PREDICTION OF ARTIFICIAL LIGTHING CONSUMPTIONS BY NEURAL NETWORKS MODELS

foulchilianΤεχνίτη Νοημοσύνη και Ρομποτική

20 Οκτ 2013 (πριν από 3 χρόνια και 9 μήνες)

60 εμφανίσεις

2
nd

European Conference on Energy Performance and Indoor Climate in Buildings

Lyon (France), November 1998



1

PREDICTION OF ARTIFICIAL LIGTHING CONSUMPTIONS BY NEURAL
NETWORKS MODELS


Alfio Galatà

CONPHOEBUS s.c.r.l.
-

Zona Industriale, Passo Martino 95030 Catania

Tel. +39 95 7489250

Fax. +39 95 291246 E
-
mail: conphoeb@tin.it



Preface

Building energy consump
tion predictions may be adopted to evaluate possible energy saving
by means of rehabilitation actions (building or plant), or to evaluate a more suitable
technological plants management. In fact after a rehabilitation action or a plant management
change, i
t is difficult to quantify building electrical consumption if the action was not
undertaken, this is due to varying boundary conditions (climatic conditions, occupancy, etc.).

Establishing these evaluations often need particularly complex simulation models
, which
require building
-
plant system detailed description. Neural Networks, on the other hand, allow
to solve this question by efficiently substituting simulation models, with no need of detailed
system description, achieving reasonable needed resource dr
opping joined to interesting
precision and reliability results.

The present paper will report about a Neural Network application for building artificial lighting
electrical consumption prediction. In particular a comparison is carried out, between electric

consumption measured and foreseen by two different lighting system management
strategies: first based upon a manual control; second based upon an automatic control
sensible to room occupancy scheduling and maximum use of natural gains (free).

Activities t
hat ENEL and CONPHOEBUS jointly carried out, finalised to the functional
performances of the CIB (Computer Integrated Buildings) system in an ENEL building,
allowed experimental data collection over two test rooms: a room (CONPHO) under
automatic control t
o maintain visual comfort as function of room occupancy, and equipped
with lighting power supply
dimming

regulation to exploit daylight, an other room (CIB) under
manual control over lights switching on and off.

CIB room has been considered the reference o
ne (building before rehabilitation) for energy
evaluations, while CONPHO room has been considered as environment building standard
with a different plants management (building after rehabilitation). Predictions were made
over artificial lighting system ele
ctric consumption within CIB room, as function of its own
occupancy scheduling, foreseen, for it, the same automatic control strategy applied within
CONPHO room. Difference, achieved by the comparison between measured and predicted
consumption over CIB roo
m, gives information on achievable energy saving for different
lighting system management.


Artificial lighting predictive controller

Control strategy, applied over CONPHO room, oversees visual comfort, upon user working
plane
(450


100 [lux]),
as functio
n of the room occupancy scheduling and free natural
lighting.

Artificial lighting are turned on in real time when personnel is within the room and the
measured illuminance value, upon user working plane, is below the prefixed comfort set
-
point value. Light
s are turned off, after a fixed delay (10 minutes), when no human presence
is achieved within the room or, when the measured illuminance value, over the user working
plane, is higher than illuminance values range established for visual comfort. When
person
nel is within the room and the internal illuminance value is within the visual comfort
fixed illuminance values range, artificial lighting power supply is tuned, acting on a
dimming

regulator, so to integrate natural lighting to maintain visual comfort.

Ne
ural Network model has three input and one output.

2
nd

European Conference on Energy Performance and Indoor Climate in Buildings

Lyon (France), November 1998



2



Input quantities
:

1.

Room occupancy scheduling (0 = Absence, 1 = presence)

2.

Illuminance upon the working plane

3.

Lights switching off delay (0 = no delay, 1 = under delay)



Output quantities
:

Artificial lighting
instantaneous electric consumption

The applied methodology was selecting a data set for the
training phase

to establish the rule
base, evaluating the learning capability over different data sets (
checking phase
) and
predicting lighting electric consumption

on CIB room as function of its own occupancy,
supposing for it the same control strategy applied on CONPHO room.

Several Neural Networks topology has been tested by trials, assigning, time by time, a
different number of neurones to typical layers (input,
hidden and output). Adopted selection
criteria was to verify which configuration achieves, by the same learning epochs, the
minimum absolute value of the SSE (sum of square differences between measured and
predicted values) parameter. The configuration ari
sing the minimum SSE parameter was
that with a neurones number for typical layers equals to 3:12:1.

Training phase

Training phase was leaded up to stabilise the value of SSE parameter up to relative small
numbers. Results are summarised in table 1.

Learnin
g Epochs

Electric measured
consumption

[Watt]

Electric predicted
consumption

[Watt]

SSE

1000

23882

23882

0.0294902

Table 1


Training phase results.



Figure 1


Measured and predicted consumption and residuals values for the training phase


2
nd

European Conference on Energy Performance and Indoor Climate in Buildings

Lyon (France), November 1998



3

Figure 1 re
ports graphs on measured and predicted consumption, and obtained residuals
showing differences ranging in (
-
5, +7) [Watt].

Checking phase

The model got, as inputs, quantities defined during the training phase, while no information
was specified for the out
put variable. Under these conditions, comparison between
measured and predicted values gives an evaluation item to the acceptability degree about
model prediction capabilities.

Figure 2 shows graphical trends for measured and predicted consumption, and res
idual
achieved from tested sample data.


Figure 2

Measured and predicted consumption and residuals values for the checking phase


Graphical analysis for the checking phase confirms the reasonable learning capabilities to
the chosen model. Table 2 summari
ses main statistical information, about the two days
taken as reference sample data.

Day

Measurement [ Watt ]

Prediction [ Watt ]


Consumption

Mean

Min.

Max

Consumption

Mean

Min.

Max

17 March

39146

46.60

0

288

39187

46.65

0

288

12 April

38144

45.41

0

28
8

38183

45.46

0

288

Table 2


Statistical information about input and output quantities


Results achieved from both training and checking phases, despite of sophistication of the
control algorithm and simplifications due to the neural model, show predicte
d values within
10% of measured ones.

2
nd

European Conference on Energy Performance and Indoor Climate in Buildings

Lyon (France), November 1998



4

Instantaneous predictions

Model predictive capabilities, after training and checking phases, were furthermore verified
through instantaneous predictions on the same sample data.

In particular, Neural Network model run
ning on an hourly loop predicted the lighting
consumption instantaneous values as function of instantaneous values about illumination on
the working plane, occupancy and lights switch off time delay.

In the following figures most explanatory instances are
shown, as a demonstration of the
optimum prediction capabilities achieved by the model. Each figure shows on the top
measured and predicted consumption trends, on the bottom trends for occupancy and
illumination on the working plane and set
-
points for visu
al comfort.

Figure 3 shows predicted lighting instantaneous consumption as function of the room
occupancy, showing also how lights are switched on during the following ten minutes [
1
]
after the human presence has been achieved in the room.


Figure 3


I
nstantaneous predictions at nine o’clock

F
igure 4
shows

the control strategy efficiency for real time lighting switch
ing

on
,

and
the
delayed switch off, and the tuning effect to the power supply realised by a dimming regulator.


Figure 4


Instantaneous
predictions at fifteen o’clock




1

fixed delay within the ap
plied control strategy

2
nd

European Conference on Energy Performance and Indoor Climate in Buildings

Lyon (France), November 1998



5

Electric consumption prediction for CIB room

Simulating the behaviour of CIB room, which is subjected to the same control strategy of
CONPHO room, the difference between measured and predicted consumption points out
energy sa
ving potential, with comfort levels adopted for CONPHO room.

Figure 5 reports on top graphic trends for measured and predicted consumption, on bottom
integral values for the same quantities, obtained for CIB room during the two days with
sample data.


Fig
ure 5


Measure and predicted consumption, and integral values for CIB room

Table 3 summarises the main statistical information on the two reference.

Day

Measured [ Watt ]

Predicted [ Watt ]


Consumption

Mean

Min.

Max

Consumption

Mean

Min.

Max

17 March

139680

166.28

0

288

40781

48.54

0

288

12 April

155088

184.62

0

288

61469

73.17

0

288

Table 3


Statistical information on input and output


It is evident that if the control strategy applied on CONPHO room was adopted in the CIB
room as well and visual
comfort conditions were maintained for the whole period with its
room occupancy, energy saving potential could be about 70% for the first sample data set
and about 60% for the second sample data set.

Long period analysis: seasonal consumption prediction

St
arting from seasonal prediction of artificial lighting electric consumption in the two sample
rooms, it is to be confirmed long period indications (analysis with hourly data) on achievable
energy saving, when in the CIB room is adopted the same control str
ategy applied in the
CONPHO room.

About evaluations on electric consumption predictions, it is to be noted the use of hourly data

can hide, in some circumstances, the rules of the adopted control, this is due to:

2
nd

European Conference on Energy Performance and Indoor Climate in Buildings

Lyon (France), November 1998



6



Delay on lights switch off (it can happen t
hat against a presence of one only minute lights
stay on for twenty
-
one minutes, or it is verified an electric consumption for an unoccupied
room)



Effect due to the hourly data mean, which can hide instantaneous effects of control
parameters.

Moreover, nat
ural lightning contribution from windows (windows of the same size for both
rooms) depends exclusively from building exposure (rooms are exposed to West) and from
solar position over the day. It was not possible to measure this contribution, so it has been

supposed that it depends from the time of the day (less illuminance during morning, more
illuminance during afternoon).

