发电技术2026，Vol.47Issue(3)：449-483,35.DOI:10.12096/j.2096-4528.pgt.260301
演化博弈与联邦学习融合驱动的新型电力系统多元主体协同优化
Multi-Agent Collaborative Optimization in New Power Systems Driven by Integration of Evolutionary Game and Federated Learning
摘要
Abstract
[Objectives]To address the collaborative optimization difficulties arising from participation of multiple market entities,deep integration of distributed resources,and multi-level decision coordination in new power systems,where traditional centralized optimization models struggle to adapt in terms of privacy protection,distributed computation,and interest coordination,this study aims to clarify the theoretical connotation,generalized framework,and applicability boundaries of integrating evolutionary game theory(EGT)and federated learning(FL)in power systems,thereby resolving the key scientific problem of multi-agent strategy coordination under privacy constraints.[Methods]Following the main thread of"problem-driven-theoretical support-method construction-scenario implementation",the respective mechanisms of EGT in characterizing the strategy evolution of bounded rational agents and of FL in achieving privacy-preserving distributed modeling are systematically reviewed.The coupling principles of their integration(EGT-FL)are then elaborated from four perspectives,namely dynamical isomorphism,information-theoretic consistency,learning-theoretic unification,and optimization-objective correspondence,and a generalized integration framework adapted to the physical constraints of power systems is constructed.On this basis,key methods including multi-agent game modeling,federated strategy evolution algorithm design,and incentive-privacy co-optimization are summarized,and simulation verification is conducted on a modified IEEE 33-bus system and a large-scale demand response case involving 1 050 agents.[Results]Simulation results demonstrate that the EGT-FL algorithm outperforms FL algorithms such as FedAvg and FedProx,as well as classical game-theoretic methods such as Nash equilibrium and Stackelberg game,across five dimensions:convergence,privacy protection,economic performance,computational efficiency,and robustness.The convergence speed is improved by 30%-50%compared with traditional methods.Under a differential privacy parameter of ε=1,more than 90%of the optimization performance is still maintained.The total system cost increases by only about 3.2%,while the privacy protection level reaches 95%.The system stability is maintained above 70%under a 20%proportion of malicious agents.[Conclusions]The integration of EGT and FL can effectively resolve conflicts of interest among multiple agents while preserving data privacy,and achieve system-wide benefit maximization,providing a new paradigm with both theoretical rigor and engineering feasibility for multi-agent collaborative decision-making in new power systems.The deep integration of federated reinforcement learning,cross-level collaborative optimization,and adaptive privacy mechanisms are key research directions for future breakthroughs,offering theoretical support and technical reference for the intelligent operation of new power systems.关键词
新型电力系统/演化博弈论/联邦学习/多主体协同决策/隐私保护优化/分布式能源管理Key words
new power system/evolutionary game theory/federated learning/multi-agent collaborative decision-making/privacy-preserving optimization/distributed energy management分类
能源科技引用本文复制引用
程乐峰,孙润宝,倪曼琦,邹涛,余涛,张孝顺,王怀智..演化博弈与联邦学习融合驱动的新型电力系统多元主体协同优化[J].发电技术,2026,47(3):449-483,35.基金项目
国家自然科学基金项目(52171331,U24B6010) （52171331,U24B6010）
广东省自然科学基金项目(2023A1515011311) （2023A1515011311）
广东省普通高校创新团队项目(2024KCXTD031) （2024KCXTD031）
广州市教育局高校科研项目(2024312278).Project Supported by National Natural Science Foundation of China(52171331,U24B6010) （2024312278）
Natural Science Foundation of Guangdong Province(2023A1515011311) （2023A1515011311）
Innovation Team Project for Ordinary Universities in Guangdong Province(2024KCXTD031) （2024KCXTD031）
Guangzhou Education Bureau University Research Project(2024312278). （2024312278）
