A Review of Trustworthy Deep Learning for Renewable-Rich Distribution Smart Grids: Forecasting, Resilience, and Operation

This PRISMA-informed structured review examines trustworthy deep learning for low- and medium-voltage distribution feeders and coordinated multi-feeder systems, from standard test feeders to utility networks; it does not assume one fixed bus count. Renewable generation, electric vehicles, flexible demand, storage, and digital monitoring create nonlinear, temporal, spatial, and probabilistic decision problems. We organize the evidence through a task–data–model–risk–validation pathway that links each operational task to its data structure, defensible model family, principal trustworthiness risk, and minimum deployment evidence. Common mathematical formulations provide shared evaluation criteria for forecasting, state estimation, reliability, resilience, control, and deployment. Reported results are interpreted only within their stated datasets and test conditions. Compact temporal models are defensible first-line baselines for moderate-horizon forecasting; topology-aware and physics-informed models are suited to tasks where network structure and feasibility matter; and deep reinforcement learning is a candidate for sequential control only when constraints and fallback behavior are tested. Exact database search dates and retrieval counts were not retained, so the review does not claim exhaustive systematic coverage or a quantitative meta-analysis.