Professional Documents
Culture Documents
efficiency as alternatlVeS tO grld capaclty en- mlZed hlgh-Current dual active brldge con-
pp 450-453, May2012
Verter for automotive applications,EE
hancement,eIgyPo/icy, VO1 67, PP 222231,
[36] D MaggioIo,F Picano, A. MarlOn, and M. Guar- 7Jons hd ETec/ron, VO1 59, nO. 7, PP. 2745-
Apr. 2014.
neri, AppllCatlOn O=he Lattice-Boltzmam
2760, July 2012
21] S Sundararagavan and E. Baker, Evaluating method for modeling alvanadium redox flow
energy storage technoIogies for wind power batteries,n POC /V ht Coni PartICIe-Bosed [51] R. W. A. A. De Doncker, D. M. DIVan, and M.
H. Kheraluwala`A three-Phase soft-SWitched
integration:SoIergy, VOl 86, PP 2707-2717, Methods-ndameoIs md AppIICO/10nS-
Sept. 2012 Par[!C/e 20I5, Sept 28-30, 2015, PP. 579-589.
hlgh-POWer-density DC/DC converter for high-
POWer aPPlications:/E 77S. /nd App , VO1
22] A. Zahedl, MaxlmlZlng SOlar PV energy pen- i37 P AIotto, M Guamierl, F. Moro, and A. Stella, 27, nO. 1, pP. 63-73, Jan./Feb. 199l.
Multphysic 3D dynamic modelling of poly-
etration using energy storage technoIogy:Re-
neSus!eReVO=5, pp. 866-870, Jan mer membranes with a proper generalized [52] Z. Wang and H. Li, A soft switching three-
2011 decompositlOn mOdel reduction approach Phase currenfed bldirectional dc-dc conver
er with high efficlenCy OVer a Wide input volt-
Electrochlm Ac/0, VO1 57, nO l, PP. 250-256,
[23] C. Pieper and H Rubel, Revisiting energystor- age range,7ns P0erEIectOn , VOl. 27,
age There is a business case:i Boston Consul Dec. 2011.
no. 2, PP. 669-684, Feb. 2012.
ing Group Report, Feb. 201l. [38] Q.Xu and T. S. Zhao, Fundamental models for
[53] J Hiltunen, V Vaisanen, and P. Sllventoinen, A
flow batteries,POgr Ene)gy Combus/ Sc
[24] W F PICkard, A Q Shen, and N. J. Hanslng, bldirectional currened resonant push-Pu=
Parklng the power: Strategies and physical VOl 49, PP. 40-58, Aug 2015
COnVerter for low voltage, high current appl
limitations for bulk energy storage in supply-
[39 Q. Zheng, X. Ll, Y. Cheng, G. Ning, F Xlng, and CatlOrlS,n Proc 2013Eney Conuers/On
demand matchlngOn agrld whose mPut POWer F Zhang, `Development and perspective in va- Cbngr Expos!t/On, Sept. 15-19, 2013, PP. 4770-
lS PrOVided by intermittent sources,Rene( nadium flow battery modeling,App/ Ene}gy, 4774.
5us[. EneJg Reu. vol. 13, PP. 1934-1945, Oct VO=32, PP. 254-266, Nov. 2014.
2009. 54] G. SplaZZi and S Buso, `Analysis of the lnter-
[40] X Qiu, T. A. Nguyen, J. D Guggenberger, M. leaved isolated boost converter with coupled
[25] B. Dunn, H Kamath and J TarasconElectr L. Crow, and A. C. Elmore, A field validated inductors:7ns hd EIeclron , VOl. 62,
Cal energy storage for the grld a battery of model of a vanadium redox llow battery for no. 7, PP 4481-449l, July2015
Choices,SclenCe, VOl. 334, nO. 6058, PP. 928- microgrids:7Jans. Smar( GrId, VOl. 5, nO.
