1562 IEEE TRANSACTIONS ON SMART GRID, VOL. 15, NO.

2, MARCH 2024

Smart Home Energy Management: VAE-GAN

Synthetic Dataset Generator and Q-Learning
Mina Razghandi , Graduate Student Member, IEEE, Hao Zhou , Graduate Student Member, IEEE,
Melike Erol-Kantarci , Senior Member, IEEE, and Damla Turgut, Senior Member, IEEE

Abstract—In recent years, there has been a growing interest in research include new technologies and strategies for man-
academia and industry in the analysis of electrical consumption aging energy generation, storage, supply, and demand [1].
in residential buildings and the implementation of smart home Smart homes, as a critical component of the smart grid,
energy management systems (HEMS) to reduce household energy
usage and costs. HEMS have been designed to emulate the statis- are projected to increase household energy efficiency, save
tical and functional characteristics of real smart grids. However, energy costs, and improve user comfort [2]. Utilities and
a major challenge in this research area is the limited availabil- smart home controllers can acquire information from smart
ity of publicly accessible datasets. To address this challenge and meters for use in forecasting consumption and genera-
further leverage the potential of artificial HEMS applications, it tion, demand-side management, and economic power dis-
is crucial to develop time series that accurately represent diverse
operating conditions of synthetic systems. This paper introduces patch [3]. Thus, acquiring fine-grained data about the living
a novel approach based on the combination of variational auto- environment is becoming an essential precondition for a
encoder-generative adversarial network (VAE-GAN) techniques smart home.
to generate time-series data of energy consumption in smart It has become increasingly common to use artificial intelli-
homes. Additionally, we investigate the performance of the gener- gence and statistical analysis for demand response and energy
ative model when integrated with a Q-learning based HEMS. The
effectiveness of the Q-learning based HEMS is assessed through management tasks requiring accurate short-term (second-to-
online experiments using real-world smart home data. To evaluate minute scale) representations of load behavior. These tech-
the quality of the generated dataset, we employ various metrics niques usually require large datasets of representative data for
including Kullback–Leibler (KL) divergence, maximum mean training. However, the collection of such data presents sig-
discrepancy (MMD), and the Wasserstein distance, which quan- nificant security and privacy challenges, with relatively few
tify the disparities between probability distributions of the real
and synthetic data. Our experimental results demonstrate that high-quality publicly available datasets [4].
the synthetic data generated by VAE-GAN closely aligns with the Therefore, synthetic data generation become an attrac-
distribution of real data. Furthermore, we demonstrate that the tive alternative to training machine learning algorithms that
utilization of the generated data facilitates the training of a more can decide, for example, the optimal time for implement-
efficient Q-learning based HEMS, surpassing the performance ing demand response or charging an EV [5]. A variety of
achieved with datasets generated using baseline approaches.
techniques had been used to generate such datasets, includ-
Index Terms—Synthetic data, load consumption, smart grid, ing Markov chains [6], statistical models [7], and physical
deep learning, generative adversarial network, q-learning. simulator-based methods [8]. A major drawback of these
methods can be attributed to targeting a limited number
I. I NTRODUCTION of problems, applications tailored to specific scenarios, and
the lack of scalability. Specifically, these methods require
HE DESIGN of modern smart homes must take
T into consideration energy-efficient technologies and
renewable energy sources. Recent advances in smart grid
case-by-case analyses for each device, user behavior, and
environment. But it is impractical to model all the detailed
energy consumption behaviors and state transitions when a
Manuscript received 22 August 2022; revised 29 December 2022; accepted large number of user devices are involved. Recent studies, on
14 May 2023. Date of publication 30 June 2023; date of current ver- the other hand, have utilized deep learning-based approaches
sion 21 February 2024. The work of Hao Zhou and Melike Erol-Kantarci that can be applied to large-scale datasets and can be con-
was supported in part by the Natural Sciences and Engineering Research
Council of Canada (NSERC); in part by the Collaborative Research and ducted directly on raw data, without the need for further
Training Experience Program (CREATE) under Grant 497981; and in part analysis [9].
by the Canada Research Chairs Program. Paper no. TSG-01232-2022. In our recent paper [10], our approach involved generat-
(Corresponding author: Mina Razghandi.)
Mina Razghandi and Damla Turgut are with the Computer Science ing electric load profiles and PV generation synthetic data
Department, University of Central Florida, Orlando, FL 32816 USA (e-mail: for a smart home using variational autoencoder-generative
[email protected]). adversarial networks (VAE-GAN), and we compared the dis-
Hao Zhou is with the School of Electrical Engineering and Computer
Science, University of Ottawa, Ottawa, ON K1N 6N5, Canada. tributions of synthetic data for the VAE-GAN model and
Melike Erol-Kantarci is with the School of Information Technology and vanilla GAN network to the real data distributions. According
Engineering, University of Ottawa, Ottawa, ON K1N 6N5, Canada. to the statistical metrics such as KL divergence, Wasserstein
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TSG.2023.3288824. distance, and maximum mean discrepancy, the distribution
Digital Object Identifier 10.1109/TSG.2023.3288824 of data points was extremely close between synthetic data
1949-3053 
c 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Adichunchanagiri University. Downloaded on April 13,2024 at 06:02:41 UTC from IEEE Xplore. Restrictions apply.
RAZGHANDI et al.: SMART HOME ENERGY MANAGEMENT 1563

