ARTIFICIAL INTELLIGENCE-ENABLED DIGITAL TWINS FOR ENERGY EFFICIENCY IN SMART GRIDS

Abstract

This PRISMA-guided systematic review synthesizes and critically evaluates how artificial intelligence (AI)-enabled digital twins (DTs) contribute to advancing energy efficiency in modern smart grids, spanning assets, feeders, microgrids, and system-level operations. A comprehensive database search and two-stage screening process identified 103 peer-reviewed studies that were assessed for context, twin architecture, AI methods, data pipelines, evaluation metrics, deployment maturity, and risk of bias. Evidence was systematically organized by the functional roles of DTs—monitoring, forecasting, optimization, and control—as well as by grid layers, revealing the growing integration of physics-informed surrogates, graph neural estimators, and reinforcement learning frameworks with data fabrics, semantic standards, and edge–cloud architectures. Quantitative synthesis demonstrates consistent and reproducible efficiency improvements when DTs mediate AI decisions against calibrated models: median feeder-level technical loss reductions of approximately five percent, median peak-demand reductions of nearly six percent, voltage compliance improvements of about twelve and a half percentage points, and renewable curtailment avoidance in the range of seven to nine percent relative to transparent baselines. These benefits are most concentrated when loop latencies are sub-second, particularly under control cycles closing within 300 milliseconds, and when DT deployments embed semantic interoperability, co-simulation, and uncertainty-aware decision-making with human-in-the-loop oversight. At the asset level, health-oriented DTs for transformers, breakers, cables, and wind turbines deliver measurable value through predictive maintenance that reduces inefficiencies and mitigates outages, with efficiency gains fully realized only when diagnostic outputs are integrated into scheduling, reconfiguration, and Volt/VAR optimization routines. Collectively, these findings advance the discourse beyond conceptual taxonomies by providing a reproducible, evidence-based blueprint for AI-enabled DT design: one that couples probabilistic and calibrated forecasting with latency-hardened voltage and topology control, links diagnostics to operations in a closed loop, and enforces transparency, explanation, and safety guardrails.