Grid Resilience - Session 4: Data Analytics and Artificial Intelligence for Resilient Grid Operation

The threat of high impact low probability events on power distribution system is substantial but quite unpredictable. Resilient operation of power distribution systems against such events requires solving combinatorial operational problems in stochastic spaces. Since traditional mathematical optimizations struggle with both uncertainty and the curse of dimensionality, data-driven and deep learning techniques are gaining momentum for solving those problems. In this work capabilities of artificial intelligence in making fast, scalable, optimal and reliable restorative and recovery actions are explored. In particular, potential applications of deep reinforcement learning (DRL) in finding near optimal decision trajectories for resilience enhancement of power distribution systems are investigated. Further, pre-training and post-training techniques are investigated to verify the feasibility of deep learning solutions for critical grid operations, such as fault management, reactive power dispatch, and voltage control. These techniques include finding convex feasibility regions within the decision space, as well as hierarchical combination of deep learning with mathematical optimization to utilize the advantages of both. The proposed models are implemented on a cyber-physical test bed that includes a digital real-time power system simulator, to validate the practicality of the proposed solutions.