Data-augmentation acceleration framework by graph neural network for near-optimal unit commitment

Lishen Wei

doi:10.1016/j.apenergy.2024.124332

Return

Data-augmentation acceleration framework by graph neural network for near-optimal unit commitment

2025-01-01

PRE

DOI：10.1016/j.apenergy.2024.124332

Original

Original request for help

Save

Abstract

En 中文

The acceleration of large-scale unit commitment (UC) problems has been a long-standing challenge in the power industry. Independent system operators are required to find a near-optimal solution within a short timeframe, which is a challenging task in practice. In this paper, to obtain near-optimal commitments rapidly, we develop a data-augmentation acceleration framework by neural network. First, a data augmentation technique is developed to enrich the training dataset with numerous near-optimal solutions instead of just incorporating the optimal one. Secondly, we employ the graph neural network (GNN) to learn the commitments, where the UC problems are treated as graphs to represent the relationship between variables and constraints. Last, a repair method is introduced to address potential inaccuracies caused by GNN, converting infeasible or suboptimal solutions into near-optimal ones. The proposed framework was validated on the RTS-GMLC and Jiangsu province systems. Numerical experiments with the Gurobi solver demonstrated a remarkable acceleration in finding near-optimal solutions, achieving speed-ups of 10.7x and 15.9x, respectively. The quickly obtained near-optimal solutions also greatly aid in finding optimal solutions. Furthermore, integrating the proposed framework into the opensource solver significantly outperformed the Gurobi solver in finding near-optimal commitments.

Keywords:

Unit commitment (UC)

Mixed-integer linear problem (MILP)

Machine learning (ML)

Near-optimal solutions