Performance Optimization of Machine-Learning Algorithms for Fault Detection and Diagnosis in PV Systems

Abstract

The early detection of faults in photovoltaic (PV) systems is crucial for ensuring efficiency, minimizing energy losses, and extending operational lifespan. This study evaluates and compares multiple machine-learning models for fault diagnosis in PV systems, analyzing their performance across different fault types and operational modes. A dataset comprising 2.2 million measurements from a laboratory-based PV model, covering seven fault categories—including inverter failures, partial shading, and sensor faults—is used for training and validation. Models are assessed under both Maximum Power Point Tracking (MPPT) and Limited Power Point Tracking (LPPT) conditions to determine their adaptability. The results indicate that the ensemble bagged tree classifier achieves the highest accuracy (92.2%) across all fault scenarios, while neural network-based models perform better under MPPT conditions. Additionally, the study highlights variations in model performance based on power mode, suggesting the potential for adaptive diagnostic approaches. The findings reinforce the feasibility of machine learning for predictive maintenance in PV systems, offering a cost-effective, sensor-free method for real-time fault detection. Future research should explore hybrid models that dynamically switch between classifiers based on system conditions, as well as validation using real-world PV installations.

1. Introduction

Due to the rapid growth of industrial expansion and the constant increase in energy consumption, the global demand for energy per capita continues to increase significantly. This scenario has driven intensive research into new, safe, and sustainable green energy technologies, such as solar, wind, hydroelectric, tidal, biomass, and geothermal energy. These alternatives seek not only to meet the growing demand, but to do so in a way that reduces the environmental impact and is complemented by competitive costs within the energy market. Among these options, solar power generation is recognized as one of the most mature methods of extracting energy directly from the natural environment. Thanks to advances in solar technology, the installed photovoltaic capacity worldwide reached 2 Terawatts (TW) at the end of 2023, reflecting the global commitment to this form of renewable energy [1].

However, one of the main challenges facing solar power generation is its continuous exposure to variable and often adverse environmental conditions. This prolonged exposure can reduce the optimal performance of the photovoltaic system, affecting the efficiency of solar panels over time. Common symptoms of a reduction in the power generated by a solar panel include various types of failures. Solar energy is a sustainable and pollution-free power source, but photovoltaic (PV) panel degradation remains a concern. PV modules (PVMs) degrade irreversibly at an estimated rate of 1% per year, making early fault detection crucial for maintaining efficiency and extending operational lifespan. Predictive maintenance is essential to proactively detect and address potential failures, thereby minimizing downtime and repair costs. Predictive diagnostic techniques applied explicitly to each solar panel ensure the early detection and isolation of predictive symptoms that alert to the occurrence of some degrading phenomenon that may lead to the inevitable development of different types of failures or faults [2].

Degradation modes are the processes by which PVMs lose their characteristic properties: electrical, optical, chemical, or mechanical. The degradation modes manifest themselves physically in the PVMs, in one of their elements or in several simultaneously (

Table 1. PVM degradation modes.

Mode of Degradation	Description	Effect
Delamination	Loss of adhesion between the different layers of material	Light reflection increases
Discoloration	Degradation of the encapsulant or the adhesive material between glass and the PVM	Optical transmission is modified
Corrosion	The destruction of a metal by an electrochemical reaction	Increases the electrical conductivity, causing leakage currents
Module breakage	Roof glass breakage	Cracks or even breakages
Bubbles	Chemical reactions that release gases trapped inside the PVM	Hindering the heat dissipation in the photovoltaic cells, reducing their useful life
Potential induced degradation	Increase in volts in a chain	Leakage current
Electrical discharges	Electrical discharges from storms	Direct discharge or magnetic coupling

). In addition, it is difficult to differentiate what has produced the degradation since it may be due to concurrent causes. The root cause of the failure may be due to synergies between different mechanisms influenced, both in the time of their appearance and in their severity, by a series of meteorological factors like solar radiation, UV radiation, humidity, snow, wind, hail, high temperatures, dirt, salt, or gas. External factors such as vandalism that can produce irreparable faults must also be considered.

Various studies conclude that the annual energy loss due to these types of failures is around 18.9%, which underlines the importance of having systems that can prevent and manage these problems [3]. To meet this challenge, it is essential to develop advanced techniques that allow the true potential of renewable energy sources to be accurately estimated and to detect failures in photovoltaic panels in real time. A comprehensive solution to these problems is the implementation of a monitoring system complemented by fault diagnosis techniques. This system aims to maximize the operational reliability of the PV system and minimize system costs by early detection of the causes affecting its performance. With this approach, corrective measures could be taken in time, preventing the PV system from underperforming for extended periods, thus improving its overall efficiency and reducing energy losses associated with undetected faults. Recently, efforts have been made to understand the faults of PV systems resulting in the development of new techniques to detect and localize the type of fault present in the system. These techniques have helped in improving the system reliability and lifespan of PV systems. The classification of different fault-detection techniques to identify the type and location of the fault occurring on the DC and AC sides of the PV system are shown in Table 2, Table 3 and Table 4.

