Reconciling Economy, Carbon and Resilience in Data-Center Integrated Energy Systems: A Normal-Stress Separated Planning Framework with Risk-Aware Scenarios

Abstract

: Data-center integrated energy systems require coordinated planning of cost, carbon and reliability, yet conventional methods often evaluate reliability under the same grid-connected dispatch used for annual accounting, causing risk objectives to collapse to zero. This paper proposes a risk-aware bi-level planning framework with separated normal and stress dispatch. In the scenario stage, an interpretable electricity-cooling model and a leakage-controlled random forest identify propagation sensitivity and abnormal cooling residuals. The prior-risk components are aggregated and reduced by category-preserving clustering, retaining risk representatives alongside typical days. In the planning stage, a genetic algorithm optimizes capacities of components in the upper level. In the lower level, normal dispatch computes annual cost and emissions, while stress dispatch, with a derated grid limit and component availability coefficients, evaluates reliability indices. For the grid-constrained case study, the method yields 14 non-dominated designs spanning zero to non-zero risk. Eliminating stress-state shedding from the minimum-cost design requires a 2.61% annual-cost premium; the balanced compromise achieves zero risk with a 36.10% emission reduction at an 11.25% cost premium. Zero-risk designs appear across a broad economic-carbon range, indicating that resilience alone does not determine a unique optimal capacity mix. The posterior risk is electrical rather than cooling-driven, with active components distinguishing shortage magnitude, interruption duration and tail severity. The framework restores an informative reliability dimension and quantifies the investment required to reconcile economic