935, Nov. 2011. 4, pp 1592-160l,July2014. 551 Proxhima. Online]. Available: http.//www
PrOXhima.com
[26] G L. Soloveichik, FIow batteries: Current sta-
[41 B. Xiong, J. Zhao, Z. Wel, and M. Sky=as-Kaza-
tus and trends:Chem Reu , VO=15, nO. 20, PI). 56] N. Femla, G. Petrone, G. SpagnuoIo, and M.
COS, Extended Kalman fllter method for state
Vitelll, Pou)er EIecJromCSOnd Con!ro/ 7h-
11,533-11,558, Sept 2015. Of charge estlmat10n Of vanadlum redox flow
mqueS forMoxlmumergyHaruesg m Pho-
[27] K. Knehr and E. KumburOpen circuit voltage battery using thermal-dependent electricaI
[ooI[aC fys[ems, Boca Raton, FL; CRC Press,
Ofvanadium redox flow batterleS DISCrePanCy model,/ PoOer SouCeS, VOl. 262, PP. 50-61,
2013
between models and experiments:Electro- Sept. 2014
Chem Commun, VO1 13, nO. 4, PP342-345, [57] P M Spaziante, Novel design and non-COn-
[42 M. J. WatSmith, P. Ridley, R. G. A. Wills, A. . VentlOnal appllCations for vanadium redox
Apr 2011.
Shah, and F. C. Walsh`The lmPOrtanCe Of key
technology:presented at the !nternational Re-
[28] B. Xiong, J. Zhao, Z. Wel, and M. Sky11as-Kaza- OPeratlOnal variables and electrolyte monltOr-
dox FIow Battery Forum, Viema, ustrla, June
COS, Extended Kalman filter method for state ing to the performance of an all vanadium re-
15-16, 2010
Of charge estlmation of vanadlum redox flow dox flow batlery,i / !em 7chnoI BIOteChnoI,
battery uslng thermal-dependent electrical VOL88,nO l,pp 126-138,Jan.2013
tegenBidings
Of the Future
tybenuare, DeLeaied
and Haman hemcting
SMART HOME
MILOS MANiC,
KASUN AMARASINGHE,
JUAN . RODRIGUEANDINA,
AND CRAIG RIEGER
US and Wordnergy
Consumptiom Preditions
The U.S. Department of Energys (DOE9 Energy
and other appliances. In the commercial sector, lighting accounts for about 19% of energy usage and
is the single largest user of electricity. This is because lighting includes office, institutional, Public,
and government buildings as well as pubIic street lighting. Space and water heating (10%), SPace
COOling (11%), and ventilation (6%) together make up 27% of all electricity used. While electricity
use is projected to grow sIowly, efficiency improvements and new appliance standards are expected
to contribute to sIower future growth. The EIAs `Annual Energy Outlook 2015projected a total U,S.
electricity use growth of less than l% per year from 2014 to 2040 [1].
On the world level, the EI, in its International Energy Outlook 2013,projected that world ener-
gy consumption would increase 56% between 2010 and 2040, from 524 quadrillion to 820 quadrillion
British thermal units [3]. Most of this growth will come from non-Organization for Economic Coop-
eration and Development countries, Where demand is driven by strong economic growth. Renew-
able energy and nuclear power are the worlds fastest-grOWing energy sources, eaCh increasing 2.5%
Per year. However, fossil fuels are expected to continue to supply nearly 80% of the worlds energy
through 2040. Natural gas is the fastest-grOWing fossil fuel, aS global supplies of tight gas, Shale
Source: EiA, Annuai Energy Source: EIA, Annuai Energy Source: EiA, Manufacturing Energy
Outiook 201 5, Reference Case, Out10Ok 201 5, Reference Case, Consumption Survey 2010
Table 5 (Aprii 2015) labie 5 (Aprii 2015)
FiGURE l - ,S. electicity consumption in 2014 in the (a) esidentiai, (b) commeCiai, and O manufactuing sectos [1] (Coutesy of EiA.)
