晨光
暗夜
晨光
极光
Bilingual Paper Reading · 中英对照精读

基于关联深度 Q 学习的微电网能量管理:让储能、光伏、负荷智能协商

准大一 · 电气工程 × 微电网 × 强化学习 —— 微电网能量管理精读材料
原文:arXiv:2103.04152 2021年3月6日发布 arXiv 预印本(eess.SY / cs.LG) 微电网 × 多智能体 × 关联深度Q学习 附英文摘要朗读音频

一、论文档案

英文标题Correlated Deep Q-learning based Microgrid Energy Management
中文标题基于关联深度 Q 学习(CDQN)的微电网能量管理
作者周浩, 梅利克·埃罗尔-坎塔尔奇(机构未在素材中标注)
发布时间2021年3月6日(v1)|分类:eess.SY(系统工程与控制)/ cs.LG(机器学习)
一句话概括把微电网里的储能、光伏、负荷都当成会「自己学习」的智能体,各自用 LSTM 深度 Q 网络做决策,再用「关联均衡」协调它们——储能智能体利润提高 40.9%,光伏智能体利润提高 9.62%。
💡 为什么选这篇给你:① 微电网是新型电力系统的「最小单元」,储能+光伏+负荷怎么协同,是电气工程最火的方向之一;② 方法上是多智能体深度强化学习的经典入门组合——DQN + LSTM + 博弈论里的「关联均衡」;③ 结果有两个干净的百分比(40.9%、9.62%),而且作者明确说是「首次把 CDQN 用到微电网」,故事线完整。

二、核心术语表(先扫一遍再读正文)

英文术语中文大白话解释
microgrid (MG)微电网由分布式电源、储能和负荷组成的小型电力系统,可并网也可孤岛运行。
energy storage system (ESS)储能系统电池等能存电、放电的装置,是微电网的「蓄水池」。
renewable energy resources (RER)可再生能源光伏、风电等清洁发电资源,出力受天气影响。
photovoltaic (PV)光伏太阳能发电,是微电网最常见的电源。
island-mode孤岛运行微电网脱离主电网独立运行的模式。
agent智能体能感知环境、自主做决策的主体;本文中储能、发电、负荷各是一个智能体。
reinforcement learning (RL)强化学习智能体通过「试错+奖励」学会最优决策的机器学习范式。
Q-learningQ 学习经典强化学习算法:估计每个「状态-动作」组合的价值 Q 值,按 Q 值选动作。
deep Q-learning (DQN)深度 Q 学习用神经网络近似 Q 值函数,解决状态空间太大、查表查不过来的问题。
correlated equilibrium (CE)关联均衡博弈论中的均衡概念:多个智能体共享信号(如 Q 值),协调出整体最优的联合动作。
CDQN关联深度 Q 学习本文提出的算法:各智能体独立跑 DQN,再用关联均衡做协调。
MARL / MADRL多智能体(深度)强化学习多个智能体同时学习、相互影响的强化学习框架。
LSTM长短期记忆网络擅长处理时间序列的循环神经网络,这里用来预测各智能体的 Q 值。
Markov decision process (MDP)马尔可夫决策过程强化学习的问题数学框架:状态、动作、转移、奖励。
decentralized去中心化各智能体本地决策、不把私有数据交给中央,保护隐私。

三、摘要中英对照(精读核心)

🎧 音频在文末,可先听一遍原文再读;每个英文句都配了逐句翻译。

摘要 Abstract

EN · 原文
Microgrid (MG) energy management is an important part of MG operation.
CN · 翻译
微电网(MG)能量管理是微电网运行的重要组成部分。
EN · 原文
Various entities are generally involved in the energy management of an MG, e.g., energy storage system (ESS), renewable energy resources (RER) and the load of users, and it is crucial to coordinate these entities.
CN · 翻译
微电网能量管理通常涉及多种主体,例如储能系统(ESS)可再生能源(RER)用户负荷协调这些主体至关重要
EN · 原文
Considering the significant potential of machine learning techniques, this paper proposes a correlated deep Q-learning (CDQN) based technique for the MG energy management.
CN · 翻译
考虑到机器学习技术的巨大潜力,本文提出一种基于关联深度 Q 学习(CDQN)的微电网能量管理技术。
EN · 原文
Each electrical entity is modeled as an agent which has a neural network to predict its own Q-values, after which the correlated Q-equilibrium is used to coordinate the operation among agents.
CN · 翻译
每个电气主体被建模为一个智能体,各自拥有预测自身 Q 值的神经网络,之后用关联 Q 均衡来协调各智能体之间的运行。
EN · 原文
In this paper, the Long Short Term Memory networks (LSTM) based deep Q-learning algorithm is introduced and the correlated equilibrium is proposed to coordinate agents.
CN · 翻译
本文引入基于长短期记忆网络(LSTM)的深度 Q 学习算法,并提出用关联均衡协调各智能体。
EN · 原文
The simulation result shows 40.9% and 9.62% higher profit for ESS agent and photovoltaic (PV) agent, respectively.
CN · 翻译
仿真结果表明,储能智能体利润提高 40.9%光伏(PV)智能体利润提高 9.62%

关键词 Keywords:Microgrid 微电网 | Energy Management 能量管理 | Deep Q-learning 深度 Q 学习 | Correlated Equilibrium 关联均衡 | Multi-Agent 多智能体 | LSTM 长短期记忆网络

四、引言精选(为什么这个问题重要)

① 微电网是什么:能并网、能孤岛的「小电网」

EN · 原文
The integration of microgrids (MGs) to the electrical distribution system will improve the efficiency, reliability and resilience of future smart grid [1]. MGs are electrical systems that can connect to the main grid or work in island-mode. In particular, they bare many opportunities for enhancing the reliability of the electricity grid when they are able to form energy sharing communities [2, 3].
CN · 翻译
微电网(MG)接入配电系统将提升未来智能电网的效率、可靠性与韧性。微电网是既能连接主电网、又能孤岛运行的电力系统;尤其当它们能形成能量共享社区时,将为提升电网可靠性带来许多机会。

② 多智能体视角:储能、发电、负荷各有各的小算盘

EN · 原文
An MG usually contains an energy storage system (ESS), one or more generators which can be in the form of renewable energy resources (RER) and the load of users. MGs can be modelled as multi-agent systems where storage, generation and load are each represented by an agent and the agents might have different owners or controllers. For example, a RER unit can sell energy to the grid or to the ESS. The decision on who to sell and the price of selling, will determine the revenue for the RER agent. Cooperation and competition may exist between agents to maximize their own utility or the overall profit[4]. In this case, the agents need to trade energy to maximize their own profit, and the competition may increase the difficulty of balancing the profit. To this end, considering the potential of machine learning techniques learning based methods are promising to help agents maximize their profit.
CN · 翻译
微电网通常包含储能系统(ESS)、一台或多台可再生能源(RER)发电机组以及用户负荷。微电网可以建模为多智能体系统:储能、发电、负荷各代表一个智能体,且它们可能分属不同的所有者或控制者。例如,RER 机组可以把电卖给电网,也可以卖给 ESS——卖给谁、什么价格,决定了 RER 智能体的收益。智能体之间既可能合作也可能竞争,以最大化自身效用或整体利润。在这种情况下,智能体需要通过交易能源来最大化自身利润,而竞争会加大利润平衡的难度。考虑到机器学习的潜力,基于学习的方法有望帮助智能体最大化利润。

