TRENDS, METHODS, AND APPLICATIONS OF PREDICTIVE ENERGY ANALYTICS IN DATA CENTER CAPACITY PLANNING: A COMPREHENSIVE LITERATURE REVIEW

The rapid expansion of data-center infrastructure has intensified the need for reliable and efficient energy management, particularly as growing computational workloads place greater demands on power, cooling, and supporting infrastructure. Predictive energy analytics has emerged as a key response, using statistical and physics-based models, machine-learning and deep-learning techniques, and optimization and digital-twin approaches to predict energy behavior and support operational decision-making. However, the literature remains fragmented across these methodological streams, with limited synthesis of how they collectively support data-center capacity planning. This review asks: How do statistical, machine-learning, and digital-twin/optimization approaches to predictive energy analytics compare for data center capacity planning? A comprehensive narrative synthesis of 22 studies published between 2014 and 2026 examines evidence across the three methodological streams and their applications to data-center energy and capacity management. Five themes emerged: methodological fragmentation; digital twins as an unproven-at-scale integration layer; the dominance of PUE and cooling targets over direct capacity-planning objectives; the prevalence of single-facility validation; and the absence of calibrated uncertainty. The findings from this synthesis indicate that predictive energy analytics is substantially more developed for optimizing existing energy infrastructure than for forecasting workload-driven capacity requirements, while cross-stream integration, multi-facility validation, and uncertainty-aware prediction remain limited. Closing these gaps would move the field toward integrated, generalizable, and decision-oriented predictive frameworks capable of supporting more reliable data-center capacity-planning decisions.