Fault detection and classification in photovoltaic arrays are important for

improving the reliability, safety, and energy production of photovoltaic systems.

This study presents a two-stage evaluation of machine learning-based PV fault

diagnosis using both simulation-generated and measured experimental data. First,

a 64-module PV array with an 8×8 configuration is modeled in

MATLAB/Simulink under ten normal and faulty operating conditions. Support

Vector Machine and Naive Bayes classifiers are evaluated before and after

Principal Component Analysis preprocessing. PCA-SVM achieves the highest

accuracy of 97.93% in the ten-class simulation study, compared with 86.50% and

43.125% for SVM and Naive Bayes, respectively. Second, external validation is

performed using twelve measured current-voltage curves obtained from six

physical PV panels under clean and partial-shading conditions. The measured

curves are processed into 3,600 observations and evaluated using leave-one-panel

out validation. Six classifier families, including SVM, Naive Bayes, Random

Forest, XGBoost, LightGBM, and MLP are investigated both with and without

PCA. Random Forest achieves the highest experimental-data accuracy of 97.53%.

PCA increases SVM accuracy from 68.14% to 79.69% and XGBoost accuracy

from 83.33% to 90.03%, while its effect is unfavorable for several other classifiers.

The results demonstrate that PCA can improve feature separability, but its

effectiveness is classifier- and dataset-dependent. The external validation confirms

the usefulness of electrical and environmental features for measured partial

shading diagnosis while also highlighting the need for broader field validation.