COmmunication concept has evoIved more ecient, SeCure, reliable, and generation and usage. The distribItiorz
toward two-Way electricity and infor- green, aCCOrding to SmartGrid.gov, a intelence (a term coined by the DOE
mation exchange between utilities DOE information gateway. One of the OE) enables inteIligent and automatic
and their customers, Offering unprec- Pivota=deas of the smart grid para- COntrOIs leading to self-healing and re-
edented levels of consumer participa- digm is distributed generation, With Silience to both malicious and benign
tion and thus creating the essence of microgrids relying on typically renew- failures. Buildings, eStimated to be re-
the modem smart grid. The develop- able energy sources such as hydro, bio- SPOnSible for about 40% of all energy
ing symbiosis of communications, COn- mass, SOIar, Wind, and geothermal, the used in the United States, Play a pivotal
troIs, and automation makes the grid COnCePt enabling autonomous power role in smart grids.
lntera`tiveness of Buiidings
and Smart Grids
One of the key elements of smart grid
-----
technoIogies is the interactive rela- e0
tionship between grid operators, utili-
"
ties, and consumers, Smart grids can
es S sea, 0
be viewed as interconnected resourc- s-0e0s, Ses.
es and consumers of energy, With
buildings as major entities of the equa-
gS ki
tion [mainly because of central hea also to help integrate variable renew- Simply referred to as s (Sam-
e
r
ing ventilation and air-COnditioning able power sources into the grid, Which Sungs home monitori
(HVAC) and lighting]. Hence, buildings With adequate financial incentives can aptly named Sma[7n they are
Play roles in both energy usage and as accelerate their market penetration. quickly entering the con mer world.
energy generation entities avith stor- Gartner Analytics predi a30% in-
age capabilities). Through interaction Cometivity (Sart = Come`ted
With smart grids, Smart buildings @s AnaltiS)
building blocks of smart cities [5], [6]) Another key concept of smart build-
Can Perform load reduction and peak ings is connectivity In fact, in acontext 2020 (Figure 3).
Shaving (reducing demand for electric- Of smart buildings and smart devices,
ity during peak usage time?) and load the words smar[ and connec[ed are est energy consumers in b ngS, are
Shifting, and can reduce blackouts (to- being frequently and interchange- being highly modemized ugh the
ta=oss of powe) and brownouts (volt- ably used to denote Intemet of Things Penetration of IoT device The key is-
age drops). The concept of islanding in (Ic/) devices. However, Certain dis-
microgrids can be reduced to the level tinctions need to be made, The mean-
Of buildings, Where one or several of ing of comected is typically reflected
them can share distributed resources by the ability of buildings to commu- 1
frequently integrated with distributed nicate with other buiIdings, With the doing at what time. IoT devices pro-
energy storage systems. Examples of
grid (for reasons of scheduIing, island-
distributed resources can entail re- ing, and peak shaving), With utilities,
newables Guch as rooftop solar, Small With energy storage unitsater tanks
turbine/wind, and hydro projectS) and for hot or cold energy storage and tht hV nt Xistd bfr, Th
be completed in 2018, aims at demon- COnneCted @.g., a thermostat needs to COnVenience (e.g., remOt OntrOl and
Strating competitively priced, multi- know about its environment through
megawatt, 6-10-h operation, 20-year an outside weather station or in-home
OPeration lifetime vanadium redox
battery energy storage systems [7].
