2020 International Conference for Emerging Technology (INCET)

Belgaum, India. Jun 5-7, 2020

Energy Management via Anomaly Detection for

Manufacturing Enterprises
Parvathy S Safni Usman T Harikrishnan K
EPK-EC EPK-EC EPK-EC
Robert Bosch Robert Bosch Robert Bosch
Bangalore, India Bangalore, India Bangalore, India
[email protected] [email protected] [email protected]

Abstract—Dominant share of Indian production sector lacks loads in a manufacturing unit can be categorized as,
knowledge in energy efficiency measures and transparency of manufacturing load, air conditioning load and utility load
energy flow. Work force employed in the unit are usually semi- (including lighting loads, lift etc). Manufacturing load turns out
skilled. As a result, erroneous operations and minor electrical to be the primary load and the rest forms the secondary load.
faults may advance to device failures and energy loss further Industrial energy optimization has been discussed in literature,
leading to reduced production quality, quaintly and enhanced considering different load types [11-13].
production cost. This paper explains the approach for the
detection of electrical and operational anomalies of connected Unfortunately the knowledge of the workforce in the
loads in the manufacturing unit under consideration via analyzing production unit on energy efficient measures is limited.
electrical parameters collected through installed energy meters at Moreover, considerable proportion of labor force driving or
load level. The gathered data is subjected to clustering and administering the loads in units is semi-skilled, which may
classification algorithms for anomaly detection. Effective accelerate toward energy wastage through erroneous operation
conclusions are drawn and economic recommendations are made of devices and device failures. Hence production output,
from the analysis and are reported to the facility management. product quality and production cost are disturbed. Such
anomalies or flaws in operation must be spotted and should be
Index Terms—Anomaly detection, classifying algorithm, communicated to the facility manager for the healthy and
clustering algorithm, energy optimization, HVAC System
economic operation of the unit.
I. INTRODUCTION The work presented reduces industrial energy consumption
The layout of electricity generation capacity is designed to and improves production through Energy Analytics and Energy
catch the peak demand to assure reliability in service. To rule Optimization. The paper is organized as follows: Section II
out the idea of building new power plants, the peak to average describes various concepts explored in the work. Section III
ratio of cumulative demand profile need to be minimized. This defines the objective of the work. Section IV details the data
target can be achieved by optimal load control, increasing analytics and anomaly detection. Section V includes the
energy efficiency and smart pricing by utilities. However smart conclusions drawn from the work.
pricing has extensively been discussed in literatures [1-3]. The
rest is gaining pre-eminent importance where determining II. CONCEPT
controllable load potential with operational efficiency is under A. Energy Analytics
spotlight. Preceding studies in the literature on optimal load
Energy Analytics involves three major steps including,
control are pertained mostly with residential [4-5] and
commercial loads [6-7]. (i) Monitoring phase: includes collecting data (captured
through installed energy meters) from the loads at the
On the other hand, industries draw up nearly 43% of net
test site, preprocessing the data i.e. organizing data for
energy consumption in India [8]. Currently the leading interest
analysis, cleaning the data to discard redundant,
of manufacturing industries is in the optimization of resource
irrelevant and unreliable data and finally press out
utilization and curtailment of production expenditure. Moreover
appropriate features from the data set for analysis
growing concern about greenhouse gas emissions and
expeditiously increasing prices of energy resources demands (ii) Analyzing phase: is exploratory data analysis. The
efficient use of these resources in production environments. gathered data is inspected, compiled and correlated with
Literature discusses either production unit optimization or the aim of developing practically useful information’s,
building energy management [9-10]. This novel archetype initiating conclusions and aiding in decision making.
insists not only optimal production but also flexible energy
management in the production unit. Such an advent incorporates (iii) Exhibiting phase: the final stage where the drawn
handling manufacturing process along with facility management conclusions about the loads in the production units
or building management. including machine health, working efficiency etc. and
recommended changes in load settings, incurred human
As majority of the manufacturing units are already errors are drawn to the attention of the facility manager.
automated, this existing infrastructure can be employed and
upgraded to aid the optimal load control. The major electrical

