Reliable Few-Shot Fault Diagnosis for Grid-Connected PV Systems: When Real-Only Training Outperforms Generative Synthesis

Abstract

We identify a reliable, low-compute few-shot diagnostic pipeline for grid-connected photovoltaic (PV) systems and quantify how far six modern generative synthesis methods fall short of it when their output is used as the classifier’s only training data. Training a 200-tree Random Forest directly on as few as five real waveform windows per fault class—what we call k-shot real-only training (k-shot TRTR)—achieves macro-F1 = 0.38 at k = 5 and 0.74 at k = 50 on the public GPVS-Faults benchmark, runs in seconds, and the entire audit executes in a few hours on a single consumer GPU. We benchmark this baseline against six generative methods—four conditional GAN baselines (cGAN, WGAN-GP, AC-GAN, FaultGAN) and two diffusion-prior hybrids (DP-cGAN, DP-AC-GAN)—plus a few-shot diffusion-only DDPM reference, across 376 trained pipelines (282 generative and 94 real-only reference runs across two datasets), three downstream classifiers (Random Forest, 1D-CNN, bidirectional LSTM), eight evaluation metrics, four few-shot regimes k ∈ {5, 10, 20, 50}, and five random seeds for the main matrix, 10 for the real-only baselines and three for selected diagnostic experiments, with 95% confidence intervals, under a train-on-synthetic, test-on-real (TSTR) protocol. None of the generative pipelines exceeds the real-only baseline at any k on Random Forest, and the same ordering replicates on a second, independently collected outdoor field dataset under episode-level splitting; a cross-operating-mode stress test within GPVS-Faults further shows that at few-shot scale every pipeline, real-only and generative alike, additionally depends on condition-consistent sampling. We attribute the generative failure to PV waveform complexity—50 Hz grid harmonics, submillisecond inverter switching transients, and slow insolation drift coexisting on a 13-channel × 512-sample window—which the generators answer with severe mode collapse (intraclass diversity at 1%–7% of real for the BCE-family models) and with fabricated high-frequency content (three to four orders of magnitude above ground truth in the inverter-side current band, inflating total harmonic distortion by more than an order of magnitude). A low-pass intervention shows that removing the fabricated high-frequency band alone does not restore TSTR performance; together with the diversity analyses, this is consistent with mode collapse being a major remaining limitation. Per-class analysis identifies one seed-robust, physically interpretable specialized win that may inform future hybrid designs. We recommend k-shot TRTR as a strong, low-compute default for sample-scarce PV diagnostic settings and as a routine sanity baseline in future few-shot fault-synthesis studies. Every result reported here is obtained under the synthetic-only TSTR protocol; real-plus-synthetic augmentation, in which the k real windows are retained alongside the synthetic ones, is a distinct question lying outside the scope of this audit.