Predictive Analytics and Machine Learning for Energy Forecasting and Load Optimisation in Hospital Systems: A Critical Narrative Review

Abstract

Hospitals occupy an unusual position within the built environment. They operate continuously, maintain tightly regulated indoor conditions for clinical reasons, house diagnostic and therapeutic equipment with sharp and irregular power draws, and carry an obligation of uninterrupted supply that few other building types share. These characteristics have made hospital energy systems an attractive target for predictive analytics and machine learning, and a distinct body of applied work has accumulated over the past decade. That literature nonetheless remains difficult to interpret. Reported accuracy rankings among algorithm families are mutually inconsistent, most studies are built on a single facility, validation protocols vary widely, predictive uncertainty is rarely quantified, and the operational consequences of improved forecasts are seldom measured. This critical narrative review examines the state of knowledge on data-driven forecasting of hospital energy demand and on the use of such forecasts for load optimisation. Literature was identified through structured searching of open scholarly indexes, supplemented by backward and forward citation searching, with a final search date of 22 June 2026. The synthesis is organised around four analytical problems: the structural features that make hospital demand different from other non-residential demand; the contested evidence on model superiority; the methodological weaknesses that limit confidence in reported performance; and the weak evidentiary chain linking forecast accuracy to operational or economic benefit. The available evidence supports several defensible conclusions. Activity-related predictors improve accuracy beyond weather and calendar variables; ensemble and hybrid methods perform robustly at daily resolution; and deep architectures show advantages mainly at sub-hourly resolution when training records are long. Confidence in claimed hierarchies among model families remains limited, and evidence that forecasting systems reduce hospital energy cost, emissions or peak demand under clinical constraints is preliminary. Research priorities include multi-site benchmarking, decision-linked evaluation, probabilistic forecasting, and reporting standards adapted to critical-care settings.