978-1-7281-6221-8/20/$31.00 ©2020 IEEE 1

Authorized licensed use limited to: Lulea University of Technology. Downloaded on July 10,2021 at 18:22:52 UTC from IEEE Xplore. Restrictions apply.
 A. HVAC System
B. Anomaly detection
Fig. 1 illustrates the approach defined for energy analytics
Anomaly detection is the identification of observations or for the test site. Critical loads in the unit are equipped with
data items which do not fit into the expected pattern in the data energy meters. Electrical parameters including line and phase
set. Hence categorized as unexpected or abnormal data points. voltages and currents, power, energy and power factor of the
Interestingly these data points are not unique but are unusual loads, supply voltages, frequency etc. are gathered at one
objects in the data set and can be declared as outliers. minute frequency through the installed energy meters. Physical
C. Categorizing Algorithms parameters regarding batch size of different process lines,
Key part of analysis lies in determining the anomalous operating modes of various loads, rated operating points of
operating points for which various categorizing algorithms are machines (i.e. rated speed, temperature limits etc.) are collected
used. The four algorithms employed in this work includes, from Facility Manager (FM).
Following lists the objectives of energy analytics to be
achieved through the detection of electrical and operational
abnormalities in the test site.
(i) Defining and Comparing Trends By analyzing the
load profiles (power and energy), trends in plant energy
consumption, production and energy intensity is defined
on daily, weekly, monthly and yearly basis.

Fig. 1. Energy analytics approach.

(i) k-means clustering algorithm: (an unsupervised

learning technique) is a segregating method and
separates data into ‘k’ mutually exclusive groups or
clusters [17].
(ii) Hierarchical clustering algorithm: (an unsupervised
Fig. 2. Yearly analysis of HVAC System.
learning technique) partitions data over different scales
by developing cluster tree or dendrogram [18].
(iii) Support Vector Machine (SVM) classifying algorithm:
(a supervised learning technique) is a separating
classifier defined by a discriminating hyper plane.
After training the classifier with labeled data set, the
original data set is classified into groups by defining a
hyper plane placed at maximum distance from all the
classes formed [19-21].
(iv) Decision tree classifying algorithm: (a supervised
learning technique) is a classification tree predicting
responses to data points. In order to conclude the
response, the decision in the tree need to be followed
from the root node down to the leaf node since leaf
node holds the response. The tree gives responses that Fig. 3. Shift wise analysis of HVAC System.
are nominal (i.e. true (1) or false (0)) or even numeric
(i.e. 1 or 2) based on learning [22]. B. DATA CLUSTERING
III. OBJECTIVE The data points are either clustered or classified into
two groups: faulty and non-faulty data. The data fields
In the work presented, a diesel injection pump considered for analysis includes, per day HVAC energy
manufacturing unit in India is the test site under observation consumption, number of pumps produced per day and
and analysis. The operation of one of the major loads is day’s average temperature (outdoor).
investigated which includes,

Authorized licensed use limited to: Lulea University of Technology. Downloaded on July 10,2021 at 18:22:52 UTC from IEEE Xplore. Restrictions apply.
 IV. METHODOLOGY day, the energy consumption may shoot up. This increment on
The contract demand of the test site is 6.5 MVA. On high temperature days and/or high production days should not
large production days, the plant energy consumption be mistaken as anomaly. However this glitch is eliminated by
shoots beyond the contract demand, levying heavy fine on the algorithm.
the industry by State Electricity Boards. To bring down the Fig. 4 represents the clustering results considering all the
energy consumption either the primary load (i.e. three data fields mentioned above.
manufacturing load) or secondary load (i.e. air condition,
lighting and other loads) can be considered. The plant Data points with optimum energy consumption and
operational data is collected and analyzed for all the months optimum pump production are clustered as non-faulty data i.e.
of the year 2016. Fig. 2 shows the total energy consumption for a data point with relatively high pump production and
of the unit and the share held by HVAC. relatively low energy consumption may be clustered as non-
faulty point, but same point with low temperature field may be
clustered as faulty point.

