AI Optimizes Grid Dispatch with Decision-Focused Scenario Generation

Power grid dispatch faces increasing uncertainty from renewable energy sources and flexible demand, making robust optimization crucial. Traditional scenario generation for distributionally robust optimization (DRO) often prioritizes statistical accuracy, potentially leading to suboptimal operational decisions. This paper introduces a novel decision-focused generative framework that creates and selects forecast scenarios specifically to minimize downstream operational costs. The framework integrates with mainstream generative models like VAEs, GANs, and diffusion models, capturing complex spatial correlations in uncertainties across the grid. Instead of merely fitting historical distributions, the generated scenarios are optimized to directly improve operational efficiency. To enhance computational tractability, a differentiable scenario selector is also developed, allowing for the selection of only decision-relevant scenarios within the same optimization pipeline. Case studies demonstrate that this approach consistently reduces operational costs by 0.80% to 2.02% compared to methods focused solely on forecast accuracy, offering a more economically efficient and robust grid management solution.