Value-Oriented Forecast Combinations for Unit Commitment
Mehrnoush Ghazanfariharandi, Robert Mieth
TL;DR
This work tackles value-oriented forecasting for a two-stage $UC$/$RT$ problem under uncertainty in net-load $L$, proposing progressive hedging (PH) to train forecast combinations without altering the underlying UC/RT structure. A push-forward variant (PFPH) further speeds up training by selectively re-solving high-deviation days, enabling at-scale use with full network-constrained unit commitment. Compared to traditional single-level reformulations using KKT conditions or their convex hull relaxations (ST-N/ST-M), PH-based training yields high-quality forecast weights and demonstrably reduces operational costs, achieving up to 1.8% improvements in the tested cases and scalability to a 2736-bus system within about 20 hours of computation. The results highlight a practical pathway to incorporate multiple forecast sources into value-aware decisions and motivate future work on context-aware forecast purchasing and richer forecast models.
Abstract
Value-oriented forecasts for two-stage power system operational problems have been demonstrated to reduce cost, but prove to be computationally challenging for large-scale systems because the underlying optimization problem must be internalized into the forecast model training. Therefore, existing approaches typically scale poorly in the usable training data or require relaxations of the underlying optimization. This paper presents a method for value-oriented forecast combinations using progressive hedging, which unlocks high-fidelity, at-scale models and large-scale datasets in training. We also derive one-shot training model for reference and study how different modifications of the training model impact the solution quality.