Model-based and Model-free Approaches for Power System Security Assessment

Continuous security assessment of a power system is necessary to insure a reliable, stable, and continuous supply of electrical power to customers. To this end, this dissertation identifies and explores some of the various challenges encountered in the field of power system security assessment. Accordingly, several model-based and/or model-free approaches were developed to overcome these challenges. First, a voltage stability index, named TAVSI, is proposed. This index has three important features: TAVSI applies to general load models including ZIP, exponential, and induction motor loads; TAVSI can be used for both measurement-based and model-based voltage stability assessment; and finally, TAVSI is calculated based on normalized sensitivities which enables identification of weak buses and the definition of a global instability threshold. TAVSI was tested on both the IEEE 14-bus and the 181-bus WECC systems. Results show that TAVSI gives a reliable assessment of system stability. Second, a data-driven and model-based hybrid reinforcement learning approach is proposed for training a control agent to re-dispatch generators’ output power in order to relieve stressed branches. For large power systems, the agent’s action space is highly dimensioned which challenges the successful training of data-driven agents. Therefore, we propose a hybrid approach where model-based actions are utilized to help the agent learn an optimal control policy. The proposed approach was tested and compared to the generic data-driven DDPG-based approach on the IEEE 118-bus system and a larger 2749-bus real-world system. Results show that the hybrid approach performs well for large power systems and that it is superior to the DDPG-based approach. Finally, a Convolutional Neural Network (CNN) based approach is proposed as a faster alternative to the classical AC power flow-based contingency screening. The proposed approach is investigated on both the IEEE 118-bus system and the Texas 2000-bus synthetic system. For such large systems, the implementation of the proposed approach came with several challenges, such as computational burden, learning from imbalanced datasets, and performance evaluation of trained models. Accordingly, this work contributes a set of novel techniques and best practices that enables both efficient and successful implementation of CNN-based multi-contingency classifiers for large power systems