and real data for both PV power generation and energy The main contributions of this paper are as follows:
consumption load profiles. Nevertheless, it is a key defin- • We propose a VAE-GAN-based scheme to generate high
ing characteristic of a realistic time series to maintain an temporal resolution synthetic time-series data for energy
awareness of temporal differences between synthetic and real- consumption in a smart home.
time series. For instance, If PV power production shifted • We compare the performance of the proposed approach
earlier or later in the day rather than following sunrise with techniques based on Gaussian Mixtures and vanilla
and sunset patterns, the distribution would still be similar GANs.
to the real data, but it would not be realistic. The addi- • We further investigate the quality of the synthetic data
tion of electric vehicles (EV) charging load consumption by using it to train a Q-learning-based HEMS model.
synthetic data could provide additional insights and value Evaluating the HEMS agent online in environments based
in understanding smart home electricity consumption pat- on real-world data, we find that the VAE-GAN method
terns. Also, no experiments were conducted to investigate allows higher HEMS profit than other baselines.
if the synthetic dataset is applicable in real-life practice. The rest of this paper is organized as follows. Section II
To address these shortcomings, we propose a revised archi- introduces related work. Section III shows the proposed smart
tecture for the variational autoencoder-generative adversarial home synthetic data generation model and Q-learning-based
network (VAE-GAN) technique for generating time-series data HEMS model. Section IV presents the simulation settings and
of smart homes from a variety of synthetic data sources results, and Section V concludes the paper.
and investigate the performance of synthetic data genera-
tive models in the presence of control operations. In contrast
to the schemes mentioned above, this strategy enables the II. R ELATED W ORK
learning of various types of data distributions in a smart The majority of previous work in generating synthetic
home, including electric load profiles, PV power generation, energy consumption data for smart homes can be classi-
and EV charging load consumption, and can subsequently fied into model-based or data-driven approaches. Model-
generate plausible samples without carrying out any prior based approaches describe the features of household devices
analysis before the training phase. The VAE-GAN architec- with hand-crafted features and mathematical equations. For
ture is favored in this model since it allows us to fine-tune instance, [14] proposes a bottom-up residential building energy
the regulated latent space that influences the generated out- simulation, which simulates occupant behavior patterns with
put, as opposed to vanilla GAN, in which the generator a Markov chain clustering algorithm. Reference [6] combined
maps the input noise to the generated output and makes non-intrusive load decomposition and Markov chain meth-
it susceptible to mode collapse, in which the generator ods for user energy consumption simulation. Reference [8]
repeatedly produces the same optimal input sequence to mis- defines a smart residential load simulator based on MATLAB-
lead the discriminator [11]. Furthermore, we compare one Simulink, and it includes dedicated physical models of various
model-based and two data-driven generative models, includ- household devices. References [7] and [15] utilize Gaussian
ing Gaussian Mixture Model (GMM), vanilla GAN, and Mixture Modeling to estimate the distribution of the arrival
VAE-GAN. and departure time of electric vehicles (EVs) and to perform
Taking advantage of the synthetic data, we propose a temporal modeling of the charging sessions.
Q-learning-based smart home energy management system On the other hand, recent developments in using GANs
(HEMS). The generated data is used for the offline training have achieved great success in producing synthetic time-series
of Q-learning HEMS agents, which aims to maximize long- data. Reference [16] coupled non-intrusive load decomposi-
term management profit. Then, the trained agents are tested tion with a conditional GAN to generate synthetic labeled
online in an environment based on real-world data. Finally, (e.g., appliances) load for a smart home. For smart grid appli-
we compare online profits to further investigate how the data cations, [17] proposes a GAN-based approach that captures
generation method will affect the HEMS performance. The spatial and temporal correlations between renewable energy
idea behind this scheme is that we assume good-quality syn- sources. Reference [18] applies a GAN to the features derived
thetic data in the offline training can better prepare the agent by ARIMA and Fourier transform for generating realistic
for real-world operations [12]. Thus, the performance of the energy consumption data. Reference [4] uses a GAN to learn
Q-learning-based HEMS is used to demonstrate the quality of the level and pattern features of smart home time-series data,
the synthetic data. both of which are defined by household consumption and
Compared with conventional model-based optimization activity, and generate synthesized data with a distribution sim-
algorithms, such as convex optimization, the reinforcement ilar to the real data. Reference [19] clusters daily load profiles
learning (RL) based method can avoid the complexity of defin- with known clustering techniques and uses a GAN to synthe-
ing a sophisticated optimization model since the optimization size daily loads for each cluster. Reference [20] proposed a
problem can be transformed into a unified Markov decision GAN with a Wasserstein loss function to learn temporal and
process (MDP) scheme. On the other hand, most existing power patterns of EV charging sessions and create a synthetic
HEMS models assume a perfect environment for data col- load dataset.
lection and algorithm training [13]. By contrast, the proposed Reinforcement learning-based energy management models
Q-learning HEMS uses synthetic data for training, overcoming have been extensively investigated in the smart grid. For
the data availability bottleneck. example, a correlated Q-learning-based approach is proposed

Authorized licensed use limited to: Adichunchanagiri University. Downloaded on April 13,2024 at 06:02:41 UTC from IEEE Xplore. Restrictions apply.
1564 IEEE TRANSACTIONS ON SMART GRID, VOL. 15, NO. 2, MARCH 2024

Fig. 1. Overall architecture of the proposed scheme.

in [5] for microgrid energy management, and a multi-agent space. Then, the supervisor module trains the encoder module
reinforcement learning method is introduced in [21] to min- to approximate the next time step. In the generator module,
imize the energy cost of different smart home devices. A the input sequence is reconstructed from the latent space in
self-learning HEMS model is presented in [22], which includes an attempt to fool the discriminator. On the other hand, the
price forecasting, price clustering, and power alert systems. discriminator module trains the generator to create realistic
Reference [23] introduced a Bayesian deep reinforcement sequences by identifying fake samples.
learning method for microgrid energy management under In the following, we first introduce the autoencoder (AE)
communication failure. and variational autoencoder (VAE), then we present generative
However, most aforementioned make the strong assumption adversarial network (GAN) and VAE-GAN generative models.
of a perfect data collection and agent training environment. In Finally, we include the Q-learning-based HEMS control.
a real-world environment, fine-grained and good-quality data
may be inaccessible due to privacy concerns or measurement A. Autoencoder (AE)
errors. As such, synthetic data-based HEMS is much more An autoencoder is a neural network architecture that uses an
realistic since the agent can use pre-generated and fine-grained unsupervised learning technique to compress the input dimen-
data for training. sions into a compressed knowledge representation, called
latent space, and then reverses the compressed knowledge
III. M ETHODOLOGY representation into the original input dimensions. This archi-
tecture has two essential modules: the encoder and the decoder.
The overall architecture of the proposed approach is
The encoder module maps the input sequence (x) into the
described in Fig. 1, and the organization of this work is
meaningful latent space (z), and based on z the decoder mod-
introduced as followings:
ule outputs a reconstruction of the original input sequence (x̂).
• Synthetic data generation: Given a real-world smart
The original input is unlabeled, but the decoder is trained
home dataset, we first apply the VAE-GAN-based method
to make reconstructions as close as possible to the original
to produce synthetic data, including the smart home
input by minimizing the reconstruction error, Lreconstr , which
load profiles, PV power generation, and EV charging
is the distance between the original input and the subsequent
load consumption, which is introduced in the following
reconstruction.
Sections III-A–III-D.
• Q-learning based HEMS training: Next, the generated Lreconstr = x̂ − x2 (1)
synthetic data is used for the offline training of Q-learning
based HEMS. The intelligent HEMS agent will interact where x and x̂ are original input and the reconstruction,
with a synthetic data-based environment and produces respectively.
HEMS strategies to maximize long-term profit, which is
included in Section III-E. B. Variational Autoencoder (VAE)
• Real-world HEMS operation test: Finally, the trained After training the autoencoder model, the decoder module
HEMS agent will be tested in a real-world data-based can produce new content given a random latent space. Due to
environment. The test is designed to evaluate the real- the unregulated latent space, which can lead to severe over-
world performance of the synthetic data-based HEMS, fitting, the decoder may not be able to interpolate indefinitely
providing an in-depth evaluation of the synthetic data in the absence of data points included in the input sequence.
quality. Specifically, it is expected to demonstrate that Latent space regularity is reliant on the distribution of the
our synthetic data can be applied to HEMS training, and input sequence and the dimension of the latent space, hence
the agent can obtain a satisfying real-world performance it is not always guaranteed. As a solution to this limitation,
without touching real-world data. variational autoencoders include a regularization parameter,
Among the well-known deep learning-based generative such as Kullback–Leibler (KL) divergence, in the learning pro-
models, GAN and Variational Autoencoders are two of the cess to achieve the Gaussian distribution of the latent space
best known. In this network, the encoder module first encodes and avoid overfitting. In this architecture, the VAE encoder
the input sequence as a Gaussian distribution over the latent (E) encodes the input sequence (x) as a distribution over

