Decentralized Federated Learning Architecture for Networked

Microgrids

Ilyes Naidji

1 a

, Chams Eddine Choucha

2 b

and Mohamed Ramdani

3 c

RLP laboratory, Mohamed Khider University of Biskra, Algeria

LSSD laboratory, University of Science and Technology of Oran Mohamed-Boudiaf, Algeria

Linﬁ laboratory, Mohamed Khider University of Biskra, Algeria

Keywords:

Federated Learning, Networked Microgrids, Decentralized Architecture, Energy Management.

Abstract:

The expansion of large-scale distributed renewable energy drives the emergence of networked microgrids

systems, necessitating the development of an efﬁcient energy management approach to minimize costs and

maintain energy self-sufﬁciency. The use of smart systems that are based on deep learning algorithms has be-

come prevalent while addressing the energy management problem due to its real-time scheduling capabilities.

However, training deep-learning algorithms requires substantial energy operation data from these microgrids,

which raises concerns regarding privacy and data security when collecting data from various microgrids. To

address this challenging problem, this article proposes a decentralized federated learning architecture for net-

worked microgrids. The architecture incorporates a distributed federated learning (FL) mechanism to guaran-

tee data privacy and security and prevent the system from signle point of failure. A decentralized networked

microgrids model is constructed, where each participating microgrid has an energy management system re-

sponsible for managing its energy. The goal of the EMS is to minimize economic costs and maintain energy

self-sufﬁciency. Initially, MGs independently undergo self-training using local energy operation data to train

their individual models. Subsequently, these local models are regularly exchanged, and their parameters are

aggregated to create a global model. This approach allows sharing of experiences among the microgrids with-

out transmitting energy operation data, thereby safeguarding privacy and ensuring data security and preventing

from single point of failure.

1 INTRODUCTION

Efﬁcient energy management plays a crucial role in

mitigating environmental impacts and addressing the

growing demand for energy, even in the face of in-

creasing energy consumption (Rehman et al., 2021).

To achieve a balance between demand and supply, es-

tablish robust power infrastructure, optimize gener-

ation schedules, and minimize the repercussions of

renewable energy production, a smart energy man-

agement should be addressed (Chouikhi et al., 2019),

(Naidji et al., 2019). Utility providers can now mea-

sure and record energy usage for buildings or indi-

vidual residences within an hour or less, thanks to

advanced metering infrastructure and the widespread

adoption of smart meters (Naidji et al., 2018). This

https://orcid.org/0000-0001-8747-0766

https://orcid.org/0000-0003-0194-4890

https://orcid.org/0000-0002-8723-5827

technology has gained signiﬁcant traction in the UK,

where over 15 million smart meters are currently em-

ployed in homes and businesses. As a result for the

large adoption for smart meters, different research

work were published discussing the utilization of ma-

chine learning in smart grids (Massaoudi et al., 2021).

The work in (Feng et al., 2020) detect real-time build-

ing occupancy from Advanced Metering Infrastruc-

ture (AMI) data based on a deep learning architec-

ture. The developed deep learning model consists

of a convolutional neural network (CNN) and a long

short-term memory (LSTM) network. The simulation

results show that the developed model outperforms

the existing state-of-the-art ML classiﬁers and other

deep learning architectures with around 90% occu-

pancy detection accuracy.

A deep learning load prediction model is proposed

in (Zhu et al., 2020). A daily time step-by-step load

prediction method is employed where a deep cycle

neural network (DCNN) model is established for the

Naidji, I., Choucha, C. and Ramdani, M.

Decentralized Federated Learning Architecture for Networked Microgrids.

DOI: 10.5220/0012215200003543

In Proceedings of the 20th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2023) - Volume 1, pages 291-294

ISBN: 978-989-758-670-5; ISSN: 2184-2809

291

total daily load and hourly load of users. The sim-

ulation results show the superiority of the proposed

model.

The work in (Li et al., 2022) presents a novel ap-

proach for detecting FDIA (False Data Injection At-

tacks) by leveraging the concepts of federated learn-

ing, secure federated deep learning, and the Trans-

former model. The proposed method combines the

multi-head self-attention mechanism of the Trans-

former, which is deployed as a detector in edge nodes,

to explore the intricate relationships among individual

electrical quantities. By adopting a federated learning

framework, our approach allows collaborative train-

ing of a detection model using data from all nodes

while ensuring data privacy by keeping the data lo-

cally during the training process. To enhance the se-

curity of federated learning, we have designed a se-

cure federated learning scheme that incorporates the

Paillier cryptosystem into the federated learning pro-

cess.

The work proposed in (Jithish et al., 2023) intro-

duces a smart grid anomaly detection scheme that uti-

lizes Federated Learning (FL). The scheme involves

training machine learning (ML) models locally within

smart meters, without the need to share data with a

central server. Initially, a global model is obtained

from the server and deployed on the smart meters for

on-device training. Following local training, the up-

dated model parameters are transmitted to the server,

contributing to the enhancement of the global model.

The work in (Wen et al., 2022) proposes a novel

privacy-preserving federated learning framework for

energy theft detection, namely, FedDetect. In this

framework, the authors consider a federated learning

system that consists of a data center (DC), a control

center (CC), and multiple detection stations. In this

system, each detection station (DTS) can only ob-

serve data from local consumers, which can use a lo-

cal differential privacy (LDP) scheme to process their

data to preserve privacy. To facilitate the training of

the model, the authors design a secure protocol so that

detection stations can send encrypted training param-

eters to the CC and the DC, which then use homomor-

phic encryption to calculate the aggregated parame-

ters and return updated model parameters to detection

stations. In this study, the authors prove the security

of the proposed protocol with solid security analy-

sis. To detect energy theft, a deep learning model

based on the state-of-the-art temporal convolutional

network (TCN) is designed.

However, none of the above studies consider the

potential single point failure of the server which can

block the entire process. To address this challenge,

we propose a decentralized federated learning archi-

tecture for networked microgrids which consists of a

decentralized model exchange and a decentralized ag-

gregation model.

2 SYSTEM MODEL

2.1 Problem Formulation

Energy management in networked microgrids system

requires a real-time control and monitoring of renew-

able energy sources which is intermittent and need to

be predicted. For this reason, load forecasting and en-

ergy price forecasting should be addressed in order to

reduce energy cost and improve energy availability.

2.2 Distributed Federated Learning

The optimization problem of federated learning for K

devices can be formulated with objective function as

follows:

min l(w) =

∑

K=1

(w) (1)

Where

(W ) =

∑

i∈p

(w) (2)

l(w) is the loss function of the global model, LK(x)

is the loss of the kth device, and l

(w) is the loss for

sample i ∈ p

Decentralized federated learning builds upon the

principles of federated learning but takes the con-

cept a step further by removing the need for a cen-

tral server or coordinator. Instead, it distributes the

model training process across multiple devices or en-

tities in a peer-to-peer manner. This approach further

enhances privacy and eliminates the single point of

failure introduced by a central server.

Here’s an overview of how decentralized feder-

ated learning works:

2.2.1 Network Formation

A network of microgrids is established in a decen-

tralized manner. These microgrids can communicate

with each other directly or through a peer-to-peer

network infrastructure. Here in our case, we con-

sider the mechanism of coalition formation between

networked microgrids developped in (Naidji. et al.,

2019), (Naidji et al., 2020). This mechanism allow

the network to autonomously self organize in coali-

tions to achieve better performance in terms of energy

exchange.

ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics

292

Microgrid 2

Microgrid n

Microgrid 1

Microgrid 3

Coalition 1

Coalition n

Coalition 2

Coalition Formation

Figure 1: Network Formation.

Figure 1 shows the process of network formation

which is based on an merge and split coalition for-

mation algorithm (MSCF) proposed in (Naidji. et al.,

2019). The microgrids autonomously self-organize in

coalitions in order to achieve an optimal energy man-

agement. Note that, here in our case, decentralized

federated learning is conducted for every coalition.

The reason behind that is that microgrids belonging to

the same coalition shares multiple features, and thus,

sharing their learning models will bring more beneﬁts

for each one.

2.3 Model Initialization

Initially, each microgrid initializes its local model pa-

rameters, either randomly or based on a pre-trained

model.

2.4 Local Model Training

Each microgrid independently trains its local model

using its local data. This process is similar to tradi-

tional federated learning, where each device performs

local computations on its data to update the model.

2.5 Model Exchange

After local training, microgrids exchange model up-

dates with their neighboring microgrids. The ex-

change can be performed directly between microgrids

or through a decentralized network infrastructure.

2.6 Model Aggregation

Each microgrid receives model updates from its

neighbors and aggregates them with its own local

model to create an updated model. The aggregation

process can involve averaging the models. Here we

use the decentralized federated averaging algorithm

proposed in (Sun et al., 2021).

To begin, we provide a brief overview of the

this algorithm. This algorithm follows the following

steps:

1) Each microgrid, denoted as microgrid i, pos-

sesses an approximate copy of the parameters, x(i).

2) Communication takes place among the micro-

grids. During this phase, microgrid i updates its lo-

cal parameters, x(i), by taking a weighted average of

its neighboring microgrids’ parameters. The updated

value is denoted as x(i) =

∑

l∈N(i)

i,l

.x(l).

3) The training process starts. Microgrid i up-

dates its parameters using the expression x(i) = x(i)−

αg(i), where α is a positive learning rate.

Microgrid 1

Microgrid n

Coalition

Figure 2: Model aggregation.

Figure 2 shows the process of model aggregation

in each microgrid that belongs to a given coalition.

The microgrid receives multiple model parameters

Decentralized Federated Learning Architecture for Networked Microgrids

293

that will be aggregated to form its ﬁnal model.

2.7 Iterative Process

Steps 3-6 are repeated for multiple rounds or itera-

tions. In each round, microgrids train their models

locally, exchange model updates with neighbors, ag-

gregate the updates, and broadcast their updated mod-

els.

2.8 Convergence

Over iterations, the models on each microgrid be-

come more reﬁned as they learn from the combined

knowledge of neighboring microgrids. The decen-

tralized federated learning process continues until the

desired performance or convergence is achieved. De-

centralized federated learning allows for collaborative

model training across multiple microgrids in a peer-

to-peer manner, without relying on a central server. It

enables microgrids to exchange information directly

with their neighbors, reducing latency and potential

privacy risks associated with transmitting data to a

central authority.

It’s important to note that decentralized federated

learning can take various forms depending on the spe-

ciﬁc network architecture, communication protocols,

and consensus algorithms used. These aspects can

vary based on the requirements and constraints of the

decentralized system being implemented.

3 CONCLUSION

In this paper, we have proposed a decentralized fed-

erated learning architecture for networked microgrids

system to improve energy management. The pro-

posed architecture guarantees the data privacy and se-

curity of microgrids by exchanging only their models.

Furthermore, the proposed decentralized architecture

prevents from signle point of failure by eliminating

the central server and exchanging models in a peer-

to-peer manner.

REFERENCES

Chouikhi, S., Merghem-Boulahia, L., Esseghir, M., and

Snoussi, H. (2019). A game-theoretic multi-level en-

ergy demand management for smart buildings. IEEE

Transactions on Smart Grid, 10(6):6768–6781.

Feng, C., Mehmani, A., and Zhang, J. (2020). Deep

learning-based real-time building occupancy detec-

tion using ami data. IEEE Transactions on Smart

Grid, 11(5):4490–4501.

Jithish, J., Alangot, B., Mahalingam, N., and Yeo, K. S.

(2023). Distributed anomaly detection in smart grids:

A federated learning-based approach. IEEE Access,

11:7157–7179.

Li, Y., Wei, X., Li, Y., Dong, Z., and Shahidehpour,

M. (2022). Detection of false data injection at-

tacks in smart grid: A secure federated deep learn-

ing approach. IEEE Transactions on Smart Grid,

13(6):4862–4872.

Massaoudi, M., Abu-Rub, H., Refaat, S. S., Chihi, I., and

Oueslati, F. S. (2021). Deep learning in smart grid

technology: A review of recent advancements and fu-

ture prospects. IEEE Access, 9:54558–54578.

Naidji, I., Ben Smida, M., Khalgui, M., Bachir, A., Li, Z.,

and Wu, N. (2019). Efﬁcient allocation strategy of en-

ergy storage systems in power grids considering con-

tingencies. IEEE Access, 7:186378–186392.

Naidji., I., Mosbahi., O., Khalgui., M., and Bachir., A.

(2019). Cooperative energy management software for

networked microgrids. In Proceedings of the 14th

International Conference on Software Technologies -

ICSOFT, pages 428–438. INSTICC, SciTePress.

Naidji, I., Mosbahi, O., Khalgui, M., and Bachir, A. (2020).

Two-stage game theoretic approach for energy man-

agement in networked microgrids. In van Sinderen,

M. and Maciaszek, L. A., editors, Software Technolo-

gies, pages 205–228, Cham. Springer International

Publishing.

Naidji, I., Smida, M. B., Khalgui, M., and Bachir, A.

(2018). Non cooperative game theoretic approach

for residential energy management in smart grid. In

The 32nd Annual European Simulation and Modelling

Conference, pages 164–170, Ghent, Belgium.

Rehman, A. U., Hafeez, G., Albogamy, F. R., Wadud, Z.,

Ali, F., Khan, I., Rukh, G., and Khan, S. (2021). An

efﬁcient energy management in smart grid consider-

ing demand response program and renewable energy

sources. IEEE Access, 9:148821–148844.

Sun, T., Li, D., and Wang, B. (2021). Decentralized feder-

ated averaging.

Wen, M., Xie, R., Lu, K., Wang, L., and Zhang, K.

(2022). Feddetect: A novel privacy-preserving fed-

erated learning framework for energy theft detection

in smart grid. IEEE Internet of Things Journal,

9(8):6069–6080.

Zhu, W., Zeng, Y., Kang, Z., and Fu, J. (2020). Deep learn-

ing based short term load prediction in smart grids.

In 2020 IEEE 3rd International Conference on Elec-

tronic Information and Communication Technology

(ICEICT), pages 674–678.

ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics

294