Learning-based stochastic multi-objective optimizer for uncertain power system scheduling

Abstract

Power system scheduling with renewable energy sources poses significant challenges due to high computational complexity and uncertainty in operating conditions. Multi-period and multi-scenario modeling further escalates these issues, creating large-scale optimization problems that overwhelm traditional Stochastic Optimization Algorithm (SOA) with slow convergence and limited solution diversity. To tackle these challenges, we propose the Feature-Driven Multi-Objective Group Search Optimizer (FDMOGSO), a novel SOA for large-scale power systems scheduling. FDMOGSO employs the Self-Learning Method of Solution Space Feature (SLMSPF) to extract key features, reducing computational complexity by focusing exploration on promising regions. A Multi-Block Network (MBNet) classifier further enhances robustness by prioritizing high-quality solutions under uncertainty, while an enhanced Multi-Objective Group Search Optimizer (EMOGSO) adapts search strategies to improve convergence and solution diversity. Experimental results on IEEE 9-bus and IEEE 118-bus systems show that FDMOGSO significantly outperforms classical SOAs, including MOGSO, NSGA-II, MOPSO, and EMOGSO, on the Cumulative IGD Efficiency (CIGDE) metric, with improvements of 96.20%, 95.61%, 98.83%, and 94.68%, respectively. This demonstrates that FDMOGSO can find high-quality solutions for large-scale optimization problems with limited evaluations, enhancing the practical application potential of SOAs in complex power system scheduling.