FedTrees A Novel Computation-Communication Efficient Federated Learning Framework Investigated in Smart Grids_2

2025-04-27 0 0 2.35MB 21 页 10玖币

侵权投诉

FedTrees: A Novel Computation-Communication Eﬃcient Federated Learning

Framework Investigated in Smart Grids

Mohammad Al-Quraana,∗

, Ahsan Khana, Anthony Centenoa, Ahmed Zohaa, Muhammad Ali Imrana,

and Lina Mohjazia

aJames Watt School of Engineering, University of Glasgow, Glasgow, G12 8QQ, UK,

(e-mail: {m.alquraan.1, a.khan.9}@research.gla.ac.uk,

{Anthony.Centeno, Ahmed.Zoha, Muhammad.Imran, Lina.Mohjazi}@glasgow.ac.uk).

Abstract

Smart energy performance monitoring and optimisation at the supplier and consumer levels is essential to

realising smart cities. In order to implement a more sustainable energy management plan, it is crucial to

conduct a better energy forecast. The next-generation smart meters can also be used to measure, record, and

report energy consumption data, which can be used to train machine learning (ML) models for predicting

energy needs. However, sharing ﬁne-grained energy data and performing centralised learning may compromise

users’ privacy and leave them vulnerable to several attacks. This study addresses this issue by utilising

federated learning (FL), an emerging technique that performs ML model training at the user level, where

data resides. We introduce FedTrees, a new, lightweight FL framework that beneﬁts from the outstanding

features of ensemble learning. Furthermore, we developed a delta-based early stopping algorithm to monitor

FL training and stop it when it does not need to continue. The simulation results demonstrate that FedTrees

outperforms the most popular federated averaging (FedAvg) framework and the baseline Persistence model for

providing accurate energy forecasting patterns while taking only 2% of the computation time and 13% of the

communication rounds compared to FedAvg, saving considerable amounts of computation and communication

resources.

1. Introduction

Fifth-generation (5G) and beyond wireless technologies are expected to unlock the full potential of the

Internet of Things (IoT), the key enabler of the smart city model [1]. The smart city concept combines

several elements such as a smart environment, mobility, living, and energy to improve citizens’ quality of life.

Fulﬁlling smart energy, in the form of smart grids and smart buildings, has been the focus of many bodies in

industry and academia through monitoring and predicting energy consumption patterns [2]. Accurate short-,

mid-, or long-term energy forecasting is the ultimate goal that helps managers or consumers to prepare better

future plans, thus improving energy performance.

∗Corresponding author

arXiv:2210.00060v1 [cs.LG] 30 Sep 2022

Energy suppliers need to maintain an equilibrium point between supply and demand, since producing ex-

cessive amounts of energy will result in energy wastage. In contrast, failure to meet consumers’ demands may

lead to the need to purchase energy at higher rates; otherwise, frequent blackouts will happen. Therefore,

various load forecasting techniques have been considered for electricity networks’ eﬃcient and reliable opera-

tion. Statistical forecasting methods, e.g., multiple linear regression (MLR), autoregressive (AR), and moving

average (MA) techniques, were used to project past and present load proﬁles into future predictions. Later,

the introduction of smart metering and the evolution of artiﬁcial intelligence (AI) technology paved the way

for replacing traditional prediction techniques with various machine learning (ML) algorithms, due to their

ability in analysing large amounts of datasets in short periods of time while providing impressive accuracy

levels [3]. Advanced metering infrastructure (AMI), a system of smart meters connected to a communication

network for two-way communications between customers and utility companies, is the ﬁrst step toward smart

energy, which helps collect and analyse smart-meter data. However, collecting consumers’ load proﬁles into

a central entity to conduct energy forecasting raises privacy concerns. Individuals’ load information could be

misused by revealing consumer habits and household occupancy.

To address this issue, the ML community recently introduced a new learning paradigm termed as federated

learning (FL) [4]. The FL is analogous to the concept of distributed learning in terms of handling enormous

datasets and developing eﬃcient and scalable systems. However, maintaining data privacy is the goal of FL

as it does not involve the collection of data in a central location, but instead it sends the model to the clients

where the data is generated. The FL framework is orchestrated by a server placed in a central entity, i.e.,

an energy supplier, to train and improve a shared model with many clients collaboratively. Two typical FL

architectures exist based on the scale of the federation. The ﬁrst is cross-device, where the number of clients

may be massive, for example, consumers’ smart meters. The second is cross-silo, which considers relatively

limited and reliable clients, for example, substations. The FL process starts by initialising a global model

in the server and then sending it to the clients to conduct model training. Once completed, the clients send

back the model updates to the server, which will aggregate them, resulting in an updated model. Then, the

updated model is sent to the clients for another training round. This process is repeated until the limit of

communication rounds is reached, or the model achieves the desired accuracy.

The use of FL in energy forecasting is still in very early stages, and few studies have considered this

approach [5, 6, 7, 8, 9]. In these studies, the authors focus on utilising long short-term memory (LSTM)

architectures, a type of recurrent neural network (RNN) used in the ﬁeld of deep learning (DL), due to their

remarkable performance in predicting time-series data sequences. However, the mentioned works overlook a

critical issue: DL models are extremely resource-consuming (energy, memory, processor, etc.), and the lengthy

and extensive underlying mathematical operations demand resource-rich hardware. Considering such schemes

of combining FL with LSTM models requires extended computation time to reach the desired precision and

impede their scalability. Furthermore, individual households are the focal point of the abovementioned studies

to be used as FL clients. Conversely, our study applies FL at the substation level, i.e., a part of the power

system dedicated to servicing local dwellings in a speciﬁc area, allowing for the use of FL algorithm within

one or more energy suppliers, hence attaining more generalised ML models.

In this paper, we propose FedTrees; a novel light aggregation algorithm developed to utilise decision trees

(DTs) under the FL setting. Speciﬁcally, we use the light gradient boosting model (LGBM) [10], one of the

boosting techniques in ensemble learning, to be sent and trained across the clients of the FL framework.

The main reasons for considering the LGBM models are due to their rapid training speed, less memory

consumption, higher eﬃciency, and accurate predictions. FedTrees attempts to minimise computations and

the number of communication rounds while guaranteeing high training performance; hence it is envisioned to

play a crucial role in a wide range of FL-based applications in several ﬁelds, such as smart energy. Moreover,

this work considers the common challenge of optimising the number of communication rounds that may

lead to suboptimal performance or excessive rounds of unnecessary training, thus consuming computation

and communication resources. Therefore, we developed a delta-based early stopping patience technique,

a dynamic algorithm that monitors the FL training process and stops it when no further enhancement is

possible. Also, this paper examines the importance of each feature used in the training process and oﬀers a

study of the eﬀect of using a diﬀerent number of features on the ﬁnal training performance. Additionally, the

performance of the FedTrees algorithm is benchmarked against the popular LSTM-based federated averaging

(FedAvg) and the naïve Persistence model. Finally, the proposed framework is evaluated based on state-of-

the-art metrics and by conducting extensive simulations. The main contributions of the paper are summarised

as follows:

•First, we introduce FedTrees, a novel, light algorithm that employs DT-based models within the FL

setup. FedTrees using LGBM models shows improved performance in terms of the required number of

communication rounds and computation time compared to LSTM-based FedAvg aggregation scheme

when employed for energy forecasting.

•Unlike other FL-based energy forecasting framework that rely on ﬁxing the number of communication

rounds, this study develops a delta-based early stopping technique to reach the best possible accuracy

while reducing the computation and communications costs.

•A feature importance evaluation study is conducted against individual load proﬁles for optimal fore-

casting performance.

•Finally, we compared FedTrees performance with LSTM-based FedAvg and a naïve Persistence model

as a benchmark using state-of-the-art metrics. The results reveal a signiﬁcant improvement in overall

performance using the FedTrees framework.

The remainder of this paper is structured as follows. Section 2 recalls the related work, while Section 3

presents some background notions. Section 4 describes the proposed FedTrees algorithm versus FedAvg and

represents the delta-based early stopping patience technique. Simulation setup, used dataset, and numerical

results are presented in Section 5. Finally, the concluding remarks are drawn in Section 6.

2. Related Work

This section ﬁrst reviews the works that use diﬀerent load forecasting methodologies performed either

locally or centrally; then, it presents the state-of-the-art research that takes into account the FL framework

for load forecasting.

2.1. Local and Centralised Approaches

Load proﬁles store energy/power consumption information in the form of time-series data, which can

be predicted using several approaches, such as statistical and computational intelligence. For a complete

view of the energy forecasting methods, we refer the reader to [11]. Statistical methods have been used

in the literature for the short and medium temporal forecasting ranges and show quite good performance.

For instance, the study in [12] uses the MLR technique to verify its reliability in energy demand forecasting

instead of traditional and more complex methods. The authors in [13] rely on the traditional AR methodology

to predict electrical energy in Lebanon. Using univariate historical time series data, the second-order MA

model is combined with another statistical method for forecasting gas consumption in China from 2009 to

2015 [14]. Moreover, combining AR and MA resulted in many variants that enhance the forecasting process

such as the autoregressive moving average (ARMA) [15], ARMA model with exogenous inputs (ARMAX)

[16], AR integrated MA (ARIMA) model [17], and seasonal ARIMA (SARIMA) [18].

Later, the focus on ML methods became dominant owing to the advantages of AI in analysing large

amounts of data, which was made available by the introduction of smart metering. The contribution in

[19] highlights the use of support vector machine (SVM) to forecast the monthly electrical load of Taiwan.

On the other hand, many studies use NNs in the energy forecasting domain by virtue of their impressive

performance in various areas. In [20], the work uses electricity time-series data from three countries to

compare the performance of convolutional NN (CNN) and multilayer perceptron (MLP) algorithms. Similarly,

the authors in [21] demonstrate that a three hidden layer CNN model performs well in energy forecasting

compared to statistical and other simple ML methods. RNN is also widely used in smart energy, particularly

LSTM and its variants [22, 23].

Apart from NNs, DTs have also been used in the energy forecasting task. Although the DT approach

is simple, it shows desired performance when predicting future energy consumption in Hong Kong [24].

Following the same DT approach, the authors in [25] demonstrate that employing DT for predicting energy

use intensity in Japanese residential buildings can give very well accuracy. However, DT alone has not been

widely used because it suﬀers from several drawbacks, such as instability, easily losing generalisation, and

performing poorly with noisy and non-linear data. Later, ensemble-based algorithms gained popularity and

were explored in the energy research domain. In general, ensemble techniques possess attractive features that

draw the research community’s attention, such as simplicity, ease of use, interpretability, and computational

eﬃciency. For instance, random forest forecasting performance is compared with NN in [26], and the results

demonstrated that both are feasible and eﬀective in building energy applications. The gradient-boosted

decision tree (GBDT) and extreme gradient boosting (XGBoost) algorithms are also utilised in predicting

future electricity loads and have shown to be eﬀective in [27] and [28], respectively.

The proposed studies and frameworks mentioned above are generally based on a centralised model training

where the energy consumption information is transmitted across the network and combined in a central

location. However, this training scheme raises privacy concerns. Consumers’ load proﬁles hold sensitive

information that can be used in various dimensions like inferring occupancy and usage patterns of households,

government surveillance, data selling, and illegal data use among others.

2.2. FL-based Load Forecasting

Recently, the ML community has investigated a new research direction to address the privacy concerns

associated with traditional training methods. As a distributed learning algorithm, FL can perform model

training at the edge of the network rather than a central location without the need to share sensitive load

information. Few studies have begun to consider the use of FL in the context of smart energy; for example,

the aim of the work in [5] is to evaluate the use of federated settings in predicting electrical load consumption

patterns that will assist in load monitoring and energy demand response. The FedAvg technique is applied

to aggregate the parameters of LSTM models while conducting training, and to produce a generalised global

model. Applying this framework to several households in the USA, the study demonstrated that FL has a

great potential in smart grids through training a powerful future energy forecasting model. Similarly, FedAvg

and LSTM are adopted in [6] to provide a generalised electrical load forecasting model. The authors use

complementary features related to calendar and weather conditions along with the sequences of previous

electrical load to improve the forecast model.

Briggs et al. [7] present a study on the importance of using smart meters in residential areas to forecast

energy consumption using FL, which helps move towards renewable energy generation. LSTM algorithm

is considered to perform short-term forecasting tasks. Similarly, the contribution in [8] evaluates the per-

formance of FL versus centralised and local training methods when using LSTM models in the context of

electrical load forecasting. In addition, clustering is considered to group the clients based on similar model

hyperparameters and, accordingly, similar data characteristics. The study concluded that local learning is

better for predicting individual energy consumption than FL. However, FL is needed when a generalised

forecasting model is required and access to aggregated data is impossible. Very recently, the research work

in [9] adopted the LSTM algorithm under the FL setting to forecast energy proﬁles. Two strategies are con-

sidered, namely, federated stochastic gradient descent (FedSGD) and FedAvg to perform models’ parameter

aggregation. Experimental results demonstrated that FedAvg achieves better accuracy and requires fewer

communication rounds.

The aforementioned FL-based energy forecasting studies rely on DL algorithms, speciﬁcally LSTM net-

works. Although the LSTMs have shown to achieve excellent prediction accuracy, they require intensive

processing duties that yield a heavy computational burden. The following section provides a background to

the main elements considered for conducting this study.

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：

FedTrees:ANovelComputation-CommunicationEcientFederatedLearningFrameworkInvestigatedinSmartGridsMohammadAl-Quraana,,AhsanKhana,AnthonyCentenoa,AhmedZohaa,MuhammadAliImrana,andLinaMohjaziaaJamesWattSchoolofEngineering,UniversityofGlasgow,Glasgow,G128QQ,UK,(e-mail:{m.alquraan.1,a.khan.9}@research.gl...

展开>> 收起<<

FedTrees A Novel Computation-Communication Efficient Federated Learning Framework Investigated in Smart Grids_2.pdf

共21页,预览5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！

FedTrees A Novel Computation-Communication Efficient Federated Learning Framework Investigated in Smart Grids_2

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: