A broad power-flow test finds physics solvers can match learned models
In plain words
Power-flow calculations describe how electricity moves through a grid under a given pattern of generation and demand. This paper introduces a common test of physics-based solvers and learned models, and a new GPU solver that handles many calculations together.
Across the tested grids and task, a parallel CPU solver or the new GPU solver matched or exceeded the learned models’ speed. An independent check also found the GPU solver’s results met the stated accuracy tolerance after five iterations.
The research
Learned models are often presented as a faster way to calculate power flow, but their speed comparison can depend on which CPU solver is used. Does that advantage remain when parallel CPU and batched GPU physics solvers are included?
The authors compare learned models with CPU and GPU physics solvers on a shared power-flow task, finding parallel physics solvers match or exceed their speed.
The authors compare learned power-flow models with CPU and GPU physics solvers because a model’s apparent speed advantage can depend on its CPU baseline. They build a shared test and introduce gpusim2grid, a GPU solver that batches calculations.
On a fixed-topology task, the tested parallel CPU solver and gpusim2grid matched or exceeded the learned models’ best-observed throughput. An independent KCL check found gpusim2grid’s five-iteration result below the stated mismatch tolerance, supporting accuracy alongside speed.
These results cover the tested graph neural network surrogates, grids, machines, and synthetic data. Throughput is reported as a single best-observed value for each solver and grid; the learned models were not retrained or resized for the larger grids.
Analysis
Deep
- Original paper authors
- Benjamin Donnot, Sami Tazi, Eva Boguslawski, Nourredine Henka
- Submitted to arXiv
- License
- CC BY 4.0