Under these conditions, the chosen model presents four input variables and one output
variable, as shown in the following:



Input quanti
ties
:

1.

Hours of the day;

2.

Solar radiation (measured on the horizontal plane)

3.

Illuminance on the working plane

4.

Occupancy



Output quantities
:

Artificial lighting electric consumption

The adopted methodology was the same of that used for instantaneous prediction
s;
predictive model capabilities have been verified through training and checking phases. Then
they were applied to CIB room so to evaluate possible energy gain achieving from use of
control strategy for artificial lighting within this room. The adopted me
thod was to select, for
each climatic season, in turn, a data day taken from sample data set used for the training
phase, and the day after in the sample data set used for the checking phase.

Table 4 summarises main statistical information on sample data s
et for training phase, while
table 5 summarises main statistical information on sample data set for checking phase.

Climatic
season

hours

occupancy

[minutes]

Solar Rad.

[Watt/sqm]

illuminance

[lux]

Consumption

[Wh]


min

max

min

max

min

max

min

max

min.

ma
x

Spring

0

23

0

60

0

844

0

784

0

288

Summer

0

23

0

60

0

800

0

547

0

288

Autumn

0

23

0

60

0

604

0

485

0

288

Winter

0

23

0

60

0

578

0

432

0

288

Table 4


Statistics on input/output quantities for the training phase


Climatic
season

hours

occupancy

[minu
tes]

Solar Rad.

[Watt/sqm]

illuminance

[lux]

Consumption

[Wh]


min

max

min

max

min

max

min

max

min.

max

Spring

0

23

0

60

0

873

0

659

0

288

Summer

0

23

0

60

0

831

0

498

0

288

Autumn

0

23

0

60

0

633

0

456

0

288

Winter

0

23

0

60

0

574

0

426

0

288

Table
5
-

Statistics on input/output quantities for the
checking

phase


2
nd

European Conference on Energy Performance and Indoor Climate in Buildings

Lyon (France), November 1998



7

Table 6 summarises, for each climatic
period
, measured and predicted seasonal
consumption values.

Season

Epoch

SSE

Training

Free Predictions




Measured

[Wh]

Predicted

[Wh]

Measured

[Wh
]

Predicted

[Wh]

Spring

2600

0.060765

26394

26395

25510

25198

Summer

310

0.008201

24853

24865

27600

27546

Autumn

1430

0.022293

47293

47330

45203

45224

Winter

1410

0.044284

28596

28574

29141

28895

Table 6


Measured and predicted consumption for train
ing and checking phases


Model predictive capabilities have been verified through weekly consumption predictions.
Results are reported on table 7 and, in graphical form, on the following figures.


Spring

Summer

Autumn

Winter

Week

Meas.

Pred.

Diff.

Meas.

P
red.

Diff.

Meas.

Pred.

Diff.

Meas.

Pred.

Diff.


[Wh]

[Wh]

[%]

[Wh]

[Wh]

[%]

[Wh]

[Wh]

[%]

[Wh]

[Wh]

[%]

1

5349

5380

0,58

505

505

-
0,06

5181

5218

0,71

7032

7002

-
0,43

2

5877

5907

0,51

4142

4164

0,53

7677

7693

0,21

408

406

-
0,49

3

4879

4906

0,55

5217

520
4

-
0,25

7596

7579

-
0,22

378

367

-
2,91

4

5119

5108

-
0,21

5679

5664

-
0,26

9731

9716

-
0,15

9297

9276

-
0,23

5

5199

5234

0,67

3686

3685

-
0,03

8068

8067

-
0,01

6344

6304

-
0,63

6

6279

6259

-
0,32

5909

5879

-
0,51

5945

5962

0,29

4503

4460

-
0,95

7

4987

5019

0,64

6
21

623

0,37

7610

7617

0,09

6983

6898

-
1,22

8

6463

6437

-
0,40

2410

2411

0,04

8198

8215

0,21

7085

7060

-
0,35

9

5136

5113

-
0,45

3732

3735

0,08

6856

6855

-
0,01

8240

8247

0,08

10

2236

2231

-
0,22

5025

5015

-
0,20

9722

9730

0,08

5643

5637

-
0,11

11







6057

6
053

-
0,07

7443

7445

0,03

1824

1810

-
0,77

12







9470

9470

0,00

8469

8452

-
0,20







Total

51524

51594

0,14

53453

52408

-
0,09

92496

92549

0,06

57737

57467

-
0,47

Table 7


Weekly consumption for the CONPHO room


Figure 6
-

Summer: Weekly measured and
predicted consumption for CONPHO room

2
nd

European Conference on Energy Performance and Indoor Climate in Buildings

Lyon (France), November 1998



8

Seasonal prediction for CIB room

To compare measured and predicted electric consumption, it is needed to know the rooms
occupant behaviour on lights switching on. In CONPHO room the two light groups were
switched on, t
uned or switched off at the same time, in function of the room occupancy and
daylight. In CIB room the two light groups were switched on/off singularly and manually,
without visual comfort control. Hourly consumption in CIB room varies between the
maximum
values of 144 [W] (just one group turned on) and 288

[W] (two groups turned on),
while predictions trend to a maximum of 288 [W], following rules learned for CONPHO room.

