Constraints on OPF Surrogates for Learning Stable Local Volt/Var Controllers
arXiv:2306.04464 · doi:10.1109/LCSYS.2023.3287985
Abstract
We consider the problem of learning local Volt/Var controllers in distribution grids (DGs). Our approach starts from learning separable surrogates that take both local voltages and reactive powers as arguments and predict the reactive power setpoints that approximate optimal power flow (OPF) solutions. We propose an incremental control algorithm and identify two different sets of slope conditions on the local surrogates such that the network is collectively steered toward desired configurations asymptotically. Our results reveal the trade-offs between each set of conditions, with coupled voltage-power slope constraints allowing an arbitrary shape of surrogate functions but risking limitations on exploiting generation capabilities, and reactive power slope constraints taking full advantage of generation capabilities but constraining the shape of surrogate functions. AC power flow simulations on the IEEE 37-bus feeder illustrate their guaranteed stability properties and respective advantages in two DG scenarios.
References in corpus (8)
- Adam: A Method for Stochastic Optimization
- Convex Relaxation of Optimal Power Flow, Part I: Formulations and Equivalence
- Data-driven Local Control Design using Optimization and Machine Learning Techniques
- Optimal Distributed Feedback Voltage Control under Limited Reactive Power
- Lyapunov Theory for Discrete Time Systems
- Reverse and Forward Engineering of Local Voltage Control in Distribution Networks
- On the Properties of the Compound Nodal Admittance Matrix of Polyphase Power Systems
- Learning Provably Stable Local Volt/Var Controllers for Efficient Network Operation