Artificial Intelligence and Optimization for Distributed Generation Planning in Unbalanced Distribution Networks: A Systematic Critical Review of Engineering Transferability and Validation

Distributed-generation (DG) planning is shaped by two complementary methodological streams: artificial-intelligence (AI) methods that search or learn efficiently in difficult non-convex spaces, and mathematical-optimization methods that expose structure, bounds, recourse, and constraint logic. Reported objective improvements cannot be compared directly when studies differ in feeder physics, uncertainty models, protection assumptions, or validation settings. This review critically examines peer-reviewed DG-planning evidence published from 2016 to March 2026 using PRISMA 2020 reporting guidance, with review-specific deviations disclosed explicitly. The historical search ledger contains 122 records. Following decision-variable re-audit and source verification, the harmonized planning matrix contains 26 studies; sixteen previously scored records provide contextual evidence, and thirteen post-closure sources update the narrative only. The synthesis separates computational-intelligence and learning-assisted AI from convex, mixed-integer, decomposition, stochastic, and robust optimization, then evaluates the strength of the planning claim supported by each study. Of the 26 planning studies, 10 use explicit unbalanced or multiphase models, eight treat uncertainty non-deterministically, and two embed protection centrally. The heterogeneous evidence does not support a statistically adjusted ranking of solver families, and reported objective improvements should not be read as proof of general superiority. Credible DG-planning conclusions require objective design, phase-aware network physics, uncertainty treatment, protection feasibility, and validation to be aligned with the decision being supported.