Privacy Constrained Load Profiling using

2023 7th International Conference on Computer Applications in Electrical Engineering-Recent Advances (CERA) | 979-8-3503-0500-5/23/$31.00 ©2023 IEEE | DOI: 10.1109/CERA59325.2023.10455436

Smart Meter Data

Ritika Mahajan Kailash Chand Sharma, SMIEEE Deven Vatsal
Center for Artificial Intelligence Department of Electrical Engineering Department of Electrical Engineering
NIT, Jalandhar NIT, Jalandhar NIT, Jalandhar
Jalandhar, India Jalandhar, India Jalandhar, India
[email protected] [email protected] [email protected]

Abstract—Smart meters have transformed energy management an impressive 96 million. India has ambitious plans to install
by enabling precise measurements, real-time data collection, and around 250 million smart meters by 2026 [1], showcasing
enhanced communication capabilities. The availability of detailed its commitment to modernize the power system and promote
smart meter data presents significant opportunities for load
profiling, a vital component of electricity consumption analysis. smarter energy management practices.
However, privacy concerns arise due to the sensitive nature of The deregulation of the power industry has led to competitive
individual energy consumption patterns. Moreover, challenges electricity retail markets globally. In this context, detailed
related to bottleneck, data heterogeneity, and communication smart meter data plays a crucial role for businesses operating
overhead occur because of the need to aggregate and analyze in the retail market by enabling optimization of pricing models
data from diverse sources, transmit large amounts of data
efficiently, and preserve privacy by minimizing data transmission [2-3], customized energy efficiency solutions, efficient demand
and ensuring secure processing. To address these challenges, response programs [4], and improved customer engagement.
this research paper proposes an unsupervised federated learning Clustering and data mining techniques are widely used for
approach that incorporates collaborative model training and load profiling to understand consumer behavior and facilitate
autoencoder. The proposed approach aims to overcome the effective energy management. These techniques leverage smart
privacy concerns by encoding data locally and reducing com-
munication overhead through the transmission of compressed meter data to identify and group similar load patterns, provid-
representations. Furthermore, it addresses challenges related to ing valuable insights for decision-making and contributing to
bottleneck and data heterogeneity by leveraging collaborative various energy management applications.
model training with Irish (C.E.R) dataset which is particitioned Several studies have conducted comparisons and evaluations
into weekdays and weekend. Numerical attributes are used of clustering algorithms for load profiling. For example,
for evaluation, comparing the proposed approach with model
averaging and gradient sharing methods. The results demonstrate hierarchical clustering, K-means, fuzzy K-means, and self-
the effectiveness of the proposed approach in extracting mean- organizing maps were assessed for electricity customer clas-
ingful features and patterns from unlabeled smart meter data sification. Techniques such as Sammon map, PCA, and CCA
while ensuring privacy, optimizing communication efficiency, and were explored to reduce data size for efficient classification
improving load profiling accuracy. in [5]. Deep learning-based clustering algorithms, such as
Index Terms—Keywords: Load profile analysis, federated
learning, bottleneck problems, gradient sharing, model averag-
deep embedded clustering (DEC) and quantize regression
ing, privacy-preserving, scalability, efficient,Collabrative Model forest models, were introduced in resulting improved baseline
Training with Auto-Encoder estimation performance [6].
A combination of deep convolutional neural network (CNN)
I. I NTRODUCTION and support vector machines was employed for consumer
characterization in [7]. A fuzzy clustering-based model for
Smart meters have rapidly gained popularity in the power mining household monthly electricity consumption patterns
system due to technological advancements and the increasing is propesed for understanding consumer behavior and energy
demand for efficient energy management. These advanced usage through the analysis of electricity consumption data in
devices, serving as modern versions of traditional electricity [8]. Subspace clustering and multi-resolution clustering meth-
meters, are equipped with enhanced communication capa- ods were proposed with subspaces or local shape variability
bilities and real-time data collection features. By accurately within load profiles using subspace projection methods, and
measuring and monitoring energy consumption, smart meters incorporating a finite mixture model (FMM) clustering algo-
empower utilities and consumers to make well-informed deci- rithm was aimed to capture and analyze the distinct variations
sions regarding energy usage, load management, and billing. and sources of variability within the load profiles, respectively,
The global adoption of smart meters has been remarkable, with addressing challenges related to load profile analysis, including
leading countries such as U.K., U.S., China, and India driving distinguishing load shapes and handling large sample sizes [9-
the deployment. For instance, by 2016, the U.K. had installed 10]. The impact of temporal resolution on clustering quality
2.9 million smart meters, U.S. had 70 million, and China had and efficiency was examined in [11]. While these algorithms

979-8-3503-0500-5/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on February 23,2025 at 05:37:18 UTC from IEEE Xplore. Restrictions apply.
 TABLE I: Comprehensive Analysis of Clustering Algorithms and Critical Reviews
Ref.No Clustering Algorithms DataSet Distribution Framework
[5],[11] Hierarchical clustering, K-means, fuzzy K-means, and self-organizing maps Synthetic smart meter data, Metered data Centralized Python/TensorFlow
[6],[7] DEC, CNN, selecting optimal clusters, regression forest mode Residential energy data, smart meter Centralized Python, PyTorch
[8] Fuzzy c-means clustering model Household energy consumption data Centralized Python, Scikit-fuzzy
[9],[12] Subspace projection method based clustering, Temporal resolution Load profiling data, Residential energy load profiles Centralized PySpark, Apache Flink, Tensor-
Flow
[10] Multi-resolution load profile clustering, Gaussian mixture models Smart meter data Centralized Python, scikit-learn
[13] Privacy-preserving accelerated average (PP-aac) Retailer combines load dataset Decentralized R
[16] Model averaging, gradient sharing clustering algorithm Household dataset Decentralized Python
[17] DBSCAN Residential data Decentralized Python, Matlab
[18] Density-based clustering algorithm GPS-AMI dataset Decentralized Python, PyTorch

primarily focused on centralized approaches, recent research Collaborative Federated Learning with Auto encoders for
has emphasized the need for privacy-preserving techniques due Extraction of Consumption Patterns, Section IV conducts
to distributed data ownership and privacy concerns. Evalution criteria, Section V includes Case Study of data set
Protecting the privacy of smart meter data is of utmost and Results and Conclusion is derived from the results in
importance. Differential privacy, which offers a high level of Section VI.
privacy when utilizing smart meter data for K-means clus-
tering [12]. However, not all privacy methods are optimized II. P ROBLEM STATEMENT
for load pattern extraction. A proposed distributed clustering Load profile analysis plays a crucial role in gaining insights
framework was used to ensure the privacy preservation while into electricity consumption patterns and optimizing energy
performing joint clustering on the combined load dataset of management strategies. However, conventional methods en-
retailers which was created to leverages the privacy-preserving counter bottlenecks due to communication overhead, privacy
accelerated average consensus (PP-AAC) algorithm and adapts concerns, and the decentralized nature of data sources. In
clustering methods like k-means, fuzzy c-means, and Gaussian response to these challenges, federated learning has emerged
mixture models [13]. The first proposed Federated learning, as a promising approach to address them while ensuring
a privacy-preserving machine learning framework was intro- data privacy. Nonetheless, significant bottleneck issues arise
duced in 2016 which was widely studied for many application in techniques like k-means clustering, gradient sharing, and
[14]. Different variations, including horizontal, vertical, and model averaging employed in federated learning for load
transfer learning, have been explored based on data distribution profile analysis. These bottlenecks impede scalability, privacy
for federated learning approach [15]. While most existing preservation, and accuracy, thereby limiting the comprehensive
federated learning algorithms focus on supervised learning, and effective achievement of results in load profile analysis
A federated learning approaches was introduced utilizing k- through federated clustering.
means clustering for load profiling privacy. These approaches 1) K-Means: In k-means clustering, the bottleneck arises
employed model averaging and gradient sharing strategies, during the iterative centroid update step, where the com-
demonstrating their effectiveness in preserving data privacy putation of cluster centroids becomes computationally
and achieving load profiling accuracy [16]. expensive as the dataset grows. Mathematically, this can
The mentioned papers in Table I do not address the crucial bot- be expressed as:
tleneck issue of privacy concerns, which limits comprehensive
clustering analysis due to the inability to directly aggregate 1 X
ck = xi (1)
sensitive consumer data from multiple sources. This research |Sk |
xi ∈Sk
paper proposes a novel solution using a collaborative training
model and an autoencoder approach. The collaborative train- where ck represents the updated centroid of cluster k, Sk
ing model enables data aggregation while preserving privacy is the set of data points assigned to cluster k, and |Sk |
through federated learning techniques, eliminating the need denotes the cardinality of Sk .
for data sharing. Simultaneously, an autoencoder learns low- 2) Gradient Sharing: In gradient sharing, each participant
dimensional representations of electricity consumption pat- computes local gradients based on their local data and
terns, facilitating effective clustering analysis. This approach shares them with a central server for aggregation. The
overcomes traditional limitations by enabling clustering on a bottleneck arises as the number of participants or the size
larger scale, incorporating data from diverse sources while of their local gradients increases, the aggregation process
ensuring data privacy. Unlike supervised federated learning becomes more time-consuming and resource-intensive,
methods like gradient sharing and model averaging, the pro- resulting in the bottleneck problem, as represented by the
posed approach allows unsupervised extraction of consump- following equation:
tion patterns, enhancing scalability, privacy, and applicability N
1 X
The subsequent sections of this research paper restructured as ∇F (Θ) = ∇Fi (θi ) (2)
follows: N i=1
It include a technical insights on the Problem Statement where Θ = (θ1 , θ2 , ..., θN ) represents the set of local
in section II, Section III has the proposed methodology of model parameters, ∇Fi (θi ) is the gradient of the objec-

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on February 23,2025 at 05:37:18 UTC from IEEE Xplore. Restrictions apply.
 tive function computed on participant i’s local data, and The problem can be formulated as follows: Given a dataset
∇F (Θ) denotes the aggregated gradient. of load profiles from N participants, D = D1 , D2 , ..., DN ,
3) Model Averaging: The model averaging bottleneck in where Di represents the load profile data of participant i,
federated learning arises due to the extensive communi- we aim to find the optimal model parameters, denoted as Θ,
cation and synchronization required between participants that minimize the overall loss across the participants’ data.
and the central server. Let N denote the total number of Mathematically, this can be expressed as:
participants. During the model averaging process, each N
participant computes a local model θi based on their
X
min L(Di , f (Di , Θ)) (5)
local data, and these models are shared with the server Θ
i=1
for aggregation. The aggregated model θ̄ is obtained by
Here, Θ = θ1 , θ2 , ..., θN represents the set of model param-
averaging the local models:
eters for the autoencoders, and f (Di , Θ) denotes the output
N obtained by applying the corresponding autoencoder model to
1 X
θ̄ = θi (3) participant i’s load profile data. The loss function L measures
N i=1
the discrepancy between the predicted outputs and the ground
The communication overhead, denoted by C, is the total truth.
amount of communication required for model averaging To address privacy concerns, we further aim to minimize
and can be expressed as the sum of the sizes of the local the reconstruction loss for each participant’s load profile data,
models: denoted as Xi , using their individual autoencoder models
N
X parameterized by θi . This can be formulated as:
C= size(θi ) (4)
i=1 min L(Xi , g(f (Xi , θi ))) (6)
θi
This process involves substantial communication over-
Here, f (Xi , θi ) represents the encoding function that maps
head, leading to increased latency and longer convergence
the input load profile Xi to a latent representation, and
time.
g(·) denotes the decoding function that reconstructs the load
To address these bottlenecks and communication over head, profile from the latent representation. The reconstruction loss
this paper proposes a novel approach in load profile analysis measures the dissimilarity between the original load profile
using federated learning. The approach utilizes collaborative and its reconstructed version.
model training with an autoencoder, which enables privacy- In the collaborative model training process, participants
preserving analysis of electricity consumption patterns which share their trained autoencoder models or model parameters
improves scalability, privacy, and accuracy compared to tradi- θi with a central server for aggregation. The server aggregates
tional clustering methods. these models by computing the average of the parameters:
III. C OLLABORATIVE F EDERATED L EARNING WITH N
1 X
AUTOENCODERS FOR E XTRACTION OF C ONSUMPTION θ̄ = θi (7)
N i=1
PATTERNS
Collaborative federated learning with the integration of By utilizing collaborative model training with autoencoders,
autoencoders introduces a novel approach for extracting we can effectively address the challenges of bottleneck and
consumption patterns in load profile analysis. By leveraging communication overhead. The approach allows for the
the power of autoencoders, this method enables the optimization of model parameters while ensuring data
collaborative training of models on decentralized data privacy, enabling accurate extraction of consumption patterns
sources while preserving data privacy. The autoencoder from load profiles. This leads to improved load profiling
architecture facilitates the unsupervised learning of latent accuracy, reduced communication overhead, and efficient
representations, capturing meaningful patterns and features utilization of computational resources across the participants,
in the load profiles. Collaborative federated learning enables facilitating the deployment of privacy-constrained load
participants to collectively gain insights from each other’s data profiling systems at scale.
while safeguarding sensitive information. This collaborative
approach enhances both the accuracy and scalability of In Fig. 1 participants (Participant 1 to Participant p) have
consumption pattern extraction. their load profile data (D1 to Dp) and individual auto-encoder
models (f (X1 , θ1 ) to f (Xp , θN )). The central server receives
their model parameters (θ1 to θN ) and aggregates them
A. Problem Formulation by computing the average parameters (θavg ). The resulting
aggregated model (favg (X1 , θavg ) to favg (XN , θavg ) cap-
We propose a collaborative model training approach using
tures collective knowledge while preserving privacy. This ap-
autoencoders. The objective is to optimize the model param-
proach improves load profiling accuracy, addresses bottleneck
eters while guaranteeing data privacy and enabling accurate
and communication overhead, and enables scalable privacy-
extraction of consumption patterns from load profiles.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on February 23,2025 at 05:37:18 UTC from IEEE Xplore. Restrictions apply.
 algorithm’s performance and facilitate the comparison of
various clustering outcomes or algorithms. The evaluation
provides six valuable insights for informed decision-making
in energy management, optimizing load profiling while
ensuring data privacy.

1) Average Inertia: It is an evaluation metric that quantifies

clustering quality by computing the average sum of
squared distances between samples and their respective
cluster centers. It indicates the compactness and separa-
tion of clusters
N
Fig. 1: Collaborative Model Training with Auto-encoders 1 X X
A.I = |x − centroid(x)|2 (8)
N i=1
x∈Xi

constrained load profiling systems.

Algorithm 1 Collaborative Model Training with Autoencoders 2) Average Davies-Bouldin Index: is an evaluation metric
for Load Profile Analysis that measures the quality of clustering by considering
the ratio of the average distance between clusters to the
Input : Set of participants {X1 , X2 , . . . , XN } with load within-cluster scatter. It provides an indication of how
profile data well-defined and separated the clusters are, with lower
Output: Trained autoencoder models {θ1 , θ2 , . . . , θN } values indicating better clustering results.
Initialization: Initialize participant models: θi ←
N
random initialization for each participant i
 
1 X si + sj
while not converged do D.B.I = max (9)
N i=1 j̸=i d(ci , cj )
for each participant i in parallel do
Compute the reconstruction output: Ri ← f (Xi , θi )
end
3) Ratio Within Cluster Sum of Squares to Between Cluster
Exchange and Aggregation: for each participant i in
Variation: It calculates the ratio between the sum of
parallel do
squared distances of samples within clusters and the sum
Participant i sends their local model θi to the
of squared distances between the overall centroid and
central server S
individual sample centroids. This ratio provides insights
end into the compactness and separability of clusters in a
Central serverPaggregates the participant models: clustering analysis.
N
θglobal ← N1 i=1 θi PN P 2
for each participant i in parallel do x∈Xi |x − centroid(x)|
WCBSC = PN i=1 (10)
Participant i receives the aggregated global
P 2
i=1 x∈Xi |x − overall centroid|
model θglobal from the central server S
end
for each participant i in parallel do 4) Silhouette Index: It calculates the average silhouette co-
Update: Update participant model θi efficient for each sample, which represents the difference
by minimizing the reconstruction loss: between the average distance to samples within the same
θi ← arg minθi L(Xi , Ri ) Participant cluster (cohesion) and the average distance to samples in
i sends their updated model θi to the the nearest neighboring cluster (separation).
central server S N
1 X bi − ai
Silhouette Index = (11)
end N i=1 max(ai , bi )
end
Return: Trained autoencoder models θ1 , θ2 , . . . , θN .
5) Reconstruction Loss: The reconstruction loss is typically
calculated as the average or sum of a chosen loss function,
such as mean squared error (MSE) or binary cross-
IV. E VALUATION C RITERIA
entropy, applied to each individual input-output pair.
Evaluation criteria are utilized to determine the caliber N
and efficacy of clustering algorithms or techniques. These 1 X
R.L = L(Xi , f (Xi , θi )) (12)
criteria furnish a quantitative assessment of the clustering N i=1

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on February 23,2025 at 05:37:18 UTC from IEEE Xplore. Restrictions apply.
 6) Calinski-Harabasz Index: It calculates the ratio of
between-cluster dispersion to within-cluster dispersion,
providing a higher score for well-separated and compact
clusters.
B N −k
CH = × (13)
W k−1
Fig.(a) Fig.(b)
Note that the notations used are as follows: N represents
the number of clusters, Xi denotes the data points in cluster
i, centroid(x) is the centroid of data point x, si and sj
represent the average distances between data points and their
respective cluster centroids, d(ci , cj ) represents the distance
between centroids ci and cj , L is the chosen loss function,
and θi denotes the parameters of the autoencoder model for
participant i.

V. C ASE S TUDY
Fig.(c) Fig.(d)
A. DataSet Fig. 2: Clustering Centers of Week Days Data: (a) K-Means, (b) M.A., (c)
G.S. (d) CMT-AE
The evaluation of our collaborative model training with
autoencoder methods involves testing them on the publicly
available Irish CER dataset. This dataset, obtained from the TABLE III: Evaluation Metrics for Clustering Algorithms of Week-
Commission for Energy Regulation (CER), contains detailed end days (CER) dataset
load profiles of 6435 consumers, allowing us to analyze Evaluation Metric K-means M.A G.S CMT-AE
electricity consumption patterns. By dividing the dataset into (Avg-I) 0.8909 3.5973 2.5973 0.7118
weekdays and weekends, we can tailor our approaches to better (DBI) 1.4856 1.2012 1.4055 1.2567
(WCBSC) 78.1256 68.7954 31.4987 25.0820
capture specific consumption behaviors during different time (SI) 0.2236 0.2630 0.5830 0.6290
periods. Six indices are used for evaluate the performance of (CHI) 491.3620 347.8050 449.5749 1234.9500
our clustering analysis. (RL) 1.1071 1.1240 1.1190 0.7118
During the clustering process, the data is standardized into
Z- Score normalization ensures that the load profiling is scaled
and centered around a mean of zero with a standard deviation
of one and specific parameters are used including the number
of clusters (k) =8, alpha = 4,it randomly selecting active
clients, batch size = 20, learning rate =0.8, and Q = 20, Q
value for local iterations. For reconstruction, a neural network
with three dense layers and activation functions is utilized,
along with mean squared error (MSE) as the reconstruction
loss metric. Collaborative model training (CMT-AE), Model
Averaging (M.A), and Gradient Sharing (G.S) are performed Fig.(a) Fig.(b)
over multiple epochs, The paper uses different parameter com-
binations on the performance of federated clustering specifi-
cally for weekdays and weekends, as depicted in the following
context.

TABLE II: Evaluation Metrics for Clustering Algorithms of Week-

days (CER) dataset
Evaluation Metric K-means M.A G.S CMT-AE
(Avg-I) 0.9779 1.2278 1.7886 0.4577
(DBI) 1.5738 1.4668 1.1813 1.1896
(WCBSC) 96.4480 93.5465 71.2530 20.8030
(SI) 0.2058 0.2168 0.0456 0.6084 Fig.(c) Fig.(d)
(CHI) 567.3340 410.3590 977.4360 1075.9895 Fig. 3: Clustering Centers of Week-End Days Data: (a) K-Means, (b) M.A.,
(RL) 1.7720 1.7995 1.7821 0.7134 (c) G.S. (d) CMT-AE

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on February 23,2025 at 05:37:18 UTC from IEEE Xplore. Restrictions apply.
 B. Results improved accuracy, efficiency, scalability, and reduced com-
The study assessed various clustering algorithms for load munication overhead. This approach enables comprehensive
profiling, with a particular emphasis on addressing bottlenecks, analysis of large-scale smart meter data while ensuring data
communication overhead, and privacy preservation. The algo- privacy, leading to enhanced energy management strategies
rithms examined included k-means, federated supervised al- and informed decision-making in the power sector. Further-
gorithms (M.A and G.S), and unsupervised federated learning more,future research directions include integrating hierarchical
(CMT-AE). Six indices were utilized for evaluation. clustering algorithms, exploring privacy-preserving techniques
The evaluation matrices for clustering of weekdays data set for post-processing, and advancing collaborative model train-
is shown in Table II. From this table, it is observed that ing for improved scalability and effectiveness in load profiling.
CMT-AE(0.4577) outperformed k-means(0.9779) with a lower R EFERENCES
inertia score,indicating superior clustering performance. It ef-
[1] Y. Wang, Q. Chen, T. Hong, and C. Kang. ”Review of smart meter data
fectively addressed the Communication Bottleneck challenges analytics: Applications, methodologies, and challenges.” IEEE Trans.
(WCBSC) by achieving a score of 20.803, significantly lower Smart Grid, vol.10, pp 3125-3148, 2018.
than G.S’s score of 71.253, thereby mitigating communication [2] R. Li, Z. Wang, C. Gu, F. Li, and H. Wu, “A novel time-of-use tariff
design based on Gaussian mixture model,” Appl. Energy, vol. 162,pp.
overhead. CMT-AE exhibited a Silhouette Index of 0.6084, 1530–1536, Jan. 2016.
indicating well-defined separation within the clusters. The [3] C. Feng, Y. Wang, K. Zheng, and Q. Chen, “Smart meter data-driven
Cluster Heterogeneity Index (CHI) values were 1075.9895 for customizing price design for retailers,” IEEE Trans. Smart Grid, vol.
11,no. 3, pp. 2043–2054, May 2020.
CMT-AE and 977.436 for G.S, demonstrating enhanced cluster [4] J. Kwac and R. Rajagopal, “Data-driven targeting of customers for
separation. Moreover, the Davis-Bouldin index values were demand response,” IEEE Trans. Smart , vol. 7, no. 5, pp. 2199–2207,
1.4668 for M.A, 1.1813 for G.S, and 1.1896 for CMT-AE, de- Sep. 2016.
[5] G. Chicco, R. Napoli, and F.Piglione,“Comparisons among clustering
noting satisfactory levels of similarity and dissimilarity among techniques for electricity customer classification,” IEEE Trans. Power
the clusters. These findings are represented in Fig. 2, which Sy,vol. 21,pp 933 – 940, 2006.
presents the clustering centers of various clustering methods, [6] M. Sun, Y. Wang, F. Teng, Y. Ye, G. Strbac, and C. Kang, ”Clustering-
based residential baseline estimation: A probabilistic perspective,” IEEE
showcasing the distinctive patterns and characteristics of each Trans. Smart Grid., vol 10, pp 6014 – 6028, 2019.
method. [7] Y. Wang, Q. Chen, D. Gan, J. Yang, DS. Kirschen and C. Kang, “Deep
These findings are evident in the weekend dataset as well learning-based socio-demographic information identification from smart
meter data,” IEEE Trans. Smart Grid., vol 10, pp 2593 - 2602, 2018.
as shown in Table III. K-means achieved an inertia score [8] K. Zhou, S. Yang and Z. Shao, “Household monthly electricity con-
of 0.8909, while CMT-AE performed better with a lower sumption pattern mining: A fuzzy clustering-based model and a case
value of 0.7118, indicating improved clustering effectiveness. study.” J. Clean. Prod., vol 141, pp 900 - 908, 2017.
[9] M. Piao, H. S. Shon, J. Y. Lee, and K. H. Ryu, “Subspace projection
CMT-AE successfully mitigated communication overhead by method based clustering analysis in load profiling,” IEEE Trans. Power
achieving a WCBSC score of 25.082, compared to M.A’s score Syst., vol. 29, no. 6, pp. 2628–2635, Nov. 2014.
of 68.7954. The Silhouette Index of 0.629963 demonstrated [10] R. Li, F. Li, and ND. Smith. ”Multi-resolution load profile clustering
for smart metering data.” IEEE Trans. Power Sys., vol. 31, pp 4473 –
significant separation within the clusters for CMT-AE. Addi- 4482, 2016.
tionally, the CHI value of 1234.952 for CMT-AE indicated [11] R. Granell, CJ. Axon and D. Wallom. ”Impacts of raw data temporal
improved cluster separation. The Davis-Bouldin index values resolution using selected clustering methods on residential electricity
load profiles.” IEEE Trans. Power Sys., vol. 30, pp 3217 – 3224, 2015.
were 1.201234 for M.A and 1.25672 for CMT-AE, reflecting [12] N. Ravi, A. Scaglione, S. Kadam, R. Gentz, S. Peisert, B. Lunghino, E.
a satisfactory level of both cluster similarity.The results are Levijarvi, and A. Shumavon. ”Differentially Private-Means Clustering
corroborated by the data presented in Fig.3, which displays Applied to Meter Data Analysis and Synthesis.” IEEE Trans. Smart
Grid., vol. 13, pp 4801 – 4814, 2022.
the unique clustering centers obtained through the utilization [13] M. Jia, Y. Wang, C. Shen, and G. Hug, “Privacy-preserving distributed
of different clustering methods. This visual representation fur- clustering for electrical load profiling,” IEEE Trans. Smart Grid, vol.
ther reinforces the observed distinctions between the clusters 12, no. 2, pp. 1429–1444, Mar. 2021.
[14] T. Li, A. K. Sahu, A. Talwalkar, and V. Smith, “Federated learn-
generated by each method. CMT-AE outperformed other algo- ing:Challenges, methods, and future directions,” IEEE Signal Process.
rithms in clustering effectiveness and addressing communica- Mag.,vol. 37, no. 3, pp. 50–60, May 2020.
tion overhead in both datasets, demonstrating distinct cluster [15] Q. Yang, Y. Liu, T. Chen, and Y. Tong, “Federated machine learning:
Concept and applications,” ACM Trans. Intell. Syst. Technol., vol. 10,
separation and desirable similarity. Additionally, it excelled in no. 2, pp. 1–19, 2019.
load profiling while maintaining stringent privacy constraints, [16] Y. Wang, M. Jia, N. Gao, L. Von Krannichfeldt, M. Sun, and G.
ensuring the protection of sensitive user information. Hug,”Federated clustering for electricity consumption pattern extrac-
tion,” IEEE Trans.Smart Grid, vol. 13, no. 3, pp. 2425-2439, 2022.
VI. C ONCLUSION [17] Q. Wanrong, F. Zhai, Z. Bao, B. Li, Q. Yang andY. Cao,“Clustering
approach and characteristic indices for loadprofiles of customers using
In conclusion, this paper presented a novel approach for data from AMI,” InternationalConference on Electricity Distribution, pp
Privacy Constrained Load Profiling using Smart Meter Data. 1-5, 2016.
[18] E. Cook, M. Saleem, Y. Weng, S. Abate,K. Pitou, and
By utilizing Collaborative Model Training with Autoencoders, B.Grainger,“Density-based clusteringalgorithm for associating
the proposed method effectively tackled the challenges of transformers with smart meters via GPS-AMIdata,” Int. J. Electr.
bottleneck and communication overhead. The results from Power Energy Syst, vol 142, 108291, 2022.
[19] “Commission for Energy Regulation (CER) Smart Metering Project.”
the case studies demonstrated the superiority of the unsuper- Irish Social Science Data Archive. 2012. [Online]. Available:
vised federated learning approach over other methods, offering http://www.ucd.i.e.,/issda/data/commissionforenergyregulationcer/

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on February 23,2025 at 05:37:18 UTC from IEEE Xplore. Restrictions apply.