Applying Artificial Neural Networks to Promote
Behaviour Change for Saving Residential Energy
Yaqub Rafiq
1
, Shen Wei
2
, Robert Guest
3
, Robert Stone
3
and Pieter de Wilde
2
1
School of Marine Science and Engineering, Plymouth University, Plymouth, U.K.
2
Building Performance Analysis Group, Plymouth University, Plymouth, U.K.
3
School of Electronic, Electrical & Computer Engineering,
University of Birmingham, Birmingham, U.K.
Abstract. In this paper Artificial Neural Networks (ANNs) is used to model ef-
fects of various human behaviour on energy consumption of the residential
buildings in the UK. A virtual model of a bungalow has been developed in
which various aspects of the, physical changes in the building such as wall and
floor insulation, single and double glazing combined by the human behaviour
aspects such as thermostat setting, various periods of door and/or window open-
ing etc. are modelled using EnergyPlus software for evaluating energy con-
sumption for a combination of scenarios. ANNs were then used to learn the ef-
fects of various human behaviours on energy consumption. The results demon-
strated that the ANN is capable of learning the effects that changes in the hu-
man behaviour have in evaluating energy saving in residential buildings and it
generated results very quickly for unseen cases.
1 Introduction
Reducing residential energy consumption is essential for achieving the UK Govern-
ment’s 2050’s target of lowering CO
2
emissions. This is due to a high contribution to
the CO
2
emissions as the total energy consumption of the residential buildings in the
UK is about 30% [1]. Occupant behaviour plays an important role in energy usage in
residential buildings [2, 3]. Due to high importance of occupant behaviour on the
energy consumption, the UK government has funded several projects in the past 10
years, such as the green deal [4] and the 22 (Build)TEDDI projects [5], exploring
how to improve occupant behaviour in residential buildings for energy saving.
Currently, dynamic building performance simulation (DBPS) is being explored to
help occupants make informed decisions on adopting ways to save energy in their
houses [6-8]. An advantage of using DBPS is its flexibility in conducting parametric
studies by changing one aspect with respect to the definition of either the building,
building system or occupants’ building operation in DBPS, while keep other aspects
constant. Therefore, the impact of that particular change on the building energy con-
sumption can be reflected clearly by the simulation results. In this process, the energy
saving potential of possible behaviour change options was evaluated as the difference
between the predicted energy consumption before changing the behaviour with the
Rafiq Y., Wei S., Guest R., Stone R. and de Wilde P..
Applying Artificial Neural Networks to Promote Behaviour Change for Saving Residential Energy.
DOI: 10.5220/0005049600030010
In Proceedings of the International Workshop on Artificial Neural Networks and Intelligent Information Processing (ANNIIP-2014), pages 3-10
ISBN: 978-989-758-041-3
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
one after the change.
In reality, behaviour change options, especially those relevant with occupants’
building operations, are complex [9, 10]. For example, the thermostatic setting in
winter can be set at any values within a reasonable range, i.e. between 14°C and
22°C. Therefore, using DBPS to predict the building energy consumption of all pos-
sible scenarios defined by these options or adding other options to this can make it a
complicated combinatorial problem, which would be unfeasible using DBPS. There-
fore, this paper has preliminarily tested the use of ANNs to model the energy con-
sumption with respect to behaviour change in buildings, and the performance of using
the ANN model for the prediction of changes in energy consumption has been vali-
dated against prediction results from DBPS.
2 Methodology
2.1 Case Study Building
The simulation model was developed based on a virtual UK bungalow building, as
shown in Figure 1a, by the teammates at the University of Birmingham. This house
has two bedrooms, a living room, a kitchen, a bathroom and a main corridor. The
main façade of the house faces the North.
(a) Case study house (b) Simulation model
Fig. 1. Case study Building and Simulation model.
2.2 Dynamic Building Performance Simulation
Generally, in residential buildings there are two ways for reducing energy consump-
tion [11]. The first way is by upgrading the insulation of the building façade or the
efficiency of the building systems, so that the building itself can be more energy-
efficient; the second method is by improving occupants’ operation of the building,
e.g. opening/closing windows and setting the temperature values for the heating sys-
tem etc., so that the building can be used more energy-efficiently. Based on this clas-
sification, the behaviour change options, investigated in this study, are listed in Table
1 with corresponding simulation scenarios defined for the DBPS.
4
Table 1. Simulation scenario definitions.
Behaviour change
options
Simulation
scenario (before)
Simulation scenario (after)
Upgrading external wall insulation No 50mm insulation
Upgrading ground floor insulation No 50mm insulation
Upgrading pitched roof insulation No 50mm insulation
Upgrading ceiling insulation No 50mm insulation
Upgrading window layers Single glazing Double glazing
Upgrading door layers Single glazing Double glazing
Lowering thermostatic setpoint 23°C
22°C, 21°C, 20°C, 19°C,
18°C, 17°C, 16°C or 15°C
Reducing window opening time 4 hours 3 hours, 2 hours, 1 hour or 0
Opening curtains during the day No Yes
Based on the information presented in the above table, 5760 possible scenarios can
be used to assess energy consumption in the building.
