Discover then Refine: A Joint Multiple Choice Learning and Flow Matching Framework for Heat Demand Forecasting

District heating optimization requires forecasting models that characterize the inherent multimodality of demand. While generative approaches address this, they typically rely on heavy, undirected sampling and require post-hoc clustering to yield actionable results. We propose a two-stage framework integrating Multiple Choice Learning (MCL) with Flow Matching to generate forecasts with explicit probabilities. Our approach offers three primary benefits: First, MCL identifies diverse hypotheses that guide targeted sampling, ensuring comprehensive coverage of the outcome space, including low-probability events. Second, because these hypotheses act as an aligned prior, the flow model requires fewer integration steps, reducing inference time compared to global generative baselines. Finally, by providing explicit probabilities for each hypothesis, the framework equips operators with a set of weighted scenarios for risk assessment. Evaluated on real-world public data from 3,021 buildings in Aalborg, Denmark, our method achieves competitive accuracy and calibration while allowing for real-time operational decision-making.