Federated Reinforcement Learning for Real-Time Electric Vehicle Charging and Discharging Control Zixuan Zhangy Yuning Jiangx Yuanming Shiy Ye Shiy and Wei Chenz

2025-04-27 0 0 1.56MB 7 页 10玖币

侵权投诉

Federated Reinforcement Learning for Real-Time

Electric Vehicle Charging and Discharging Control

Zixuan Zhang†, Yuning Jiang§, Yuanming Shi†, Ye Shi†, and Wei Chen‡

†School of Information Science and Technology (SIST), ShanghaiTech University, China

§Automatic Control Laboratory, ´

Ecole Polytechnique F´

ed´

erale de Lausanne (EPFL), Switzerland

‡Department of Electronic Engineering, Tsinghua University, China

Email: zxzhang@gmail.com, yuning.jiang@ieee.org, {shiym,shiye}@shanghaitech.edu.cn, wchen@tsinghua.edu.cn

Abstract—With the recent advances in mobile energy storage

technologies, electric vehicles (EVs) have become a crucial part

of smart grids. When EVs participate in the demand response

program, the charging cost can be signiﬁcantly reduced by taking

full advantage of the real-time pricing signals. However, many

stochastic factors exist in the dynamic environment, bringing

signiﬁcant challenges to design an optimal charging/discharging

control strategy. This paper develops an optimal EV charg-

ing/discharging control strategy for different EV users under

dynamic environments to maximize EV users’ beneﬁts. We ﬁrst

formulate this problem as a Markov decision process (MDP).

Then we consider EV users with different behaviors as agents

in different environments. Furthermore, a horizontal federated

reinforcement learning (HFRL)-based method is proposed to

ﬁt various users’ behaviors and dynamic environments. This

approach can learn an optimal charging/discharging control

strategy without sharing users’ proﬁles. Simulation results illus-

trate that the proposed real-time EV charging/discharging control

strategy can perform well among various stochastic factors.

I. INTRODUCTION

In the past decades, the advent of electric vehicles (EVs)

has signiﬁcantly mitigated air pollution and fossil energy

depletion [1]. When EVs are connected to the power grid, they

can serve in the discharging mode as vehicle-to-grid (V2G)

devices or in charging mode as grid-to-vehicle (G2V) devices

[2]. As a new type of mobile and adjustable load, through

switching their working modes alternatively, a ﬂeet of EVs

connected to the grid can work in the G2V mode at the valley

time to reach valley ﬁlling and in the V2G mode at the peak

time to achieve peak shaving [3]. In addition, users’ charging

costs can be reduced by responding to the electricity price

signals and changing the working pattern in time [4]. The

V2G concept and its beneﬁts are shown in Fig. 1.

With the aim of maximizing users’ beneﬁts, the EV

charging/discharging control strategy [5] is always supposed

to coordinate the charging/discharging action, including the

charging/discharging decision and the charging/discharging

rate. However, due to plentiful stochastic factors lying in

the dynamic environment [6], like time-varying electricity

prices and uncertain behaviors from a variety of users, it is

challenging to design an optimal charging/discharging control

strategy considering many kinds of EV users.

Many day-ahead approaches, such as robust optimiza-

tion (RO) [7] and stochastic optimization (SO) [8] have

been proposed to handle the price uncertainty. Although the

methods above achieved great success in day-ahead charg-

ing/discharging control, it may be hard to depict the complex,

real-time scenarios with more uncertain factors. Generally,

the real-time charging/discharging control strategy considering

uncertain electricity prices and users’ demand satisfaction

can be formulated as an optimization problem with a known

state transition. Then it can be solved by model-based ap-

proaches like dynamic programming [9], model predictive

control (MPC) [10], model-based RL [11], etc. Nevertheless,

it is tough to establish an accurate system model or estimate

the state transition when considering various EV users’ inde-

terminate charging/discharging behaviors.

As a technique that can directly learn the optimal policies

without establishing or estimating the environment model [12],

Model-free RL has been applied to numerous smart grid

issues [13], [14] and obtains good control performances. A

model-free RL approach in [15] can avoid grid congestion by

coordinating EV charging/discharging. [16] took the dynamic

electricity price, non-EV residential load consumption, and

drivers’ behaviors into consideration to construct the dynamic

environment. However, in these above papers, it is assumed

that agents’ state transitions follow the same distribution, i.e.,

the environments of different agents are IID. However, in

actual scenarios, the situations faced by EV users may differ

slightly, resulting in the non-IID environments and different

state transitions.

As a novel type of distributed machine learning, federated

learning (FL) [17], [18], [19] has received considerable interest

from academia and industry. FL allows the use of isolated data

from multiple devices without violating the privacy protection

policy [20], [21], and it has been applied in many areas

[22], [23], [24]. Recently, an emerging ﬁeld called federated

reinforcement learning (FRL) [25] combines the advantages

of both FL and RL. It can not only provide agents with

the experience to learn to make good decisions in unknown

and dynamic environments but also train a global model

collaboratively without sharing their own experiences. As a

branch of FRL, horizontal federated reinforcement learning

(HFRL) ﬁts well for agents who are likely to isolate from

each other but face similar decision-making issues and have

fewer interactions [26].

This paper considers EV users with different behaviors as

agents in different environments. Motivated by [27], we aim

arXiv:2210.01452v1 [eess.SY] 4 Oct 2022

to collaboratively learn a real-time EV charging/discharging

control strategy that can perform uniformly well in different

kinds of environments. We ﬁrst formulate this problem as

a Markov decision-making process (MDP), then a HFRL-

based approach is proposed to deal with the dynamic charg-

ing/discharging environments and the users’ various behaviors.

In our approach, Soft Actor-Critic (SAC) algorithm [28] can

alleviate the sample-efﬁciency problem in RL as the local

training method, and the FedAvg algorithm [29] is utilized for

global aggregation to help each agent quickly learn the optimal

policy while considering privacy preservation. Moreover, our

simulation results demonstrate that the proposed real-time EV

charging/discharging control strategy can make a good trade-

off between dynamic electricity prices and uncertain behaviors

from different EV users.

Transformer

Control

Unit

Battery

Building

Control Unit

Bidirectional

Power Converter

EV Charging Station

Electric Vehicle

Smart Grid

Short-Term Value

Peak shaving and valley

filling

Reduce charging cost and

make discharging profits

Long-Term Value

EV users participate in

electricity transactions, and

benefit from power dispatch

and auxiliary modes

Fig. 1. V2G concept and its values

II. SYSTEM MODEL AND PROBLEM FORMULATION

This section ﬁrst introduces a model to describe the dynamic

changes of EVs’ batteries. Then, the EV charging/discharging

problem is formulated as a Markov decision process (MDP).

Finally, we formulate the objective of the optimal charg-

ing/discharging control policy.

A. EV Battery Model

In this paper, we consider NEVs equipped with the same

batteries indexed by i∈ {1, ..., N }and we assume that the

charging infrastructures are the same for each EV. We deﬁne

the times by ti

aat which EV iarrives at the charging station

and by ti

dat which it departs from the station. If the State

of Charge (SoC) of EV iat time tand t+ 1 are denoted by

SoCi

tand SoCi

t+1, respectively, then the dynamic of i-th EV’s

battery between time instants tto t+ 1 can be modeled as

SoCi

t+1 =(SoCi

tt<ti

a, t ≥ti

SoCi

t+η·ai

tti

a≤t<ti

d,(1)

where ai

tis the i-th EV’s total charging/discharging rate during

the time interval [t, t + 1). We assume the EV is under either

V2G (charging mode, ai

t≥0) or G2V mode (discharging

mode, ai

t≤0). Here, we also assume charging/discharging

has the same efﬁciency η∈(0,1] in this paper. Besides, SoCi

satisﬁes SoCi

t∈[0,1] for all iand t, and the input ai

tis

constrained by the charging infrastructure.

B. MDP Formulation

The EV charging/discharging control problem has the same

form as the sequential decision-making problem, such that it

can be regarded as a Markov decision process (MDP) with

discrete steps. Let us consider EV users having different charg-

ing/discharging behaviors such that Nagents, respectively,

interact with Nindependent environments. The environments

have different state transitions {Pi}N

i=1 but the same state

space S, action space A, reward function Rand discount

factor γ. Then the MDP of this problem can be denoted by

Mi=hS,A,Pi,R, γi, for all i={1,2,··· , N}.

1) State: As the input of the charging/discharging control

strategy, the environment state is used to generate a real-time

charging/discharging action. For agent i, the state si

t∈ Rn+6

at time tincludes the current and the past nhours’ electricity

price (ψt−n, ψt−n+1,··· , ψt)∈ Rn+1, the departure time ti

the anxious time ti

x, the current SoCi

t, the expected SoCi

xat

the anxious time and the departure time SoCi

d, that is

t={ψt−n, ψt−n+1, . . . , ψt, ti

d, ti

x,SoCi

t,SoCi

x,SoCi

d}.(2)

2) Action: The action ai

tdenotes the charging/discharging

rate of EV i’s battery during the state transition step [t, t +

1) with the given state si

t. Due to the limitation of charging

infrastructures, the action is restricted as follows

a≤ai

t≤a, (3)

where aand aare the minimum and maximum rate for this

EV charging/discharging problem.

3) Reward: The reward represents the immediate system

feedback after the state si

tchanges to si

t+1 with ai

t. We

proposed a reward settlement scheme integrating EV’s demand

response factor into this reward function. A mathematical

model in [30] is used to quantify the effect of anxiety on

SoC, and the anxious time ti

xis deﬁned in this model. The

model can be denoted as

SoCi

x=di

1e−di

2(t−ti

a)/(ti

d−ti

a)−1

e−di

2−1,(4)

where tsatisﬁes t∈[ti

x, ti

d). It can map the user’s anxiety to

the expected SoC properly. di

1∈[0,1] and di

2∈(−∞,0) ∪

(0,∞)are both shape parameters of the SoC curve. A larger

1leads to a higher SoC at ti

dand a larger di

2determines a

higher SoC during the charging/discharging duration.

We assume that the price of selling and the purchasing elec-

tricity are the same, the reward can be deﬁned as rt(si

t, ai

t) :=











−σp·ψt·ai

tti

a≤t<ti

−σp·ψt·ai

t−σx·max(SoCi

x−SoCi

t,0) ti

x≤t<ti

−σd·max(SoCi

d−SoCi

t,0) t=ti

Here, the factors σp,σx, and σddepict the user’s sensitivity to

price, anxiety, and demand response, respectively. When ti

a≤

t<ti

x, the reward fully considers the inﬂuence of electricity

price ﬂuctuations to EV charging/discharging decision. Then

the EV user’s anxiety is taken into consideration during ti

x≤

t<ti

d. As the EV leaves the charging station, i.e., t=ti

d, the

reward calculates the SoC gap between expected and current

SoC to meet the EV user’s demand as well as possible.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

FederatedReinforcementLearningforReal-TimeElectricVehicleChargingandDischargingControlZixuanZhangy,YuningJiangx,YuanmingShiy,YeShiy,andWeiChenzySchoolofInformationScienceandTechnology(SIST),ShanghaiTechUniversity,ChinaxAutomaticControlLaboratory,´EcolePolytechniqueF´ed´eraledeLausanne(EPFL),Switzerl...

展开>> 收起<<

Federated Reinforcement Learning for Real-Time Electric Vehicle Charging and Discharging Control Zixuan Zhangy Yuning Jiangx Yuanming Shiy Ye Shiy and Wei Chenz.pdf

共7页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Federated Reinforcement Learning for Real-Time Electric Vehicle Charging and Discharging Control Zixuan Zhangy Yuning Jiangx Yuanming Shiy Ye Shiy and Wei Chenz

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: