Leveraging random forest cross-validation and k-means algorithms with optimal feature extraction combinations for photovoltaic fault diagnosis and classification / Siti Nor Azlina Bt. Mohd Ghazali

Faculty of Electrical & Electronics Engineering Technology

Thesis (Doctor of Philosophy) -- Universiti Malaysia Pahang – 2025

Includes bibliographical references

Uneven solar radiation distribution poses significant challenges for fault detection in photovoltaic (PV) systems using conventional protection devices (CPD). CPD detects only large current faults, leaving undetected faults reduces system output and poses safety risks. Recent studies have focused on advanced fault detection techniques such as power loss and I-V curves (PIV), machine learning (ML) algorithms, output signal analysis techniques (OST), and statistically-based techniques (SBT). However, these methods have limitations, such as high costs, the need for precise threshold settings, and a lack of validation on multi-scale PV arrays. Diagnosing PV faults becomes more challenging, especially for larger PV systems. This research proposes a multi-scale equation (MsE) for designing various sizes of PV array models using MATLAB/Simulink and aims to develop a smart PV fault diagnosis and classification (SFDC) model for ground fault (GF), open circuit fault (OCF), line-line fault (LLF), and partial shading (PS). The SFDC model employs Random Forest with Cross-Validation (RF-CV) algorithm. The Random Forest (RF) is a powerful ML technique known for its ability to handle complex and time-series datasets. Cross-validation (CV) is employed to assess the model's performance and ensure its robustness. In addition, optimal feature extraction (FE) combinations have been determined by selecting the most relevant features, and have proven enhanced the fault classification accuracy. The proposed multi-scale equations, with minimal tuning, successfully developed small, medium, and large-scale PV array models. Using real irradiance data as input to generate simulated data, the simulation results for Pmax, VOC, and ISC matched the detailed specifications of the Panasonic VBMS250AE04 datasheet, confirming the accuracy of the SFDC model under both normal and faulty conditions. Moreover, the optimal FE combination of Root Mean Square (RMS), Skewness (Skew), Maximum Power (Pmax), Mean (M), Standard Deviation (STD), Waveform length (WL), and Autoregressive (AR) significantly improved fault classification accuracy. Three major contributions have been achieved in this research. First, new multi-scale equations developed reduced simulation and processing times without sacrificing accuracy, making the model adaptable to real-world PV applications. Second, optimal feature extraction combinations identified have significantly improved PV fault classification accuracy. The SFDC model achieved 100% accuracy during training for GF, OCF, LLF, and PS across all PV array models. In testing, nearly 100% accuracy was attained for OCF and PS, 99.6% for GF, and 96.9% for LLF. Lastly, the SFDC model's performance was validated through a comparison of fault classification accuracies using RF-CV, KNN, and SVM algorithms. RF-CV demonstrated good overall accuracy across all scales. Although the RF-CV model's average accuracy (99.0%) was slightly lower than in previous studies (99.4% and 99.5%), it remained surpass and reliable. While validation using the real PV, data demonstrates its dependability in real-world application. Ultimately, this research can serve as a reference that contributes to effective corrective and predictive maintenance strategies for PV systems.