 |
Center for Nonlinear Studies |
Recent work on stochastic control (for e.g. OPF) has handled stochasticity by relying on chance constraints, based on distributional assumptions on the stochasticity, and functional constraints on the control policies available. However the optimal control policy is typically nonlinear, so affine policies that arises from solving conic approximations to the problem are suboptimal to the underlying problem. We provide a different approach by incorporating techniques from machine learning to learn the optimal control policy, and demonstrate empirical improvements over existing methods in reducing total costs, while obeying generation and limit flow limits.
Host: Sidhant Misra/Line Roald