1 Northwest Branch, State Grid Corporation of China, Xi’an 710001, China
2 China Electric Power Research Institute Co., Ltd., Beijing 100192, China
3 School of Automation, Beijing Institute of Technology, Beijing 100081, China
Show less
文章历史+
Received
Published
Revised
2025-11-02
2026-08-18
2025-12-08
Online Date
2026-09-04
PDF (3206K)
摘要
为了解决多类型储能的协同调度问题,提出了一种基于多智能体深度确定性策略梯度算法(multi-agent deep deterministic policy gradient,MADDPG)的新型综合能源系统(integrated energy systems,IES)优化调度方法。随着全球能源转型进程加快及“碳达峰、碳中和”目标的提出,大规模可再生能源接入使综合能源系统中多类型储能的协同调度成为提升系统经济性与可再生能源消纳能力的关键问题。系统采用“源-汇-荷”三段式结构:源端整合电网、光伏及风电;汇端通过电能转换设备实现多源并网;荷端部署电化学储能(electrochemical energy storage,EES)、压缩空气储能(compressed air energy storage,CAES)和超级电容(supercapacitors,SC)三类储能设备,结合用户负荷形成多能互补体系。针对传统优化方法难以处理高维非线性约束的局限,设计多智能体合作博弈模型,通过集中训练-分布执行机制实现储能设备的协同控制。各智能体基于Actor-Critic网络独立决策充放电策略,Critic网络共享全局状态信息以优化联合动作奖励(含电能交易成本、弃风弃光惩罚等)。仿真表明,MADDPG在动态电价与可再生能源波动下显著提升系统经济性及可再生能源消纳率,为多类型储能系统的高效调度提供了新思路。
Abstract
A novel optimization scheduling method for integrated energy systems (IES) based on the multi-agent deep deterministic policy gradient (MADDPG) algorithm was proposed to address the cooperative scheduling of multi-type energy storage. With the accelerating global energy transition and large-scale integration of renewable energy, the coordinated dispatch of multi-type energy storage in IES has become a key issue for improving both system economy and renewable energy utilization. The system was configured with a three-segment “source-conversion-load” structure: on the source side, the grid, photovoltaic, and wind power were integrated; on the conversion side, multi-source grid integration was enabled through power conversion devices; on the load side, three types of energy storage devices—electrochemical energy storage (EES), compressed air energy storage (CAES), and supercapacitors (SC) were deployed, forming a multi-energy complementary system in combination with user loads. To overcome the limitations of traditional optimization methods in handling high-dimensional nonlinear constraints, a multi-agent cooperative game model was constructed, through which coordinated control of energy storage devices was achieved by a centralized training and decentralized execution mechanism. Each agent was enabled to determine its charging and discharging strategy independently via an Actor-Critic network, while global state information was shared within the Critic network to optimize the joint action reward, including electricity transaction costs, penalties for wind and solar curtailment, and other relevant terms. Simulation results show that the MADDPG-based method significantly enhances system economic performance and renewable energy consumption under dynamic electricity prices and fluctuating renewable generation, thereby providing a promising approach for the efficient scheduling of multi-type energy storage systems.
WANG Wen-zhuo, GUO Shu-yu, GUO Xiang-yang, LI Wu-jing, ZHANG Ya-gang, JING Yi-ran, HE Li-hao, ZOU Su-li.
Optimization of Multi-type Storage Integrated Energy Systems Based on MADDPG[J].
