An ensemble generative framework for fault diagnosis in data-limited photovoltaic systems

Photovoltaic diagnostic systems frequently encounter the fundamental limitation of representative fault signature scarcity. The acquisition of comprehensive fault datasets is impeded by significant constraints including temporal demands, potential equipment damage, and operator safety considerations. This research presents the Ensemble Particle Swarm Optimization Generative Adversarial Network, an innovative computational architecture engineered for the generation of high-fidelity fault data via an optimized ensemble configuration. The methodology achieves synergistic integration of multiple distinct Generative Adversarial Network variants within a unified paradigm. Outputs are collectively optimized utilizing MOPSO. Implementation incorporates rigorous I-V curve correction protocols and multidimensional feature extraction preprocessing. EPSOGAN leverages MOPSO to ascertain optimal sampling coefficients for constituent GAN based upon four complementary quantitative metrics: Maximum Mean Discrepancy, Coverage, Wasserstein Distance, and Mode Score. Empirical validation confirms that EPSOGAN surpasses individual GAN implementations in synthesizing data possessing superior statistical fidelity while preserving distributional diversity. This enhancement directly translates to improved diagnostic accuracy, achieving up to 99% accuracy.