@article{LUO2026, 
author = {Shoukang LUO and Kehuan WEN and Yubo ZHANG and Xiaojia SUN and Yang LI and Wenjing SUN and Tianfeng LI},
title = {A Random Forest Fault Diagnosis Method for Photovoltaic Arrays Based on Multi-feature Optimization Mechanism and OOB Optimization},
year = {2026},
journal = {Distributed Energy},
volume = {11},
number = {5},
pages = {120-128},
keywords = {photovoltaic array, fault diagnosis, random forest (RF) algorithm, out-of-bag (OOB) error, multi-feature optimization},
url = {https://www.sciopen.com/article/10.16513/j.2096-2185.DE.26100426},
doi = {10.16513/j.2096-2185.DE.26100426},
abstract = {To address the frequent faults in distributed photovoltaic (PV) arrays and the limitations of traditional random forest (RF) methods in diagnostic accuracy and generalization capability, this paper proposes a fault diagnosis method based on RF that integrates multi-feature optimization and dynamic out-of-bag (OOB) error optimization. First, a PV array simulation model is constructed in MATLAB/Simulink based on the single-diode equivalent circuit of a PV cell. The model simulates five typical fault types, including short circuit, open circuit, partial shading, aging degradation, and bypass diode failure, and extracts multi-dimensional fault features covering voltage, current, and power etc. Furthermore, the OOB error is introduced as a dynamic evaluation metric during model training to monitor and optimize the ensemble strategy of base learners in real time, thereby improving convergence performance and generalization capability. Simultaneously, a two-stage multi-feature optimization strategy—“permutation importance preliminary screening + principal component analysis secondary dimensionality reduction”—is adopted to screen the most critical fault feature subset from the original high-dimensional features, enhancing model robustness. Experimental results demonstrate that the proposed OOB-RF model achieves a fault diagnosis accuracy of 95.44% on the test set, which is 8.12 percentage points higher than that of the traditional RF model, while the overfitting index is significantly reduced by 51.26%. The proposed method exhibits superior classification accuracy and stability, providing an effective solution for the intelligent operation and maintenance of distributed PV systems.}
}