In recent years, the use of machine learning in fault diagnosis has increased significantly, enabling more accurate, automated detection and classification of anomalies by leveraging complex data patterns and features. The study [4] proposes an improved approach for fault analysis in bearings by extracting time and frequency domain features, enhanced through Cepstrum pre-whitening. Using machine learning and deep learning techniques, the authors improve fault classification in bearings, demonstrating that this methodology enhances diagnostic accuracy. Ref. [5] introduces the Latent Nestling Method based on Hybrid Colored Petri Nets for diagnosing intermittent faults in IGBTs. This methodology advances the field by enhancing fault detection and classification, offering a more accurate and efficient approach for predictive maintenance in power electronics. Various studies have also been conducted to predict or classify faults in photovoltaic systems. The paper [6] examines predictive maintenance in PV systems, focusing on challenges, current techniques, and the importance of machine learning and real-time monitoring for improving performance and reducing failures. In [7], the authors present the development and in situ testing of SmartPV, a cost-effective monitoring and diagnostic system for individual PV panels, demonstrating its effectiveness in tracking performance variations and detecting faults in real time. The article [8] analyzes the performance of various photovoltaic (PV) array configuration series (S), parallel (P), series-parallel (SP), total-cross-tied (TCT), and bridge-linked (BL) under partial shading and faulty conditions. Using MATLAB/Simulink R2022a, it evaluates seven key indicators to optimize PV system design and management in variable operating conditions. The same program is used in [9] to investigate the impact of partial shading on PV installations by modeling array performance with an equivalent circuit approach and simulating different shading scenarios. Ref. [10] proposes a predictive fault diagnosis approach for ship photovoltaic module systems, leveraging sensors and advanced algorithms for early fault detection. The research conducted in [11] presents a novel predictive diagnosis approach based on predictor symptoms for isolated photovoltaic systems using MPPT charge regulators. On the subject of storage methods, [12] proposes a reliability-based sizing method for stand-alone residential PV systems using Monte Carlo simulations to optimize solar panel and battery storage capacities under varying conditions. Some studies focus on identifying a single type of fault. For example, Ref. [13] investigates different levels of partial shading under MPPT conditions using I–V data. An artificial neural network was employed for this task, achieving accuracy between 97% and 98%. Similarly, Ref. [14] also relies exclusively on I–V data but decomposes them to extract key features. A two-stage SVM classifier reached an accuracy between 91% and 95% when identifying mismatch faults. Other research efforts aim to develop robust and cost-effective systems [15]. By applying principal component analysis to the I–V data provided by the inverter, this study achieved over 97% accuracy in detecting partial shading. Following a similar approach, Ref. [16] detected faults by analyzing the I–V curve. Some studies, such as [17], utilize additional sensors to gather more information about the system, aiming for higher fault-detection accuracy. A recent study [18] presents a predictive energy management model for small-scale residential PV systems, combining a Markov prediction chain with a dynamic decision algorithm. The work addresses optimization under varying levels of renewable energy penetration, evaluating system performance through metrics such as energy output, fossil fuel dependency, and financial indicators. By incorporating predictive intelligence into operational decision-making, the study demonstrates how advanced data-driven models can enhance both efficiency and reliability in PV systems. This aligns with the current research focus on leveraging machine-learning techniques to improve fault detection and overall system performance.

Table 2. PV DC-side fault-detection techniques based on electrical characterization.

Technique	Description	Detectable Faults	References
Climate Data Independent (CDI)	Uses external devices like LCR meters and signal generators. Analyzes the PV system’s response to injected signals. Does not require climate data.	Breakpoints, impedance changes, degradation, electrical faults.
Electrical Current and Voltage Measurement (EM)	Measures voltage and current at the output terminals. Used for real-time monitoring and automatic fault detection.	General electrical faults, voltage and current anomalies, inverter failures.	[19]
Comparison of Measured and Modeled Outputs (CMM)	Compares real vs. expected PV system outputs. Uses predictive models and threshold values to detect deviations.	Power output deviations, system inefficiencies, performance degradation.	[20]
Heat and Temperature Exchange (HET)	Monitors temperature variations in PV modules. Identifies thermal anomalies due to electrical faults.	Hot spots, thermal degradation, module failures.
Power Loss Analysis (PLA)	Compare monitored power data with simulated results. Identifies excessive power losses.	Power loss in PV modules and inverters, energy inefficiencies.	[21]
Ground Fault Detection and Interruption (GFDI)	Monitors ground current. Shut down the inverter if the current exceeds a safe limit.	Ground faults, leakage currents, electrical safety issues.	[22]
Machine Learning (ML)	Uses trained models to recognize fault patterns. Adapts to different weather conditions.	Partial shading faults, inverter failures, array defects.	[23,24,25]
Insulation Monitoring Devices (IMDs)	Measures resistance between conductors and the earth. Triggers an alarm if resistance falls below a set value.	Ground faults, insulation degradation.
Frequency Spectrum Analysis (FSA)	Analyzes voltage and current waveforms. Focuses on specific frequency ranges to detect arc faults.	Arc faults, frequency noise interference.	[26]
Estimating Randomness in Voltage (ERV)	Uses statistical filters to measure randomness in voltage signals. Compares results to a threshold to detect anomalies.	Arc faults, signal instability.	[27]
Spread Spectrum Time-Domain Reflectometry (SSTDR)	Uses high-frequency signals to generate an autocorrelation plot. Detects faults without disconnecting the inverter.	Transmission line faults, wiring defects, ground faults.	[28]
Residual Current Monitoring Devices (RCDs)	Measures the difference between incoming and outgoing current in PV conductors. Helps to detect leakage currents.	Line-to-line (L-L) and line-to-ground (L-G) faults, leakage currents.