③ 为什么用关联均衡:协调多智能体,又不靠中央集权

EN · 原文
Most of the works in the literature [6, 7, 8, 9] mainly focus on one-agent case. However, the decentralized method is more suitable and scalable for MG application [10]. A multi-agent reinforcement learning (MARL) method is proposed in [11], where each agent has the learning ability to maximize its own profit. In addition, the cooperative reinforcement learning is proposed in [12] for the MG economic dispatch, where the cooperation is guaranteed by the diffusion strategy. In MARL, each agent is expected to learn the cooperation strategy. Naturally, MARL can also be generalized to multi-agent deep reinforcement learning (MADRL). Nash based method are generally used to distribute profits among cooperative agents [13]. However, it needs to exchange information iteratively, which may lead to a heavy computation burden. In this paper, we use the correlated equilibrium (CE) to harmonize agents. In CE, the agents exchange their Q-values, and make an optimal joint action [14]. The CE is conducted by a linear program, which is easy to be implemented in a decentralized way.
CN · 翻译
文献中大多数工作只研究单智能体情形;而去中心化方法更适合微电网应用、扩展性更好。多智能体强化学习(MARL)让每个智能体自主学习以最大化自身利润;协同强化学习则用于微电网经济调度。MARL 可以自然推广为多智能体深度强化学习(MADRL)。分配合作利润常用纳什(Nash)方法,但它需要迭代交换信息,计算负担重。本文改用关联均衡(CE)来协调智能体:智能体之间交换 Q 值,做出最优联合动作;CE 通过线性规划求解,易于去中心化实现

④ 本文的定位:首次把 CDQN 用到微电网领域

EN · 原文
However, the main difference between this paper and other related work is that, we propose a correlated deep Q-learning method for MG energy management, which is a MADRL algorithm. To the best of our knowledge, this is the first time CDQN is proposed and applied in MG field. The main contribution of this paper is that we generalize MARL to multi-agent deep reinforcement learning. We propose a correlated deep Q-learning method for the MG energy management, where each agent runs the DQN independently and the CE is used for coordination. The decentralized manner aims to protect agents' privacy. Our simulation results demonstrate the success of CDQN by having 40.9% and 9.62% higher profit for ESS agent and PV agent, respectively.
CN · 翻译
本文与其他工作的主要区别在于:我们提出一种用于微电网能量管理的关联深度 Q 学习方法(CDQN),它是一个 MADRL 算法。据我们所知,这是 CDQN 首次被提出并应用于微电网领域。主要贡献是把 MARL 推广到多智能体深度强化学习:每个智能体独立运行 DQN,用 CE 进行协调;去中心化的方式旨在保护智能体隐私。仿真结果证明了 CDQN 的成功——储能智能体与光伏智能体的利润分别提高 40.9%9.62%
💡 这是全文最有味道的一句“In CE, the agents exchange their Q-values, and make an optimal joint action.”——多智能体协调的两种路径:要么「纳什」式反复讨价还价(算得慢),要么「关联均衡」式共享信号一步到位(算得快)。选对协调机制,比加大算力更重要。

五、论文贡献(3 个要点)

EN · 原文
1. First CDQN in microgrids. However, the main difference between this paper and other related work is that, we propose a correlated deep Q-learning method for MG energy management, which is a MADRL algorithm. To the best of our knowledge, this is the first time CDQN is proposed and applied in MG field.
CN · 翻译
1. 微电网领域首次提出 CDQN。提出用于微电网能量管理的关联深度 Q 学习方法(CDQN)——一个 MADRL 算法;据作者所知,这是 CDQN 首次被提出并应用于微电网领域。
EN · 原文
2. LSTM-based DQN agents with correlated equilibrium coordination. Each electrical entity is modeled as an agent which has a neural network to predict its own Q-values, after which the correlated Q-equilibrium is used to coordinate the operation among agents. In this paper, the Long Short Term Memory networks (LSTM) based deep Q-learning algorithm is introduced and the correlated equilibrium is proposed to coordinate agents.
CN · 翻译
2. LSTM-DQN 智能体 + 关联均衡协调。每个主体用基于 LSTM 的深度 Q 网络预测自身 Q 值,再用关联均衡协调智能体间的联合动作——协调通过线性规划求解,易于去中心化实现
EN · 原文
3. Demonstrated profit gains. The simulation result shows 40.9% and 9.62% higher profit for ESS agent and photovoltaic (PV) agent, respectively.
CN · 翻译
3. 可量化的利润提升。仿真显示,储能智能体利润提高 40.9%光伏智能体利润提高 9.62%——「协调」的价值被直接量化。