Authorized licensed use limited to: Adichunchanagiri University. Downloaded on April 13,2024 at 06:02:41 UTC from IEEE Xplore. Restrictions apply.
RAZGHANDI et al.: SMART HOME ENERGY MANAGEMENT 1565

Fig. 2. VAE-GAN model architecture. In this network, the encoder mod-

ule encodes the input sequence as a Gaussian distribution over the latent
space, defined by mean and variance vectors. The supervisor module trains
the encoder module to approximate the next time step closely in the latent
space. In the generator module, the input sequence is reconstructed from the
latent space in an attempt to fool the discriminator so the generated sequence
is considered real. The discriminator module trains the generator module to
create realistic sequences by identifying fake samples from real ones.

latent space defined by Gaussian distribution first moment,

mean, and second moment, standard deviation, and parameters.
Fig. 3. VAE-GAN encoder module structure.
Ultimately, the encoder is trained to minimize the Lprior loss
and output a latent space with a Gaussian distribution. In the
training process, the VAE minimizes a loss function (LVAE )
that consists of two terms: the reconstruction term and the
regularization term, which is the KL divergence between the
latent space distribution and standard Gaussian distribution.
Lprior = DKL (E(x)||N (0, 1)) (2)
LVAE = Lprior + Lreconstr (3)
where DKL is the KL divergence, E is VAE encoder, N (0, 1)
is the Gaussian distribution, and Lprior is the loss.

C. Generative Adversarial Network (GAN)

Generative models, such as GAN, are used to generate
new data samples. This architecture consists of two neural Fig. 4. VAE-GAN generator, supervisor, and discriminator modules
networks: one attempts to detect and learn patterns in the input structure.
data and produce a new sample, while the other aims to dis-
tinguish between real input data and the synthesized sample. shortcomings of the vanilla GAN or VAE. For example, vanilla
These two networks play against each other and actively seek GAN is prone to mode collapse, a situation where the gener-
equilibrium. ator finds an input sequence that fools the discriminator and
• Generator (G): Takes an input noise z, which usually has repeatedly produces it over and over again. The VAE-GAN
a normal distribution, and maps that into data samples architecture, introduced by Larsen et al. [24], generates new
G(z; θg ), with θg as the network parameter. data samples based on a regulated latent space rather than
• Discriminator (D): D(x; θd ) returns the likelihood of producing new data samples from a noise input. The discrim-
x being classified as a member of the original data, where inator will further classify the generated sample. Considering
0 represents fake data and 1 represents original data. that the data for smart homes is time-series data, it is essen-
G and D engage in an adversarial game, equation (4), where tial that the encoder be able to not only detect the underlying
G attempts to fool the D and maximize the final classification pattern and distribution in the input data but also retain the
error, while D is trained to classify fake data more accurately realistic order in which events occur. As a result, we took
and minimize the error. TimeGAN’s [25] concept of incorporating a supervisor module
  into our VAE-GAN architecture. We will discuss each module
min maxLGAN (D, G) = Ex log(D(x))
G D in more detail.
 
+ Ez 1 − log(D(G(z))) (4) As opposed to mapping the input sequence into a meaning-
ful latent space, The encoder module (E) here translates the
where G and D has been defined as generator and discrimina- original input sequence x into two vectors that represent the
tor, x and z are input sequence and noise, respectively. mean (μ) and variance (σ ) of a standard Gaussian distribution.
Minimizing Lprior (equation (2)) forces the encoder to com-
D. Data-Driven Generative Model press the data over a standard Gaussian distribution. μ and σ
To generate synthetic smart home data we adopted a VAE- are mapped into the latent space z using the reparameterization
GAN architecture, as presented in Fig. 2 that avoids the technique. The encoder uses five layers of Dilated One
Authorized licensed use limited to: Adichunchanagiri University. Downloaded on April 13,2024 at 06:02:41 UTC from IEEE Xplore. Restrictions apply.
1566 IEEE TRANSACTIONS ON SMART GRID, VOL. 15, NO. 2, MARCH 2024

