Learning-Enhanced Column Generation for Large-Scale Bulk Dispatch Optimization

The Balancing Mechanism plays a critical role in modern electricity markets by ensuring real-time balance between generation and demand through the coordinated dispatch of a massive fleet of Balancing Mechanism Units (BMUs). The resulting Bulk Dispatch Optimization (BDO) problem is a large-scale mixed-integer program with a block-angular structure, naturally amenable to Dantzig--Wolfe decomposition and column generation (CG). However, standard CG can require many iterations due to poor initial columns and greedy column selection that considers only immediate reduced-cost improvement. We propose a learning-enhanced CG framework that addresses both bottlenecks. First, a \emph{column pre-population} module predicts high-quality initial dispatch profiles and projects them onto the corresponding BMU feasible sets to warm-start the Restricted Master Problem (RMP). Second, we formulate iterative \emph{column selection} as a Markov Decision Process over a heterogeneous bipartite representation of the RMP, enabling a reinforcement learning agent to select promising candidate columns for each BMU. Experiments on historical Great Britain Balancing Mechanism data show that the combined framework reduces CG iterations by 21.03\% and total solving time by 15.77\% relative to standard CG.