Reliability-Constrained Behind-the-Meter BESS Dispatch for Data Centers: Co-Optimizing Utility Costs and Critical-Load Continuity Under Stochastic Outages
Article
Iqbal, H, Sarwat, A. (2026). Reliability-Constrained Behind-the-Meter BESS Dispatch for Data Centers: Co-Optimizing Utility Costs and Critical-Load Continuity Under Stochastic Outages
. IEEE ACCESS, 10.1109/ACCESS.2026.3696227
Iqbal, H, Sarwat, A. (2026). Reliability-Constrained Behind-the-Meter BESS Dispatch for Data Centers: Co-Optimizing Utility Costs and Critical-Load Continuity Under Stochastic Outages
. IEEE ACCESS, 10.1109/ACCESS.2026.3696227
Behind-the-meter (BTM) battery energy storage systems (BESS) in data centers are often dispatched for bill reduction, yet the same dispatch decisions determine the state of charge (SOC) available when a grid outage occurs. This paper presents an offline evaluation framework that quantifies the cost-reliability tradeoff created by this coupling. To the best of our knowledge, this is the first framework for behind-the-meter BESS in data centers that jointly addresses economic dispatch, explicit dispatch-to-outage SOC coupling, stochastic generator start behavior, separate UPS and BESS resource modeling, tail-risk reporting, and an explicit planning-versus-control scope. A two-timescale architecture links 15-minute mixed-integer linear dispatch co-optimizing energy cost, demand-charge exposure, and battery degradation with optional reserve enforcement...to a 1-second contingency simulator that models generator start stochasticity, UPS ride-through with load-dependent depletion of a finite energy bucket sized to NFPA 110 class boundaries, and BESS discharge under stochastic outages. Three dispatch policies (reliability-aware, naive economic, and standby-only) are compared under variance-reduced sequential Monte Carlo with common random numbers, reporting both mean and tail-risk metrics including CVaR and SLA exceedance probability. Results for a 20 MW data center with a 25 MW/30 MWh BESS reveal a sharp reserve threshold at 30 minutes, where per-outage failure probability drops from 0.975 to 0.044 and mean downtime decreases tenfold, at an annual demand-cost premium of approximately 3.3M. A per-event binding-resource classifier confirms that BESS depletion, not generator insufficiency, is the dominant failure mechanism across all duration regimes under economic-only dispatch, providing direct empirical support for the dispatch reliability coupling. All tail metrics (P95, P99, CVaR95) are reported with 95% percentile-bootstrap confidence intervals, and a convergence study justifies the choice of N = 80 trials for the design-point results. A load-shape robustness sweep across four workload classes (hyperscale, enterprise-diurnal, AI training, and a mixed reference) and a one-at-a-time sensitivity analysis over modeling assumptions and reliability parameters confirm that the policy ranking, reserve threshold, and qualitative conclusions are preserved across the tested parameter space.