Table 2. PV DC-side fault-detection techniques based on electrical characterization.

Technique	Description	Detectable Faults	References
Climate Data Independent (CDI)	Uses external devices like LCR meters and signal generators. Analyzes the PV system’s response to injected signals. Does not require climate data.	Breakpoints, impedance changes, degradation, electrical faults.
Electrical Current and Voltage Measurement (EM)	Measures voltage and current at the output terminals. Used for real-time monitoring and automatic fault detection.	General electrical faults, voltage and current anomalies, inverter failures.	[19]
Comparison of Measured and Modeled Outputs (CMM)	Compares real vs. expected PV system outputs. Uses predictive models and threshold values to detect deviations.	Power output deviations, system inefficiencies, performance degradation.	[20]
Heat and Temperature Exchange (HET)	Monitors temperature variations in PV modules. Identifies thermal anomalies due to electrical faults.	Hot spots, thermal degradation, module failures.
Power Loss Analysis (PLA)	Compare monitored power data with simulated results. Identifies excessive power losses.	Power loss in PV modules and inverters, energy inefficiencies.	[21]
Ground Fault Detection and Interruption (GFDI)	Monitors ground current. Shut down the inverter if the current exceeds a safe limit.	Ground faults, leakage currents, electrical safety issues.	[22]
Machine Learning (ML)	Uses trained models to recognize fault patterns. Adapts to different weather conditions.	Partial shading faults, inverter failures, array defects.	[23,24,25]
Insulation Monitoring Devices (IMDs)	Measures resistance between conductors and the earth. Triggers an alarm if resistance falls below a set value.	Ground faults, insulation degradation.
Frequency Spectrum Analysis (FSA)	Analyzes voltage and current waveforms. Focuses on specific frequency ranges to detect arc faults.	Arc faults, frequency noise interference.	[26]
Estimating Randomness in Voltage (ERV)	Uses statistical filters to measure randomness in voltage signals. Compares results to a threshold to detect anomalies.	Arc faults, signal instability.	[27]
Spread Spectrum Time-Domain Reflectometry (SSTDR)	Uses high-frequency signals to generate an autocorrelation plot. Detects faults without disconnecting the inverter.	Transmission line faults, wiring defects, ground faults.	[28]
Residual Current Monitoring Devices (RCDs)	Measures the difference between incoming and outgoing current in PV conductors. Helps to detect leakage currents.	Line-to-line (L-L) and line-to-ground (L-G) faults, leakage currents.

Table 3. Other PV DC side faults detection techniques.

Technique	Description	Detectable Faults	References
Visual Inspection	Manual examination of PV modules under proper lighting conditions. Used to identify visible defects.	Cracks, discoloration, broken panels, physical damage.	[29]
Ultrasonic Inspection	Uses ultrasonic vibrations to detect defects. Can be performed using pulse-echo or transmission methods.	Cracks, micro-cracks, delamination in PV modules.	[30]
Lock-in Thermography (LIT)	Uses pulsed current injection to identify power losses and temperature variations. Detects local defects.	Shunt defects, localized heat losses, module degradation.	[31]
Infrared Thermal Imaging (IR)	Uses thermal cameras to detect temperature variations in PV modules. Identifies faults based on heat distribution.	Hot spots, poor contacts, short circuits, degradation.
Electroluminescence Imaging (EL Imaging)	Uses emitted photons from solar cells to visualize defects. Excitation can be conducted via an electrical current.	Cracks, poor connections, non-uniform current flow.	[32]

Table 3. Other PV DC side faults detection techniques.

