Incorporating extreme scenarios into multi-stage planning of multi-timescale energy storage

City-scale integrated energy systems (CIESs) couple electricity and gas networks with heterogeneous multi-timescale energy storage (MTES) to balance renewable variability from hourly to seasonal timescales. Conventional representative-scenario methods, however, select days primarily from exogenous source-load profiles and may omit operation-critical conditions whose importance changes as multi-stage investments evolve, while directly solving the full-year electricity-gas mixed-integer linear program is also computationally demanding. This paper therefore develops a model-based extreme scenario aggregation (ESA) framework for multi-stage MTES planning in CIESs. A scenario-driven equivalent MTES model preserves the inter- and intra-scenario states of battery, compressed-air, and gas storage. Moreover, full-year posterior operation simulations provide load shedding, renewable curtailment, and upstream energy purchases for identifying operation-critical days. Based on the operational data, the spectral residual-convolutional neural network adaptively separates extreme and regular days, while a stage influence coefficient (SIC) transfers anticipated later-stage operating stress to earlier-stage scenario updating. Then, sequential rolling and iterative simultaneous aggregation reconstruct both scenario sets within a fixed representative-scenario budget, and iteration stops according to both objective change and scenario-set stability. In the 3-bus-3-gas-node case, rolling ESA reduces the average benchmark-relative cost gap from 13.42% for k-means and 7.51% for k-medoids to 2.41%, while reducing the annual and conditional extreme-period load-shedding ratios from 1.30% and 3.03% to 0.61% and 1.15%, respectively. Iterative ESA with SIC achieves a 2.77% average gap and annual and conditional extreme-period ratios of 0.69% and 1.36% in 10.57 h. The 39-bus-20-gas-node case achieves corresponding values of 2.57%, 0.23%, and 0.95% in 7.23 h, confirming scalability. Multi-day sensitivity analysis further shows that 5-day block provides the most favorable accuracy-efficiency tradeoff. These results show that posterior cross-stage scenario updating improves planning accuracy and operational adequacy while preserving the complementary hourly, cross-day, and seasonal roles of MTES.