Core Concepts
GraPhyR, a physics-informed graph neural network framework, can learn to optimize dynamic reconfiguration of power distribution grids to minimize grid losses and satisfy operational constraints.
Abstract
The paper proposes GraPhyR, a physics-informed graph neural network (GNN) framework for dynamic reconfiguration (DyR) of power distribution grids. DyR involves optimizing the grid topology by selecting the open/closed status of switches to minimize power losses while satisfying operational constraints.
The key components of GraPhyR include:
- Modeling switches as gates in the GNN message passing layers to control information flow and represent the physics of power flow through switches.
- Using local predictors to make scalable predictions of power flows, voltages, and switch statuses, rather than a global predictor.
- Embedding the discrete open/closed switch decisions directly within the neural framework using a physics-informed rounding layer.
- Taking the grid topology as an input to the GNN, allowing the framework to adapt to changing grid conditions.
The authors demonstrate that GraPhyR outperforms prior methods in learning to predict near-optimal and feasible solutions for the DyR problem. It also shows the ability to adapt to unseen grid conditions, such as switch failures or maintenance, without retraining. The results highlight the benefits of incorporating domain knowledge and physical constraints directly into the neural architecture.
Stats
The grid topology is described by the number of nodes N, lines M, and switches Msw.
The objective function linearly approximates electric losses as the sum of squared real and reactive power flows.
Power balance, Ohm's law, and switch constraints are enforced as equality and inequality constraints.
Quotes
"To maintain a reliable grid we need fast decision-making algorithms for complex problems like Dynamic Reconfiguration (DyR)."
"The DyR problem is a mixed integer program (MIP) due to the discrete nature of switch decisions. It is well known that MIPs are NP-hard (i.e. cannot be solved in polynomial time) and thus may be computationally intractable for large-scale problems."