightarrow{trend}$

c. Else:

Move according to group behavior (passive or active motion):

i. Passive motion (exploration):  $X_i \left(t+1\right)=X_i \left(t\right)+\gamma ·rand(0,1)\times (U_b-L_b)$

ii. Active motion (interaction with another jellyfish):

❖ $\mathrm{Compare} \mathrm{fitness} \mathrm{of} X_i$$\mathrm{and} \mathrm{randomly} \mathrm{selected} X_j$.

❖ Update position using:

${fitness \left(X_i\right)<X}_i \left(t\right) if fitness \left(X_j\right)-X_j \left(t\right)\times rand(0,1)+\times X_i \left(t\right)=X_i \left(t+1\right)$

5. Update the global best solution if a better solution is found.

6. Increase the iteration counter.

End While

$\mathrm{Return} \mathrm{the} \mathrm{best} \mathrm{solution} {\mathit{X}}^{*}$.

Figure 3 shows that a random function plays a key role in updating the jellyfish’s position. This function generates a random value that determines the type of movement: passive movement with the ocean current or active interaction with other members for exploration. This random value establishes the balance between exploration and exploitation in the JFOSA method.

Figure 4 provides an overview of the workflow of the new framework, in which the feature extraction process starts using the VGG19 network, then feature selection is performed using the JFOSA, and finally, the classification is performed by an NN. Figure 4 shows how JFOSA enhances the discriminative power of deep features by selecting a subset that maximizes inter-class separability.

4. Results

In this section, the quantitative results obtained from the implementation of the new approach based on VGG19-JFOSA are compared with baseline models and various optimization methods. Evaluation metrics such as sensitivity, specificity, precision, F1 score, and AUC were used to evaluate the effectiveness of the model. Experimental measurements have indicated that typical PV defects significantly reduce energy yield, with hot spots causing to around a 17% loss, shadowing about 23%, and micro-cracks close to 12%. These losses highlight the critical importance of accurate fault classification to maintain PV system efficiency.

4.1. Classification Report

Table 2. Classification report result.

	Sensitivity	Specificity	Accuracy	Precision
Non-Defective	97.34	97.79	98.21	97.49
Defective	98.03	98.24	98.26	98.39

displays the classification report, including sensitivity, specificity, accuracy, precision, and F1 score for both “defective” and “non-defective” categories. Table 2 displays the classification report, including sensitivity, specificity, accuracy, precision, and F1 score for both “defective” and “non-defective” categories. The high scores in all the criteria indicate the power of the proposed model in accurately detecting and classifying defective solar panels.

The performance metrics shown in Table 2 were computed based on the confusion matrix results using the following standard formulas:

Accuracy = (TP + TN)/(TP + FP + TN + FN)

(11)

Sensitivity (Recall) = TP/(TP + FN)

(12)

Precision = TP/(TP + FP)

(13)

4.2. ROC Curve Analysis

The Receiver Operating Characteristic (ROC) curve, illustrated in Figure 5, depicts the model’s true positive rate (TPR) against the false positive rate (FPR) across different thresholds. The proposed method achieved an area under the curve (AUC) score of 94.03%, indicating excellent discrimination capability between the defective and non-defective classes.

4.3. Comparative Results on DL Models

Figure 6 presents a bar graph comparing the AUC metrics for four different DL models: DenseNet121, MobileNetV2, VGGNet-16, and Xception. This figure clearly shows that for all four deep learning models, using the JFOSA method for feature selection results in a significant improvement in AUC performance. This indicates that JFOSA effectively selects more relevant features, resulting in better discrimination capabilities for the classification tasks for which these models are designed. VGGNet-16 shows the largest improvement in AUC when applying JFOSA.

Ablation Study

To assess the robustness of the proposed model, an ablation study was conducted by removing the JFOSA feature selection module. This resulted in a 9.1% reduction in total classification accuracy, which confirms the significant contribution of JFOSA to generalization and feature relevance. This demonstrates the model’s dependence on this valuable component and supports the necessity of integrating JFOSA into the framework.

4.4. Comparison Among Optimization Algorithms

Figure 7 presents a comparative evaluation of the five optimization algorithms—Ant Colony Optimization (ACO), Genetic Algorithm (GA), Gray Wolf Optimization (GWO), Particle Swarm Optimization (PSO), and the new approach (JFOSA)—based on sensitivity, accuracy, and AUC. The results show that JFOSA consistently outperforms the others in all criteria, achieving a sensitivity of 98.71%, an accuracy of 98.4%, and a leading AUC of 94.03%. Although GWO slightly outperforms JFOSA in sensitivity and ACO performs relatively well, both GA and PSO show poorer results overall. These findings confirm the effectiveness and balanced performance of JFOSA within the proposed framework.

