Energy 151 (2018) 729e739

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

A zone-level, building energy optimisation combining an artificial

neural network, a genetic algorithm, and model predictive control
Jonathan Reynolds*, Yacine Rezgui, Alan Kwan, Sole

ne Piriou
BRE Trust Centre for Sustainable Engineering, Cardiff University, Cardiff, UK

a r t i c l e i n f o a b s t r a c t

Article history: Buildings account for a substantial proportion of global energy consumption and global greenhouse gas
Received 14 August 2017 emissions. Given the growth in smart devices and sensors there is an opportunity to develop a new
Received in revised form generation of smarter, more context aware, building controllers. Therefore, in this work, surrogate, zone-
26 January 2018
level artificial neural networks that take weather, occupancy and indoor temperature as inputs, have
Accepted 20 March 2018
been created. These are used as an evaluation engine by a genetic algorithm with the aim of minimising
Available online 20 March 2018
energy consumption. Bespoke 24-h, heating set point schedules are generated for each zone in a small
office building in Cardiff, UK. The optimisation strategy can be deployed in two modes, day ahead
Keywords:
Building energy management
optimisation or as model predictive control which re-optimises every hour. Over a February test week,
Artificial neural network the optimisation is shown to reduce energy consumption by around 25% compared to a baseline heating
Genetic algorithm strategy. When a time of use tariff is introduced, the optimisation is altered to minimise cost rather than
Model predictive control energy consumption. The optimisation strategy successfully shifts load to cheaper price periods and
HVAC control reduces energy cost by around 27% compared to the baseline strategy.
Heating set point scheduler © 2018 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).

1. Introduction zones are heated when they are not required to be.
Energy infrastructure is also undergoing substantial changes.
Buildings account for a considerable proportion of global energy Energy is becoming increasingly decentralised with the concept of
consumption and greenhouse gas emissions [1]. Therefore, microgrids and the smart grid gaining traction [4]. Large scale,
improving energy efficiency in this sector is gaining increased focus centralised, fossil fuel power plants are giving way to more local
from research and industry. The recent growth in popularity of the renewable resources and smaller scale local generation. This sub-
Internet of Things, IoT, means that future buildings will be equipped stantially reduces energy transmission losses and allows genera-
with a wealth of potential sensing devices. This additional infor- tion to be far more efficient as waste heat from power generation
mation provides an exciting opportunity to reduce building energy can be utilised in local heating systems through cogeneration units.
consumption as it could be leveraged by a new generation of smart Energy is also no longer unidirectional, increased use of small scale,
building controllers to manage building energy consumption in a residential PV solar panels have given rise to the concept of the
more efficient way. It is estimated that buildings have the potential ‘prosumer’, one who both consumes and produces energy. Given
to reduce their energy consumption by 20e30% whilst using that the share of controllable energy production is decreasing
existing building components [2]. Dynamic information such as through use of stochastic renewable generation, the system must
occupancy and outdoor weather conditions are not currently transition from a demand led network to one that considers both
considered in the internal logic of traditional building management supply and demand as partially controllable. This could come in the
systems, BMS, which largely employ a reactive, rule based approach form of direct demand response, DR, controls or through encour-
[3]. Furthermore, many older, smaller buildings have one central aging consumer behavioural change through dynamic time of use,
thermostat that controls the temperature set point throughout the TOU, tariffs. Therefore, the next generation of smart building
entire building. This leads to large energy wastage as unoccupied controller must not only take into consideration aspects such as
predicted occupancy and weather conditions, it must also be
adaptable enough to maximise the use of local renewable re-
sources, the use of energy storage, and schedule consumption
* Corresponding author.
around low energy price periods [5].
E-mail address: ReynoldsJ8@Cardiff.ac.uk (J. Reynolds).

https://doi.org/10.1016/j.energy.2018.03.113
0360-5442/© 2018 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
730 J. Reynolds et al. / Energy 151 (2018) 729e739