Fig. 4. K-means clustering results considering per day HVAC energy

consumption, per day pump production and day’s average temperature.

HVAC sharing nearly 10% of the total plant consumption Fig. 5. K-means clustering results considering per day HVAC energy
and being categorized as secondary load, justifies the selection consumption and per day pump production.
of HVAC over the production load to shoot down the plant
energy consumption.
A. HVAC System
HVAC system is installed in the production hanger of the
manufacturing unit. The unit works in three shifts. The first
shift operates from 6:00 AM – 2:00 PM, second shift from 2:00
PM – 10:00 PM and third shift from 10:00 PM – 6:00 AM. Fig.
3 shows the preliminary analysis of the HVAC system where
the shift wise energy consumption is investigated. Since
operating pattern of the system varies with temperature, the
energy consumption during the first shift should be higher than
the consumption during the second shift which should be
higher than the third shift. However this pattern is not followed
during the months of March, May, July and August, which
proves the existence of anomalies in the system.(The operating
pattern of the system also depends on the pump production, i.e. Fig. 6. K-means clustering results considering per day HVAC energy
during high production days, the required cooling load will be consumption and day’s average temperature.
more and hence more energy consumption). Further, the entire
data set of the system (parameters including, voltage, current, Fig. 5 represents the clustering results with two fields
power, energy, speed of blower, frequency of VFD) is chosen (i.e. per day energy consumption and per day
analyzed for the entire year, 2016. Classifying and clustering pump production). Here classification algorithm is
algorithms are employed to categorize the data points. executed after grouping the data points based on different
temperature ranges. Hence the data points with high
1) K Means Clustering Algorithm energy consumption are clustered as faulty data. Fig. 6
Due to the dependence of HVAC system on external represents the clustering results with energy consumption
factors (like temperature, pump production etc.), the energy and temperature as the two fields chosen. It’s evident that
consumption may vary with respect to these factors. For there exist both low and high consumption points when
instance, on a high temperature day or on a large production the outdoor temperature is high, which may be credited to

Authorized licensed use limited to: Lulea University of Technology. Downloaded on July 10,2021 at 18:22:52 UTC from IEEE Xplore. Restrictions apply.
 either variation in pump production or operational shows the final classification results of the data set. The final
anomalies. Most importantly, the date and time related to hyper plane separating faulty and non-faulty data points is
data points grouped into faulty data is identified. determined. The date and time related to data points grouped
into faulty data is identified.
2) Hierarchical Clustering Algorithm
Fig. 7 represents the clustering results by considering per 4) Decision Tree Classification Algorithm
day pump production and per day energy consumption as the Fig.9 illustrates the classification by decision tree
fields (the data points are grouped with reference to the algorithm, feeding per day energy consumption and per day
outdoor indices and link height respectively). The height pump production as the input fields. The data points are
represents the distance linkage computed between different grouped with reference to the outdoor temperature prior
data points. The data points are clustered into two groups- clustering. At each node of the tree the condition is being
faulty and non-faulty. checked and the data is classified as either faulty (1) or non-
faulty (2). The data and time related to data points are
identified.

Fig. 7. Dendrogram developed by hierarchical clustering considering per day

HVAC energy consumption and per day pump production. Fig. 9. Decision tree classification results considering per day HVAC energy
consumption and per day pump production.

