Due to prolonged exposure to natural environmental conditions, photovoltaic (PV) arrays are prone to various degradation mechanisms, including module aging, open-circuit faults between strings, localized shading of modules, inter-module short circuits, and simultaneous multi-failure occurrences. These issues can lead to reduced power generation efficiency and potential safety risks. Consequently, extensive research has been conducted on photovoltaic fault diagnosis. Artificial intelligence algorithms have emerged as effective tool in this field due to their high computational speed and superior diagnostic accuracy.

In recent years, artificial intelligence technology has gained widespread popularity among industry professionals due to its efficiency, accuracy, and strong adaptability, and has been extensively adopted across various sectors. The process begins with data collection, where sensors are used to gather various operational parameters from photovoltaic systems, which are then preprocessed. Next, an appropriate AI diagnostic model is developed and trained using the segmented dataset, followed by optimization. Finally, the model’s performance is evaluated.

Zain et al. (2023) proposed a method using multi-output deep learning algorithms to detect, classify, and locate short-circuit, ground-fault, and open-circuit faults. Compared with existing methods, this approach significantly reduces the number of sensors required per string—by up to 50%—while maintaining a high diagnostic accuracy rate. Lin et al. (2024) proposed a photovoltaic array open-circuit fault diagnosis model based on a one-dimensional VoVNet-SVDD (Variety of View Network with Support Vector Data Description). The model automatically extracts fault features from the input raw current-voltage curve data, then combines these extracted features with environmental parameters to construct an SVDD model, thereby enabling diagnosis. Sridhar Patthi et al. (2024) proposed a multi-layer neural network (MLNN) technique capable of identifying issues of any size or mismatch severity in photovoltaic arrays. Gao et al. (2024) proposed a PV array fault diagnosis method based on Variational Mode Decomposition (VMD) and the Subtractive Average Optimizer (SABO) to optimize the Kernel Extreme Learning Machine (KELM). By utilizing VMD to extract nine feature vectors from fault data and employing the SABO algorithm to optimize the kernel and regularization parameters of the KELM, they established a PV array fault diagnosis model based on VMD-SABO-KELM. Zhong et al. (2024) proposed a lightweight PV array fault diagnosis model based on an improved SqueezeNet. This model significantly reduces computational complexity by replacing traditional convolutions with deep convolutions, while incorporating residual connections to facilitate effective information flow between layers, thereby enhancing diagnostic efficiency and accuracy. Ali et al. (2025) developed a simple and accurate one-dimensional convolutional neural network model that classifies faults based on selected features. Lu et al. (2023) proposed an accurate FDD model for photovoltaic arrays based solely on small datasets, utilizing a Wasserstein-based generative adversarial network to generate more labeled samples to improve the performance of the CNN-based classifier. Liu et al. (2021) proposed a PV array fault diagnosis method based on stacked autoencoders and clustering algorithms, which can automatically extract features and mine data sample characteristics for fault diagnosis using a small number of labeled data samples. Fu et al. (2025), to address the issue of data imbalance in photovoltaic arrays affecting the performance of fault diagnosis models, proposed integrating an improved slime mold algorithm with a CatBoost model based on a polyloss function to achieve accurate composite fault diagnosis. Zahra Yahyaoui et al. (2024) proposed a novel Bayesian optimization-based diagnostic framework using interval-gated recurrent units to reduce the computational and storage costs associated with sensor uncertainty. Shiue-Der Lu et al. (2024) were able to rapidly and accurately measure and locate short-circuit and open-circuit faults in bypass diodes by integrating voltmeters and support vector machines into solar photovoltaic array modules. Mansour Hajji et al. (2021) developed a fault diagnosis method based on principal component analysis (PCA) to extract and select the most relevant multivariate features, and applied supervised machine learning classifiers to address the challenge of selecting more relevant and sensitive features.

The NCA-CNN model proposed by Umit Cigdem et al. (2025) achieves a fault recognition rate of up to 99% for both aging and series resistance faults. The 1D-CNN model proposed by Belqasem (Aljafari et al. 2024) achieves a 98.15% accuracy in identifying shadow faults in open- circuit conditions. Aziz, F et al. (2020) proposed a fault diagnosis method based on a 2D-CNN. The results show that the proposed fine-tuned pre-trained CNN model achieves a fault detection accuracy of 73.53% and 70.45% in noisy scenes, outperforming existing machine learning and deep learning methods.

In summary, four typical types of AI-detected photovoltaic (PV) faults are identified for single fault scenarios, including module surface shading, aging, open circuits, and short circuits. The diagnostic accuracy ranges from 93% to 99%. However, although the aforementioned models demonstrate relatively high accuracy in diagnosing several types of single faults, the diagnosis of multiple failures still requires further improvement.

In this paper, an ensemble learning model is proposed based on weighted probabilistic averaging. This integrated model effectively combines the strengths of DBN-ELM, CNN-SVM, and CNN-BiGRU. Seven types of faults, encompassing both single and multiple faults, were identified. Corresponding training and testing datasets were then constructed, and the performance of the model was computed and evaluated.

As the foundation of fault diagnosis research, the construction and application of photovoltaic system fault datasets have a decisive impact on the development and optimization of diagnostic models. Voutsinas et al. (2022) used a single-diode model to calculate the photocurrent generated by given irradiance and temperature settings, creating values for a dataset to be used in the neural network training process. Karmacharya et al.(2018) modeled and simulated an ungrounded photovoltaic system in a real-time digital simulator (RTDS). Signals from the model outputs were collected to construct the dataset. Wang et al. (2019) set up a fault data acquisition system and, based on this, used MATLAB to establish a 4×3 PV array model. They simulated the PV array’s normal, open-circuit, short-circuit, inter-string, and multi-fault states, collecting 779 data samples across these five states. Based on the fundamental principles of the various components of off-grid photovoltaic systems, fault simulations are performed using a simulation model in this chapter, and fault labels are established and a fault dataset is constructed from the results of the simulation analysis.

To conduct an in-depth analysis of the output performance of a photovoltaic array under various fault conditions, this paper uses MATLAB/Simulink to build a 3×3 photovoltaic array model with a total output power of 2 kW. The study examines the output characteristics of four single faults—shading on the module surface, short circuits within module strings, open circuits between strings, and array aging—as well as three types of concurrent faults. Figure 1 shows a schematic diagram of a photovoltaic array. The detailed parameters of each photovoltaic module under standard test conditions are shown in Table 1. The schematic illustrates the behavior of several types of single faults.

By establishing a simulation model of the output characteristics of photovoltaic modules, multi-condition simulations were conducted by varying irradiance (200–1000 W/m²) and temperature (15–55 °C) to reveal the dynamic relationship between the modules’ electrical parameters and environmental factors. Simulations show that when the module temperature is held constant at 25 °C, a decrease in irradiance from 1000 W/m² to 200 W/m² causes the short-circuit current (Isc) to decrease approximately linearly from 41 A to 8 A (a reduction of 80.5%), resulting in a drop in maximum power point from 3700 W to 600 W; The open-circuit voltage (Voc) decreased from 110 V to 103 V, a reduction of approximately 6.3%. As shown in the figure 2 to figure 5.

Under conditions of constant irradiance (1000 W/m²), simulation data shows that as the module temperature rises from 15°C to 55°C, the open-circuit voltage (Voc) exhibits a significant negative temperature coefficient, decreasing from 115 V to 101 V at a rate of -0.35%/°C, with a total reduction of 12.2%. In contrast, the short-circuit current (Isc) shows only slight changes. The combined effect of these factors causes the maximum power point to decrease from 3700 W to 3100 W at a rate of -0.41%/°C, representing a reduction of approximately 16.2%.

These regular patterns provide important evidence for fault diagnosis in photovoltaic arrays. Under normal operating conditions, changes in output parameters should strictly follow the aforementioned radiation-temperature coupling laws, and any deviation from these patterns may indicate a potential fault.

Simulations were conducted to obtain the I-V and P-V characteristic curves for a single photovoltaic module under surface shading conditions. These simulations were performed under STC (Standard Test Conditions) with an irradiance of 1000 W/m² and a module temperature of 25 °C. Analysis shows that when the module is shaded, the array output current decreases in a “step-like” manner with voltage, and the output power first increases, then decreases, and then increases again, exhibiting multiple peaks; the open-circuit voltage and short-circuit current do not change significantly, while the maximum output power decreases by approximately 1000 W as shown in the figure 6 and 7.

Analysis of the I-V and P-V characteristic curves of a photovoltaic array under STC conditions during a short-circuit fault, as obtained through simulation, reveals that when a shorting wire is used to short-circuit a single PV module in the array, the open-circuit voltage of the array drops significantly compared to normal operating conditions—by approximately 30 V—resulting in a reduction in maximum power output of about 1,100 W. This reduction is directly proportional to the number of affected modules as shown in the figure 8 and 9.

Analysis of the I-V and P-V characteristic curves of the photovoltaic array under STC conditions following a simulation of an open-circuit fault reveals that when a single PV string is disconnected from the array, the array’s open-circuit voltage remains largely unchanged, while the short-circuit current decreases significantly, leading to a drop in output power. Furthermore, the maximum output power decreases by approximately 1,200 W compared to normal operating conditions, as shown in figure 10 and 11.

