[Paper Review] Machine Learning for AC Optimal Power Flow

Neel Guha, Zhecheng Wang+2 more|arXiv (Cornell University)|Oct 19, 2019

Electric Power System Optimization11 references59 citations

TL;DR

The paper presents two ML formulations for ACOPF: end-to-end prediction of optimal generator settings and active-constraint prediction to speed up solvers, evaluated on IEEE 30- and 118-bus grids.

ABSTRACT

We explore machine learning methods for AC Optimal Powerflow (ACOPF) - the task of optimizing power generation in a transmission network according while respecting physical and engineering constraints. We present two formulations of ACOPF as a machine learning problem: 1) an end-to-end prediction task where we directly predict the optimal generator settings, and 2) a constraint prediction task where we predict the set of active constraints in the optimal solution. We validate these approaches on two benchmark grids.

Motivation & Objective

• Motivate real-time approximate solutions for ACOPF in power grids with growing demand and renewable variability.
• Explore two ML formulations to handle ACOPF: end-to-end prediction of generator settings and active-constraint prediction to aid warm-starts.
• Assess feasibility and reliability of ML-based ACOPF solutions on benchmark grid topologies.

Proposed method

• Formulate ACOPF with fixed grid topology  N buses, L branches, G controllable generators.
• End-to-end prediction: train a model to map load distributions to optimal generator settings (P^G, V^G) using supervised learning.
• Incorporate soft penalties in the loss to encourage legal solutions that respect constraints (P^G, Q^G, V, delta bounds).
• Optimal constraint prediction: train a model to predict which constraints are active in the optimal solution (multi-label classification).
• Generated dataset by sampling load distributions around the IEEE baselines and solving OPF with Matpower; 90/10 train-test split; cases 30-bus and 118-bus.
• Evaluation metrics include legality rate (for end-to-end) and avg. cost deviation; constraint prediction evaluated by accuracy.

Experimental results

Research questions

• RQ1 Can ML models accurately predict optimal generator settings from load distributions for ACOPF?
• RQ2 Can ML models reliably identify the active constraint set to enable faster warm-starts for traditional solvers?
• RQ3 What are the legality rates and cost deviations of ML-predicted solutions on standard IEEE test grids?
• RQ4 How does model performance differ between smaller (30-bus) and larger (118-bus) grids?

Key findings

• End-to-end prediction achieves legality rates of 0.51 (case30) and 0.70 (case118) with average cost deviation around 0.002 for legal grids.
• Constraint prediction achieves high accuracy overall, 0.99 for case30 and 0.81 for case118.
• Optimal-cost predictions on legal grids are within about 1% of the true optimal cost.
• The best end-to-end models varied by grid: case30 used a two-layer tanh network without a constraint loss; case118 used a three-layer network with 512 units and a constraint loss penalty.
• Learning active constraint sets is generally easier and provides potential speedups by informing solver warm-starts.