TABLE I. COMPARISON FAULTY DATA AND NON FAULTY

DATA

DISCUSSION AND RECOMMENDATION

The learning algorithms have categorized the data points as
faulty and non-faulty. The date and time of occurrence of these
anomalies have been identified. The dates on which physical
Fig. 8. SVM algorithm (a) training results (b) classified data set. operational error has occurred in the unit, is gathered from FM
(i.e. increased operating time of the unit or other systems,
The date and time related to data points grouped into power supply problems, device malfunction etc.)
faulty data is identified. Table I presents the comparison of the account of faulty
3) SVM Classification Algorithm data generated by the learning algorithms and the data received
from FM. The comparison of 150 samples are presented. It’s
Classification is performed by considering two fields, per been observed that the number of faulty points determined by
day energy consumption and per day pump production (the categorizing algorithms are unusually high comparing to the
data points are grouped with reference to the outdoor physical operational error encountered in the HVAC system.
temperature prior clustering). The data set is scaled and fed to This demands modification in the present operating points of
the algorithm. Fig. 8(a) shows the results of the classifier the HVAC system as the power profile shows anomalies quite
after being trained with labeled data and depicts the initial high number of times.
hyper plane defined by the algorithm after training. Fig. 8(b)

Authorized licensed use limited to: Lulea University of Technology. Downloaded on July 10,2021 at 18:22:52 UTC from IEEE Xplore. Restrictions apply.
 TABLE II. ENERGY SAVINGS COMPARISON

Month Days Interval: 4 hour Interval: 2 hour Interval: 1 hour

E_Present E_Proposed E_Savings E_Proposed E_Savings E_Proposed E_Savings

(kWh) (kWh) (kWh) (kWh) (kWh) (kWh) (kWh)
01-Jun-16 1 6216 4938 1278 5388 828 5469 747
June, 2016 26 161616 128390 33228 140090 21528 142190 19422
1st June – 15th Aug 71 441336 350600 90738 382550 58788 388300 53037
15th Aug, 2016 1 6216 5298 918 5100 1116 5136 1080
15th Aug- 15th Sept 27 167832 143050 24786 137700 30132 138670 29160
15th Aug- 15th Oct 54 335664 286090 49572 275400 60264 277340 58320
Net 125 777000 662250 114750 637500 139500 642000 135000
Savings (Rs) 125 573750 697500 675000

(4)
Of the twelve blowers present in the system, nine are
equipped with VFD control. In the present operational settings,
blower VFDs (nine blowers) are set at 47 Hz from 8:00 AM to
10:00 PM and at 35 Hz from 10:00 PM to 8:00 AM. It’s
evident that the prevailing settings is thoroughly inefficient and
leads to greater energy loss as there is no optimal control over
blower speed which is a function of cooling load which is
again a variable of time. The suggested solution is to alter the
duration and frequency of VFD setting. The three different
intervals recommended are of four hour, two hour and one hour
Fig. 10. HVAC System block diagram.
duration. Cooling load calculation and ݂ܿ݉ estimation are
completed for entire day at required time intervals and the VFD
To modify system settings, knowledge about the system
setting is defined. The comparison between the present and
model is necessary. Fig. 10 illustrates the HVAC system
recommended frequency settings and corresponding power
employing evaporative cooling. The blower pumping
consumption during specified time intervals are illustrated as
conditioned air into the hangar is equipped with variable
follows. Fig. 11 – Fig. 13 represents the comparison between
frequency drive (VFD) control. The blower speed is adjusted by
the present, 4 hour, 2 hour and 1 hour frequency settings and
VFD regulating the blower power output. Eq. 1 shows the
the corresponding active power consumption. Table II shows
relation between blower power (in kW) and its speed (in rpm)
the comparison of energy consumption in the system with
[23].
present and recommended settings for 125 days. Hence 2 hour
kW‫ן‬rpm͵ ሺͳሻ settings generate maximum savings, it’s recommended the
most to FM.
The determination of blower speed leads to ݂ܿ݉ estimation
(cubic feet per minute: air moved out by blower per minute).
Eq. 2 shows the relation between ݂ܿ݉ requirement of the
hangar and blower rpm [23].
‫݉݌ݎן‬ ሺʹሻ
For ݂ܿ݉ estimation, the required data (physical data of
hangar, occupancy, hangar load etc.) is collected from the FM
and ݂ܿ݉ requirement of the hangar is determined [23-24]. The
net specific heat gain (SHG) of the system is estimated (Eq. 3)
and the ݂ܿ݉ requirement is calculated from Eq. 4 [24-25].
ܵൌሺܳͳ൅‫ڮ‬൅ܳ͹) (3)

Fig. 11. Comparison of blower frequency settings and active power

consumption of present settings and 4 hour settings.