Technique	Description	Detectable Faults	References
Visual Inspection	Manual examination of PV modules under proper lighting conditions. Used to identify visible defects.	Cracks, discoloration, broken panels, physical damage.	[29]
Ultrasonic Inspection	Uses ultrasonic vibrations to detect defects. Can be performed using pulse-echo or transmission methods.	Cracks, micro-cracks, delamination in PV modules.	[30]
Lock-in Thermography (LIT)	Uses pulsed current injection to identify power losses and temperature variations. Detects local defects.	Shunt defects, localized heat losses, module degradation.	[31]
Infrared Thermal Imaging (IR)	Uses thermal cameras to detect temperature variations in PV modules. Identifies faults based on heat distribution.	Hot spots, poor contacts, short circuits, degradation.
Electroluminescence Imaging (EL Imaging)	Uses emitted photons from solar cells to visualize defects. Excitation can be conducted via an electrical current.	Cracks, poor connections, non-uniform current flow.	[32]

Table 4. PV AC side islanding fault-detection techniques.

Technique	Description	Detectable Faults	References
Power Loss Analysis (PLA)	Compares monitored power data with simulated results to detect deviations.	Inverter failures, excessive power loss.
Machine Learning (ML)	Uses trained models to recognize fault patterns in inverter performance.	Inverter failures, abnormal inverter behavior.
System State Monitoring (SSM)	Monitors system parameters (voltage, frequency) using SCADA systems to detect grid loss.	Inadvertent islanding, grid disconnection.	[33]
Switch State Monitoring (SSM)	Tracks the status of reclosers and circuit breakers via SCADA to detect islanding conditions.	Inadvertent islanding, circuit breaker failures.	[34]
Intertripping	Uses communication between sensors and generating units to detect disconnections.	Inadvertent islanding, sudden grid disconnection.	[35]
Under/Over Voltage and Frequency	Uses protection relays to monitor voltage and frequency at the PCC.	Grid disconnection, voltage/frequency anomalies.	[36]
Voltage Phase Jump Detection	Monitors the phase difference between current and voltage to detect sudden shifts.	Inadvertent islanding, phase mismatch.
Harmonic Measurements	Measures total harmonic distortion (THD) to detect anomalies at the PCC.	Grid disconnection, harmonic distortion issues.
Voltage Unbalance Monitoring	Detects unbalanced three-phase voltage variations as an indicator of islanding.	Grid disconnection, unbalanced load conditions.	[37]
Impedance Measurement	Measures changes in system impedance due to short circuit current and reduced supply voltage.	Grid disconnection, impedance anomalies.	[37]
Slip-Mode Frequency Shift (SMS)	Uses positive feedback to introduce phase and frequency shifts for islanding detection.	Grid disconnection, islanding conditions, power quality degradation.

Table 4. PV AC side islanding fault-detection techniques.

Technique	Description	Detectable Faults	References
Power Loss Analysis (PLA)	Compares monitored power data with simulated results to detect deviations.	Inverter failures, excessive power loss.
Machine Learning (ML)	Uses trained models to recognize fault patterns in inverter performance.	Inverter failures, abnormal inverter behavior.
System State Monitoring (SSM)	Monitors system parameters (voltage, frequency) using SCADA systems to detect grid loss.	Inadvertent islanding, grid disconnection.	[33]
Switch State Monitoring (SSM)	Tracks the status of reclosers and circuit breakers via SCADA to detect islanding conditions.	Inadvertent islanding, circuit breaker failures.	[34]
Intertripping	Uses communication between sensors and generating units to detect disconnections.	Inadvertent islanding, sudden grid disconnection.	[35]
Under/Over Voltage and Frequency	Uses protection relays to monitor voltage and frequency at the PCC.	Grid disconnection, voltage/frequency anomalies.	[36]
Voltage Phase Jump Detection	Monitors the phase difference between current and voltage to detect sudden shifts.	Inadvertent islanding, phase mismatch.
Harmonic Measurements	Measures total harmonic distortion (THD) to detect anomalies at the PCC.	Grid disconnection, harmonic distortion issues.
Voltage Unbalance Monitoring	Detects unbalanced three-phase voltage variations as an indicator of islanding.	Grid disconnection, unbalanced load conditions.	[37]
Impedance Measurement	Measures changes in system impedance due to short circuit current and reduced supply voltage.	Grid disconnection, impedance anomalies.	[37]
Slip-Mode Frequency Shift (SMS)	Uses positive feedback to introduce phase and frequency shifts for islanding detection.	Grid disconnection, islanding conditions, power quality degradation.