Figure 1b shows the simulation model of the case study building developed using
DesignBuilder [12], a commercial graphical user interface of EnergyPlus [13], which
is one of the popular DBPS tools in assessing building energy usage. In this paper,
EnergyPlus V7.2 was adopted as the simulation engine and the simulation model was
exported from the DesignBuilder to EnergyPlus. The weather data used in this simu-
lation was provided by the Climate Design Data 2009 ASHRAE Handbook, for Bir-
mingham UK applications. The simulation period was chosen as from 1
st
to 31
st
Janu-
ary, based on an interval of 30 minutes. The inputs of the simulation include defini-
tions of building construction, systems, occupants’ building operations and weather
conditions, while the output is heating energy demand during the simulation period.
In this study, the energy saving potential of each intervention listed in Table 1 was
calculated as difference between the predicted heating energy demand before the
intervention applied and the one after the intervention applied.
2.3 Artificial Neural Networks
ANNs have been widely used in a number of applications in buildings [14], e.g. pre-
dicting building energy consumption [15, 16], predicting thermal comfort in buildings
[17], designing [18, 19] and managing buildings [20, 21]. A Neural Network has the
ability to learn from experience and examples and then to adapt with changing situa-
tions [22], imitating some of the learning activities of the human brain. The ANN
model used in this study consists of two hidden layers of 10 and 5 neurones using
TANSIG activation function; one output layer of one neurone using LINEAR activa-
tion. It is a back propagation Neural Network [23] that was used to model the simu-
lated energy consumption at different behavioural scenarios, presented in Table 1.
The data for the Neural Network model was generated using EnergyPlus as discussed
in section 2.2. Both Levenberg-Marquardt (trainlm) and Bayesian regulation (trainbr)
5
back-propagation learning algorithms were used. In this study MATLAB R2013b
Neural Network Toolbox was used for developing the ANN model [24]. The training
and testing data consist of 8 input and one output variables.
2.4 Training and Validation of the ANN
In this study, two sets of data were prepared using EnergyPlus, one for training the
ANN and another for validating the trained ANN model. Figure 2a presents the com-
bination of the simulation scenarios used in training the ANN and Figure 2b shows
those used in validating the trained ANN model. The input data of the ANN are vari-
ous parameter definitions with respect to occupant behaviour in buildings and the
output of the ANN is the energy consumption for that simulation scenario. In Figure
2, the most left hand side parameters define the base case scenario of the case study
building against which the energy saving potential could be calculated for comparison
purposes. The remaining parameters on the right hand side columns define the possi-
ble interventions. Changes in the energy consumption is evaluated by replacing the
house original conditions from those of base case scenario to the one or a combina-
tion of scenarios presented on the right hand side column in Figure 2.
23°C
Heating setpoint:
21°C 19°C
4H
Window opening hours:
2H
0H
No
Curtain open during the day:
Ye s
17°C 15°C
External wall insulation:
Ground floor insulation:
Roof insulation:
External window type:
No
50mm
Single
Double
External door type:
Single
Double
Ceiling insulation:
No 50mm
No 50mm
No 50mm
Heating setpoint:
22°C 20°C
Window opening hours:
3H
1H
No
Curtain open during the day:
Yes
18°C 16°C
External wall insulation:
Ground floor insulation:
Roof insulation:
External window type:
No
50mm
Single
Double
External door type:
Single
Double
Ceiling insulation:
No 50mm
No 50mm
No 50mm
(a) ANN training dataset definition (b) ANN validation dataset definition
Fig. 2. Dataset definition.
After considering the possible combination of all simulation scenarios defined in
Figures 2a and 2b, finally, 2941 possible scenarios were generated for the ANN train-
ing and testing processes (1917 for training the ANN and 1024 for validating the
ANN).
3 Results
As mentioned before the MATLAB Neural Network toolbox was used for both train-
ing and testing the Neural Network. Both ‘trainlm’ and ‘trainbr’ back-propagation
algorithms were used. At first, due to the speed of learning, ‘trainlm’ was used in the
learning process. A back-propagation network architecture having 2 hidden layers
6
and one output layer, was adopted for both learning algorithms.
Initially the data was divided into two sets for training and testing/validating pur-
poses. The training data included all data for 23°C, 21°C, 19°C, 17°C and 15°C cov-
ering upper, lower and intermediate temperature sets of data. This constituted a total
of 1917 set of data. The testing date included the rest data (1017 sets of data in total).