Johnson ControIs LlOOO In-Building
usage pattems from sensor;). On the
Other hand, COnneCted does not nec-
essarily imply smart. For example, the
trol @utonomousntelli
Distributed Energy Storage System motion sensor may not be smart even ability to control and imp Ve energy
features 40-65-kWh storage capac- though it is connected to the thermo- efficiency. Comectivity is underly-
ity increments of lithium-ion (Li-ion) Stat, but the thermostat has to be con-
Cylindrical batteries with 20-year life nected for it to depIoy learning and
expectancy [8]. Another aspect of Prediction algorithms. Remote access
distributed load and generation are alone does not imply smart aspects
POWer back into the grid, a COnCePt
Called L,ehCle Jo grid [9]. As mobile ate interactively and autonomously food that needs restocki
energy storage devices, PEVs have While being wirelessly networked
the potential to inject additional power
into the grid at critical peak times and
with other such devices or connected
to the Internet directly. Sometimes
DECEMBER 2016 EEE -NUSTR-AEECTRON MAGAZNE 35
.>zzSPuOU=S-D
u>
u>i
PPd
u
SS
Ee
qPuSpu
Su
uquO< S
d$
pON
uOEOu<
S.
uSS
<
8u
Nest, Samsung, Sleep Number, and
Others improve quality of life and per-
e deveoping sybiosis oconications
SOnal health, One Of the major benefits
is energy efficiency, energy Savings, CO0s, andtomationakes the grid mor
and peak shaving, Starting with such
ecieSere, eabe, and gee
Simple tasks as deferring refrigera-
tor defrost cydes or dishwasher and
laundry deferral untiI off-Peak hours
[10, [11]. monitoring, and reporting on the and smart lock;) or larg' @.g., large
Theskyis the limit when it comes to lighting for any space in a building for energy storage batteries ld electric
energy efficiency and smart devices to- maximum energy efficiency, COmfort, Car?). As highly integra d as they
day. Windows, doors, and skylights can and productivity [16]. are, SuCh systems run hi( l eXPOSure
gain and lose heat through conduction An example of such concepts is so- risks. The same informati l flow that
(U-factor) or radiation (Solar heat gain Called doylight zJtOnOmy, a SuStain- enables users and mana!
COefficienO. In 1990 alone, the energy able concept driven by Lutron that building complexes and idential
used to offset unwanted heat losses adapts to its environment and claims home users to monitor Of COntrOl
large
and gains through windows in residen- reducing daytime lighting energy
tial and commercial buildings cost the use by 65% or more through the use
United States US$20 billion (One-fourth Of automated shades [16]. Designing
Of all the energy used for space heat- for daylight autonomy invoIves un-
ing and cooling) [12]. Smart windows, derstanding how the entire building ample, mOnitoring occu
Smart glass, Or SWitchable glass use is affected by the dynamic nature of Can tell when a person es home,
the technoIogy called s[ISpended po
POlymers and liquid crystal materials then so too are cyberthreats to build- In the world of home
to control transparency [14], [15]. Be-
Sides their use for energy efficiency,
ing services [17]. According to Grand
View Research, a San Francisco-based
the security research firm
these technoIogies can be used in COnSulting firm, the inteIligent build-
Smart buildings for instant privacy, tO ing automation technoIogies market is devices, eValuating sim featu res
eliminate the need for curtains, tO filter expected to reach US$98,95 billion by SuCh as the encryption of
ultraviolet rays, Or aS a rear-PrOjection 2024 [18]. The technoIogy is expected
Surface for theaters or high-PrOfile cor- to grow rapidly due to benefits such as
eywell, and Johnson ControIs provide thermore, the increasing demand for exploitation, SuCh devices Suffer from
SOlutions ranging from commercial
Smart grid, Whole building, and resi-
SeCurity and life safety systems (e.g.,
SyStemS that indicate the presence
hardware security issue
Lutrons Quantum lighting controand large commercial complexes is ex- CeSS, aS Shown in 2014 Si m ilarly,
lighting output during peak demand, Cations among sensors, hubs, and nu- is the introduction of IoT de-
instantly saving energy. The solutions merous smart devices, Whether smaIl vices into a network. exfiltra-
also provide control, COnfiguration, @.g., COffee machines, POWer Outlets, tion through any devic built-in
E.EMBER 216,US,R,ARN,CAZ,NE 3,
Plans should be drawn up for periods
during which intelligence is not avail-
VAa ing systS. tgSt nrgy able (known as plfor fIee tt)OrsO. To
maintain minimum acceptable levels of
SerVice, hardwired hardware may be a
o Penetration o=odevices necessary cost for the resilience of mis-
ion-Critical building automation sys-
network co a Department of Homeland Security tems, eVen at a COSt Of sacrificing the
and requir( alert [22]. intelligent part of the automation 7].