Authorized licensed use limited to: Lulea University of Technology. Downloaded on July 10,2021 at 18:22:52 UTC from IEEE Xplore. Restrictions apply.
 2013.
[4] Faria, Pedro, and Zita Vale. "Demand response in electrical energy
supply: An optimal real time pricing approach." Energy, vol. 36, no.
8,pp. 5374-5384,Aug. 2011
[5] A. Bego, L. Li, and Z. Sun, “Identification of reservation capacity in
critical peak pricing electricity demand response program for
sustainable manufacturing systems,” Int. J. Energy Res., vol. 38, no.
6, pp. 728–736, May 2014.
[6] Vedanta Pradhan, V. S. K. Murthy Balijepalli, Shrikrishna A.
Khaparde, "An Effective Model for Demand Response Management
Systems of Residential Electricity Consumers", Systems Journal
IEEE, vol. 10, no. 2 pp. 434-445, Jun 2016.
[7] Y. Guo, M. Pan, and Y. Fang, “Optimal power management of
residential customers in the smart grid,” IEEE Trans. Parallel Distrib.
Syst., vol. 23, no. 9, pp. 1593–1606, Sep. 2012.
[8] A.-H. Mohsenian-Rad and A. Leon-Garcia, “Optimal residential load
control with price prediction in real-time electricity pricing environ-
ments,” IEEE Trans. Smart Grid, vol. 1, no. 2, pp. 120–133, Sep.
2010.
[9] Tanguy Hubert and Santiago Grijalva, "Modeling for Residential
Electricity Optimization in Dynamic Pricing Environments", Smart
Grid IEEE Transactions on, vol. 3, no. 4, pp. 2224-2231, Dec 2012.
Fig. 12. Comparison of blower frequency settings and active power [10] S. Kiliccote, M. Piette, and D. Hansen, “Advanced controls and com-
consumption of present settings and 2 hour settings munications for demand response and energy efficiency in
commercial buildings,” in Proc. 2nd Carnegie Mellon CEPS,
Pittsburgh, PA, USA, Jan. 2006, pp. 1–10.
[11] M. Ghamkhari and H. Mohsenian-Rad, “Energy and performance
man- agement of green data centers: A profit maximization
approach,” IEEE Trans. Smart Grid, vol. 4, no. 2, pp. 1017–1025,
Jun. 2013.
[12] S. A. Raziei and H. Mohsenian-Rad, “Optimal demand response
capac- ity of automatic lighting control,” in Proc. IEEE PES Conf.
Innov. Smart Grid Technol., Washington, DC, USA, Feb. 2013,
pp. 1–6.
[13] Growth of Electricity Sector in India from 1947-2015, Central
Electricity Authority of India, June 2015.
[14] Moon, Joon-Yung, Kitae Shin, and Jinwoo Park. "Optimization of
production scheduling with time-dependent and machine-
dependent electricity cost for industrial energy efficiency" The
International Journal of Advanced Manufacturing Technology,
vol. 68, no. 1-4, pp. 523–535, Sep. 2013.
[15] Groscurth, H-M., and R. Kümmel. "The cost of energy
optimization: A thermoeconomic analysis of national energy
Fig. 13. Comparison of blower frequency settings and active power system" Energy, vol. 14, no. 1, pp. 685-696, Nov 1989.
consumption of present settings and 1 hour settings. [16] Kümmel, R., H-M. Groscurth, and W. Van Gool. "Energy
optimization in industrial models" System Modelling and
V. CONCLUSION Optimization, Springer, vol. 14, no. 1, pp. 518-529, Nov 1988.
[17] Guo, Y., Loenders, J., Duflou, J. and Lauwers, B., “Optimization
This paper proposes an approach to optimize the energy of energy consumption and surface quality in finish turning”,
