Photovoltaic (PV) inverter fault prediction is critical for maintaining system reliability and minimizing energy loss. While recent studies have improved predictive accuracy using data-driven approaches, most evaluations remain focused on offline settings and do not address how probabilistic predictions are translated into operational decisions. This study investigates multi-horizon fault prediction for PV inverters under real-world constraints, with a particular focus on decision-level behavior. A modular prediction framework is implemented by combining transformer-based TimeXer embeddings with probabilistic classification using XGBoost. The model operates on sliding-window sensor data and produces fault probabilities across multiple future horizons. To support operational use, these probabilities are aggregated into a single risk score, and threshold-based alarm policies are evaluated through a systematic threshold sweep. The results show that predictive performance varies across horizons, with usable lead-time information concentrated in near-term predictions. Under severe class imbalance, imbalance-aware training significantly improves detection performance in precision–recall space, but performance remains sensitive to temporal variation. Most importantly, the threshold-sweep analysis reveals a structural trade-off between detection performance and alarm burden, where achieving moderate early-warning capability requires substantially increased alarm rates. These findings indicate that improving predictive accuracy alone is insufficient for practical deployment. Instead, decision-level behavior must be explicitly considered when designing predictive maintenance systems under operational constraints.
Kim et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: