关键词: 新型电力系统, 经济调度, 多目标优化, 分布式框架, 分叉融合注意力网络, 分布式多目标海星优化算法, 深度学习

Abstract:

The integration of high proportions of renewable energy has significantly increased the difficulty of economic scheduling optimization in new power systems， while traditional heuristic algorithms generally suffer from low computational efficiency and a tendency to fall into local optima. To address these bottlenecks， a branched fusion attention network （BFAN）-accelerated distributed multi-objective starfish optimization algorithm （BFAN-DMOSFOA） was proposed to achieve coordinated optimization of economic and environmental performance while ensuring stable system operation. A multi-objective optimization model minimizing power generation cost， carbon dioxide emissions， and sulfur dioxide emissions was constructed， and a distributed framework was adopted to divide the system into multiple autonomous regions， thereby reducing the risks of single-point failure and privacy leakage in centralized methods. Subsequently， the BFAN was designed. This network integrated the temporal feature extraction capability of gated recurrent units and the global correlation capture capability of the self-attention mechanism to accurately predict population evolution trends. BFAN was embedded into the iterative process of DMOSFOA to form a two-stage hybrid optimization mechanism of operation decision and simulation evaluation. The trained BFAN directly predicted and generated high-quality solutions within specific iteration intervals， replacing part of the random search process and greatly accelerating algorithm convergence. Simulation results on IEEE 118-bus and 1436-bus power systems showed that the proposed BFAN-DMOSFOA significantly outperformed comparative algorithms， including discrete multi-objective grey wolf optimizer （DMOGWO）， discrete multi-objective slime mould algorithm （DMOSMA）， discrete multi-objective whale optimization algorithm （DMOWOA）， a deep-neural-network-accelerated DMOSFOA （DNN-DMOSFOA）， and the unaccelerated DMOSFOA， in both solution-set quality and computational efficiency. In terms of solution-set quality， the proportions of global first-front non-dominated solutions contributed by BFAN-DMOSFOA in the two experiments reached 52.58% and 78.28%， respectively， showing absolute dominance. The average Euclidean distances from Pareto solution sets to the ideal point were 0.322 7 and 0.652 8 in the two experiments， respectively， both of which were the best among all algorithms， demonstrating excellent convergence accuracy. In terms of computational efficiency， BFAN-DMOSFOA achieves markedly shorter running times than all comparison algorithms across both test systems， while exhibiting strong scalability on the larger-scale 1436-bus system. BFAN-DMOSFOA not only protected data privacy and system robustness through the distributed framework， but also realized intelligent acceleration of the optimization process through BFAN， substantially improving solution speed while ensuring high quality and uniform distribution of Pareto-optimal solution sets. The findings confirm that this method can efficiently adapt to large-scale， multi-constraint complex scheduling scenarios， providing strong technical support and a feasible solution for real-time， economic， and environmentally friendly scheduling of new power systems.

Key words: new power systems, economic scheduling, multi-objective optimization, distributed framework, branched fusion attention network, distributed multi-objective starfish optimization algorithm, deep learning