Graph Neural Network–Driven Decomposition for Fast Production Cost Minimization Simulation

Production cost minimization (PCM) problem in power system faces critical challenges in achieving rapid solving. The conventional time-domain decomposition (c–TDD) approach is widely used to accelerate the PCM solving process, but often results in significant constraint violations due to fixed decomposition sections at regular intervals. Indeed, the placement of decomposition sections can have a considerable effect on the extent of constraint violations. This paper proposes a graph neural network (GNN)–driven TDD approach to strategically select decomposition sections. In this approach, the yearly PCM problem is represented as graph-structured data. The graph is then processed through gated graph convolutional network (GCN) to obtain enhanced feature representations, followed by a multi-layer fully connected network scoring sections. These scores are used to select decomposition sections that lead to fewer constraint violations. The simulation results demonstrate that the proposed approach achieves 1–2 orders of magnitude faster computation with near-optimal solutions over full-scale PCM approach and up to 3× reduction in constraint violations than c–TDD approach. It also demonstrates robustness and generalization to out-of-distribution (OOD) scenarios.