munications (e.g., if a SoIutions in this area need to take Resilience-a SyStems ability to
Smart rer into account various approaches. One bounce back after malicious or benign
Payment Ce ZOne) [20]. Another is- is increasing resilience by segmenta- failures or, from a business perspec-
Sue invoIve tion of the network of connected de- tive, the ability to maLintain continuity
application Vices (1imiting the exposur ), COuPled Of business operations-aSSumeS an
With the use of strong user/PaSSWOrd intelligent response that goes beyond
U.S. Federal Bureau StrOPhic events at data centers in one represent an integral component of
reported a memo on location. Cloud services must include resiIience. The interactiveness and
into the heating and multifactor authentication via various interoperability of modem buildings,
ng system of a New PrOOfs of identity-SOmething that us- along with the ubiquitous IoT devices
COmPany. Widely re- ers know, through a password; SOme- that generate large data streams, also
thing they possess, aS With a phone; bring bioinspired data mining and data
and something inherent to them, aS analytics into the purview of resilience
With fingerprints [23]. In addition to in the domain ontelligent buildings.
themselves
StateS, and identify anomalous scenar-
ios that may lead to suboptimal behav-
ior due to benign or malicious faults.
These approaches can be effectively
used for tasks ranging from building
energy management and energy effi-
Ciency, Self-healing, and adaptation to
the securitynformation assurance,
Brfef OveVfew of A/
artificialneuralnetworks@NN?),fuzzyprocessing,andcomputervisionappli-SeVeraldeepleamingar hitectures:
1ogicsystems(FLS;)(Figure5),SuPPOrtCationssuchasfacerecognition.convolutionalneuralnetwo ks(CNNi),
VeCtOrmachines(SVM?),anddiffer-Deeplearningextendsthecapabili-longshort-termmemOr (LSTM),
entBayesianprobabilisticmodelsaretiesofstandardANNsbyenablingar-andrestrictedBoltzman machines
We11-documentedmethodsandhavechitecturesofmanylayers.Therefore,(RBMi).Thesealgorithms reselected
beenextensivelyusedinawidearraydeepleamingenabIesleamingofcom-Sincetheyarerelevanttot econcrete
Ofdomainsforleamingpattemsandplexpattemsthatexistindatathroughapplicationandcasestudy iscussedin
trendsindata,Recently,theadventofmultiplelayersofabstraction.Thethenextsections"However, tshouldbe
FuzzyLogic naturallyindinedresearcherstoinves-Canbeusedinbuildingene
inganditsrecentwidespreadusehaveingandtraditionalAIalgo gysystems.
Ofintelligentbuildings.ThecapabiIityprehensivereviewofde Pleaming
OfleamingcomplexpattemsusingthetechniquesandmethodoIo ies.
multilayeredarchitectureenablesdeepANNsareAIarchitect resbased
learningalgorithmstoextractpatternsonbioIogicalneuralnet Orks(See
Fuzzy tigateitseffectivenessinthedomainReadersarereferredto[26
andtrendsfromgrowinganddiverseFigure6).Asinthelatte
foracom-
,thebasic
Ofthefuture.Therefore,throughneuronproducesanoutp tbycom-
FiGJRE5-FLS[35].theeffectiveuseofdeeplearning,theparingaweightedsumof lnPutStOa
(Ac tivationFunction)
HiddenHidden
s,/LayerLayer2
x
XnandA CtivationFunction
Wb
+1
FIGJRE6-AnANNachitectue[35L MAGAZiNE39
threshold. image classification, and time series at the grid level as well as at the in-
ics the bi neuron through a data modeling. dividual building level [33]. Modern
Similar methodoIogy of using weights LSTMs and RBMs are increasingly building design is heading in a direc-
and a thr Old value to produce an used in the literature as generative tion to incorporate energy storage de-
deep leaming algorithms. LSTM algo- Vices, SuCh as thermal energy storage
l rithms, introduced by Hochreiter and tanks [34], and/Or reneWable energy
rorlS is arranged to produce Schmidhuber [30], are a tyPe Of recur- generators, SuCh as photovoltaics or
given an input vec- rent neural network (RNN) composed Windmills. Thus, Smart buildings will
nnected network has Of memory cells with self-COnneCtions. have to make decisions on what pro-
leaming the interde- Figure 8 shows an LSTM architecture POrtions of utility energy, StOred en-
Ween the inputs and With multiple LSTM cells. These can ergy, and locally generated energy to
Well-trained ANN has be stacked in a multiIayer architecture use to attain optimal energy eficiency
to create a deep network. LSTMs have [35], Therefore, at the building level,
ltS for inputs it has not been shown to be extremely useful in accurate energy load forecasting helps
lSt. ANNs are we11 docu- generative models, eSPeCia11y in the make building-level decisions, SuCh
roven for being useful in natura=anguage processing domain. as optima=ocal energy storage con-
ions, RBMs [31] are shallow two-layer neural trol [34] or renewable energy control.
a special type of neural networks. The first layer of an RBM is From a grid perspective, Smart grids
network anq are mainly used to process Called the L)isib[e /qyer and the second have to optimally utilize the various
topoIogy (Figure 7) [29]. is thedden /qyer RBMs can learn fea- energy sources, including renewables
The CNN ning process is a great ex- tures that exist in data in an unsuper- [36]. Smart grids promise an unprec-
ample of deep learning algorithms Vised fashion. Therefore, they do not edented level of flexibility in energy
learn thr layers of abstraction. As need preamotated data to leam the management, making power genera-
features. RBMs are the building blocks tion and distribution more efficient
Of deep belief networks (DBNi) [32] and minimizing energy waste [33],
and have been used in a range of do- [38], [39]. Therefore, at the grid level,
mains, including image processing and accurate future predictions of individ-
time series data modeling. ual demand help the grid adapt to the
Variable demand, and having individ-
Concree Appat;on ofA/: De ual building-level energy forecasting
Leaming-Based Loademd helps the grid carry out the demand
tence of a erent feature. Hence, When eCOng h Bu()gS response locally [33]. In other words,
Put tOgethet, CNN layers have the capa- Load/demand forecasting is predict- local demand can be met with locaI
bility of le ing very complex patterns ing the energy consumption of an indi- distribution, Which leads to more e
and trend data. Figure 7 shows the Vidual building or an aggregate, SuCh ficient energy transmission and distr
arch itectu r Of a standard CNN. CNNs as a city or a county, for a future time bution. Therefore, it can be seen that
have been hown to be extremely use- SteP. Accurate load forecasts are ex- demand prediction can be discussed
fu=n fields SuCh as computer vision, tremely beneficial for decision making from an aggregate standpoint and an
Fu=y
Functions of a Convolutional Layer (Expanded) ConvoiutionaI Connected
Laye Laye r(S)
NN aChitecture [28].
individual standpoint. Howevern
ing [33], [39]. Regardless of the type of and smart cities of tomorrow Next, We
forecasting, tO aCCurately predict ener- look at a case study conducted by the
gy consumption, the forecasting meth- authors to compare the performance of Pared to the FCRBM met . Further-
OdoIogy should be able to accurately deep leaming techniques. more, tO keep the tests form. the
model the dependencies between build- Same training and testin data were
ing factors and energy consumption. Re- Case Study: Compahson of Deap used as in [33].
