Synthetic data-based machine learning applications for 21st-century power systems

Abstract

Driven by climate change, power system engineers are developing solutions to integrate renewable energies. Such modernization of electrical networks requires novel technologies to increase the efficiency of the electrical generation processes, making them greener and more sustainable. Developing new algorithms and methods is not agnostic to the implosion and success of data-driven strategies in science and engineering. This thesis collects different innovations to facilitate the development of new data-driven algorithms by exploiting physics-based modeling, simulation technologies, and modern computing technologies. The resulting software tool uses models built with the Modelica language and the Python programming language. First, a pipeline for synthetic data generation for electric power transmission systems is described. Deep insight is given into the structure of the different modules and the rationale behind their implementation. Next, the developed tool is used for the implementation of data-driven algorithms. Two case studies corresponding to relevant applications for power systems are presented: small-signal stability assessment and forced oscillation detection. For applications that will require the execution of Machine Learning algorithms at the edge, a low-cost hardware platform is introduced for oscillation detection, which has promising potential for education and research. Finally, to better exploit the embedded physics in electric power transmission systems, a novel recurrent neural network architecture inspired by dynamical systems is presented. Such physics-aware solution promises to play an important role in devising data-driven solutions required for the operation, planning, and control of 21st-century power systems.