Stochastic Hybrid Approximation Algorithms and Applications to Power Generation Dispatch

In many important applications, including power system operations planning, optimization problems arise where decisions need to be made in the presence of uncertainty. Solving these problems is in general a challenging task due to the computational complexity of evaluating the functions that account for the uncertainty. Typical approaches for solving problems of this type either work with deterministic approximations or with noisy versions of these functions. As a result, they either suffer from an increased problem size or from high susceptibility to noise. In this talk, I present a different approach called stochastic hybrid approximation that aims to reduce these limitations by using manageable deterministic approximations that are gradually updated using noisy observations. I begin by describing the approach, its properties, and its performance on a two-stage stochastic optimal generator dispatch problem. Then, I describe a primal-dual extension of this approach for handling expected value constraints, which are useful for bounding risk. Lastly, I show how the approach can be extended to handle multi-stage stochastic optimization problems, which can capture complex decision-making processes under uncertainty. For both of these extensions, I show experimental results that compare the performance of these approaches against that of widely-used approaches on different versions of the stochastic optimal generator dispatch problem.

Speaker: Tomas Tinoco De Rubira, ETH Zurich