SearCh has shown that load forecasting Leamhg Aohms for Demond
is an extremely difficult problem [33], forecasfg h Bui/di/)gS Data Set
[40]. Hence, a range Of methods-Phys- s mentioned, deep learning method- The data set contains el(
ics principles-based and statistics and OIogies have been shown to be ben-
AI based-have been proposed in the eficial in the domain of building-level
literature. This articIe focuses on the demand forecasting. However, the
recent advancements in AI-based build- Published studies lack compara-
ing-leve1 1oad forecasting. tive analyses among deep learning
One of the most recent advances techniques. The presented case tudy
using deep learning techniques over tures implemented on the same data forthe next 60 hwas pr
traditional techniques. Mocanu et al. Set. The architectures were: Figure 9-(o illust
tested two variations of the RBM: the the standard LSTM-based load abiIity present in the
factored conditional RBM (FCRBM) forecasting architecture ure 9(shows the
and the conditional RBM (CRBM) the LSTM-based Sequence-tO-Se- COnSumPtion across
[33]. In the study, the CRBM and the quence architecture (LSTM S2S)
FCRBM were compared to the NN, the CNN-based architecture.
SVM, and RNN. The aLuthors conclud- All three architectures were tested
ed that the FCRBM outperforms all the On a benchmark data set for a single
Other methods.
In another study, DBNs, Which are
architectures built with layered RBMs,
Were uSed for electricity load fore- O-1]iLSMCii y(t)
..EMBER 2,6,NR.A=..C,RN,MAAZ,N. ,,
the previous step and time stamp data
are fed into the model. Table l lists
1.2 the time stamp data used as inputs to
$3,500 3,000 2,500 .1 a11 three models. The output from the
LSTM network is the predicted power
08
COnSumPtion for the next time step.
Tb predict further into the future, the
06
0.4
02 Prediction of the next time step can be
0 0 2,2 2
1,8 0 used. The mode=s trained using the
Standard backpropagation through
time (BPTT) method [44] with gradient
Time(Months)Time(Day)
descent. For the elaborated methodo
(a)(b) e-Mon. ee. e-Wed. --hu. Fi.
Ogy, readers are referred to [39].
"6 ) . .
S2S architectures for neural networks
Were PrOPOSed by Sutskever et al. [45]
as a method of mapping sequences of
arbitrary lengths. The LSTM S2S archi-
tecture for load forecasting contains
two standard LSTM networks, namely,
1 the encoder and decoder. The objec-
0.8 0.2 0 - iiii tive of the encoder is to convert input
SequenCeS Of variable length into a
l-H
I STAMPINPUTSFORTHEDpEARNINGALCORITHMS,
into the architecture as inputs. The
encoder network is trained alone as
a preprocessing step, uSing the same
methodoIogy mentioned in the stan-
INPUTVARIAB DESCRIPTION VAUS
dard LSTM method. Then the encoder
Houoftheday HourofthedayfoJthefirstpediction [1,24] is plugged to the decoder, and both
networks are trained together using
Month Monthofthefistpediction [1,12!
BPTT. For the eIaborated methodol-
DyOftheweek Dayoftheweekfothefistprediction [17]
Ogy, readers are referred to [39].
DyOfthemont Dayofthemonthfo=hefirstpediction Il,]
Weekendfiag aVerageen rgy Flagthatissetifthedayofthefistpedictionisaweekend COnSumPtionacrossFurthermore,itshows IO,1] thattheweek- CNN-Based Load Forecasting
Time series energy consumption
data are viewed as a one-dimensional
Figure lO shows a sample prediction Humans in the building can be divid- the human element from the perspec-
obtained from the standard LSTM. It ed into two main categories: 1) occu- tive of the occupants shod be taken
can be observed that the LSTM is able PantS Or uSerS Of the building and 2) into consideration on the following
to follow the general trend of the pre- building managers. The behavior of two fronts to optimize d buildings
Pant COmfort:
mOdeling occupant activity
0102030405060
Human-Buing /ntemcfron for
Time(Hours)
Comfortdey Efenay
While data-driven intelligent algo-
FIGURE 12 - The pediction esuIts fo the CNN [46]. rithms can provide contro=or in-
CreaSed energy efficiency, balancing
Case Stu(m-Machhe
eCnrO)g O FIGURE 14 - The inteface of the deveioped tooI fo incopoating the building manager
behavior of the building and thus are by the authors. However, those frame- plication. The functionality was kept
CaPable odentifying behavior that WOrks are out of the scope of this ar- identical across the two versions.