To address these needs, this paper will demonstrate a zone- scheduler utilising EnergyPlus was developed in Ref. [13] with the
level, heating set point scheduler that minimises the energy con- aim of simultaneously controlling the thermal comfort, visual
sumption over the next 24 h whilst maintaining thermal comfort comfort and indoor air quality whilst minimising the energy con-
within the building. The remainder of this paper is organised as sumption. It used a genetic algorithm which used an EnergyPlus
follows. Section 2 provides a review of related literature. Section 3 model as the evaluation engine to control window blinds, venti-
discusses the modelling of the case study building in a simulation lation, and window opening operation for just a single zone. HVAC
environment, EnergyPlus, and as a series of Artificial Neural Net- operational optimisation was addressed in Ref. [14]. An EnergyPlus
works, ANN. Section 4 outlines the optimisation strategy. Section 5 model was combined with a MATLAB multi-objective GA to mini-
gives the results of the optimisation strategy, compares day ahead mise annual energy consumption, thermal discomfort and pro-
scheduling to hourly model predictive control for both a standard ductivity loss by setting the heating and cooling set point
tariff scenario and a TOU tariff scenario. Section 6 provides the temperature. However, the same set point temperatures were used
conclusion. throughout the entire year failing to adjust to variable weather of
occupancy conditions of each day. In both [15,16], Ascione has
2. Related works developed a multi-objective GA optimisation procedure to control
indoor set point temperatures using an EnergyPlus model to eval-
The optimisation of building controls is currently a popular topic uate potential solutions. Both case studies have demonstrated sig-
in the literature. This is illustrated by a recent review paper [6] nificant potential energy savings, however, the case study building
which assessed over 100 peer-reviewed papers. The review as- was relatively simple, containing just three zones. Using the Ener-
sesses the benefits of different control schemes, optimisation gyPlus model as an evaluation engine led to a computational time
techniques and prediction model methods. It determines that the of 90 min to develop an optimal schedule for the next 24-h. Such a
most popular control strategy found in the literature is Model computational period would inhibit the use of sliding-window,
Predictive Control, MPC. This is also confirmed by a further review MPC, which would have to re-optimise every hour.
[7]. MPC has proven to be valuable due to its ability to adapt to In practice, using a detailed white box simulation in conjunction
unforeseen disturbances or prediction errors, its ability to exploit a with an advanced metaheuristic optimisation strategy, such as a
buildings' thermal mass, take account of variable energy pricing GA, is not possible in most scenarios targeting operational opti-
and be able to shift loads away from peaks. Whilst MPC appears to misation. This is due to the considerable number of evaluations
be the leading control scheme, there is still debate of the most required per iteration and the computational time required to
appropriate modelling methods and optimisation techniques to complete an evaluation. The previously discussed works focus on
deploy in conjunction with the MPC control scheme. A number of very simple building energy models or just a single zone. To apply
advanced computational methods exist that can be utilised for these methods to a realistically complex building would require
optimal building control and these are discussed in Ref. [8]. These significant computational power to reduce simulation times to
include metaheuristic optimisations, multi agent systems, fuzzy acceptable limits (i.e. below 1 timestep). Thus, the focus must turn
logic controls and ANN. The paper also suggests the use of cloud to creating surrogate, black or grey box, models which can accu-
computing to achieve the optimisation and relay results to be rately replicate the output of a white box model but can compute
implemented by the existing Building Management Systems, BMS. with minimal computation expense and time allowing their use in
MPC strategies applied to building control optimise decision real time.
variables over a time horizon which usually ranges 8e24 h ahead of An example illustrating the use of surrogate models combined
the current timestep. Only the first timestep, usually 15 min to an with optimisation can be found in both [17,18]. A TRNSYS model
hour, of the optimal strategy is implemented before re-optimising was run several times to produce a representative bank of data
with updated feedback from the relevant sensors [9]. Therefore, from which an ANN was trained. The developed ANN accurately
the controller must have an internal model of the process to be able predicted annual energy consumption and thermal comfort within
to calculate the objective function over the complete time horizon. the building based on retrofit design decisions as inputs. The ANN
Li [10], reviews methods of building modelling for optimisation of was combined with a multi-objective GA to minimise energy con-
building control, these include white, grey or black box modelling sumption, discomfort and retrofit costs. Both studies showed the
techniques. White box models include full energy simulations such benefits of deploying an ANN as opposed to a white box simulation
as TRNSYS or EnergyPlus, these models are highly detailed but take model as the evaluation engine due to the dramatic decrease in
a long time to accurately calibrate and run. Black box models have reported computational time. This type of scheme was further
no understanding of the physical properties they are attempting to enhanced in Ref. [19] which developed generic ANN that accurately
model, they are simply based on extensive amounts of training data replicated entire classes of buildings (e.g. an office built from 1920
that they are provided. This includes statistical models, ANN and to 70) rather than just a single building. Once combined with an
Random Forest models. These can achieve good accuracy and very optimisation procedure, the methodology recommended the most
low calculation time but require a large amount of training data cost-effective building retrofit measures depending on budget.
[11]. Grey box models are simplified physical models such as Magalha ~es [20], developed an ANN to forecast the annual energy
Resistor-Capacitance (RC) models, they also require some historical consumption of a building based on readily available energy per-
data to set their coefficients and also have a relatively low calcu- formance certificates, EPC, and specific user defined characteristics
lation time so they can be used for online optimisation. such as the length of the heating period and the percentage of area
In Ref. [12], the authors' coupled an EnergyPlus simulation with heated. The authors' argued that providing such information to
a MATLAB, MPC procedure using the middleware software BCVTB, occupants would allow more informed decisions in relation to
Building Controls Virtual Test Bed, which is designed to facilitate energy saving measures.
data exchange between EnergyPlus and MATLAB. The MPC scheme Rather than address design or retrofit decisions, this paper aims
controlled the extent of the pre-cooling with the objective of to target operational optimisation of building energy consumption.
minimising energy cost. The various potential solutions were The authors aim to emphasise the importance of optimisation every
assessed in EnergyPlus and compared to typical control strategies. day to adjust to the specific conditions at hand. Thus, the building
However, the case study building was very simplistic due to the energy models required must have prediction granularity of an
simulation time involved in complex, realistic buildings. A 24 h hour or less. Benedetti [21], developed a methodology to automate
 J. Reynolds et al. / Energy 151 (2018) 729e739 731

