Storage in long-term energy planning models with temporal aggregation: Best practices and algorithms

Abstract Representing energy storage in long-term capacity expansion models is challenging because storage dynamics are chronological and highly sensitive to temporal aggregation. This study evaluates four storage aggregation formulations: Niet (direct link), Welsch (fixed-pattern), Kotzur (day-verified cluster), and Novo (hourly-verified cluster). We assess them in three prototypical variable renewable energy (VRE) systems dominated by wind, solar, and hydropower. Using the open-source OSeMOSYS framework [1], we compare a full-year hourly benchmark with 108 reduced-resolution variants. These variants combine 18 sets of representative days (ranging from 2 to 100) with six intra-day block sizes (from 1 to 12 hours). Fidelity is measured by total system cost error, mean absolute error, root mean squared error, Pearson correlation of state-of-charge trajectories, and rainflow cycle counts. The choice of algorithm and the temporal resolution together substantially affect model accuracy. Under coarse aggregation, cost errors exceed 10%, and storage-capacity errors reach several hundred percent for the direct-link and fixed-pattern formulations in the wind-dominated case. For seasonal storage with limited intra-day variability, as in the hydropower case, the lightweight Niet formulation performs as well as the more complex formulations at the lowest computational burden; its accuracy is governed by the number of representative days. For short-duration storage in wind- and solar-dominated systems, Kotzur offers a strong accuracy-to-efficiency trade-off, while Novo achieves the highest fidelity at a substantially higher computational cost. Under a binding storage-energy limit, Kotzur’s reconstructed state of charge leaves the physical bounds in 1.22% (wind) and 5.50% (solar) of hours and lowered its reported system cost by 39.8% and 3.1% relative to Niet. In such cases, Kotzur should be used only after verifying that its reconstructed trajectory satisfies the storage bounds. These recommendations hold for a single-node, single-storage-technology, single-year model class, and their transfer to multi-node, multi-technology or network-constrained systems is untested.