Fault Prediction and Early-Detection in Large PV Power Plants Based on Self-Organizing Maps.

Area of Science:

• Renewable Energy Systems
• Data Science and Machine Learning
• Predictive Maintenance

Background:

• Photovoltaic (PV) plants generate critical data via Supervisory Control and Data Acquisition (SCADA) systems.
• Effective fault prediction is essential for optimizing PV plant performance and minimizing downtime.
• Existing methods may lack flexibility or require extensive system-specific tuning.

Purpose of the Study:

• To present a novel, flexible, data-driven solution for generic fault prediction in PV systems.
• To develop an original Key Performance Indicator (KPI) for enhanced fault detection capabilities.
• To validate the proposed model's effectiveness across diverse PV plant and inverter technologies.

Main Methods:

• Utilized a data-driven approach leveraging historical SCADA data.
• Implemented a Self-Organizing Map (SOM) for pattern recognition in operational data.
• Defined a unique Key Performance Indicator (KPI) to quantify system status and predict faults.

Main Results:

• The model accurately predicted incipient generic faults an average of 7 days in advance.
• Achieved a high true positives rate of up to 95% in fault prediction.
• Demonstrated effectiveness across three PV plants (up to 10 MW) and over sixty inverters from three brands.

Conclusions:

• The proposed fault prediction method is effective and reliable for PV systems.
• The model is easily deployable for online anomaly monitoring in new PV plants and technologies.
• Requires only historical SCADA data, fault taxonomy, and inverter datasheets for implementation.