A Multi-Layer Spatio-Temporal Learning and Optimization Framework for Line-Loss-Oriented Distribution Networks

High DG/BESS penetration reshapes power-flow patterns in distribution networks, amplifying line-loss (LL) volatility and stressing conventional planning and operation. We present ML-STELLO (Multi-Layer Spatio-Temporal nEtwork Learning and Line-loss Optimization), a multi-layer framework that unifies: (i) data governance for 96-point feeder curves via RODDPSO-enhanced FCM imputation and coefficient-of-variation improved Isolation Forest (CV-I Forest); (ii) multi-scale spatial-temporal (ST) learning using CNN-LSTM with attention for LL estimation; (iii) operation-time control coupling BESS dispatch and local VAR; and (iv) hierarchical planning and control across five time-scales—from multi-year investment to 5–15 min MPC—extending two-layer models driven by WGAN-GP scenarios and solved with IWOA. On IEEE-33 and a provincial feeder, ML-STELLO reduces LL and voltage violations relative to two-layer baselines while retaining robustness to missing/noisy data. Our design distills and extends advances in LL analysis, ST modeling, and uncertainty-aware planning.