This dissertation was developed within the scope of the Master Course in Electrical and Computer Engineering of the Faculty of Engineering of the University of Porto. It had Vladimiro Miranda as supervisor and Ricardo Bessa as co-supervisor – both researchers at CPES, the first of whom is also member of the Board of INESC TEC.

The work focuses on the analysis of the dynamic behaviour of power systems, namely on the recognition of critical events through the observation of frequency signals captured by PMUs (Phasor Measurement Units) at a rate of 60 phasors per second.

The use of Deep Learning techniques concerning these signals allowed training temporal pattern classification systems in 20-second event windows. In addition to the classifiers based on MLPs (Multilayer Perceptrons) and on DBNs (Deep Belief Networks), one of the innovations consisted on the transformation of temporal signals into 2D images, thus allowing the application of CNNs (Convolutional Neural Networks), whose architecture represents the structure of the human visual cortex. The construction of these 2D images through pixel-intensity association allowed CNNs to obtain the best results: 100% accuracy in the identification of events.

The use of a GPU (Graphics Processing Unit) as calculation hardware was also an innovative strategy, enabling the real-time practical application of the developed technique in electrical grid operation control centres.

APRP aims to promote theoretical and practical advancements in the scientific discipline of pattern recognition. Accordingly, the Best Master’s Thesis Award has the goal of awarding high merit works in the field of Pattern Recognition, encouraging young Portuguese researchers to publish their work.

The award was presented during the closing session of RecPad 2017 – 23rd Portuguese Conference on Pattern Recognition, which took place at the Military Academy in Amadora on the 27th of October.

The INESC TEC researchers mentioned in this article are associated with UP-FEUP and INESC TEC.