AIoT Seminar | Learning-augmented Decision-making and Control: Theory and Applications in Smart Grids

Recently, augmenting classical methods in real-world cyber-physical systems such as smart grids with black-box AI tools, forecasts and ML algorithms attracts growing interests. Integrating AI techniques into smart grids, on the one hand, provides a new approach to handle uncertainties caused by renewable resources and human behaviors, but on the other hand, creates practical issues such as reliability, privacy, and scalability, etc. to the AI-integrated algorithms. 

In this talk, I will present novel problems raised in designing learning-augmented decision-making algorithms. First, I will introduce a problem in linear quadratic control, where untrusted AI predictions of system perturbations are available. We show that it is possible to design an online algorithm with performance guarantees even if the prediction error goes to infinity. Second, I will introduce a scalable learning-based MPC scheme, which is proven to have guaranteed performance in a two-controller problem, with a central controller decides aggregate actions and a local controller disaggregates the actions and observes high-dimensional states. The two controllers minimize a total cost coordinatively using aggregated low-dimensional feedback that can be learned by deep RL algorithms. I will show how these problems relate to practical applications in smart grids and demonstrate that the learning-augmented methods help improve previous classical results. Finally, I will discuss interesting potential research directions down the road.