the generation a of sub-hourly building energy consumption ANN. The importance of considering occupancy was shown in
In a Rome based case study, they found that a minimum of two Ref. [31]. A distributed MPC strategy is developed where each room
months of historical data was required to accurately predict the has an independent controller that can exchange data with
next sixty days of electricity consumption. Three measures of ac- neighbouring controllers to ensure heat gains from adjacent zones
curacy were used and once all three measures fell below a pre- are considered. Pisello [32], reviewed the difference in building
defined threshold the ANN was deemed inaccurate and re- performance between the design stage assumptions and the actual
trained. Similarly [22], tests a sliding window approach or accu- post occupancy reality. By studying the real occupancy patterns
mulative training of an ANN to predict sub-hourly electricity con- throughout a large multipurpose building in New York, alterations
sumption. The sliding window approach retrains the ANN every day can be made to controller schedules to achieve potential savings of
using the previous four weeks data. Accumulative training uses all up to 20.5%. Erickson [33], also developed a HVAC control strategy
the available data to train the ANN. Both models performed equally based on occupancy. A Markov Chain occupancy model is devel-
well with an average percentage error around 5%. Afram [23], oped to allow the building control strategy to take advantage of
developed a new algorithm for training ANN which was applied to sporadically occupied zones to save up to 20% on an EnergyPlus,
modelling several HVAC components. The ANN were integrated simulation-based, case study.
into a MPC platform to control the ventilation rate, buffer tank set From a review of the relevant literature it is clear building
point temperature and indoor set point temperature. The control controllers need to become more context aware, considering both
scheme showed aptitude for reducing the energy costs of the house predicted weather conditions and occupancy profiles. Furthermore,
by shifting the load to cheaper time periods. However, the building predictive control needs an accurate yet simple enough prediction
only has one set point temperature rather than zone level and oc- model to be able to deploy in (near) real time. Optimisation stra-
cupancy was not considered in the MPC formulation. tegies could make significant energy savings if they are focussed at
Papantoniou [24], optimised the operation of fan coil units in a a zone or room level, ensuring that energy is only consumed when
Greek hospital. An ANN predicted the outdoor temperature and the necessary. Therefore, the main contribution this paper makes is
indoor temperature also taking the HVAC operation as an input. A summarized as follows:
genetic algorithm was used in conjunction with a fuzzy controller
to minimise the cost of the energy consumption and ensure ther-
Zone level ANN have been developed to accurately forecast the
mal comfort for the occupants. However, the optimisation time indoor temperature and energy consumption by considering
horizon was limited to only 8 h. Lee [25], used an ANN based MPC variable weather, occupancy and temperature set points.
strategy to control a zone AHU. It aimed to calculate the optimal
This is combined with a genetic algorithm to optimise the
AHU cooling operation over the next 24 h to minimise the energy temperature set point to minimise either energy consumption
cost and maintain thermal comfort using Mixed Integer Non-Linear or energy cost within a computationally short period.
Programming, MINLP. The ANN an accurately predicted indoor
The effect of deploying the optimisation as day-ahead optimi-
temperature and energy consumption, but the application was sation or hourly, sliding window MPC was assessed.
limited to only a single zone within a building. An ANN based
The control scheme was demonstrated to be adaptable to time
controller was also developed in Ref. [26]. The ANN predicts the varying energy prices.
change in indoor conditions including temperature, relative hu-
midity and the Predicted Mean Vote, PMV. These predictions are
subsequently used to control heating, cooling, humidifying and 3. Modelling methodology
dehumidifying devices to minimise over or undershoots often
found in non-predictive, conventional control. Whilst this approach The methodology involved a case study building, i.e. a small
provided better thermal comfort compared to conventional con- office building in Cardiff, UK. The building was scanned using a Faro
trollers, it did not consider the minimisation of energy consump- 3D laser scanner to generate a point cloud of the building. From this
tion as an objective in its control scheme. an as-built representation was created in the BIM software Auto-
MPC using grey box modelling techniques were applied to a desk Revit. The relevant floor plans and building sections could
Czech university building in Refs. [27,28]. Blocks of the building then be exported and used to draw an accurate representation of
were modelled using an RC model taking weather predictions as the buildings' geometry in the energy simulation software Design
inputs. The optimisation was set up as a linear quadratic pro- Builder, the resulting model is shown in Fig. 1. Construction and
gramming problem and the objective was to minimise energy material properties, occupancy profiles, lighting and electronic
consumption by controlling the supply water temperature set equipment specifications were inputted to the model based on a
point. This strategy was implemented on the real building for over 2 building survey. The building contains 6 conditioned zones
months and was shown to reduce energy consumption by 15%e including 3 office spaces, a reception, a kitchen and a meeting
28%. Whilst this optimisation considers occupancy as a disturbance, room. The building is naturally ventilated and cooled, and an
it does not include predicted occupancy as a model input. electrical heating system was modelled with separate zone ther-
Furthermore, only block level supply water temperature is mostat controls assumed. In this paper, we have not considered day
controlled rather than the desired set point temperature in each by day occupancy prediction and simply assume the same occu-
zone. Oldewurtel [29], adapted traditional MPC to Stochastic MPC. pancy schedule for each working day as office building occupancy
Essentially this means the MPC strategy took into consideration patterns are fairly consistent throughout the working week. We
uncertainties in forecasts when carrying out the optimisation. This have assumed the 3 office zones and the reception occupied from
resulted in a slightly more cautious optimisation that did not go so 08:00 until 19:00. The kitchen occupied from 12:00 until 14:00 and
close to the comfort boundaries whilst still achieving good energy the meeting room from 10:00 to 11:00 although if deployed in
savings. Molina [30], produced an MPC strategy to control heating reality, meeting room occupancy patterns would be available from
and cooling in a residential building using a state space model as an the electronic booking system used for this zone.
evaluation engine for a GA. However, this work considered unre-
alistically simplified ideal heating and cooling and the control 3.1. Modelling using and artificial neural network
strategy only considers a 1-h prediction horizon which is not long
enough to be able to effectively utilise pre-heating or pre-cooling. For the optimisation utilised in this paper, it was necessary to be
732 J. Reynolds et al. / Energy 151 (2018) 729e739