As a conclusion, it can be stated that numerous works have explored machine-learning (ML) approaches for fault detection in PV systems. Some models focus on single-fault detection, such as partial shading, using artificial neural networks (ANNs), while others apply support vector machines (SVMs) to classify mismatch faults with high accuracy. Other research highlights the potential of I–V curve analysis and principal component analysis (PCA) for robust and cost-effective fault diagnosis. However, few studies comprehensively compare multiple machine-learning models across various fault types and operating modes. This study aims to fill this gap by evaluating multiple machine-learning models for PV system fault diagnosis, distinguishing between Maximum Power Point Tracking (MPPT) and Limited Power Point Tracking (LPPT) conditions, and analyzing each model’s performance for different fault types. The objective is to develop a simple, robust, and cost-effective solution that relies solely on inverter-provided data, eliminating the need for additional sensors. This work provides an extensive evaluation of various ML models for fault diagnosis in PV systems. By exploring seven distinct types of faults, ranging from inverter failures to sensor faults and partial shading, it offers a well-rounded comparison of the models’ performance in different scenarios. The study uses a substantial dataset of 2.2 million measurements, which includes fault conditions under both MPPT and LPPT modes. These diverse data enhance the validity and generalizability of the results, providing insights into how ML models perform under varying operational conditions. The paper carefully evaluates the models using several metrics, including accuracy, training time, and confusion matrices. This comprehensive assessment allows for a nuanced understanding of each model’s strengths and weaknesses for PV fault diagnosis. The inclusion of specific fault types (e.g., grid anomalies and partial shading) helps to demonstrate the practical utility of these models for real-time fault detection. By leveraging inverter data for fault detection instead of requiring additional sensors, this study proposes a cost-effective solution, which is particularly useful for large-scale PV farms where additional hardware can increase maintenance costs.

The results of this study show that certain families of models, such as neural networks, perform exceptionally well in diagnosing different faults in PV systems. Furthermore, when analyzing each model’s performance based on specific fault types, it was observed that models do not perform equally across all faults, and accuracy and training time variate for every method depending on the type of fault. Another key finding is the impact of power mode on diagnostic accuracy. There is a noticeable difference in performance when the system operates under Maximum Power Point Tracking (MPPT) versus Limited Power Point Tracking (LPPT).

This manuscript is organized as follows: Section 2 summarizes the experimental setup, the data treatment necessary for the study, and explains the fault-detection system proposed in this work. Section 3 presents the results of the study, including the performance of the different algorithms tested. Section 4 discusses the results by comparing different methods and failures. Finally, Section 5 concludes by presenting some conclusions and points that need to be considered for the future.

2. Materials and Methods

To train and test different machine-learning models, a dataset obtained from a programmable solar array emulator was used [38]. This emulator allows for the simulation of varying environmental conditions, such as irradiance and temperature, to replicate real-world photovoltaic (PV) system behavior. The system employs a Voltage-Oriented Control (VOC) strategy combined with Space Vector Pulse Width Modulation (SVPWM) for active and reactive power regulation. Grid synchronization is achieved using a Phase-Locked Loop (PLL), which ensures proper alignment of the inverter’s output voltage with the grid voltage.

Real-time measurements were collected for key electrical parameters, including PV array voltage (V_pv), current (I_pv), DC voltage (V_DC), and three-phase grid voltages and currents (V_a, V_b, V_c, I_a, I_b, I_c). These signals were recorded at a sampling rate of 100 μs to ensure high-resolution data acquisition. The dataset also included estimated positive sequence voltage and current magnitudes, along with their respective frequencies.

Given the complexity of PV system behavior under different power conditions, data was collected under both Maximum Power Point Tracking (MPPT) and Limited Power Point Tracking (LPPT) modes. This distinction was essential for assessing how various machine-learning models adapt to different operating conditions. For each case, the database contains the data shown in

Table 5. Description of measured variables.

Variable	Measurement Description
Ipv	PV array current
Vpv	PV array voltage
VDC	DC voltage
Ia	Phase A current
Ib	Phase B current
Ic	Phase C current
Va	Phase A voltage
Vb	Phase B voltage
Vc	Phase C voltage
Iabc	Positive sequence estimated current magnitude
If	Positive sequence estimated current frequency
Vabc	Positive sequence estimated voltage magnitude
Vf	Positive sequence estimated voltage frequency

.

This study addresses the detection of seven realistic faults in various types and locations within grid-connected photovoltaic (PV) systems to ensure a comprehensive analysis (

Table 6. PV system faults.

Fault	Type	Description
F1	Inverter fault	Complete fault in one of the six IGBTs
F2	Feedback sensor fault	One phase sensor fault 20%
F3	Grid anomaly	Intermittent voltage sags
F4	PV array mismatch	10 to 20% nonhomogeneous partial shading
F5	PV array mismatch	15% open circuit in PV array
F6	MPPT/LPPT controller fault	−20% gain parameter of PI controller in MPPT/LPPT controller of the boost converter
F7	Boost converter controller fault	+20% in time constant parameter of PI controller in MPPT/LPPT controller of the boost converter

). Each fault was manually introduced in separate experiments lasting approximately 10 to 15 s, with the fault occurring around the 7th or 8th second. Data acquisition, both fault-free and faulty, was performed with a sampling time of Ts = 100 μs. Unlike simulation studies, the algorithm did not have prior knowledge of the exact fault occurrence time.

