A comprehensive framework for accurate estimation of performance loss rates in large photovoltaic systems using machine learning

Accurate quantification of long-term Performance Loss Rate in photovoltaic systems is critical for ensuring system reliability, financial forecasting, and asset management across the global PV fleet. Conventional methods for estimating the performance loss rate, however, are often constrained by their sensitivity to environmental variability and reliance on rigid filtering heuristics that can introduce bias. This paper introduces a novel, data-driven framework that transcends these challenges by integrating unsupervised filtering, predictive modeling, and advanced trend analysis. The methodology employs Density-Based Spatial Clustering of Applications with Noise to adaptively isolate anomalous operational data while preserving approximately 80% of the core performance data. Subsequently, a Light Gradient Boosting Machine model, trained on early-life system data, establishes a weather-normalized performance baseline to generate a Performance Ratio Index—a high-fidelity time-series signal representing the system's intrinsic health. Finally, the degradation pathway is characterized via Seasonal-Trend decomposition combined with the Pruned Exact Linear Time algorithm, which robustly identifies change points and non-linear aging phases. The framework was validated across 8 distinct locations comprising 84 inverters, including commercial fleets and authoritative public benchmark datasets from Eurac Research and the FOSS Research Centre. While the broad fleet analysis captured a wide distribution of trend estimates (−4%/year to +3%/year) reflecting the method's sensitivity to data duration and sensor quality, the detailed primary case study demonstrated the framework's high precision, in identifying non-linear, multi-phase degradation. This analysis revealed complex aging dynamics that differed by device, including sharp initial deceleration and instances of mid-life performance acceleration. The resulting degradation rates, with both phase-specific and time-weighted averages ranging from −0.78%/year to −0.20%/year, were found to be physically plausible and consistent with reported industry benchmarks. These findings confirm the framework's utility as a scalable tool for automated performance loss rate assessment that separates non-linear degradation trends from environmental noise.