Dimensional Convolutions (dilated Conv1D) stacked one on

top of the other (see Fig. 3). This architecture promises simi-
lar performance but lower complexity and faster convergence
compared to the Long Short-Term Memory (LSTM) networks
typically used for time-series analysis. The Conv1D layer
architecture resembles the WaveNet architecture [26], as each
layer has a stride length of one and kernel size of two, but the
dilation rate varies in each layer. As the second layer dilates
at any second of the input sequence (skipping every other
timestep), the third layer dilates at any fourth, etc., the lower
layers are more focused on short-term dependencies while the
higher layers capture long-term dependencies.
The supervisor module computes the distance between the
latent representation at the current time step (z) and the next
time step (ẑ), minimizing Lsupervisor . Added to the encoder Fig. 5. The proposed smart home energy management system.
loss function, equation (6), allows the encoder to approximate
the next time step in the latent space, retaining the original
sequence of events. used for ESS charging or selling to the energy trading
market for profit.
Lsupervisor = z − ẑ2 (5) • EV and other smart home loads: EV and other smart
LE = Lprior + Lsupervisor (6) home loads are pure energy consumers. They will first
receive the power supply from PV or ESS for operation
The generator module (G) is trained to reconstruct the or buy electricity from the market if the internal power
original input sequence, given latent space z as the input supply is insufficient.
by minimizing the Mean Squared Error (MSE) between the • ESS: Finally, the ESS can be a power supplier when dis-
reconstructed sequence and original input sequence (Lreconstr + charging, or a consumer when charging. For charging,
Lprior ). The generator loss also contains the term LdG , com- it uses the surplus PV power or buys electricity from
puted as in equation (8), quantifying the likelihood of the the market. For discharging, it supplies energy to EVs
discriminator classifying the reconstructed sequence as a fake and smart home devices or sells electricity to the energy
sequence. The goal is to make the reconstructed sequence so trading market for profit.
realistic that it can mislead the discriminator. The HEMS model takes advantage of the flexibility of ESS
  to minimize the total energy cost and maximize profit. The
LdG = Ex log(D(G(z))) (7) optimization task of the centralized HEMS model is described
Lgenerator = Lprior + Lreconstr + LdG + Lsupervisor (8) as follows:
   
Lreal = Ex log(D(x)) (9) T buy
  max Ptotal γ psell
t + (1 − γ )pt (13)
Lfake = Ez 1 − log(D(G(z))) (10) t=1

Lnoise = Ez 1 − log(D(N (0, 1)))] (11) s.t. Ptotal = PESS + PPV − PLt − PEV (13a)
LD = Lreal + Lfake + Lnoise (12)  t t t
γ =1 P total
>0 (13b)
The discriminator module (D) is responsible for classifying PESS =P q (13c)
the original input sequence and fake data samples. D mini- t ⎧ ch t
mizes LD , equation (12), in the training process, which has ⎨ −1 ESS charges
qt = 0 ESS unchanged (13d)
three terms. Lreal , equation (9), is the likelihood of original ⎩
1 ESS discharges
input data being classified as fake data, Lfake , equation (10),
is the likelihood of reconstructed data sample being classified PESS
t
Soct+1 = Soct − (13e)
as real, and Lnoise , equation (11), is the likelihood of classi- CESS
fying a random noise input as a real data which is added to Socmin ≤ Soct ≤ Socmax (13f)
LD to improve the convergence of the discriminator. buy
where T is the total optimization period, psell
t and pt repre-
sent the price of selling energy to the energy trading market
E. Q-Learning Based HEMS Control and buying energy from the market, respectively. PESS PV
t , Pt ,
In this section, we introduce the Q-learning-based HEMS L EV
Pt and Pt represent the power of ESS, PV, smart home
model. As shown in Fig. 5, the components of the smart home load, and EV charging load at time slot t, respectively. Ptotal
involved in the energy production and consumption include the denotes the total power consumption/demand of the smart
PV panels, loads, EV, and the ESS: home, and (13a) is the energy balance constraint. 1 is an indi-
• PV: The PV is considered a pure energy supplier. The cator function. 1{Ptotal > 0} = 1 when Ptotal > 0, which
PV power will first serve the energy demand of EVs and means the agent will sell surplus energy for profit; other-
other smart home devices, then the surplus energy can be wise, 1{Ptotal > 0} = 0 if Ptotal < 0, which means buying

Authorized licensed use limited to: Adichunchanagiri University. Downloaded on April 13,2024 at 06:02:41 UTC from IEEE Xplore. Restrictions apply.
RAZGHANDI et al.: SMART HOME ENERGY MANAGEMENT 1567

energy from the market. Pch is the fixed ESS charging power. Algorithm 1 Q-Learning for HEMS
qt = −1, 0, 1 when ESS charges, remain unchanged and dis- 1: Initialize: Q-learning and smart home parameters
charges, respectively. CESS is the fixed capacity of ESS, and 2: Phase 1: Offline training using synthetic data:
Soc is the ESS state of charge (SOC). Equations (13c) to (13e) 3: Input: generated synthetic data of smart home load, EV
are ESS operation constraints, and (13f) is the SOC upper and load, and PV power generation.
lower bound constraint. 4: for episode = 1 to E do
To transform this optimization task into the context of 5: With probability  choose action a randomly.
Q-learning, we define a Markov decision process (MDP) as Otherwise, a = arg max(Q(s, a))
follows. 6: Agent calculates reward based on equation (13).
• State: The agent state is defined as 7: Update agent state {t, St , wg,t } and Q-value:
st = {Soct , PPVt , Pt
Load }, where PLoad = PL + PEV
t t t 8: Q(s, a) = (1 − α)Q(s, a) + α(r + γ maxQ(s , a ))
represents the total energy consumption of smart home. 9: end for
• Action: Based on the total energy demand and PV power 10: Output: Trained Q-learning based HEMS strategy.
generation, the agent decides on the action at = qt , 11: Phase 2: Online test using real-world data:
which indicates the charging, discharging, or remaining 12: Input: Real-world data.
unchanged status of ESS. 13: Deploy the pre-trained Q-learning agent for an opera-
• Reward: By selecting actions intelligently, the agent tion test.
intends to maximize the total profit in the optimization 14: Output: HEMS profit under real-world data.
period, and the reward function is defined by:
 
buy
rt = Ptotal γ psell
t + (1 − γ )pt , (14)
Finally, the Q-learning-based HEMS algorithm is summa-
which is the objective function of our problem formula- rized in Algorithm 1, which consists of two phases. In the
tion equation (13). offline training phase, given the generated synthetic data, the
In Q-learning, the agent aims to maximize the long-term intelligent agent is trained to maximize the total profit. Then,
expected reward in the online test phase, we apply real-world data for online
∞ operation, which aims to test the performance of agents that

V π (s) = Eπ γ n r(sn , an )|s = s0 , (15) are trained on synthetic data from various datasets.
n=0
where s0 is the initial state, n is the number of iterations, Eπ IV. E VALUATION S TUDY
is the expected value under action selection policy π , r(sn , an ) A. Dataset
is the reward of selecting action an under state sn , and γ is the 1) Smart Home Dataset: The iHomeLab RAPT
reward discount factor. A lower γ means focusing on immedi- dataset [27] is a real-world dataset for residential power
ate rewards, and a high γ indicates that future reward is more traces. Five households in Switzerland have been surveyed for
important. 1.5 to 3.5 years with a sampling frequency of 5 minutes. The
Then, we define the state-action value dataset contains residential electricity consumption and PV

Qnew (sn , an ) = Qold (sn , an ) + α rn + γ max Q(sn+1 , a) energy production, including appliance-level and aggregated
a
 household consumption data. Studies of PV power generation
− Q (sn , an ) ,
old
(16) show recurring patterns with corresponding noises originating
from rain and clouds. Consumption patterns show weekly
where Qold (sn , an ) and Qnew (sn , an ) are old and new Q-values, trends, but with occasional irregular spikes.
respectively, indicating the accumulated reward of selecting The experiment was conducted on residential house D from
action an under state sn . sn+1 is the state of n+1 iteration, and the dataset. Data cleansing was done by choosing days with
α is the learning rate. A high learning rate will lead to a fast full 24-hour records, resulting in 594 days in total. As part of
learning process, but the results can be unstable; otherwise, a the training process, electrical load and PV power production
lower learning rate may result in very slow convergence and data were downsampled to a resolution of 15 minutes. For the
a long training time. train and test datasets, a split ratio of 80:20 was adopted.
In addition, we use the -greedy policy for the action 2) Residential Electric Vehicle Charging Dataset:
selection. Sørensen et al. [28] presented a residential electric vehicle

arg max Q(sn+1 , a), rand > , charging dataset recorded from apartment buildings. From
π(s) = a (17) December 2018 to January 2020, 97 users in Norway reported
random action selection, rand ≤ .
real-world EV charging experiences. Each charging session
where rand represents a random number (0 ≤ rand ≤ 1), includes plug-in time, plug-out time, and charged energy.
and  < 1. When rand > , the agent takes greedy policy This dataset provides a synthetic charging load for level-1
to select the action with maximum Q-value; otherwise, the charging and level-2 charging assuming 3.6 kW or 7.2 kW
action is selected randomly for exploration. -greedy policy of charging power. In our experiments, we interpolate the
can balance the exploration and exploitation of the Q-learning data for a 15-minute resolution. We note that a higher time
agent to maximize the long-term expected reward. scale resolution such as 5-minute will increase the simulation

Authorized licensed use limited to: Adichunchanagiri University. Downloaded on April 13,2024 at 06:02:41 UTC from IEEE Xplore. Restrictions apply.
1568 IEEE TRANSACTIONS ON SMART GRID, VOL. 15, NO. 2, MARCH 2024

 
running time, while a lower resolution such as 20 minutes −x − y2
K(x, y) = exp (20)
can not capture the environment dynamics. In addition, the 2σ 2
proposed synthetic data generator scheme can adapt to any
time resolution without loss of generality. 3) Wasserstein Distance: Intuitively, the Wasserstein dis-
Our analysis revealed that the user with the largest number tance, also called the earth mover’s distance, models a proba-
of available records has only 62 days of 24-hour data. This data bility distribution as a pile of soil and computes how much soil
contains only 11 charging sessions, which is insufficient for needs to be moved to transform one probability distribution to
the training of a deep learning model. To address this issue we the other.
generated a larger synthetic dataset by combining the charging 
data from several users and assuming that the data belonged to l1 (p, q) = inf |x − y|dπ (x, y) (21)
π ∈(p,q) R×R
an indoor charging station shared by the building’s residents,
thus accumulating 263 days of historical data, We are thereby Equation (21) represents the formal definition of
able to have more significant charging events for the model, Wasserstein distance, where p(x), q(y) are probability
resulting in a more efficient training process. distributions, (p, q) is the set of probabilistic distributions
on R × R whose marginals represent p and q on the first and
B. Baseline Models and Performance Metrics second moments, respectively.
In our experiments, we compared the performance of our 4) HEMS Models: The previous techniques measured the
VAE-GAN-based approach with two other state-of-the-art quality of the generated data based on its distance from the
approaches: probability distribution of the real data. In this technique, we
• Gaussian Mixture Model (GMM): is a probabilis- measure the quality of the synthetic data by its ability to
tic approach that partitions data into groups using be used in the training of a Q-learning-based smart home
soft clustering based on multiple multidimensional energy management system (HEMS). We implemented the
Gaussian probability distributions. The mean and variance Q-learning model in the MATLAB platform. The ESS fixed
of the distributions are calculated using Expectation- charging/discharging power is 4 kW, and the ESS capacity is
Maximization. Upon fitting the GMM to some data, a 16 kW · h. The learning rate is 0.8, the discount factor is 0.7,
generative probabilistic model can sample synthetic data, and the initial  value for the -greedy exploration algorithm
following the same distribution. is 0.05. The energy trading price follows a fixed pattern as
• Vanilla GAN: As described in Section III-C. in [5]. We recorded the average results over 10 randomized
• VAE-GAN: As described in Section III-D runs.
In the following, we introduce the metrics that we will
employ to evaluate the quality of the generated synthetic data.
C. Distance Metrics Evaluation Results
1) Kullback–Leibler (KL) Divergence: The Kullback-
Leibler divergence (KLD) is one of the most often used We used the KL-divergence, MMD, and Wasserstein dis-
measures for evaluating the similarity of two probability tance metrics to measure how close is the distribution of
distributions. Equation (18) is the formal definition of KL synthetic data to the real data. Table I shows these statisti-
divergence, where x and y are sampled data points, and p(x) cal metrics for smart home electrical load consumption, PV
and p(y) are their respective probability distributions. The KL production, and EV charging load consumption synthetic data
divergence ranges from 0 when two probability distributions generated by the GMM, GAN, and VAE-GAN generative
almost match everywhere, to ∞ for completely different models. For all metrics, the lower the errors are the better,
distributions. corresponding to a closer match between the synthetic and
 N   real data.
p(xi ) The results allow us to draw several conclusions. First, we
DKL (pq) = p(xi )log (18)
q(yi ) find that all three models have relatively close errors for syn-
i=1
thetic EV charging datasets, which can be explained by the
2) Maximum Mean Discrepancy (MMD): The MMD
comparative regularity of EV charging datasets versus PV gen-
approach represents the distance between distributions by the
eration, which is influenced by the weather or load, depending
distance between the mean embeddings of features into a
on many personal choices. Another possible explanation is that
reproducing kernel Hilbert space. Given the distributions p(x)
PV generation and load consumption training data were almost
for {xi }N
i=0 and q(y) for {yj }j=0 , MMD is calculated as follows:
M
twice the size of EV charging data, suggesting that rich his-
1   torical training data could have a positive impact on accuracy.
N N

MMD(p, q)2 = 2
K xi , xj Second, we find that for the KL-divergence and Wasserstein
N
i=1 j=1 distance, our VAE-GAN-based approach almost always pro-
vides the best performance. For the MMD, in contrast, the best
2  
N M

− K xi , yj performance is provided by the GAN approach. This reversal
MN in the relative order of the generator approaches for MMD
i=1 j=1
is likely a result of the choice of mean embedding features.
1  
M M

+ 2
K yi , yj (19) Understanding the exact mechanism of this difference requires
M future work.
i=1 j=1

Authorized licensed use limited to: Adichunchanagiri University. Downloaded on April 13,2024 at 06:02:41 UTC from IEEE Xplore. Restrictions apply.
RAZGHANDI et al.: SMART HOME ENERGY MANAGEMENT 1569

TABLE I
D ISTANCE B ETWEEN R EAL AND S YNTHETIC S MART G RID DATA D ISTRIBUTION U SING KL-D IVERGENCE , WASSERSTEIN D ISTANCE ,
AND MMD. B EST R ESULTS A RE H IGHLIGHTED IN BOLD

Fig. 6. Electrical load consumption real and synthetic data probability density for (a) GMM, (b) GAN, and (c) VAE-GAN generative models. The orange
line shows the real data PDF, the blue line shows the synthetic data PDF.

Fig. 7. PV power production real and synthetic data probability density for (a) GMM, (b) GAN, and (c) VAE-GAN generative models. The orange line
shows the real data PDF, the blue line shows the synthetic data PDF.

Fig. 8. EV charging load consumption real and synthetic data probability density function for (a) GMM, (b) GAN, and (c) VAE-GAN generative models.
The orange line shows the real data PDF, the blue line shows the synthetic data PDF.

While single numerical metrics quantifying the distance Fig. 6(d) shows the PDF for normalized real and synthetic
between the probability density functions are useful, exam- electrical load consumption for each generative model. While
ining the distributions as a whole provides us with a better both the GMM and VAE-GAN distributions are relatively
understanding of the quality of the generated data. Figs. 6–8 close to the real data, the synthetic data generated by VAE-
show the Probability Density Functions (PDF) for real and GAN provides the best matches with the real data distribution
synthetic smart home data. in terms of the standard deviation from the mean value, as

Authorized licensed use limited to: Adichunchanagiri University. Downloaded on April 13,2024 at 06:02:41 UTC from IEEE Xplore. Restrictions apply.
1570 IEEE TRANSACTIONS ON SMART GRID, VOL. 15, NO. 2, MARCH 2024

Fig. 9. A sample of 10 days of synthetic data generated from GAN, GMM, and VAE-GAN models compared with the real test data.

supported by the distance metrics. Fig. 9(a) represents the pat- consumption data and producing samples that reflect the same
tern of real and synthetic electrical load consumption during characteristics.
the test data. It is evident from the close similarity of these Fig. 7(c) illustrates that the VAE-GAN model performs
patterns that the VAE-GAN network is capable of learning better compared to two other generative models in generating
the distribution and patterns of smart home aggregated load synthetic PV production data, a conclusion that is supported

Authorized licensed use limited to: Adichunchanagiri University. Downloaded on April 13,2024 at 06:02:41 UTC from IEEE Xplore. Restrictions apply.
RAZGHANDI et al.: SMART HOME ENERGY MANAGEMENT 1571

Fig. 10. HEMS test performance comparison in 40 days by using real-world data (16 kWh ESS Capacity) The the total profit of optimal based, VAE-GAN,
GAN, and GMM based methods are 97.54$, 86.31$, 85.18$, and 79.36$, respectively.

Fig. 11. Smart home HEMS performance analyses.

by Fig. 9(b), where the pattern of synthetic PV production D. HEMS Performance Evaluation Results
data generated by VAE-GAN is reasonably close to the pat- In this section, we investigate the performance of the
tern of real PV production data. We note that although PV Q-learning-based HEMS. For each type of synthetic data, we
production follows sunrise and sunset patterns, it is also highly trained a corresponding policy, which was then evaluated using
affected by unpredictable environmental factors, therefore it is real-world data (see Algorithm 1). We compare these poli-
not expected for them to be the same equivalent. cies against an optimal baseline that uses real-world data for
Finally, Fig. 9(c) presents the EV charging load consump- both offline training and online operation testing, indicating
tion real and synthetic data PDF for GMM, GAN, and that the Q-learning agent can best learn the hidden data pat-
VAE-GAN generative models. In Fig. 8 it is shown that syn- terns and prepare for the test. The optimal baseline is expected
thetic EV charging load consumption generated by the GMM to achieve the best overall performance by using the same
model is larger in power range compared to the real EV real-world dataset for both training and testing, but it cannot
charging load consumption. On the other hand, the GAN guarantee that the optimal baseline can obtain higher profit
model in Fig. 8(b), generates data centered around the aver- every single day.
age. Although the VAE-GAN model, Fig. 8(c), generates a Fig. 10 shows the 40 days HEMS profit of different data
slightly smaller power consumption than actual EV charging generation methods. While the profit varies with the envi-
data, the distribution of load consumption is comparable to ronmental conditions each day, as expected, the strategy
that of real data. Accordingly, each generative model shows trained on real-world data achieves the highest profit on most
the same characteristics in the sample EV data presented in days. However, we found that the HEMS trained on the
Fig. 9(c). data generated by the proposed VAE-GAN model achieves

Authorized licensed use limited to: Adichunchanagiri University. Downloaded on April 13,2024 at 06:02:41 UTC from IEEE Xplore. Restrictions apply.
1572 IEEE TRANSACTIONS ON SMART GRID, VOL. 15, NO. 2, MARCH 2024

a comparable performance. The GAN-based model achieves a approaches, but the availability of find-grained data sets may
somewhat lower performance, while the HEMS trained on the prevent the applications. Therefore, synthetic data generation
GMM-generated data shows the least profit. has become a critical enabler for smart home management.
The performance of the HEMS also depends on the capacity In this paper, we present a variational autoencoder GAN
of the energy storage system (ESS) of the smart home. The (VAE-GAN) approach for the synthetic data generation of
average profit of the differently trained HEMS systems for smart home datasets. We have shown that the proposed
various ESS capacities is shown in Fig. 11(a). As expected, VAE-GAN method can generate high-quality data that better
a larger ESS improves the profit of the HEMS, as the agent matches the statistical properties of real-world data compared
has higher flexibility in making decisions. We found that for to benchmarks. In addition, we use the synthetic data to train
all ESS values, the VAE-GAN-based HEMS performs best, a Q-learning-based smart home energy management system
followed by the GAN and GMM-based methods. The higher (HEMS), achieving a higher profit than other data-generation
the ESS capacity increases the advantage of the VAE-GAN methods. The simulations prove that the generated synthetic
compared to the other methods. data can be used for HEMS offline training, and the agent
Finally, for practical applications, it is essential to establish obtains a satisfying real-world online performance without
that the learning process is stable. Fig. 11(b) shows the reward touching the real-world data.
of the Q-learning process function of the learning iterations.
The figure shows that the agent has a stable convergence after
the exploration phase. R EFERENCES
[1] M. S. Aliero, K. N. Qureshi, M. F. Pasha, and G. Jeon, “Smart
home energy management systems in Internet of Things networks for
E. Complexity and Feasibility Analysis green cities demands and services,” Environ. Technol. Innov., vol. 22,
Based on the simulation results, this section will discuss the May 2021, Art. no. 101443.
[2] B. Sovacool and D. Rio, “Smart home technologies in Europe: A critical
complexity and feasibility of the proposed scheme, including review of concepts, benefits, risks and policies,” Renew. Sustain. Energy
a) feasibility and scalability, b) implementation complexity, Rev., vol. 148, pp. 1–22, Mar. 2020.
c) management costs, and d) economic benefits. [3] A. Al-Ali, I. Zualkernan, M. Rashid, R. Gupta, and M. Alikarar, “A
• Feasibility and scalability: Compared with conventional smart home energy management system using IoT and big data analytics
approach,” IEEE Trans. Consum. Electron., vol. 63, no. 4, pp. 426–434,
model-based methods such as Markov chain and simula- Nov. 2017.
tors, the applied data-driven method has higher scalability [4] C. Zhang, S. R. Kuppannagari, R. Kannan, and V. K. Prasanna,
and can be easily generalized to other scenarios. In the “Generative adversarial network for synthetic time series data
generation in smart grids,” in Proc. IEEE Int. Conf. Commun.
experiment, note that the only requirement for applying Control Comput. Technol. Smart Grids (SmartGridComm), 2018,
the proposed methods is the dataset, indicating that it pp. 1–6.
can be easily generalized from smart homes to smart [5] H. Zhou and M. Erol-Kantarci, “Decentralized microgrid energy man-
agement: A multi-agent correlated q-learning approach,” in Proc.
builds and communities with proper datasets. By con- IEEE Int. Conf. Commun. Control Comput. Technol. Smart Grids
trast, it is impractical to build very huge simulators or (SmartGridComm), 2020, pp. 1–7.
Markov chains to describe all the detailed behaviors of a [6] S. Ge, J. Li, H. Liu, X. Liu, Y. Wang, and H. Zhou, “Domestic energy
consumption modeling per physical characteristics and behavioral fac-
large area. tors,” Energy Procedia, vol. 158, pp. 2512–2517, Feb. 2019.
• Organizational complexity: The main complexity of the [7] M. Lahariya, D. Benoit, and C. Develder, “Synthetic data generator for
proposed method is the neural network architecture and electric vehicle charging sessions: Modeling and evaluation using real-
world data,” Energies, vol. 13, p. 4211, Aug. 2020.
design, which is a well-known issue for ML-based
[8] J. López, E. Pouresmaeil, C. Cañizares, K. Bhattacharya, A. Mosaddegh,
methods. However, in this work, note that no dedi- and B. Solanki, “Smart residential load simulator for energy man-
cated models are designed for any household devices or agement in smart grids,” IEEE Trans. Ind. Electron., vol. 66, no. 2,
energy consumption behaviors, significantly reducing the pp. 1443–1452, Feb. 2019.
[9] I. Goodfellow et al., “Generative adversarial nets,” in Proc. 27th Int.
organizational complexity. Conf. Neural Inf. Process. Syst., Dec. 2014, pp. 1–18.
• Management costs: As above-analyzed, the dataset is [10] M. Razghandi, H. Zhou, M. ErolKantarci, and D. Turgut, “Variational
the only requirement of the proposed method. When the autoencoder generative adversarial network for synthetic data generation
in smart home,” Jan. 2022, arXiv:2201.07387.
physical system changes, e.g., adding or removing new [11] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and
devices, our model can be easily updated by feeding new X. Chen, “Improved techniques for training GANS,” in Proc. Adv.
datasets to the neural networks, which means a very low Neural Inf. Process. Syst., vol. 29, 2016, pp. 2226–2234.
[12] M. Razghandi, H. Zhou, M. Erol-Kantarci, and Turgut, “Smart
system management cost. home energy management: Sequence-to-sequence load forecasting and
• Economic benefits: The simulation results in Fig. 10 Q-learning,” in Proc. IEEE Globecom, Dec. 2021, pp. 1–6.
and 11(a) have demonstrated that the proposed method can [13] C. Tu, X. He, Z. Shuai, and F. Jiang, “Big data issues in smart
obtain satisfying profit for the smart home operation. The grid–A review,” Renew. Sustain. Energy Rev., vol. 79, pp. 1099–1107,
Nov. 2017.
benefit can be greater when considering larger application [14] L. Diao, Y. Sun, Z. Chen, and C. Jiayu, “Modeling energy consumption
scenarios such as smart buildings and communities. in residential buildings: A bottom-up analysis based on occupant behav-
ior pattern clustering and stochastic simulation,” Energy Build., vol. 147,
pp. 47–66, Jul. 2017.
V. C ONCLUSION [15] J. Quirós-Tortós, A. Navarro-Espinosa, L. F. Ochoa, and T. Butler,
“Statistical representation of EV charging: Real data analysis and
In recent years, smart home management increasingly applications,” in Proc. Power Syst. Comput. Conf. (PSCC), 2018,
takes advantage of advanced data-demanding machine learning pp. 1–7.

Authorized licensed use limited to: Adichunchanagiri University. Downloaded on April 13,2024 at 06:02:41 UTC from IEEE Xplore. Restrictions apply.
RAZGHANDI et al.: SMART HOME ENERGY MANAGEMENT 1573

[16] S. Kababji and P. Srikantha, “A data-driven approach for generating syn- Hao Zhou (Graduate Student Member, IEEE)
thetic load patterns and usage habits,” IEEE Trans. Smart Grid, vol. 11, received the B.Eng. degree from the Huazhong
no. 6, pp. 4984–4995, Nov. 2020. University of Science and Technology, China,
[17] Y. Chen, Y. Wang, D. Kirschen, and B. Zhang, “Model-free renewable in 2016, and the M.Eng. degree from Tianjin
scenario generation using generative adversarial networks,” IEEE Trans. University, China, in 2019. He is currently pursuing
Power Syst., vol. 33, no. 3, pp. 3265–3275, May 2018. the Ph.D. degree with the University of Ottawa. His
[18] M. Fekri, A. Ghosh, and K. Grolinger, “Generating energy data research interests include microgrid energy trading,
for machine learning with recurrent generative adversarial networks,” 5G network slicing and mmWave, and reconfig-
Energies, vol. 13, p. 130, Dec. 2019. urable intelligent surfaces. He is devoted to apply-
[19] Z. Wang and T. Hong, “Generating realistic building electrical load pro- ing machine-learning techniques for smart grid and
files through the generative adversarial network (GAN),” Energy Build., wireless network applications. He won the Best
vol. 224, Oct. 2020, Art. no. 110299. Paper Reward at 2023 IEEE ICC conference.
[20] R. Buechler, E. Balogun, A. Majumdar, and R. Rajagopal, “EVGen:
Adversarial networks for learning electric vehicle charging loads and
hidden representations,” 2021, arXiv:2108.03762.
[21] X. Xu, Y. Jia, Y. Xu, Z. Xu, S. Chai, and C. S. Lai, “A multi-agent
reinforcement learning-based data-driven method for home energy man- Melike Erol-Kantarci (Senior Member, IEEE) is
agement,” IEEE Trans. Smart Grid, vol. 11, no. 4, pp. 3201–3211, the Canada Research Chair of AI-Enabled Next-
Jul. 2020. Generation Wireless Networks and an Associate
[22] W. Li, T. Logenthiran, V.-T. Phan, and W. L. Woo, “Implemented Professor with the School of Electrical Engineering
IoT-based self-learning home management system (SHMS) for and Computer Science, University of Ottawa. She
Singapore,” IEEE Internet Things J., vol. 5, no. 3, pp. 2212–2219, is the Founding Director of the Networked Systems
Jun. 2018. and Communications Research (NETCORE)
[23] H. Zhou, A. Aral, I. Brandic, and M. Erol-Kantarci, “Multi-agent Laboratory. She is the co-editor of three books on
Bayesian deep reinforcement learning for microgrid energy management smart grids, smart cities, and intelligent transporta-
under communication failures,” IEEE Internet Things J., vol. 9, no. 14, tion. She has over 200 peer-reviewed publications.
pp. 11685–11698, Jul. 2022. She has delivered over 70 keynotes, plenary talks,
[24] A. B. L. Larsen, S. K. Sønderby, H. Larochelle, and O. Winther, and tutorials around the globe. Her main research interests are AI-enabled
“Autoencoding beyond pixels using a learned similarity metric,” in Proc. wireless networks, 5G and 6G wireless communications, smart grid, and
33rd Int. Conf. Mach. Learn., vol. 48, 2016, pp. 1558–1566. [Online]. Internet of Things. She has received numerous awards and recognitions.
Available: https://proceedings.mlr.press/v48/larsen16.html She is on the editorial board of the IEEE T RANSACTIONS ON C OGNITIVE
[25] J. Yoon, D. Jarrett, and M. Van der Schaar, “Time-series generative C OMMUNICATIONS AND N ETWORKING, IEEE I NTERNET OF T HINGS
adversarial networks,” in Proc. Adv. Neural Inf. Process. Syst., vol. 32, J OURNAL, IEEE C OMMUNICATIONS L ETTERS, IEEE N ETWORKING
2019, pp. 5509–5519. L ETTERS, IEEE Vehicular Technology Magazine, and IEEE ACCESS.
[26] A. Oord et al., “WaveNet: A generative model for raw audio,” Sep. 2016, She has acted as the general chair and the technical program chair for
arXiv:1609.03499. many international conferences and workshops. She is an IEEE ComSoc
[27] P. Huber, M. Ott, M. Friedli, A. Rumsch, and A. Paice, “Residential Distinguished Lecturer and an ACM Senior Member.
power traces for five houses: The iHomeLab RAPT dataset,” Data,
vol. 5, no. 1, p. 17, Feb. 2020.
[28] Å. L. Sørensen, K. B. Lindberg, I. Sartori, and I. Andresen, “Residential
electric vehicle charging datasets from apartment buildings,” Data Brief,
vol. 36, Jun. 2021, Art. no. 107105. Damla Turgut (Senior Member, IEEE) is a
Professor and the Chair of the Computer Science
Department, University of Central Florida. Her
research interests include wireless ad hoc, sensor,
underwater and vehicular networks, cloud comput-
ing, smart cities, smart grid, IoT-enabled health-
Mina Razghandi (Graduate Student Member, IEEE) care and augmented reality, as well as considera-
received the Ph.D. degree in computer science tions of privacy in the Internet of Things. She is
from the University of Central Florida in 2022. also interested in applying big data techniques for
She has dedicated her academic pursuit to explor- improving STEM education for women and minori-
ing the intersection of advanced AI technologies ties. Her recent honors and awards include the
and sustainable, and smart infrastructure. Her Ph.D. NCWIT 2021 Mentoring Award for Undergraduate Research Award, the UCF
research was primarily focused on the application of Research and Teaching Incentive Awards, the UCF Women of Distinction
deep-learning techniques to enhance the capabilities Award, and the University Excellence Award in Professional Service. She
and efficiency of smart homes and smart grids. This serves on several editorial boards and program committees of prestigious ACM
intellectual journey has instilled in her a keen interest and IEEE journals and conferences. She is an IEEE ComSoc Distinguished
in AI, Internet of Things, and energy microgrids. She Lecturer, the Chair of IEEE Technical Committee of Social Networks, an
is continually driven by the desire to leverage these cutting-edge tools and ACM Senior Member, and the Chair of the IEEE Technical Community on
methodologies to create a more connected and energy-efficient world. Computer Communications.

Authorized licensed use limited to: Adichunchanagiri University. Downloaded on April 13,2024 at 06:02:41 UTC from IEEE Xplore. Restrictions apply.