TY  - EJOU
AU  - Wang, Lixiao 
AU  - Li, Jiaqi 
AU  - Li, Haifeng 
AU  - Li, Hongfeng 
AU  - Wen, Yuhan 
AU  - Lin, Runhuan 
AU  - Zhou, Huansheng 

TI  - End-to-End Constraint Injection Neural Optimization for Flexible Load Scheduling in Low-Voltage Distribution Networks
T2  - Energy Engineering

PY  - 
VL  - 
IS  - 
SN  - 1546-0118

AB  - Economic scheduling of flexible loads in low-voltage distribution networks requires a balance between the accuracy of nonlinear load modeling and the efficiency of scheduling optimization. To address this challenge, this paper proposes a constraint-injected neural warm-start optimization framework termed End-to-End Constraint Injection Neural Optimization (ECINO). The framework comprises four stages: scenario data generation and simulation, end-to-end neural network model design, constraint-injection self-supervised learning, and scheduling optimization and constraint verification. Compared with the no-optimization benchmark, cold-start algorithm, relaxation-based warm-start algorithm, and supervised learning warm-start algorithm, ECINO achieves the lowest operating cost in the fixed-seed base case. Across the base scenario and five controlled perturbation scenarios, ECINO achieves a mean cost-reduction rate of 42.90% relative to the no-optimization benchmark in the corresponding scenario, and constraint verification detects no constraint violations. The results indicate that ECINO balances economic performance, solution efficiency, and constraint handling, while exhibiting favorable cross-scenario generalization to variations in photovoltaic output, load level, outdoor temperature, and peak-period electricity prices.
KW  - Low-voltage distribution network; flexible load scheduling; neural warm-start optimization; constraint-aware optimization; self-supervised learning

DO  - 10.32604/ee.2026.086358