consumption by manufacturing units and thus curtails the Procedia CIRP, vol. 1, pp. 512-517, Jul. 2012
negative environmental impacts and improves energy [18] Fumo, Nelson, Pedro Mago, and Rogelio Luck. "Methodology to
efficiency. In this work, a secondary utility load (HVAC estimate building energy consumption using EnergyPlus
System) is analyzed. For the HVAC System avenues of Benchmark Models." Energy and Buildings, vol. 42, no. 12, pp.
2331-2337, Dec 2010.
energy wastage has been detected through anomaly detection. [19] Nguyen, Anh-Tuan, Sigrid Reiter, and Philippe Rigo. "A review
Revised operating conditions granting optimum system on simulation-based optimization methods applied to building
performance are recommended. A small increment in performance analysis." Applied Energy, vol. 113, pp. 1043-1058,
operational efficiency results in significant energy savings. Dec 2010.
Hence the proposed methodology improves the production [20] Ashok, S., and Rangan Banerjee. "An optimization mode for
quality, quantity with reduced operational cost. industrial load management." IEEE Transactions on Power
Systems, vol. 16, no. 4, pp. 518-529, Nov 2001.
REFERENCES [21] X. Li, C. P. Bowers, and T. Schnier, “Classification of energy
consumption in buildings with outlier detection,” IEEE
[1] P. H. Jayantilal and N. Shah, “A review on electrical energy Transactions on Industrial Informatics, vol. 57, no. 11, pp. 3639-
management techniques for supply and consumer side in industries,” 3644, Nov. 2010.
Int. J. Sci. Eng. Technol. Res., vol. 3, pp. 550–556, Apr. 2014. [22] D. Wijayasekara, O. Linda, M. Manic, and C. Rieger, “Mining
[2] C. Joe-Wong, S. Sen, S. Ha, and M. Chiang, “Optimized day-ahead building energy management system data using fuzzy anomaly
pricing for smart grids with device-specific scheduling flexibility,” detection and linguistic descriptions,” IEEE Transactions on
IEEE J. Sel. Areas Commun., vol. 30, no. 6, pp. 1075–1085, Jul. 2012. Industrial Informatics, vol. 10, no. 3, pp. 1829-1840, Aug. 2014.
[3] P. Samadi, H. Mohsenian-Rad, V. Wong, and R. Schober, “Tackling [23] John E. Seem, “Using intelligent data analysis to detect abnormal
the load uncertainty challenges for energy consumption scheduling in energy consumption in buildings,” Elsevier Journal on Energy and
smart grid,” IEEE Trans. Smart Grid, vol. 4, no. 2, pp. 1007–1016, Jun. Buildings, vol. 39, no. 1, pp. 52-58, Jan. 2007.

Authorized licensed use limited to: Lulea University of Technology. Downloaded on July 10,2021 at 18:22:52 UTC from IEEE Xplore. Restrictions apply.
[24] Eskin E., Arnold A., Prerau M., Portnoy L., Stolfo S. ”A geometric [25] Görnitz N., Kloft M. M. , Rieck K. and Brefeld U., ”Toward
framework for unsupervised anomaly detection: Detecting supervised anomaly detection”, Journal of Artificial Intelligence
intrusions in unlabeled data”, Applications of data mining in Research, vol. 46, pp. 235-262, Feb. 2013.
computer security, vol. 6, no. 1, pp. 77-102, May. 2002.

Authorized licensed use limited to: Lulea University of Technology. Downloaded on July 10,2021 at 18:22:52 UTC from IEEE Xplore. Restrictions apply.