Analysis of the I-V and P-V characteristic curves of the photovoltaic array, obtained through simulation, when aging-related failures occur under STC conditions reveals that, prior to the maximum power point, the output current decreases more rapidly as the output voltage increases compared to normal operating conditions. This results in a reduction in maximum power output relative to normal conditions, with the extent of this reduction being closely related to the degree of array aging, as shown in figure 12 and 13.

This study addresses the issue of array fault diagnosis in off-grid photovoltaic systems by constructing an 8-label classification system comprising seven typical fault modes and one normal operating state, as listed in Table 2. Specifically, Label 1 represents the normal operating state of a 3×3 photovoltaic array and serves as the baseline reference for subsequent fault identification. Label 2 simulates shading on the module surface by randomly selecting 1 to 3 modules in the array and reducing their irradiance by 30% to 90% to characterize varying degrees of shading effects. Label 3 corresponds to aging-related faults in the PV array, achieved by connecting a 5–10 Ω resistor in series to represent performance degradation caused by increased series resistance after long-term operation. Label 4 simulates an open-circuit fault between strings by removing an entire row of PV strings, replicating open-circuit scenarios such as connector detachment or cable breakage. Label 5 describes a short-circuit fault between modules, using ultra-low-resistance resistors of 1×10⁻⁶ Ω to randomly short-circuit several modules, reflecting short-circuit phenomena caused by hidden cracks in solar cells or connection failures during actual operation. Furthermore, Labels 6–8 focus on more complex concurrent faults: Label 6 combines an inter-string open-circuit fault with shading on the module surface, simultaneously removing one string and applying 30%–90% irradiance attenuation to the remaining modules; Label 7 represents the combination of array aging and shading, superimposing random radiation attenuation on the array in addition to series-connected 5–10 Ω aging resistors; Label 8 combines inter-module short circuits with shading, applying both ultra-low-resistance short circuits and radiation density attenuation simultaneously to the array, thereby comprehensively covering common composite failure scenarios encountered in actual operation. The simulation model covers a solar irradiance range of 100–1000 W/m² and a module temperature range of 10–70°C. To enhance the randomness and representativeness of the samples, the intervals between changes in irradiance and module temperature were not fixed, thereby fully simulating the output characteristics under various operating conditions. For each operating scenario, 500 samples were collected, ultimately constructing an array failure dataset comprising 4,000 samples. The dataset was divided into a training set and a test set in a 7:3 ratio, with the training set containing 2,800 samples and the test set containing 1,200 samples (as shown in Table 2). Additionally, Table 3 lists the sample values for various array failures under Standard Test Conditions (STC), providing a baseline reference for subsequent diagnostic analysis.

To identify a high-performance fault diagnosis model for photovoltaic arrays, this study developed and compared five diagnostic models based on different principles. Convolutional neural network (CNN) models possess strong feature extraction capabilities and can effectively extract key information from photovoltaic array fault data; the CNN-LSTM model combines the feature extraction strengths of CNN with the temporal data processing advantages of LSTM, enabling better handling of time-dependent fault sequences; CNN-SVM uses SVM to classify features extracted by CNN, further improving diagnostic accuracy; CNN-BiGRU leverages bidirectional gated recurrent networks to enhance the learning and understanding of temporal features in fault data; DBN-ELM combines the deep feature learning capabilities of deep belief networks with the rapid convergence characteristics of extreme learning machines to improve the model’s diagnostic efficiency.

Subsequently, these five diagnostic models were systematically trained using a training dataset comprising 2,800 sets of fault samples. Next, each model was applied to the fault diagnosis of off-grid photovoltaic arrays using a test dataset containing 1,200 fault samples. Table 4 details the overall accuracy and accuracy rates for each fault category. The overall accuracy rates for these five models were 84.3%, 86.4%, 90.8%, 92.3%, and 94.7%, respectively. Among them, the DBN-ELM, CNN-SVM, and CNN-BiGRU models achieved relatively high overall accuracy rates, demonstrating excellent fault diagnosis performance. However, there are shortcomings in identifying single-point failures caused by array aging and concurrent failures involving both component shadowing and array aging.

To effectively address this issue, this paper proposes the use of a weighted probabilistic averaging ensemble algorithm to perform ensemble optimization on the DBN-ELM, CNN-SVM, and CNN-BiGRU models, thereby constructing a novel diagnostic model.

This paper addresses the fault diagnosis problem of photovoltaic arrays by constructing an 8-label classification system and analyzing corresponding output characteristics. A fault diagnosis method is proposed for photovoltaic arrays based on a weighted probabilistic averaging ensemble model. This method achieves a high recognition rate for photovoltaic array faults as validated by experiments. The main conclusions are as follows: