Smart Meter Data Analytics for Building

2023 IEEE PES Conference on Innovative Smart Grid Technologies - Middle East (ISGT Middle East) | 978-1-6654-6543-4/23/$31.00 ©2023 IEEE | DOI: 10.1109/ISGTMiddleEast56437.2023.10078592

Monitoring System: A Case Study

Potugunta Rajesh ∗ , Vikram Cherala † , ,Pradeep Kumar Yemula ‡
Department of Electrical Engineering
Indian Institute of Technology, Hyderabad (IITH)
ee18mtech11037@iith.ac.in ∗ , ee18resch11013@iith.ac.in † , ypradeep@ee.iith.ac.in ‡ ,

Abstract—Decarbonization, decentralization, and digitalization Analytics can be inferred as the scientific process which
are the prominent paths for the energy sector in the future. transforms the raw data into valuable insights for making better
The rise of smart meters across consumers, and industries led conclusions and decisions. It is commonly divided into four
to a massive collection of fine-grained energy and electricity
consumption-related data. A data science challenge is to analyze stages: descriptive analytics, diagnostic analytics, predictive
the Smart Meter data for the benefit of both the energy providers analytics, and prescriptive analytics.
and the consumers. In this paper, An attempt has been made
to analyze the smart meter collected from the IIT Hyderabad (i) Descriptive Analytics [5]: As the name suggests, it
campus and presented the analysis into descriptive, predictive, describes things. In this, we address what happened in the
and prescriptive analytics. The data collected from more than past and what is happening now, statistically in the form
50 meters over a period of one year have been analyzed and of visualization like plots, charts, and reports. Descriptive
results obtained. Interesting trends such as the impact of COVID- analytics tells us this is happening and can also provide real-
19 on campus energy consumption have been examined. The
framework for energy data analytics presented in this paper will time data analytics. Some of the examples are email alerts,
be useful for any campus in general, and the recommendations data stoppage alerts, heat map visualizations, etc. which we
presented will save energy expenses. see in section II.
Index Terms—Solar power plants; rooftop solar (ii) Diagnostic Analytics: Provide an insightful understanding
of particular questions raised in descriptive analytics. It is also
referred to as root cause analysis. Most of the time we get the
I. I NTRODUCTION answer and sometimes we have to do further research for the
cause. Diagnostics involve summarizing the graphs and charts,
Globally, smart meters are being installed with the main
critical analysis of all reports, and arriving at conclusions
goal of improving the efficiency and economical operation of
statistically. One of the examples is how COVID-19 impacted
electrical energy networks. For example, the Indian govern-
energy consumption behavior [6].
ment has intensified its push to roll out 250 million smart
(iii) Predictive Analytics: In this analytics, we will take
meters to all residential consumers by 2025 in order to cut
historical data and build statistical or machine learning or
losses and increase operational and financial efficiencies of the
deep learning models that can detect trends and patterns.
distribution sector [1]. The widespread use of smart meters
The prediction accuracy hugely depends on data quality and
makes it possible to gather a vast amount of precise infor-
stability of external situations. This model is then used for the
mation on electricity consumption. Smart meters no longer
prediction of what might happen next. Examples like the day
simply serve the purpose of billing. High-resolution data from
ahead forecasting, solar forecasting, etc [7].
smart meters offer detailed information on customer habits
(iv) Prescriptive Analytics: Based on the information avail-
and lifestyles related to electricity use. Smart meter data
able from the above, prescriptive analytics decide the course
is frequently collected in large volumes, therefore efficient
of action and steps to be taken to avoid a future problem while
analytics is essential to maximizing its value. To give a
addressing existing problems. One of the examples is to use
variety of electricity consumers/prosumers an in-depth insight
load scheduling to reduce peak demand [8].
into energy consumption behavior, sophisticated analytics are
needed [2]. The objective of this paper is to present case studies
Among several applications, smart meter data can be used for covering the above range of analytics using the practical data
designing new policies such as dynamic electricity pricing, in collected from more than 50 smart meters over a period of
which tariffs are high during peak times and, in turn, helps one year from an academic campus building in IITH. The rest
the grid by reducing peak demand. Smart meter analytics is of the paper is organized as follows. Section II discusses data
providing tremendous support for the consumers and utilities collection using automated metering infrastructure (AMI) in
to understand the behavior and patterns in electricity consump- IIT Hyderabad campus. Section III is devoted to a case study
tion [3]. Smart meter analytics has become an active research with different descriptive data analytic methods. Section IV
area. In a recent study, utility data analysis has become a $1 presents diagnostic analytics. Section V deals with predictive
billion market and is expected to become nearly $4 billion analytics and Section VI presents prescriptive data analytics.
market by the year 2020 [4]. Finally, section VII concludes the paper.

978-1-6654-6543-4/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: Andhra University College of Engineering. Downloaded on September 25,2024 at 05:58:29 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1. Building Monitoring System(BMS)

meters to the centralized server systems with RS-485 serial

communication. The extracted data are stored in a database,
and we then carried out data analytics. The detailed procedure
with sample code for extracting data from the smart meter is
explained in [10].
Typically for any commercial building, the power is sup-
plied to the entire building from the service main. The net
building power is measured with the help of the main meter,
which is installed at the main power supply. However, if
the building has to be energy efficient and save on energy
Fig. 2. Hardware setup for extraction of data bills, the building users can install several meters assigned to
individual departments/floors of the building, through which
monitoring of energy can be done with detailed energy usage.
Those meters’ principal goal is always to give us tips and
valuable insights into the data for the building users about
consumption patterns, and also, this idea is known as sub-
metering. The system built using all the data obtained from
submeters for analytics, visualization, and decision-making is
called a building monitoring system(BMS) [10]. Fig. 1 is the
block diagram of the BMS, which shows how data is extracted
Fig. 3. smart meters installed at panel roof of A block from various sub-meters and communicated to the database. In
the power layer, the power from the main panel is distributed
to the whole building, and this main meter records the power
II. AUTOMATED M ETERING I NFRASTRUCTURE (AMI) consumption of the building. The main meter reading is used
The availability of datasets is one of the biggest problems for electricity billing and for any accounting purposes by the
that researchers in Smart Meter data analytics encounter. utility. The distribution panel has many dedicated circuits for
Power utilities are hesitant to make their measurement data the distribution of electricity to other parts of the building.
publicly available for data analytic research purposes for a These circuits are provided with sub-metering to monitor the
variety of reasons, including security, privacy, and political detailed power flows in various other areas in the building.
concerns [9]. Electricity consumption is measured with the Fig 1 also shows the process for data acquisition; It shows
help of meters, with the evolution and advancements from in detail each sub-task involved in collecting and storing the
Induction to digital to now smart meters we can now even data in the database. For collecting the data, there should be
record the consumption data with the help of the latest smart multi-function meters pre-installed. If those meters are
smart meter technologies at desired intervals. Most of the not there, then new smart meters should be installed. These
commercial buildings are now equipped with smart meters, meters are connected to CT’s secondary. The next step is
which are installed in the panel room (centralized), as shown commissioning the meters for doing this; we have to enter into
in Fig 3, and all floors/departments (decentralized). the programming mode of the meter and adjust the CT ratings
The extraction of the data from the meters is the first step for and assign a slave id. Now with two wires, connect to D+ and
any data analytics project. The smart meters measure energy, D- data lines in series with every meter, and finally, we get
power, voltages, power factor, etc. we can communicate these two wires that two wires are connected to RS-485 to the USB

Authorized licensed use limited to: Andhra University College of Engineering. Downloaded on September 25,2024 at 05:58:29 UTC from IEEE Xplore. Restrictions apply.
 Fig. 4. Email alerts snapshot

Fig. 6. Clustering of Academic block-A for 100 days from Jan-1st to Apr-
9th 2020

• Rest of the loads are plug loads and lighting loads spread
across all floors of the building

C. Clustering-based load profiling

Electricity smart meter data clustering is regarded as a data
mining technique. Smart meter data can be clustered based
on features like density, Euclidean distance, or self-organizing
maps using unsupervised techniques including hierarchical
Fig. 5. Percentage of energy usage of A block for 100 days clustering, k-means, fuzzy c-means, and self-organizing maps
[11]. For instance, real load profile data has been classified by
the k-means clustering method in [12] to provide a realistic
converter. The converter is then connected to a Raspberry-Pi segmentation of industrial clients that reflect production oper-
as shown in Fig 2. ations. The typical load profiles were generated by segmenting
the meters into categories corresponding to their activities.
III. D ESCRIPTIVE ANALYTICS Hence, the objective was to generalize the information from
In this section, we present descriptive analysis (what hap- a set of meters to the entire pool of various profiles in the
pened?) using the smart meter historical data. The descriptive most accurate manner. Hypotheses were formulated about the
techniques used include email alerts, heat map analysis(energy stationary of the load over time and homogeneity inside each
report), and K-means clustering of various loads, and classifi- category of meters. Deployment of smart meters has drastically
cation of loads based on time of operation are studied. changed the paradigm of load profiling: from estimating the
demand using a few metered customers, to now summarizing
A. Email Alerts the information of streams of metering data. The problem of
summarizing a large amount of data into a limited number of
Email alerts are created by an automatic process through
representative points is a well-known problem in data analysis,
which we can send valuable information to designated recipi-
which has been studied extensively to produce clustering
ents. They can be generated at desired frequencies as per our
algorithms [13]. Hence, research on improving load profiling
needs. These actions consist of the standard text (description
techniques using metering data explores what has been done
of data) and list of recipients for an email. You can associate
in the field of data analysis, and clustering stands out as a
email alerts with daily reports of energy consumption, peak
well-suited approach to produce typical load profiles.
demand, and HVAC on and off time. Fig. 4 email snap shot
In this, we consider two cases of clustering one is a
which was created using python.
clustering of daily load curves of the academic block A of
IIT Hyderabad and the second we use clustering for the load
B. Heat-maps curve of HVAC to get the insight of usage and also provide
Using heatmaps, one can quickly identify the unusual pat- recommendations to provide savings in the campus.
terns visually. The variations in the intensity of colors give K-Means algorithm was applied on all the meters of A-block
an overview of the magnitude of data. However, they do not for 100 days of data i.e., from June 1st , 2019 to September
offer in-depth analysis. Observations based on heatmap shown 9th , 2019. For each meter data is re-sampled at 15-minute
in the Fig. 5 intervals and a daily load curve of 96x1 vectors of a total of
• Total 100 Days consumption of A block is 6,39,519 kWh 100 was fed to the algorithm. The number of clusters is fixed
• Major Loads: Data Center 18.7%, UPS load 13.21%, by the elbow method by calculating Silhouette scores ranging
HVAC,AC Ventilation 17.93%.These loads constitute from 2 to 10 clusters a graph was plotted and the knee point
50% of load is taken as the optimum number of clusters.

Authorized licensed use limited to: Andhra University College of Engineering. Downloaded on September 25,2024 at 05:58:29 UTC from IEEE Xplore. Restrictions apply.
 semester break COVID lockdown

2019 2020
Fig. 7. Energy heat map from June 2019 to May 2020 for ground floor

Fig. 9. Violin plot Interpretation

Fig. 8. With Load Shifted on May 1, 2021 implement it.

Fig. 6 is the output of K-Means clustering algorithm applied

for daily load curves from Jan-1st 2020 to Apr-9th 2020 of the IV. D IAGNOSTIC A NALYTICS
Academic block-A. The green color refers to the working day
pattern, in which the peak is close to 350kW. Orange color The objective of diagnostic analytics is to find the root cause
refers to holidays and weekends, and the peak demand is close of unusual patterns in the data, as presented in this section.
to 300kW because of reduced loads in the building floors. Red Fig 7 is the heat map from June 2019 to May 2020. We can
color refers to missing data that may be due to power cuts see that there is more anomaly data in November; the color
or communication failure or malfunctioning of smart meters. intensity is low, which implies that energy demand is low,
Finally, the violet color is the new pattern because of the which may be due to a power outage or communication failure.
institute’s shutdown due to COVID-19 on March 20. The peak During the beginning of June month, the color intensity is
demand has drastically fallen to around 200kW, and load shape thick, which implies that energy demand is high due to the
and base load are also changed from then. peak summer season. We can see that there is more anomaly
Fig. 8 is the clustering of daily load curves of the HVAC data from march 20; the color intensity is low, which implies
system in our campus, and the rightmost figure is the calendar that energy demand is low due to the institute lockdown
plot of the obtained results, which represent the cluster it because of the COVID-19 pandemic. At the beginning of the
belongs to on a particular day. Violet, orange, green, and red Feb month, the color intensity is thick, which implies that
are cluster-0,cluster-1,cluster-2, and cluster-3, and the dotted energy demand is high due to the beginning of summer.
lines are centroids of the clusters. The violet cluster implies Violin plots are a combination of box plots and kernel
that the HVAC system operates for a reduced time during density estimation(KDE) plots. What makes it unique is that
weekends and closely from 10 .m to 6 p.m, orange is a it also shows the entire distribution of the data. Violin plots
typical pattern during the summer season which is a very long are very useful for exploring large data sets. We can get a
duration closely from 7 a.m to 10 p.m, and Green cluster is general idea of how data is distributed. They can also help to
the result of the reduced time of operation for saving bill by visualize the number of outliers in the data. Fig.9is the month-
our estate electrical department people which is closed from wise violin plot for the average power demand of A-block. We
8 am to 8 pm and Red color implies that data is stopped on can see a clear trend of a decrease in average demand from
that day either due to power or internet failure. This is done June to December because of the winter season. In March
because, if suppose we want to reduce the peak let’s say at 2019, due to lockdown, there are two Gaussian curves, and in
5 pm, then we can target the particular meters to efficiently April clearly, we can see reduced demand.

Authorized licensed use limited to: Andhra University College of Engineering. Downloaded on September 25,2024 at 05:58:29 UTC from IEEE Xplore. Restrictions apply.
 Fig. 11. Shifted vs actual Load curve

Fig. 10. Day ahead prediction

campus. Solutions such as email alerts, heat maps, k-means
clustering of load curves, and also HVAC load curves have
V. P REDICTIVE A NALYTIC DATA A NALYTICS
been presented. These analytical methods helped us understand
In predictive analytics, we forecast the things that happen in the various load consumption patterns and also helps to
the future. In this section, the day ahead forecasting with deep provide economic savings are achieved.
neural networks, LSTM’s, and their accuracy performances
are studied. These models are built using historical data. Load
ACKNOWLEDGMENT
forecasting means predicting future loading. It is widely used
because of its importance and functionality of the power grid, Thanks to TCS Foundation for providing funding under TCS
as well as for power system planning [14], Fig. 10 shows the Research Scholarship program.
predicted day ahead load using LSTM.
R EFERENCES
VI. P RESCRIPTIVE DATA A NALYTICS [1] J. S. Jones, “India’s smart meter rollout –
250 million meters by 2025,” Sep 2021. [On-
In Prescriptive Data Analytics, Solutions for the problems in line]. Available: https://www.smart-energy.com/industry-sectors/
descriptive, diagnostic, and predictive analytics are proposed. smart-meters/indias-smart-meter-rollout-250-million-meters-by-2025/
In this section, a solution is proposed to minimize campus [2] Z. Niu, J. Wu, X. Liu, L. Huang, and P. S. Nielsen, “Understanding
energy demand behaviors through spatio-temporal smart meter data
electricity bills by scheduling loads. Load scheduling can be analysis,” Energy, vol. 226, p. 120493, 2021.
inferred as managing action (shifting) of electricity load to [3] P. Jokar, N. Arianpoo, and V. C. Leung, “Electricity theft detection
save electricity bills. Load shifting is considered to be one in ami using customers’ consumption patterns,” IEEE Transactions on
Smart Grid, vol. 7, no. 1, pp. 216–226, 2015.
of the best load management techniques for manipulating [4] D. J. Leeds. (2013, Dec.) The soft grid 2013-2020: Big data utility
peak-to-peak loads. The aforementioned load-shifting model analytics for smart grid. [Online]. Available: https://www.iea.org/
is derived from demand-side management (DSM) technology, reports/energy-technology-perspectives-2012
[5] Y. Wang, Q. Chen, T. Hong, and C. Kang, “Review of smart meter
which aims to reduce peak loads and costs. A load-balancing data analytics: Applications, methodologies, and challenges,” IEEE
approach distributes the electrical load in order to minimize Transactions on Smart Grid, vol. 10, no. 3, pp. 3125–3148, 2018.
cost functions associated with energy prices. Generally, The [6] Y. Wang, Q. Chen, and C. Kang, Smart Meter Data Analytics. Springer,
2020.
electricity is distributed to the campus by Telangana state [7] K. Hopf, Predictive analytics for energy efficiency and energy retailing.
southern power distribution corporation limited (TSSPDCL). University of Bamberg Press, 2019, vol. 36.
The HT users have 4 time-of-day tariffs (ToD’s). They are [8] S. Sulaiman, P. A. Jeyanthy, and D. Devaraj, “Smart meter data analysis
issues: a data analytics perspective,” in 2019 IEEE International Con-
from 6 AM to 9 AM unit price is 8 |, 9 AM to 6 PM unit ference on Intelligent Techniques in Control, Optimization and Signal
price is 7 |, 6 PM to 9 PM unit price is 8 |, 9 PM to 6 Processing (INCOS). IEEE, 2019, pp. 1–5.
AM tariff is 6 |. So, to reduce the electricity bill we need to [9] J. Hu and A. V. Vasilakos, “Energy big data analytics and security:
challenges and opportunities,” IEEE Transactions on Smart Grid, vol. 7,
identify shiftable loads and schedule them during low price no. 5, pp. 2423–2436, 2016.
times. [10] C. Sandhya, S. V. Tirunagari, S. Sahoo, and P. K. Yemula, “Extraction of
Fig. 11 is the output of the algorithm. The red color plot data from an rs-485 enabled multi function meter for building monitoring
systems,” in 2018 20th National Power Systems Conference (NPSC).
is the actual load curve of a water pumping plant, and the IEEE, 2018, pp. 1–6.
green curve is shifted load curve such that overall energy [11] G. Chicco, R. Napoli, and F. Piglione, “Comparisons among clustering
consumption is the same. The green curve reduces the overall techniques for electricity customer classification,” IEEE Transactions on
power systems, vol. 21, no. 2, pp. 933–940, 2006.
electricity bill as most of the operation is on a low tariff [12] I. Khan, J. Z. Huang, M. A. Masud, and Q. Jiang, “Segmentation of
zone. The overall savings are calculated for 13 days and are factories on electricity consumption behaviors using load profile data,”
calculated with a similar operation, which came to be 5434 |, IEEE Access, vol. 4, pp. 8394–8406, 2016.
[13] Cluster Analysis. John Wiley Sons, Ltd, 2008, ch. 1, pp. 1–13.
And the average savings per day is 418 |. [14] A. R. Khan, A. Mahmood, A. Safdar, Z. A. Khan, and N. A. Khan,
“Load forecasting, dynamic pricing and dsm in smart grid: A review,”
VII. C ONCLUSION Renewable and Sustainable Energy Reviews, vol. 54, pp. 1311–1322,
2016.
This paper presented different descriptive analytics models
performed on smart meter data collected at the IIT Hyderabad

Authorized licensed use limited to: Andhra University College of Engineering. Downloaded on September 25,2024 at 05:58:29 UTC from IEEE Xplore. Restrictions apply.