Employing adversarial robustness techniques for large-scale stochastic optimal power flow

Abstract

High penetrations of renewables and extreme weather phenomena are driving the need for large-scale stochastic optimization for power grid operations and planning. However, large-scale stochastic optimization remains computationally challenging due to the wide range of potential probabilistic outcomes and the non-convexity of the AC network constraints. Using a minimax formulation rooted in robust optimization, we define a generic methodology to obtain a feasible dispatch that accommodates the probabilistic nature of loads and renewable generation sources. We introduce a method to efficiently solve this minimax formulation, adopting techniques from the literature on adversarial robustness in machine learning to improve scalability and convergence. We demonstrate that our method maintains AC feasibility over a wide range of probabilistic scenarios, and demonstrate the scalability of our method by determining a dispatch for a synthetic 11,000 bus system.