is anomalous and potentially harm- ticle. The developed visualization tool Figure 14 shows the interface for the
ful. Similarly, linguistic summaries of focuses on increasing the building Windows and the Android versions.
data are generated using an AI-based managers situational awareness by Furthermore, the tooI was extend-
framework [52]. In this framework, the PrOViding a comprehensive overview ed to a web-based version to make
data from the building are summarized Of building performance. it platform independent. Figure 15
lighting, temPerature, and ventilation
(increasing loads at times of excess gen- SuCh as the personification of smart tinue their way into the buildings of
eration). Grid-interactive hot and cold building components @s with thera- the future. From automation and occu-
thermal energy storage systems (tanks) Peutic robots like ParD, Or queStions Pant behavioral pattem learning and
Will further advance load shifting and in- Of human changes in response to con- Prediction to processing of the big data
CreaSe the sustainability and autonomy tinuous interactions with buildings. generated by buildings and smart grids,
Of smart buildings. Advanced power Software will play an increasing role AI techniques have already been estab-
electronics wi11 be able to provide sub- in the buildings of the future. Similar 1ished as an approach to the autono-
minute load management [54]. to software virtualization (hard disk mous execution of routines on behalf of
When it comes to human-building drives or software-defined network?), users. With regard to energy efficiency
interactions, the increase of modali- the virtualization of power plants that as one of the most important aspects
ties and platforms of heterogeneous COmbines geographically dispersed Of future smart cities, aPPrOaChes such
ecosystems that live and breathe with energy-management SyStemS from 2] US Energy Informationdministration.
(2010) Manufacturing energy consump-
their surroundings. Intelligent build- multiple research projects funded by tlOn SurVey U.S. Dept. Energy, WashlngtOn,
ings adapt to external (ieasonal) and the U.S" Department of Energy and in- D.C. [Online]. Available: http:www eia.govI
COnS umPt lOn/manufactu r lng
intemal (OCCuPanCy and usage pat- dustry leaders. He is an IEEE Student [3] U.S. Energy Information Administration. Ouly
tem9 changes and are doing so with Member and a member of the IEEE In_ 2013). !nternationaLl energy outIook 2013 U S.
Dept. Energy, Report DOE/EIA-0484. [Online.
lnCreaSing autonomy and sustainabil- dustrial Electronics Society. Available http:www.eia.gov/todayinenergy/
PrOfessor at the Advanced Diagnosis, dedSysVOl. 201l, PP. 1-8, Jan. 2011.
Biographies Automation, and ControI Laboratory, [7] D. Alderton, Vionx Energy Corporation distrib-
uted energy storage system demonstration:in
Ms Mic Gnisko@ieee.org) is a Electrical and Computer Engineering Proc. DOE OEeI 5tomge Pkmm fieeRe!e
and (drMeeg, Portland, OR, Sept. 2015.
PrOfessor in the Computer Science De- Department, North Carolina State Uni-
[8] Johnson ControIs. (2016, July). Distrlbuted
Partment and director of the Modern VerSity, Raleigh. His research interests energystorage system: L1000 In-Bulldlng [On-
Heuristics Research Group at Virginia include the implementation of compIex 1ine]. Avable: http:www.johnsoncontroIs
COm/-/medla/jcl/des/files/Pdfs-SePt-2016/
Commonwealth University, Richmond. COntrOl and processing algorithms 11000seIIsheeL_uPdatedO716.pdf