[Paper Review] Confidence-Aware Graph Neural Networks for Learning Reliability Assessment Commitments

Seonho Park, Wenbo Chen+3 more|arXiv (Cornell University)|Nov 28, 2022

Power System Reliability and Maintenance4 citations

TL;DR

This paper proposes RACLearn, a confidence-aware graph neural network framework that accelerates Reliability Assessment Commitment (RAC) optimization in power systems by predicting generator commitments and active line constraints with uncertainty quantification. By selecting high-confidence predictions, applying a polynomial-time feasibility restoration, and seeding a MIP solver, RACLearn achieves 2x–4x speedups with negligible optimality gap loss on real-world transmission networks including MISO’s DA-FRAC and LAC systems.

ABSTRACT

Reliability Assessment Commitment (RAC) Optimization is increasingly important in grid operations due to larger shares of renewable generations in the generation mix and increased prediction errors. Independent System Operators (ISOs) also aim at using finer time granularities, longer time horizons, and possibly stochastic formulations for additional economic and reliability benefits. The goal of this paper is to address the computational challenges arising in extending the scope of RAC formulations. It presents RACLearn that (1) uses a Graph Neural Network (GNN) based architecture to predict generator commitments and active line constraints, (2) associates a confidence value to each commitment prediction, (3) selects a subset of the high-confidence predictions, which are (4) repaired for feasibility, and (5) seeds a state-of-the-art optimization algorithm with feasible predictions and active constraints. Experimental results on exact RAC formulations used by the Midcontinent Independent System Operator (MISO) and an actual transmission network (8965 transmission lines, 6708 buses, 1890 generators, and 6262 load units) show that the RACLearn framework can speed up RAC optimization by factors ranging from 2 to 4 with negligible loss in solution quality.

Motivation & Objective

• Address the growing computational burden of Reliability Assessment Commitment (RAC) formulations in power systems due to increased renewable integration and longer time horizons.
• Overcome the limitations of traditional MILP solvers in handling large-scale, time-sensitive RAC problems with high forecast uncertainty.
• Leverage structural and temporal correlations in historical RAC instances to train a machine learning model that accelerates optimization.
• Develop a framework that combines predictive modeling with uncertainty quantification and feasibility restoration to ensure high-quality, feasible solutions.
• Enable faster decision-making for Independent System Operators (ISOs) by reducing RAC solving times while maintaining reliability and economic performance.

Proposed method

• Design a novel GNN-based architecture with encoder-decoder layers to jointly predict generator commitments and active transmission line constraints from power network topology and forecast data.
• Integrate Monte Carlo dropout (MC-dropout) into the GNN to estimate epistemic uncertainty and assign confidence scores to each commitment prediction.
• Select a subset of high-confidence predictions (e.g., top 30–100%) to fix in the optimization problem, reducing the search space for the MIP solver.
• Apply a polynomial-time feasibility restoration procedure to transform potentially infeasible solutions arising from fixed predictions into feasible ones with minimal optimality gap increase.
• Seed a state-of-the-art MIP solver with the repaired feasible solution and predicted active constraints to accelerate convergence.
• Train the entire RACLearn framework end-to-end using historical RAC instances to minimize optimality gap while maximizing speedup.

Experimental results

Research questions

• RQ1 Can a GNN-based model effectively predict generator commitments and active line constraints in RAC problems with high accuracy and low computational cost?
• RQ2 How well do confidence scores derived from MC-dropout correlate with prediction accuracy in real-world RAC scenarios?
• RQ3 To what extent can fixing high-confidence predictions accelerate MIP solving without degrading solution quality?
• RQ4 Can a polynomial-time feasibility restoration procedure reliably convert infeasible solutions into high-quality feasible ones after prediction fixation?
• RQ5 What is the trade-off between solution quality and speedup when varying the number of fixed commitments in the RACLearn framework?

Key findings

• RACLearn achieved speedups of 2x to 4x on the MISO DA-FRAC and LAC formulations, with the highest speedup (4x) observed when allowing a 0.77% optimality gap loss.
• For LAC, fixing up to 100% of high-confidence commitments resulted in only a 0.09% optimality gap increase and a 50% reduction in CPU time.
• The feasibility restoration step was critical: without it, only 31.17% of DA-FRAC instances remained feasible when all commitments were fixed.
• The confidence scores were strongly indicative of prediction accuracy, with 100% of high-confidence predictions being correct in the LAC case when 90% were fixed.
• The GNN-based architecture outperformed both deep and shallow baselines in terms of prediction accuracy and parameter efficiency, enabling scalability to large networks (e.g., 8965 lines, 6708 buses).
• The framework maintained near-optimal solution quality across diverse scenarios, with geometric mean optimality gaps within 0.5% of the baseline for most configurations.