Priya L. Donti
LightNews LightNews
priyald17.bsky.social
Priya L. Donti
@priyald17.bsky.social
150 followers 350 following 19 posts
Assistant Professor, MIT | Co-founder & Chair, Climate Change AI | MIT TR35, TIME100 AI | she/they
Posts Replies Media Videos
priyald17.bsky.social
PF provides (ctd.):

* Novel data generation pipeline implemented in Julia that builds on OPFlearn.

* User-friendly PyTorch InMemoryDataset class

5/
Data generation workflow for PF∆. * Step 1: Sample load from presumed-feasible load convex set, create component outage, permute generator costs * Step 2: Run modified ACOPF with limits on PF output variables removed * Step 3: Store sample if converged. Else, adjust presumed-feasible load convex set, and re-sample.
November 17, 2025 at 5:42 PM
priyald17.bsky.social
PFΔ provides (ctd.):

* Open-source PyTorch implementation of CANOS (originally developed for the optimal power flow problem, but now adapted for PF)

* Evaluations of several state-of-the-art models, including CANOS-PF, PFNet, and GraphNeuralSolver

4/
Plot of experimental results for all selected tasks
November 17, 2025 at 5:42 PM
priyald17.bsky.social
PFΔ provides:

* 859,800 solved PF instances spanning 6 power system sizes and incl. N, N-1, & N-2 contingencies, alongside multiple evaluation tasks

* Close-to-infeasible cases near steady-state voltage stability limits that enable stress-testing of ML models under edge-cases

3/
Tasks in PF∆ benchmark. * 1.1: Unperturbed training topology * 1.2: N-1 perturbed training topology * 1.3: N-2 perturbed training topology * 2.1: Low data efficiency (same set-up as 1.3) * 2.2: Medium data efficiency * 2.3: High data efficiency * 3.1: Fixed training grid size * 3.2: Small grid size training group * 3.3: Large grid size training group * 4.1: Training with hard power flow cases * 4.2: Training with augmented hard power flow cases * 4.3: Training only with hard power flow cases
November 17, 2025 at 5:42 PM
priyald17.bsky.social
⚡Excited to share our work "PFΔ: A Benchmark Dataset for Power Flow under Load, Generation, & Topology Variations," to be presented as part of the #NeurIPS 2025 Datasets & Benchmarks Track

Led by my students Ana K. Rivera, Anvita Bhagavathula, & Alvaro Carbonero

1/
Title, author list, and abstract for PF∆. Title: PFΔ: A Benchmark Dataset for Power Flow under Load, Generation, and Topology Variations Authors: Ana K. Rivera, Anvita Bhagavathula, Alvaro Carbonero, Priya Donti Abstract at: https://arxiv.org/abs/2510.22048
November 17, 2025 at 5:42 PM
priyald17.bsky.social
Beyond humbled to be on this year's #TIME100AI.

AI can be an asset for climate & energy -- but only if its development is guided by actual climate needs & planetary limits. Shoutout to those in the community working to shape a responsible, equitable, climate-aligned AI future 🌍💪
Priya Donti's picture in the TIME100 AI frame
August 28, 2025 at 12:50 PM