PV array mismatches, such as faults F4 and F5, present detection challenges due to significant variability in sensor data on the DC side. Fortunately, these faults are less severe, primarily causing power losses. In contrast, faults F1 and F3, which occur on the grid side of the PV system, are easier to detect because they affect only the AC side, where data exhibit minimal variability. However, due to their severity, early detection of these faults is crucial. The research also examines the parametric faults F6 and F7 in the MPPT/LPPT Proportional-Integral (PI) controller on the DC side, as well as a feedback current sensor fault, F2. Fault F7 indicates an increased time-constant parameter, while F6 involves a biased gain in the PI controller, resulting in reduced MPPT/LPPT trajectory tracking performance without affecting the closed-loop system’s stability. While these F6 and F7 faults do not pose an immediate risk to the system, they may damage the converter and cause power losses if left undetected over time. In contrast, the closed-loop current feedback sensor fault F2 is more severe, as it leads to a non-zero steady-state error and incorrect configuration of feedback controllers. These types of faults are common in practice, and impact on the performance of PV systems.

Faults in PV systems exhibit diverse characteristics under MPPT/LPPT controllers and varying environmental conditions. Consequently, fault classes display varying characteristics, making data-driven fault diagnosis approaches challenging. The study [39] highlights that detecting faults with 20% to 40% mismatch levels is difficult and challenging for mismatch levels below 20%. The method presented in this study effectively addressed these challenges for fault detection.

The PV system-faults data files are organized and labelled as Fxy, where

-

x ∈ {0, 1, …, 7} represents the fault scenario, where ‘0’ stands for fault-free experiment. ‘1’, …, ‘7’ are the seven types of faults conducted in the experiments.

-

y ∈ {‘L’, ‘M’} represents the operation mode, where ‘L’ stands for limited power mode (LPPT). ‘M’ stands for maximum power mode (MPPT).

For example, F4M is a data file for an experiment including fault F4 during MPPT mode, and F1L includes fault F1 during LPPT mode.

To facilitate data processing and model implementation, MATLAB R2022a was used along with the Classification Learner toolbox. This tool provided a streamlined approach for training and comparing multiple classification models. A script was developed to automate the process, ensuring consistency in data preprocessing and model evaluation.

The dataset was structured to include labeled measurements, differentiating between normal and faulty operation. Each fault type was manually injected into the system during controlled experiments, ensuring a comprehensive representation of different failure modes. The data were then reduced to a manageable subset of 16,000 measurements, ensuring an even distribution between normal and faulty cases to prevent overfitting.

Machine-learning models were trained and validated using a five-fold cross-validation technique. This approach was chosen to optimize model performance while minimizing overfitting, as it allows each model to be trained on different subsets of data and tested on unseen samples. The primary goal was to evaluate the accuracy and reliability of each model in detecting and classifying faults under various operating conditions.

The final evaluation involved comparing model accuracy across different power modes and fault scenarios. Additionally, performance metrics such as training time were analyzed to determine the feasibility of deploying these models in real-time fault-detection systems for PV installations.

3. Results

Three experiments were performed to evaluate which model works better with the different fault types and operating modes. Once the models are trained, their accuracy and training time are compared to check if any family of models works better when predicting and classifying all the data, independently of the type of fault (

Table 7. Model accuracy and training time.

Model Name	Accuracy (%)	Training Time (s)
Bagged Trees	92.2%	17.311
Wide Neural Network	92.0%	189.37
Medium Neural Network	91.8%	91.132
Quadratic SVM	90.7%	1635.9
Bilayered Neural Network	89.4%	78.093
Trilayered Neural Network	89.2%	88.667
Narrow Neural Network	88.8%	75.468
Cubic SVM	87.2%	4306.7
Fine Tree	85.9%	28.544
Boosted Trees	82.5%	31.108
Quadratic Discriminant	77.3%	1.851
Kernel Naïve Bayes	77.0%	341.370
Weighted KNN	70.7%	71.199
SVM Kernel	57.2%	119.68
Logistic Regression Kernel	53.6%	42.408

).

In choosing the most accurate model of each family, the family that predicts the best is the Ensemble classifier, which uses bagged trees to achieve an accuracy of 92.2%. The next best family is neural networks using a wide neural network, obtaining an accuracy of 92% (Figure 1). The result is quite different regarding training time and only considering the most accurate models. The fastest one is the family from the discriminant classifiers with the model quadratic discriminant. It only took 1.85 s to train, as seen in Figure 2. On the other hand, the support vector machine was the slowest, with the model Quadratic SVM taking over 27 min to train.

Since the used database consists of measurements under MPPT and LPPT, the second test’s purpose is to verify whether there is a difference in fault-detection accuracy in both conditions. Choosing the data is performed as before, randomly and maintaining a fifty-fifty ratio between good and faulty data to avoid overfitting. In this case, the faulty data are only of one power mode or the other. Their performance is compared once the models have been trained twice, first with MPPT and then with LPPT data. This comparison allows us to see that some prediction models work better for different power modes and whether faults under any power mode are easier to predict (Figure 3).

There is a slight difference between the models’ accuracy trained using MPPT data versus LPPT data. When using MPPT data, the models are more accurate. Also, the most accurate model for MPPT is not the most accurate for LPPT. When the system is under MPPT, it may be worth using a medium neural network to detect faults, since it has an accuracy of 98.1% (

Table 8. Model’s accuracy under MPPT mode.

