How Far Ahead Must a Storage Fleet See? Certified Value of Foresight and the Limits of Information-Relaxation Bounds
Storage fleets are scheduled by rolling lookahead under forecast errors that grow with lead time. How far is such scheduling from the best achievable, and how much would better foresight at each lead time be worth? The cost under perfect information (PI) answers neither, because it mixes policy suboptimality with the value of information. We answer both with certificates. Block information relaxations, in which the scheduler sees each block of B hours in advance but remains nonanticipative across blocks, form a monotone family between PI and the optimum, and feasible block policies turn them into certified upper bounds on the value of knowing the future beyond B hours. With PI, a strong policy—scenario lookahead with stochastic dual dynamic programming (SDDP) terminal values—sandwiches the optimum. We also show why the lower bound resists tightening at system scale: when renewable availability saturates at zero, every lower bound that is convex in the forecast-error state, SDDP's included, is vacuous, and penalties built from approximate value functions are arbitraged at first order, with losses that grow with the fleet's energy capacity. On a modified RTS-GMLC system with wind capacity equal to peak load, the best policy is within 0.62% of PI. Rolling deterministic lookahead is therefore certified to lose 2.6–3.2% of weekly cost (14–15% with RTS-GMLC-level wind errors), so most of its gap to PI is policy loss, not value of information. Foresight beyond the current day is worth at most 0.11%, and multiplier errors of a few $/MWh erase the entire remaining gap. Preprint (submitted version) of a manuscript submitted to IEEE Transactions on Power Systems. Pages 9–15 contain the supplementary material. Code and stored results: see the related software record.
Authors
- Yuan Shen
Institutions
- University of British Columbia, Okanagan Campus (CA)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24299254
- Primary Topic
- Electric Power System Optimization
- Type
- preprint