Data center cooling in a hot and humid climate: physics-informed optimization,explainable monitoring, and the limits of Power Usage Effectiveness

Power Usage Effectiveness is the default efficiency benchmark for data centers, yet how much of it is recoverable at a given facilityis rarely tested against measured data, and hot, humid, grid-constrained sites are almost absent from the published record. Usingthree months (8,736 fifteen-minute records) of Data Center Infrastructure Management telemetry from a Tier-III colocation facilityin Dhaka, Bangladesh, cross-checked against ERA5 reanalysis, this study asks what governs cooling energy at such a site and howmuch can be recovered. Ambient temperature explains chiller power far better than information-technology load (R2 = 0.44 against0.11), and return humidity significantly depresses the cooling-system power ratio, inverting the load-centric narrative drawn fromhyperscale, temperate-climate studies.

Monotonicity-constrained surrogates, embedded as policy evaluators inside a non-dominatedsorting genetic algorithm, yield a policy that cuts on-peak cooling power by 23.3% yet improves Power Usage Effectiveness byonly 0.83%. This shortfall is not a modeling failure but a structural ceiling: the attainable gain is the product of the cooling-loadshare (28.6% of facility power) and the fractional cooling reduction the levers can deliver (at most 5.8%), bounding it at 1.65%,or 1.2–2.1% across the ASHRAE sensitivity range. Where cooling is a minority load, Power Usage Effectiveness is an incompletetarget and tariff-weighted peak power a more actionable one.

Two further mechanisms emerge: photovoltaic output cannot relieveon-peak cost because the tariff peak opens near sunset, and an unmetered on-site generator decouples metered grid import fromfacility load, which would silently break demand response built on utility-meter data. All gains carry block-bootstrap intervals fromfull-record replay.