六、结论中英对照

EN · 原文
The increasing complexity of energy trading within a microgrid and among several microgrids, call for novel approaches that benefit from the recent advances in machine learning. In this paper, we propose a correlated deep Q-learning (CDQN) technique for the MG energy management. In CDQN, each agent uses an LSTM network to predict their own Q-values and make decisions autonomously. Then, the correlated equilibrium is applied for coordination. Compared with the deep Q-learning without coordination, the CDQN performs 40.9% and 9.62% higher profit for ESS agent and PV agent, respectively.
CN · 翻译
微电网内部及多个微电网之间的能量交易日益复杂,呼唤受益于机器学习最新进展的新方法。本文提出用于微电网能量管理的关联深度 Q 学习(CDQN)技术:每个智能体用 LSTM 网络预测自身 Q 值并自主决策,再用关联均衡进行协调。与无协调的深度 Q 学习相比,CDQN 使储能智能体和光伏智能体的利润分别提高 40.9% 和 9.62%

七、编者解读:这篇论文到底讲了什么(大白话版)

  1. 问题:微电网里有储能、光伏、用户负荷,它们可能分属不同老板,各自想多赚钱。储能想低买高卖、光伏想高价卖电、用户想便宜买电——怎么让这帮「各怀心思」的智能体达成默契,让整体也划算?
  2. 做法:给每个主体配一个「会学习的脑袋」——LSTM 深度 Q 网络,让它自己学会「什么情况下做什么决定最赚钱」;但各学各的会互相打架,所以再引入博弈论里的「关联均衡」:大家把各自的 Q 值(对每个选择的打分)拿出来交换,用一个小线性规划算出「大家都满意的联合行动方案」。
  3. 结果:仿真显示,这套协调机制比「各干各的深度 Q 学习」强得多——储能智能体利润提高 40.9%,光伏智能体利润提高 9.62%。
  4. 最值钱的观点:多智能体协调不必「纳什式」反复博弈(慢),可以「关联均衡式」共享信号一步到位(快)。而且去中心化设计——每个智能体只交换 Q 值、不交出原始数据——顺便保护了隐私。
  5. 工程意义:微电网/虚拟电厂的自动能量管理是新型电力系统的刚需;「多智能体+强化学习」的思路可以直接迁移到社区微电网、楼宇能量管理、电动汽车聚合充电等场景。
🎯 对保研的启示:这篇论文示范了「跨学科搬运」的科研范式——把博弈论的均衡概念(关联均衡)搬进电力系统,还强调「第一次」。复试时若能讲出「原有方法哪里不够好(纳什迭代太重)、我借来了什么工具(CE+线性规划)、为什么适配(去中心化+隐私)」,就是一段完整的创新叙事。

八、给准大一的阅读路线图 & 延伸方向

📖 怎么读这篇论文(三遍法)

  1. 第一遍(10 分钟):只读摘要和术语表,回答三个问题——问题是什么(微电网多主体能量管理)?方法是什么(CDQN = DQN + LSTM + 关联均衡)?结果是什么(40.9%、9.62%)?
  2. 第二遍(20 分钟):读引言 + 结论,重点体会「为什么单智能体方法不够」「纳什均衡和关联均衡的区别」,以及「去中心化为什么重要」。
  3. 第三遍(30 分钟):读方法文字部分(智能体能量模型、Q 学习→深度 Q 学习→关联均衡的演进、线性规划求解),跳过所有公式和编号,只看文字描述;遇到不懂的术语回查术语表。

🚀 这个方向你能延伸做什么

九、英文摘要朗读(练听力用)

先盲听一遍→再看对照稿→再听一遍。目标是听出数字(40.9%、9.62%)和术语(energy storage system、renewable energy resources、correlated deep Q-learning、correlated Q-equilibrium、LSTM、photovoltaic)。