PhD defence of Tayeb Meridji – Power System Stability Assessment Frameworks Using Machine-Learning Techniques

Abstract

The widespread adoption of renewable energy is causing unpredictability in power networks, making it hard to predict critical operating points. This new reality challenges traditional stability assessment methods still used by most grid operators.

It becomes essential to adopt a systematic approach that encompasses all hourly operating points over the course of a study year. This thesis introduces novel assessment frameworks that leverage machine learning techniques to allow rapid, deterministic time-series assessments of angular transient stability in the context of high renewable penetration. The frameworks not only offer an evaluation of the transient stability of the grid at a high level but also provide the possibility of performing meticulous analyses of emerging trends in the dynamic responses of individual synchronous generators within systems experiencing reduced inertia.

The heavy computational burden associated with such time-series stability assessments are substantially reduced through the strategic use of supervised and unsupervised learning algorithms. A modified version of the Affinity Propagation clustering algorithm is proposed to cluster the subset of all operating points of a given study year and derive a representative subset of these points. In addition, Gradient Boosting Regressors are also used to predict transient stability indices for all hours of the studied year. And finally, an agglomerative hierarchical clustering algorithm is proposed to cluster synchronous generators based on their dynamic response.

The proposed frameworks are demonstrated to be ideal for grid planners in identifying pathways to achieve reliable integration of renewable energy resources. Through a series of case studies, these frameworks were evaluated to determine the transient stability performance of an IEEE-39 test system augmented with renewable energy resources.