Model Name	Accuracy (%)
Medium Neural Network	98.1%
Bilayered Neural Network	98.0%
Wide Neural Network	97.9%
Trilayered Neural Network	97.8%
Narrow Neural Network	97.7%
Quadratic SVM	97.5%
Cubic SVM	96.5%
Bagged Trees	96.4%
Fine Tree	94.3%
Linear SVM	93.3%

). But if the system is under LPPT, using the bagged trees model is more accurate, diagnosing with an accuracy of 91.5% (

Table 9. Models’ accuracy under LPPT mode.

Model Name	Accuracy (%)
Bagged Trees	91.5%
Wide Neural Network	91.3%
Cubic SVM	90.3%
Medium Neural Network	89.8%
Bilayered Neural Network	86.9%
Trilayered Neural Network	86.9%
Quadratic SVM	85.8%
Narrow Neural Network	84.8%
Fine Tree	83.9%
Medium Gaussian SVM	80.8%

). Depending on how the PV system works, one model or another could be used. Another option is to develop an algorithm that switches between both depending on the power mode. Since bagged trees was the most accurate model when being trained with both data types, that is the model used to study its performance under both conditions separately.

There are differences in the average accuracy in each case (Figure 4 and Figure 5). There is a gap of 4.77%, caused by the difference in current levels between modes. Another fact to consider is that when under MPPT, the model cannot predict fault 3 correctly, which corresponds to the grid anomaly. This issue in the model’s prediction capacity could also be caused by the nature of MPPT, adapting to the resistance of the load.

The last test is performed to analyze which model achieves a higher accuracy with each fault type. In addition, the performance of each model with different fault codes is analyzed. To do this, all the models are trained using 16,000 measurements, with half representing fault-free data and the other half representing the specific fault type. Like previous tests, the data selection is randomized for each fault type.

Figure 6 displays each model’s accuracy for every fault code. Almost every model achieved an accuracy of over 99% to forecast a good working PV system (Code 0). On average, bagged trees and various NNs were the ones that worked best in all cases. We observe that for fault 3, there is a significant drop in prediction accuracy. This gap with the other faults could be caused by the nature of the fault. In this case, it is a grid anomaly, which is more complex to forecast using only the measurements of the modules.

Some methods see their accuracy fall below 50%, and in the case of the boosted trees method, accuracy even drops to 0% for this specific fault. The only method that maintains an acceptable range in this scenario is the bilayered neural network, achieving nearly 80% accuracy. Additionally, as mentioned in previous analyses, the bagged trees method delivers the best overall performance, reaching 100% accuracy for fault 4, which involves partial shading, being one of the most frequent faults in general PV systems.

4. Discussion

When comparing the variety of available models, significant differences between classification methods can be observed under realistic conditions. The advantages and disadvantages of the different tested machine-learning methods are presented in

Table 10. Models’ performance comparison.

Model	Accuracy (%)	Training Time (s)	Advantages	Disadvantages	Application in PV Fault Diagnosis
Fine Tree	85.9%	28.5	Fast inference, simple model.	Less accurate than ensembles.	Suitable for real-time diagnosis.
Medium Tree	77.8%	4.85	Fast training and inference.	Moderate accuracy.	Useful when computational resources are limited.
Coarse Tree	64.1%	3.51	Very fast inference.	Low accuracy.	Not recommended for precise fault detection.
Linear SVM	82.0%	224.2	Good balance of accuracy and speed.	Struggles with non-linear problems.	Suitable for basic fault classification.
Quadratic SVM	90.7%	1635.9	High accuracy in complex data.	Long training time.	Good for fault detection where precision is critical.
Cubic SVM	87.2%	4306.8	Handles non-linear data well.	Extremely high training time.	May be useful if optimized for specific conditions.
Fine KNN	68.3%	14.79	Simple implementation.	Slower in inference.	Limited use due to lower accuracy.
Medium KNN	68.7%	7.26	Moderate accuracy.	High computational cost.	Less suitable for real-time applications.
Coarse KNN	60.3%	9.16	Fast training.	Low accuracy.	Not recommended.
Boosted Trees	82.5%	31.1	Improves accuracy over simple trees.	Poor performance in specific faults.	Less reliable for comprehensive fault detection.
Bagged Trees	92.2%	37.1	High accuracy, robust.	Can be slow.	Best overall choice for PV fault diagnosis.
RUSBoosted Trees	72.7%	12.2	Handles imbalanced data well.	Lower accuracy than bagged trees.	Only useful for specific failure scenarios.
Narrow Neural Network	88.9%	75.4	Handles complex relationships.	Computationally expensive.	Best for high-accuracy applications where resources are available.
Medium Neural Network	91.8%	91.1	High accuracy, adaptable to variations.	High computational cost.	Ideal for MPPT conditions.
Wide Neural Network	92.0%	89.3	Very high accuracy.	Needs significant computational power.	Useful in large-scale PV monitoring systems.

