PV Fault Diagnosis in Large-Scale Power Plants

CIFRE PhD research project on fault diagnosis, monitoring and predictive maintenance for photovoltaic power plants.

Overview

This project corresponds to my doctoral research conducted within a CIFRE collaboration between LAAS-CNRS and Feedgy. The work focused on the development of advanced methodologies for fault detection, diagnosis and performance loss assessment in utility-scale photovoltaic power plants.

The project addressed one of the major challenges faced by photovoltaic operators: identifying faults, degradation mechanisms and performance losses early enough to reduce maintenance costs, increase availability and improve energy production.

The research combined artificial intelligence, signal processing, photovoltaic modeling, feature engineering and diagnosis-oriented instrumentation to create a complete framework for photovoltaic health monitoring and decision support.

Funding

CIFRE PhD Program

Industrial Partner: Feedgy

Host Institution: LAAS-CNRS

Université Toulouse III Paul Sabatier

Scientific Context

The rapid growth of photovoltaic installations worldwide has increased the need for reliable monitoring and diagnostic systems capable of identifying faults before significant production losses occur.

Although conventional monitoring platforms provide large amounts of operational data, they generally remain limited to visualization and basic supervision functions. They rarely provide interpretable fault diagnosis or predictive maintenance capabilities.

Furthermore, several photovoltaic faults exhibit weak electrical signatures and may remain undetected for long periods while continuously degrading plant performance.

The project therefore aimed to develop new diagnostic methodologies capable of operating under real industrial conditions while remaining compatible with large-scale photovoltaic plants.

Main Objectives

The project pursued several scientific objectives:

  • Develop a comprehensive fault diagnosis framework for photovoltaic systems.
  • Build a formal photovoltaic fault dictionary.
  • Design a diagnosis-oriented data acquisition platform.
  • Develop feature engineering methodologies for fault detection.
  • Create machine-learning-based diagnostic algorithms.
  • Improve interpretability of diagnostic decisions.
  • Support predictive maintenance strategies.
  • Quantify performance losses associated with photovoltaic faults.

Methodology

The project combined instrumentation, signal processing, machine learning and photovoltaic system modeling.

Global framework developed during the PhD for photovoltaic fault detection, diagnosis and performance-loss assessment.

Axis 1 – Photovoltaic Fault Knowledge Base

A comprehensive review of photovoltaic faults was conducted to establish a structured fault dictionary.

The work included:

  • Fault classification.
  • Fault occurrence analysis.
  • Performance-loss characterization.
  • Safety-risk assessment.
  • Diagnostic signature identification.

The resulting framework formalized degradation and sudden faults occurring in photovoltaic systems.

Axis 2 – Artificial Intelligence for Fault Diagnosis

Several machine learning methodologies were investigated and evaluated.

The work included:

  • Supervised learning.
  • Unsupervised learning.
  • Semi-supervised learning.
  • Ensemble learning.
  • Hybrid diagnostic approaches.

A large-scale review of artificial intelligence techniques for photovoltaic fault diagnosis was also conducted using bibliometric analysis and topic modeling.

Axis 3 – Diagnosis-Oriented Instrumentation

A dedicated monitoring platform named Solar Vitality was designed and developed.

The platform integrates:

  • Current measurements.
  • Voltage measurements.
  • Irradiance sensors.
  • Ambient temperature sensors.
  • Wind-speed measurements.
  • Embedded data acquisition capabilities.

The platform was specifically designed for photovoltaic diagnosis rather than conventional monitoring.

Axis 4 – Feature Engineering and Signal Analysis

Several signal-processing methodologies were developed to improve fault detectability.

The proposed framework included:

  • Wavelet decomposition.
  • Statistical feature extraction.
  • Feature selection.
  • Dimensionality reduction.
  • Time-series characterization.

These techniques allowed the identification of subtle fault signatures hidden within photovoltaic electrical signals.

Axis 5 – Diagnostic Algorithms

Three major diagnostic methodologies were developed:

EB-Diag

An ensemble-learning diagnostic approach combining:

  • Support Vector Machines.
  • Decision Trees.
  • K-Nearest Neighbors.
  • Majority-voting strategies.

Serial-Diag

A hierarchical diagnostic framework integrating:

  • Dynamic Time Warping.
  • Hierarchical clustering.
  • PLS regression.
  • PLS-LDA classification.

Adaptive-Diag

A diagnosis framework capable of adapting to different photovoltaic plants through:

  • Signal normalization.
  • Knowledge-based modeling.
  • Maintenance prioritization.
  • Plant-to-plant transferability.

Main Scientific Contributions

The project generated several major scientific contributions:

  • Formal photovoltaic fault dictionary.
  • Large-scale AI review for PV diagnosis.
  • Solar Vitality monitoring platform.
  • Diagnosis-oriented instrumentation methodology.
  • Novel feature engineering framework.
  • Ensemble-learning diagnostic architecture.
  • Hybrid diagnostic methodologies.
  • Adaptive diagnosis framework.
  • Predictive maintenance concepts for photovoltaic systems.

Experimental Validation

The methodologies were validated using:

  • Real photovoltaic power plants.
  • Multiple climatic conditions.
  • Different photovoltaic technologies.
  • Utility-scale operating environments.
  • Long-term monitoring campaigns.

Fault scenarios included:

  • Snail trails.
  • Broken glass.
  • Shading.
  • Performance degradation.
  • Multiple health states.

Expected Impact

The project contributes to:

  • Improved photovoltaic reliability.
  • Reduced maintenance costs.
  • Increased energy production.
  • Better understanding of photovoltaic degradation.
  • Development of explainable diagnostic systems.
  • Advancement of predictive maintenance strategies.

The methodologies developed during this PhD constitute the scientific foundation of subsequent projects such as:

  • SOL-DAQ
  • FREE-PV-FIT
  • ALARMES
  • SOL-MIND

Publications

The project generated multiple journal articles, conference papers, software developments and international collaborations in photovoltaic diagnostics and artificial intelligence.

Keywords

Photovoltaic Systems · Fault Diagnosis · Artificial Intelligence · Machine Learning · Predictive Maintenance · Solar Vitality · Feature Engineering · Performance Loss Assessment · Explainable Diagnostics · Renewable Energy

Project Status

Completed (2023)

References