Science Technology and Engineering, 2026, 26(23): 10013-10020 DOI:10.12404/j.issn.1671-1815.2507598
随着全球能源转型进程的加速推进和“碳中和”目标的提出,综合能源系统(integrated energy system,IES)作为实现能源高效利用和可再生能源大规模消纳的重要载体,正受到学术界和工业界的广泛关注[1-3]。IES通过有机整合多种能源形式与储能设备,实现能源的梯级利用和优化配置,有效提高了能源利用效率。然而,可再生能源固有的间歇性和波动性,以及多元储能设备的特性差异,给IES的优化调度带来了严峻挑战[4-5]。特别是在以风能、光伏为代表的可再生能源渗透率不断提高的背景下,如何协调电化学储能、压缩空气储能、超级电容等多类型储能设备的运行,充分发挥各自的技术经济特性,实现系统的经济、稳定、高效运行,已成为当前IES研究领域的关键科学问题[6-8]。
传统的IES优化调度方法主要基于数学规划理论,如线性规划(linear programming,LP)、动态规划(dynamic programming,DP)、混合整数线性规划(mixed integer linear programming,MILP),随着数据驱动技术的发展,数据-模型混合驱动的方法进一步得到关注[9-10]。这些方法在特定系统类型中表现出良好的性能,Moser等[10]采用MILP方法解决了IES的最优调度问题,并评估了所提出方法的实际节能潜力。然而,这类方法对模型的线性化要求较高,难以准确刻画实际系统中复杂的非线性耦合关系与动态不确定性[11]。当系统规模扩大、设备类型增多时,传统方法易遭遇“维度灾难”,计算复杂度急剧增加,难以满足实时调度需求[12]。
此外,依据深度强化学习框架,为每个智能体定义局部观测空间和动作空间,其中,智能体 i 的局部观测空间涵盖其能够观测到的所有局部状态信息,动作空间包含其可执行的全部控制动作。在多智能体联合动作作用下,为本系统定义状态转移函数和全局即时奖励函数,其中,状态转移函数用于描述在给定任意动作的情况下,系统从当前状态转移至下一任意状态的概率分布,全局即时奖励函数用于评估在当前状态下,执行多智能体联合动作后,转移到新状态时所产生的瞬时奖励,这部分的设计是深度强化学习模型能否训练得到期望策略的关键。
在上述框架下,每个储能单元 i 持续监测其所在区域的实时运行状态,并基于其 Actor 策略网络生成对应动作,以此实现对该区域资源的调度管控。在训练初期,各储能单元i采用随机初始化的策略,该部分有助于算法的实际使用。当多储能单元的联合动作作用于环境后,系统状态发生转移。此时,系统将收集每个储能单元的状态、动作、所获全局奖励及下一状态,组成经验样本,并存入经验回放缓冲区。接着,Critic 评价网络会从经验回放缓冲区中提取历史轨迹,对当前的 Actor 策略性能展开评估,并基于优化目标通过梯度更新方式对 Actor 策略进行改进。这一过程持续进行,直至所有智能体的策略行为达到动态平衡。
为了实现多类型储能系统中各智能体的有效策略学习,本研究采用基于Actor-Critic架构的多智能体深度强化学习算法MADDPG (multi-agent deep deterministic policy gradient),构建训练流程与模型架构。该算法遵循“集中训练、分布执行”(centralized training with decentralized execution, CTDE)机制,在训练阶段集中获取所有智能体的状态与动作信息,以学习最优的联合策略;而在执行阶段,各智能体仅根据自身局部可观测信息独立作出决策,符合现实储能系统分布式控制需求。
LuoZhao, LiuDewen, ShenXin, et al. Review of research on optimal operation technology of integrated energy system[J]. Electric Power Construction, 2022, 43(12): 3-14.
[3]
GaoM, HanZ, ZhaoB, et al. Optimal planning method of multi-energy storage systems based on the power response analysis in the integrated energy system[J]. Journal of Energy Storage, 2023, 73: 109015.
XuPenghui, XiongJunhua, LiShuai. Low-carbon economic optimization scheduling of integrated energy systems based on CLHG-SOFC[J]. Science Technology and Engineering, 2025, 25(24): 10285-10293.
HuSongtao, SunTao, HeGuangwen, et al. Research on integrated energy system modeling and operation problems[J]. Scientific and Technological Innovation, 2023(26): 21-24.
WangYongli, LiuZeqiang, ChenXin, et al. Double-layer optimization model of industrial park-level integrated energy system based on cascade utilization[J]. Science Technology and Engineering, 2024, 24(29): 12595-12603.
