Electricity is a critical component in our daily life. Because it is almost always available, we take it for granted. However, given the proper conditions, blackouts do happen every once in a while and can cause discomfort at a minimum, and a catastrophe in rare circumstances. The largest blackouts typically include cascading failures, which are sequences of interdependent outages. Although timely and effective operator intervention can often prevent a cascade from spreading, such interventions require ample situational awareness.;The goals of this dissertation are twofold: to provide power system operators with insight into the risk of blackouts given the space of potential initiating outages, and to evaluate control systems that might mitigate cascading failure risk. Accordingly, this dissertation proposes a novel method to estimate cascading failure risk. It is shown that this method is at least two orders of magnitude faster in estimating risk, compared with a traditional Monte-Carlo simulation in two test systems including a large-scale real power grid model. This method allows one to find critical components in a system and suggests ideas for how to reduce blackout risk by preventive measures, such as adjusting initial dispatch of a system.;In addition to preventive measures, it is also possible to use corrective control strategies to reduce blackout sizes. These methods could be used once the system is under stress (for example if some of the elements are overloaded) to stop a potential cascade before it unfolds. This dissertation focuses on a distributed receding horizon model predictive control strategy to mitigate overloads in a system, in which each node can only control other nodes in its local neighborhood. A distributed approach not only needs less communication and computation, but is also a more natural fit with modern power system operations, in which many control centers manage disjoint regional networks. In addition, a distributed controller may be more robust to random failures and attacks. A central controller benefits from perfect information, and thus provides the optimal solution. This dissertation shows that as long as the local neighborhood of the distributed method is large enough, distributed control can provide high quality solutions that are similar to what an omniscient centralized controller could achieve, but with less communication requirements (per node), relative to the centralized approach.
展开▼