Table 8 reports seasonal prediction results for CIB room, comparing predicted consump
tion
values (column
Consumption CONPHO
) against those values measured if both neon groups
were operating at the same time, as in the CONPHO room (column
Consumption CIB
), and
indicators for measured and predicted consumption during the minutes when the roo
m was
occupied.

Climatic

Consumption CONPHO

Consumption CIB

CONPHO Indicator

CIB Indicator

Season

[Wh]

[Wh]

[Wh/min]

[Wh/min]

Spring

51625

94604

8.38

15.36

Summer

59082

84717

8.76

12.57

Autumn

113546

116016

6.96

7.11

Winter

92809

97604

6.75

7.10

TOT
. GEN.

317062

392941

7.38

9.14

Table 8


Predicted and measured consumption in the hypothesis that the two neon groups were turned on at the

same time


Values of predicted electric consumption result less than 20% against those that would be
measured with
in CIB room, in the hypothesis that the two light groups were turned on at the
same time. It is confirmed a sensible saving on lighting consumption when it is adopted a
control strategy based on room occupancy and dimming regulation.

Following figures show
, for each climatic season, measured (manual lighting management)
and predicted (automatic control strategy applied) weekly consumption.




Figure 7
-

Summer: Measured and predicted consumption for CIB room


2
nd

European Conference on Energy Performance and Indoor Climate in Buildings

Lyon (France), November 1998



9

Conclusions

Predictions performed with Neural N
etwork models for light energy consumption
demonstrates the energy saving potential achieved with experimental data analysis, when an
automatic management is performed instead of a manual one.

Neural Networks running with one
-
minute data predicted correct
values of light consumption
to establish the required visual comfort, highlighted the optimal rule base to realise a real
time control.

Neural Networks running with hourly data confirmed results obtained with one
-
minute data,
showing huge amount of energy
saving for automatic lighting control strategy based on room
occupancy and maximum exploitation of daylight.


Bibliografy

[1] "
A smart control strategy for shading devices to improve the thermal and visual comfort
".

A. Galatà, F. Proietto Batturi, R. Viada
na.

4th European Conference: SolarEnergy and Urban Planning, Berlin 1996.

[2] "
Proteus: An Intelligent Control System for Buildings
".

A. Galatà, G.Leotta, R. Sidri.

4th European Conference: SolarEnergy and Urban Planning, Berlin 1996.

[3] "
VESCO
-

Verifich
e su sistemi innovativi per il controllo Energetico
-
Ambientale degli Edifici
".

Risparmio Energetico nella climatizzazione degli Edifici.

Ordine Quadro ENEL
-
CONPHOEBUS 1995
-
1996. Edificio ENEL D.E.R. di Bologna.

[4] "
Adaptive Control of HVAC processes using

Predictive Neural Networks
".

Curtis P.S., Braunemuhel M.J., Kreider J.F., Univ of International Solar Colorado.

ASHRAE Transactions, Vol. 99, Pt.1, 1993

[5] "
Heating and Cooling of Buildings, Designfor Efficiency
".

Kreider J.F., Rabl A.

McGraw
-
Hill, New
York, 1994

[6] "
Applicazione delle Reti Neurali Artificiali ai Problemi Energetici
".

Bidini G., Grimaldi C.N., Mariani F.

49° Congresso Nazionale dell'Associazione Termotecnica Italiana, Perugia (I), 1994.

[7] "
Application of modelling, identification tech
iniques and ANNs for shading devices
characterization
".

F. Aleo, A. Galatà, M. Alabiso.

European Conference on Energy Performance and Indoor Climate in Buildings.

Lion (F), November 1994.

[8] "
Energy Management in Central HVAC plants using Neural Networks
".

Curtis P.S., Braunemuhel M.J., Kreider J.F., Univ of International Solar Colorado.

ASHRAE Transactions, Vol. 100, Pt.1, 1994

[9] "
RetroLite: an artificial intelligence tool for lighting energy
-
efficiency
".

Bo Belcher, Clay; Roddis, W.M. Kim: Univ. of Ka
nsas, Lawrence, KS.

USA Energy and Buildings v 20 n2 1993.

[10] "
Applicazioni delle Reti Neurali nelle problematiche energetiche degli edifici
".

A. Galatà, M. Alabiso.

1
a

Conferenza Italiana degli Utenti MATLAB
-

Bologna, 1995.