.

The results of this experiment have shown that certain families of models, such as neural networks, perform exceptionally well. This superior accuracy may be attributed to their training structure, which is based on interconnected nodes that allow for complex pattern recognition. Additionally, bagged trees emerged as the most accurate model, closely followed by the wide neural network. Both models demonstrated the highest predictive performance, making them the most suitable choices for a system that aims to classify and predict faults independently of its power mode.

Another key finding is the impact of power mode on diagnostic accuracy. There is a noticeable difference in performance when the system operates under Maximum Power Point Tracking (MPPT) versus Limited Power Point Tracking (LPPT). Previous research [40] has indicated that prediction systems tend to lose accuracy as irradiance decreases. This decline in irradiance reduces the current, which may explain the differences observed in accuracy between models trained on different data types. The MPPT system continuously adjusts the output impedance to match the load, maximizing the current, which may influence the model performance.

It should be noted that the most accurate model differs depending on the power mode. When diagnosing faults under MPPT conditions, the medium neural network performs best, whereas under LPPT conditions, the bagged trees model is the most reliable option. This distinction opens new possibilities for optimizing fault diagnosis. A potential approach for photovoltaic systems operating in both power modes could involve developing an adaptive algorithm capable of switching between models based on the current power mode. Such an algorithm would ensure the best performance in each case, ultimately increasing the overall prediction accuracy.

Furthermore, when analyzing each model’s performance based on specific fault types, it was observed that models do not perform equally across all faults. As seen in the first experiment, neural networks and bagged trees consistently dominated in terms of accuracy. However, certain faults were better predicted by support vector machines (SVM) and K-Nearest Neighbors (KNN). This insight is particularly useful in scenarios where a predominant fault type exists in the photovoltaic system. Instead of using a general-purpose model, a more specialized approach could enhance accuracy by selecting the best model for the dominant fault.

Moreover, the variation in predictive capability among models for different fault types suggests that an advanced predictive system could integrate multiple models operating concurrently. By assigning weighted importance to each model based on its historical performance in diagnosing specific faults, the system could improve overall reliability. However, this approach requires further study and refinement to determine the optimal method for combining multiple prediction models effectively. In the context of increasingly complex energy systems, incorporating additional layers of energy management could further enhance the robustness and applicability of photovoltaic fault-detection methods. For example, recent research [41] introduces a risk-averse energy management model for integrated electricity and heat systems, using intelligent control strategies to optimize system performance under uncertainty. Integrating similar thermodynamic considerations and risk assessment frameworks into photovoltaic system analysis could provide valuable insight for future extensions of this work, particularly when aiming to deploy fault detection within broader smart energy management architectures. Another study [42] analyzes the influence of virtual inertia control on the frequency stability of wind energy systems, emphasizing the importance of adaptive strategies under dynamic and extreme conditions. This approach highlights the potential of fault diagnosis models that can generalize beyond a single technology. Applying similar concepts to photovoltaic systems, by designing classifiers capable of adapting to varying operational contexts, may significantly improve diagnostic reliability and promote broader integration within hybrid or multi-source renewable energy networks.

5. Conclusions

In this paper, several predictive models have been successfully trained and tested using a previously existing database. Comparing them, bagged trees were the most accurate but not far from most of the neural network models. With the development of newer technology and complex computational techniques, a fine-tuned neural network will be the trend to follow for an all-case scenario model, capable of achieving the highest accuracies, independently of the fault.

This study confirms that machine learning can be a powerful tool for predictive maintenance in photovoltaic systems, with the potential to increase system efficiency, reduce operational costs, and extend the lifespan of solar panels. The choice of model depends on the fault type and operating mode, reinforcing the need for flexible and intelligent diagnostic solutions. With further development, machine-learning-based fault-detection systems could become a standard feature in modern photovoltaic farms, ensuring continuous monitoring and optimal performance with minimal human intervention. Machine learning, and AI in general, is a quickly growing area. This growth may change the models and techniques quickly, allowing to create other models or even achieve better performance using the current ones.

Despite the promising results of this study, some limitations should be acknowledged. The dataset used, while extensive, was obtained in a controlled laboratory environment, which may not fully reflect real-world conditions. Additionally, the models’ generalization ability remains uncertain when applied to different climatic conditions or less-frequent fault types. The reliance on inverter-provided data, while cost-effective, may limit early fault-detection capabilities compared to approaches integrating external sensors. Furthermore, the computational cost of training certain models, such as deep neural networks and SVM, poses challenges for real-time applications. Future work should focus on validating these models in real photovoltaic installations, incorporating additional data sources such as thermal imaging and weather conditions, and optimizing computational efficiency. Developing an adaptive hybrid model that dynamically selects the most suitable algorithm based on operational conditions could enhance diagnostic accuracy. Moreover, implementing early fault-detection mechanisms and exploring advanced neural network architectures could further improve predictive maintenance.