Solar farm voltage anomaly detection using high-resolution μPMU data-driven unsupervised machine learning

The usual means of solar farm condition monitoring are limited by the typically poor quality and low resolution data collected. A micro-synchrophasor measurement unit has been adapted and integrated with a power quality monitor to provide the high-resolution, high-precision, synchronised time-series data required by analysts to significantly improve solar farm performance and to better understand their impact on distribution grid behaviour. Improved renewable energy generation at large solar photovoltaic facilities can be realised by processing the enormous amounts of high-quality data using machine learning methods for automatic fault detection, situational awareness, event forecasting, operational tuning, and planning condition-based maintenance. The limited availability of existent data knowledge in this sector and legacy performance issues steered our exploration of machine learning based approaches to the unsupervised direction. A novel application of the Clustering Large Applications (CLARA) algorithm was employed to categorise events from the large datasets collected. CLARA has been adapted to recognise solar site specific behaviour patterns, abnormal voltage dip and spike events using the multiple data streams collected at two utility-scale solar power generation sites in England. Fourteen days of empirical field data (seven consecutive summer days plus seven consecutive winter days) enabled this analytical research and development approach. Altogether, similar to 725 million voltage measurement data points were investigated, and automatic voltage anomaly detection demonstrated.

Automated summary

The integration of a micro-synchrophasor measurement unit with a power quality monitor provides analysts with high-resolution, high-precision, synchronized time-series data to improve solar farm performance and understand their impact on the distribution grid behavior. Machine learning methods can be utilized to process high-quality data for automatic fault detection, situational awareness, event forecasting, operational tuning, and condition-based maintenance planning in large solar photovoltaic facilities. The use of the Clustering Large Applications (CLARA) algorithm for event categorization and anomaly detection in empirical field data from utility-scale solar power generation sites in England demonstrated the potential of unsupervised machine learning approaches in this sector.