Figure 8 illustrates the ROC curves of five optimization algorithms (JFOSA, GWO, ACO, PSO, and GA) used to evaluate classification performance. JFOSA’s curve (blue) is consistently closest to the top-left, indicating superior performance across thresholds. GWO follows closely, while ACO performs moderately well. In contrast, PSO and GA demonstrate weaker performance, with curves positioned lower and to the right. Overall, JFOSA achieves the best balance between TPR and FPR, confirming its effectiveness for classification tasks in this study.

4.5. Comparative Accuracy with Other Methods

The proposed VGG19-JFOSA method was evaluated against several state-of-the-art solar panel defect classifications. As summarized in

Table 3. Comparison of methods and datasets with other approaches.

Author	Dataset	Method	Dataset Focus, Strengths, and Weaknesses
Lee et al. [28]	20,000 infrared images	LIRNet	Focus on infrared images; likely strong for identifying hot spots, but may be less effective for structural defects.
Bartler et al. [29]	98,280 labeled cell images	CNN	Large dataset; covers a wide variety of defects, but details of defect types are not known.
Deitsch et al. [30]	2624 solar cell images	VGG, SVM	A moderate number of solar cell images; exact details of defects not specified.
Mayr et al. [31]	2624 solar cell images	residual network	A moderate number of solar cell images; exact details of defects not specified.
Tang et al. [32]	1800 EL images	CNN	Focusing on EL images; suitable for detecting internal cell defects, but may not cover surface defects.
Qian et al. [33]	two datasets, MCOM and ELPV	CNN	Using two datasets; increases generalizability, but the details of each dataset and the flaws being investigated are not clear.
Khosa et al. [34]	2624 solar cell images	CNN	A moderate number of solar cell images; exact details of defects not specified.
Demirci et al. [35]	2624 solar cell images	MobileNetv2, SqueezeNet	A moderate number of solar cell images; exact details of defects not specified.
Proposed Method	2624 solar cell images	VGG19-JFOSA	Dataset with healthy and damaged cells with defect probabilities; capable of detecting all types of defects, but limited in scalability and diversity of real-world images.

and illustrated in Figure 9, the proposed method outperforms all existing models, achieving a classification accuracy of 98.34%.

Compared to MobileNetV2 (75.00%), CNN (Qian) (71.89%), and a general CNN model (77.30%), the VGG19-JFOSA method shows a substantial improvement. It also surpasses deeper architectures such as Residual Net (82.00%) and specialized CNN-based models like CNN (Tang) at 83.00% and CNN (Khosa) at 83.50%.

Even higher-performing models like LIRNet (85.10%) and VGG-SVM (88.42%) fall short of the proposed method. The consistently superior accuracy of 98.34% confirms the effectiveness and robustness of the VGG19-JFOSA framework for detecting solar panel defects. To validate the generalizability of the model, we also tested its performance on various types of photovoltaic modules, such as monocrystalline, polycrystalline, and bifacial panels. The proposed model obtained consistent results, maintaining accuracy above 95% across tested panel types.

4.6. Algorithm Convergence Analysis

To further evaluate the effectiveness of the proposed JFOSA, its convergence behavior was analyzed over 100 iterations. As shown in Figure 10, the classification accuracy increases rapidly during the first 30 iterations, indicating that the algorithm quickly identifies a near-optimal subset of features. After this point, the curve gradually flattens out, indicating stable and consistent performance with minimal changes, confirming the convergence of the algorithm.

This behavior highlights the balance between exploration and exploitation in JFOSA. The initial increase in the slope indicates effective exploration of the feature space, while the subsequent stabilization indicates efficient exploitation of the most promising regions. Overall, the convergence curve visually confirms that JFOSA effectively improves classification accuracy and maintains stable performance, supporting its role as a reliable feature selection method in the proposed framework. Moreover, despite its iterative process, JFOSA demonstrated high computational efficiency. By reducing the population size and employing early stopping, the average inference time on a Jetson TX2 module was maintained below 1 s, confirming its potential for real-time applications in practical PV inspection systems.

5. Conclusions

In this study, a robust approach combining the VGG19 deep neural network with JFOSA is presented for efficient solar panel defect detection using aerial images. This new approach is rigorously evaluated and compared with baseline models and various optimization algorithms. Experimental results show that the VGG19-JFOSA framework significantly outperforms other methods, achieving exceptional classification accuracy of 98.34%, sensitivity of 98.71%, specificity of 98.69%, and AUC of 94.03%. These results highlight the effectiveness of the proposed approach in accurately detecting solar panel defects such as cracks, hot spots, and shadow effects.

Furthermore, comparative analyses show that the VGG19-JFOSA model outperforms state-of-the-art solar panel defect classification methods, showing significant improvements over models such as MobileNetV2, CNN-based approaches, and residual networks. The model exhibits rapid convergence, achieving high accuracy early and maintaining stable performance. The results confirm the potential of the VGG19-JFOSA framework as a scalable and automated solution for solar panel monitoring, ensuring increased energy efficiency and system reliability. This work contributes to the advancement of fault detection technologies in photovoltaic systems and provides a reliable and efficient tool for large-scale, real-time solar panel inspection.