[10]
YeY, PapadaskalopoulosD, YuanQ, et al. Multi-agent deep reinforcement learning for coordinated energy trading and flexibility services provision in local electricity markets[J]. IEEE Transactions on Smart Grid, 2022, 14(2): 1541-1554.
[11]
GabrielliP, FürerF, MavromatidisG, et al. Robust and optimal design of multi-energy systems with seasonal storage through uncertainty analysis[J]. Applied Energy, 2019, 238: 1192-1210.
[12]
BartoliniA, CarducciF, MuñozC B, et al. Energy storage and multi energy systems in local energy communities with high renewable energy penetration[J]. Renewable Energy, 2020, 159: 595-609.
FanHong, XuTao, JiaQingshan. Review of optimization scheduling for integrated energy systems with data centers based on hybrid data-model driven[J]. Southern Power System Technology, 2025, 19(3): 174-187.
[15]
MoserA, MuschickD, GöllesM, et al. A MILP-based modular energy management system for urban multi-energy systems: performance and sensitivity analysis[J]. Applied Energy, 2020, 261: 114342.
WenTianrun, PengDaogang, ZhaoHuirong, et al. Stochastic optimal operation of park-level integrated energy considering equipment risk and uncertainty of wind and solar[J]. Science Technology and Engineering, 2023, 23(32): 13861-13871.
[18]
LiuJ, MaL, WangQ. Energy management method of integrated energy system based on collaborative optimization of distributed flexible resources[J]. Energy, 2023, 264: 125981.
XingXueliang, JiaLimin, HuJunjie, et al. Study on comprehensive optimization operation for park-level self-contained energy systems[J]. Journal of North China Electric Power University (Natural Science Edition), 2026, 53 (2): 1-15.
[21]
WuM, DuP, JiangM, et al. An integrated energy system optimization strategy based on particle swarm optimization algorithm[J]. Energy Reports, 2022, 8: 679-691.
[22]
KhademiV, SoleymaniS, SharifiR, et al. Improving integrated energy systems with improved particle swarm optimization: co-optimizing renewables, electric vehicles, gas systems, and demand management[J]. International Transactions on Electrical Energy Systems, 2025(1): 3382601.
[23]
IbrahimO, AzizM J A, AyopR, et al. Fuzzy logic-based particle swarm optimization for integrated energy management system considering battery storage degradation[J]. Results in Engineering, 2024, 24: 102816.
ZhangQian, HeShan, HuangSong, et al. Variable pitch control of wind turbine based on DDPG algorithm[J]. Science Technology and Engineering, 2023, 23(18): 7764-7771.
RuanYingjun, HouZequn, QianFanyue, et al. Optimization of the operation of distributed energy system based on deep reinforcement learning[J]. Science Technology and Engineering, 2022, 22(17): 7021-7030.
[28]
ArulkumaranK, DeisenrothM P, BrundageM, et al. Deep reinforcement learning: a brief survey[J]. IEEE Signal Processing Magazine, 2017, 34(6): 26-38.
[29]
DingL, CuiY, YanG, et al. Distributed energy management of multi-area integrated energy system based on multi-agent deep reinforcement learning[J]. International Journal of Electrical Power & Energy Systems, 2024, 157: 109867.
PengChunhua, ZhangHaoqi, SunHuijuan, et al. Optimized scheduling of multi photovoltaics and energy storage integrated flexible direct current distribution systems based on delay-aware multi-agent deep reinforcement learning[J]. Transactions of China Electrotechnical Society, 2026, 41(11): 3742-3754.
[32]
LoweR, WuY I, TamarA, et al. Multi-agent Actor-Critic for mixed cooperative-competitive environments[J]. arXiv Preprint, arXiv: 1706. 02275, 2017.
[33]
DouJ, WangX, LiuZ, et al. Towards Pareto-optimal energy management in integrated energy systems: a multi-agent and multi-objective deep reinforcement learning approach[J]. International Journal of Electrical Power & Energy Systems, 2024, 159: 110022.
[34]
HarroldD J B, CaoJ, FanZ. Renewable energy integration and microgrid energy trading using multi-agent deep reinforcement learning[J]. Applied Energy, 2022, 318: 119151.