0 200 400 600 800 1000 1200 1400 1600 1800 2000
0
500
1000
1500
2000
2500
Number of data Points
Energy Used
Neural Net Prediction
Target
0 200 400 600 800 1000 1200
500
1000
1500
2000
Number of data Points
Energy Used
Neural Net Prediction
Target
Fig. 3. Neural Network training results for
selected data.
Fig. 4. Neural Network testing/validating
results for selected data.
0 500 1000 1500 2000 2500
0
500
1000
1500
2000
2500
Number of data Points
Energy Saved
Neural Net Prediction
Target
0 100 200 300 400 500 600 700 800
500
600
700
800
900
1000
1100
Number of data Points
Energy Saved
Neural Net Prediction
Target
Fig. 5. Neural Network training results for
extended data.
Fig. 6. Neural Network testing/validating
results for reduced data.
Figure 3 shows the results of the ANN learning 1917 sets of data. The network
successfully learned the presented data in a few minutes. Figure 4 shows results of the
validation/testing to see if the Neural Network has satisfactorily learnt and general-
ised necessary information presented to it. In all Figures the horizontal axis shows the
number of data points used in training and testing the ANN model and the vertical
axis shows energy usage. From Figure 4 it becomes clear that the Neural Network has
learned the data within the range 15°C to 19°C perfectly but failed to learn the rest of
the data outside this range. A quick fix would be to add the upper range (all data
higher than 19°C) to the training set. Figures 5 and 6 show that by doing this both
training and testing of the Neural Network processes are successful.
One problem with this quick fix is that the generalisation aspect of the ANN has
not been verified. This model can be used within the range of 15°C to 23°C but re-
sults from any other data would be doubtful. Another issue with this process is that
the data sets used in the learning process is more than that of the testing process. In
order to find an acceptable solution to this problem, the following steps have been
taken:
1. The data (combined training and testing data) for each temperature degree was
plotted using different colours in a single graph. This information is presented
in Figure 7. Form Figure 7 it becomes clear that the data is divided into dis-
tinct clusters for each °C. The more clear division is observed in the range be-
tween 15°C to 21°C with the rest of the data between 22°C and 23°C.
7
2. Based on the findings from (a), to improve the quality of training and testing
data, the training data was sampled from within each cluster and the remaining
of the data from each cluster was used for testing purposes.
Figure 8 shows results of the training process using only the training data from
each cluster. Similarly Figure 9 presents results from the testing/validating of the
trained Neural Network with the rest of the unseen data. The overall result of the
combined training and testing data is shown in Figure 10.
0 500 1000 1500 2000 2500 3000
0
500
1000
1500
2000
2500
Number of data Points
Energy Used
23 Degrees C
22 Degrees C
21 Degrees C
20 Degrees C
19 Degrees C
18 Degrees C
17 Degrees C
16 Degrees C
15 Degrees C
0 200 400 600 800 1000 1200
0
500
1000
1500
2000
2500
Number of data Points
Energy Used
Neural Net Prediction
Target
Fig. 7. Combined training and testing data
clusters.
Fig. 8. Training for selected data from each
cluster.
0 200 400 600 800 1000 1200 1400 1600 1800 2000
500
1000
1500
2000
2500
Number of data Points
Energy Used
Neural Net Prediction
Target
0 500 1000 1500 2000 2500 3000
0
500
1000
1500
2000
2500
Number of data Points
Energy Used
Neural Net Prediction
Target
Fig. 9. Combined testing extended data
clusters.
Fig. 10. Combined testing all data clusters.
From both Figures 9 and 10 it is clear that by selecting data within each cluster, the
Neural Network was able to learn and generalise the information presented to it.
Hence this trained network can be used confidently to predict the energy usage and
energy saving using particular scenarios for the case study building.
4 General Applications to Real Structure and Future Research
This paper has demonstrated that the trained Neural Networks can successfully pre-
dict the simulated building energy consumption when general behaviour change op-
tions are applied to a virtual case study building. For applications to real buildings,
there are still some important issues that need to be handled:
1. developing a representative simulation model for the real building and using
it for the preparation of data to train the Neural Networks;
2. realistically capturing occupants’ real behaviours in operating the building,
so that the base case simulation model can be developed as close as possible
to real situations;
3. realistically quantifying the behaviour change options in accordance with the
8
real applications; and,
4. implementing other factors that may influence occupants’ choice of applying
behaviour change options in the ANN, such as indoor thermal comfort and
investment payback period, rather than using energy consumption only.
An initial exploration on solving the above challenges is currently on-going by the
authors of this paper, and an Energy Efficient Education tool(s) is being developed to
help building occupants make informed decisions on changing behaviour for saving
residential energy demand.
5 Conclusions
Reducing energy consumption in residential building is a major problem universally.
