AI-Driven Energy Management and Optimal Coordination of Distributed Generation and Electric Vehicles in Smart Distribution Systems: A Systematic and Critical Review

Abstract

Distribution systems are absorbing two transitions at once: distributed generation (DG) that pushes power upstream at midday and electric vehicles (EVs) that pull it downstream in the evening. Artificial intelligence (AI) has been proposed as the coordinating layer that reconciles the two, yet the evidence for that claim has never been appraised against the standards a distribution system operator would apply before deployment. This article reports a systematic and critical review of 137 verified records, of which 60 are primary coordination studies, spanning forecasting, model-based optimisation, single-agent and multi-agent deep reinforcement learning (DRL), safe and constrained learning, and federated and graph-structured methods. Every study is scored with DEVCO-8, an eight-dimension appraisal rubric introduced here, and mapped onto a Deployment Readiness Index (DRI) that cannot be inflated by strength on one dimension alone. The corpus attains a mean DEVCO-8 score of 47.4/100; only 11.7% of primary studies report any external validity evidence (real sessions, hardware-in-the-loop or a held-out feeder), 11.7% release code or data, and 28.3% coordinate EVs or DG with no network model at all. Three analytical arguments sharpen these findings: a voltage-rise bound showing that coordinated EV absorption raises photovoltaic hosting capacity by a quantity proportional to the shifted fleet energy; a decision-space argument showing why centralised DRL cannot scale past tens of vehicles while decomposed and optimisation-primed learners can; and an exposure argument showing that unshielded exploration on a live feeder accrues constraint violations in proportion to the exploration rate. From these results the article contributes a three-timescale reference architecture, a shielded multi-agent learning algorithm, the ten-stage FAIR-GRID evaluation protocol with six reporting invariants, and a three-horizon research roadmap. The review concludes that the field has demonstrated capability but not readiness, and that the next advances will come from evaluation discipline rather than from new architectures.