AI-Powered Dynamic Fault Detection and Performance Assessment in Photovoltaic Systems

TL;DR

This work addresses the challenge of intermittent PV performance by developing a PVlib-based computational model that performs dynamic loss quantification and AI-driven fault detection using synthetic, weather-consistent data at 5-minute resolution. The approach combines a five-parameter single-diode PV model with a Sandia/NOCT-inspired inverter model, dynamically estimating losses from soiling, degradation, wiring, and inverter performance, and couples this with a PyTorch-based ANN to predict technical parameters without specialized sensors. The framework achieves a daily energy MAE of $6.0\%$, fault-detection mean accuracy of $82.2\%$ (max $92.6\%$), and competitive energy prediction against PVlib, PVWatts, and PVsyst, while enabling threshold-based fault detection robust to dynamic meteorology. The results demonstrate the value of dynamic, data-driven modeling for PV performance assessment and fault detection, with potential for real-time integration and improved decision-making in PV operations and maintenance.

Abstract

The intermittent nature of photovoltaic (PV) solar energy, driven by variable weather, leads to power losses of 10-70% and an average energy production decrease of 25%. Accurate loss characterization and fault detection are crucial for reliable PV system performance and efficiency, integrating this data into control signal monitoring systems. Computational modeling of PV systems supports technological, economic, and performance analyses, but current models are often rigid, limiting advanced performance optimization and innovation. Conventional fault detection strategies are costly and often yield unreliable results due to complex data signal profiles. Artificial intelligence (AI), especially machine learning algorithms, offers improved fault detection by analyzing relationships between input parameters (e.g., meteorological and electrical) and output metrics (e.g., production). Once trained, these models can effectively identify faults by detecting deviations from expected performance. This research presents a computational model using the PVlib library in Python, incorporating a dynamic loss quantification algorithm that processes meteorological, operational, and technical data. An artificial neural network (ANN) trained on synthetic datasets with a five-minute resolution simulates real-world PV system faults. A dynamic threshold definition for fault detection is based on historical data from a PV system at Universidad de los Andes. Key contributions include: (i) a PV system model with a mean absolute error of 6.0% in daily energy estimation; (ii) dynamic loss quantification without specialized equipment; (iii) an AI-based algorithm for technical parameter estimation, avoiding special monitoring devices; and (iv) a fault detection model achieving 82.2% mean accuracy and 92.6% maximum accuracy.

Figures (26)

• Figure 1: PV system at Universidad de los Andes. Taken from ref:C1R62.
• Figure 2: Modeling of a PV system according to PVPMC ref:C2R12ref:C2R20. The grey arrows indicates the technical and electrical specifications; in yellow the effective irradiance modeling; in green the estimation of the PV panel temperature; in purple the DC production; and in blue the AC production.
• Figure 3: Behavior of the defined performance index $PM_{norm}$. Blue represents System A and green represents System B. The color convention is used throughout the paper.
• Figure 4: Detection of cleaning events for System A (top) and System B (bottom). Each red line represents a cleaning event; the period between events is referred to as soiling interval. In orange, the Monte Carlo simulation for stochastic generation of possible soiling profiles.
• Figure 5: Comparison between $G_{POA}$ in blue and $G_{cs}$ in orange.