Occupant behaviour has shown to have an important role in reducing energy con-
sumption. Evaluation of energy consumption, using traditional analytical method
using DBPS tools is computationally expensive and very time consuming. In this
paper ANNs have been used to instantaneously evaluate the effects of various combi-
nation of human behaviour on the residential buildings in the UK. The results demon-
strated that the ANN techniques can be used to predict the simulated energy saving
potentials of various human behavioural changes which eliminates expensive and
time consuming DBPS operations.
Acknowledgements
The work reported in this paper is funded by the Engineering and Physical Sciences
Research Council (EPSRC) under the Transforming Energy Demand in Buildings
through Digital Innovation (TEDDI) eViz project (grant reference EP/K002465/1).
References
1. DECC, United Kingdom housing energy fact file, 2013.
2. Haas, R., H. Auer, and P. Biermayr, The impact of consumer behavior on residential
energy demand for space heating. Energy and Buildings, 1998. 27(2): p. 195-205.
3. Al-Mumin, A., O. Khattab, and G. Sridhar, Occupants’ behavior and activity patterns
influencing the energy consumption in the Kuwaiti residences. Energy and Buildings,
2003. 35(6): p. 549-559.
4. GOV. Green Deal. 2014; Available from: https://www.gov.uk/green-deal-energy-saving-
measures/overview.
5. TEDDINET. TEDDINET website. 2014; Available from: www.teddinet.org.
6. de Wilde, P., et al., Using building simulation to drive changes in occupant behaviour: A
pilot study, in Building Simulation 2013 Conference2013: Chambery, France.
7. Love, J., Mapping the impact of changes in occupant heating behaviour on space heating
energy use as a result of UK domestic retrofit, in Retrofit 20122012: Manchester, UK, 22-
26 January.
9
8. Kim, Y. K. and H. Altan, Using dynamic simulation for demonstrating the impact of
energy consumption by retrofit and behavioural change in Building Simulation 2013
Conference2013: Chambery, France.
9. Wei, S., R. Jones, and P. de Wilde, Driving factors for occupant-controlled space heating in
residential buildings. Energy and Buildings, 2014. 70(0): p. 36-44.
10. Fabi, V., et al., Occupants' window opening behaviour: A literature review of factors
influencing occupant behaviour and models. Building and Environment, 2012. 58(0): p.
188-198.
11. Gardner, G.T. and P.C. Stern, The Short List: The Most Effective Actions U.S. Households
Can Take to Curb Climate Change. Environment: Science and Policy for Sustainable
Development, 2008. 50(5): p. 12-25.
12. DesignBuilder. DesignBuilder website. 2014; Available from: http://www.design
builder.co.uk/.
13. DOE. EnergyPlus Energy Simulation Software 2014; Available from: v jkhttp://apps1.eere.
energy.gov/buildings/energyplus/.
14. Kalogirou, S., Artificial neural networks in energy applications in buildings. International
Journal of Low Carbon Technologies, 2006. 1(3).
15. Azadeh, A., S.F. Ghaderi, and S. Sohrabkhani, Annual electricity consumption forecasting
by neural network in high energy consuming industrial sectors. Energy Conversion and
Management, 2008. 49(8): p. 2272-2278.
16. Swan, L.G., V.I. Ugursal, and I. Beausoleil-Morrison, Occupant related household energy
consumption in Canada: Estimation using a bottom-up neural-network technique. Energy
and Buildings, 2011. 43(2–3): p. 326-337.
17. Luo, X., et al., A fuzzy neural network model for predicting clothing thermal comfort.
Computers & Mathematics with Applications, 2007. 53(12): p. 1840-1846.
18. Kalogirou, S.A., Long-term performance prediction of forced circulation solar domestic
water heating systems using artificial neural networks. Applied Energy, 2000. 66(1): p.63-74.
19. Kalogirou, S.A., S. Panteliou, and A. Dentsoras, Artificial neural networks used for the
performance prediction of a thermosiphon solar water heater. Renewable Energy, 1999.
18(1): p. 87-99.
20. Lee, W.-Y., J.M. House, and N.-H. Kyong, Subsystem level fault diagnosis of a building's
air-handling unit using general regression neural networks. Applied Energy, 2004. 77(2): p.
153-170.
21. Curtiss, P.S., M.J. Brandemuehl, and J.F. Kreider, Energy management in central HVAC
plants using neural networks. ASHRAE Transactions, 1994.
22. Rafiq, M.Y., G. Bugmann, and D.J. Easterbrook, Neural network design for engineering
applications. Computers & Structures, 2001. 79(17): p. 1541-1552.
23. Werbos, P.J., Beyond Regression: New tools for prediction and analysis in the behavioural
science, 1974, Harvard University.
24. MathWorks. Neural Network Toolbox. 2014; Available from: http://www.mathworks.
co.uk/products/neural-network/.
10