Energy Demand Forecasting Methodologies for Hyperscale Computing Facilities and Implications for Sub-National Energy Policy: A Systematic Review

Abstract

Hyperscale computing facilities, defined as data centers operating at scales of 5,000 servers or greater with standardized, highly automated infrastructure architectures, have emerged as one of the most rapidly growing and analytically challenging energy demand categories for sub-national grid operators and energy policymakers. The convergence of exponential growth in artificial intelligence training workloads, large language model inference demands, cloud computing expansion, and cryptocurrency mining operations has transformed the energy demand profile of hyperscale facilities from a predictable baseline load into a highly variable, geographically concentrated, and policy-consequential infrastructure challenge. Traditional building-sector energy demand forecasting methodologies developed for commercial and industrial facilities are inadequate for hyperscale computing environments, which combine extreme power density (up to 20 to 50 kilowatts per rack in modern artificial intelligence compute clusters), 24/7 continuous operation requirements, batch workload scheduling patterns that create sharp hourly and sub-hourly load variability, and rapid capacity expansion timelines that outpace conventional utility planning cycles. This systematic review synthesizes empirical and methodological literature on energy demand forecasting for hyperscale computing facilities published between 2010 and 2025, evaluating five primary methodological domains: engineering bottom-up models, econometric and regression-based approaches, machine learning forecasting methods, hybrid physics-informed computational approaches, and scenario-based planning frameworks. The review further examines how hyperscale energy demand characteristics complicate sub-national energy policy development, identifying analytical gaps in distribution-level grid impact assessment, renewable energy integration planning for large-scale computing loads, and the translation of facility-level energy forecasting into actionable state and regional grid planning inputs. The review concludes with a priority research agenda addressing the most consequential methodological deficiencies in hyperscale energy demand forecasting and its sub-national policy implications.