There is no leading method in the literature by which these pa-

rameters can be optimised therefore the authors have followed a
largely trial and error based approach similar to the methodology
outlined in Ref. [34]. Throughout the ANN architecture trials, the
ANN accuracy was measured using the coefficient of variation of
the root mean squared, CVRMSE, based on the 4-week testing data
described above. The parameter with the largest influence on
prediction accuracy was the selected inputs. It was found that
including the indoor temperature from hour t-2 and t-3 as well as
relative humidity made the accuracy of the ANN prediction worse,
so these were removed as potential inputs. These results suggest
that the thermal lag of the building zones is not long and only the
previous hours' indoor temperature value is required. Furthermore,
solar irradiance was found to decrease the accuracy when used as
an input for zone 1 and 6, therefore this was removed from these
zones' ANN. Varying the other ANN parameters had a much more
limited effect provided two hidden layers were used and each layer
held a reasonable number of neurons. The eventual architecture of
the ANN for zone 1 and 6 was 5-20-20-2 and for zones 2 to 5 were
6-20-20-2 (shown in Fig. 2). The selected training algorithm was
‘Levenberg-Marquardt backpropagation’ and the transfer function
Fig. 1. Design Builder model of the case study building. between each layer was ‘tansig’.
Table 1 displays the Pearson's’ correlation between ANN output
able to predict the heating energy consumption and the indoor and target values during both the training and testing stage. The
temperature of each conditioned zone at each hour of the day for results display excellent prediction from the ANN. The consistency
the entire 24-h, time horizon. This calculation needed to be between the prediction results during the training phase and the
completed quickly to be combined with a GA optimisation strategy, testing phase shows no evidence of overfitting meaning the ANN
therefore the full energy simulation could not be used as an eval- has learned the general trends in the data rather than merely
uation engine. Hence, an ANN surrogate model for each zone was finding the best fit to the training data set. However, note that
trained using the simulation data produced by the energy model so during the testing phase the previous indoor temperatures were
it could accurately replicate it during the real-time optimisation. To assumed perfectly predicted. When deployed in the optimisation it
produce the ANN training data set, 27 separate simulations were is expected that prediction accuracy will decay slightly throughout
completed from the 1st of January to the 31st of March, each time the 24-h period due to prediction errors in Tit-1. The CVRMSE and
with a different heating set point schedule in the 6 conditioned the mean bias error (MBE) have also been reported graphically for
zones. Some of these training set point schedules were realistic, both energy consumption (Fig 3) and indoor temperature (Fig 4).
‘typical’, schedules whilst others varied randomly between 12  C Fig 3 shows that the CVRMSE for the testing data is around 30% for
and 24  C. This was designed to generate 58237 h of diverse each zone whilst the MBE remains within ±10%. This is higher than
training data which would allow the ANN to produce reasonable the comparative measures for the prediction of indoor temperature
accurate results throughout the entire range of possible set point for which the ANN performs very well (CVRMSE around 2%).
schedules. However, the poorer statistical performance of predicting energy
Inputs to the ANN needed to be known in advance to allow
prediction for the entire 24-h time horizon. The variables consid-
ered as inputs in this study were weather variables including out-
door temperature, relative humidity and solar irradiance. These
variables could reasonably be retrieved from local weather stations
with good forecasting accuracy. Additional variables include the
hour of the day, the set point temperature (the decision variable),
and a binary occupancy profile. Furthermore, given that thermal
inertia in a building is a considerable factor, the indoor temperature
from the previous 3 timesteps was also considered to be an input.
However, given that the requirement is to predict for the next 24 h,
the prediction of indoor temperature at time t is used as the input
to predict at time tþ1. These predictions are rolled over until the full Fig. 2. ANN architecture for zone 2 to 5.
24-h time horizon has been completed. For example, the prediction
of energy consumption and indoor temperature at timestep 3 could
use the predicted indoor temperature from timestep 1 and 2 as well Table 1
Pearson correlation between input and output during training and testing.
as the initial measured temperature at the start of the optimisation.
The ANN were trained using the MATLAB ANN toolbox and once Zone number Zone name r Value
completed were tested against a 4-week long EnergyPlus simula- Training Testing
tion with variable set point temperatures and using an alternative
1 Downstairs Office 0.99964 0.99955
weather file. When configuring an ANN, there are many tuneable 2 Kitchen 0.99987 0.99982
parameters which can influence the quality of the resulting pre- 3 Reception 0.99980 0.99985
diction model. These include the selected inputs, the number of 4 Meeting Room 0.99984 0.99981
hidden neurons in the hidden layers, the number of hidden layers, 5 PhD Office 0.99908 0.99941
6 Researchers Office 0.99971 0.99968
the training function and the transfer functions between layers.
 J. Reynolds et al. / Energy 151 (2018) 729e739 733

40 value for each hour. Each individual solution is evaluated to assess

CVRMSE
its ‘fitness’ to the objective function. From this fitness, the in-
MBE
dividuals are ranked in order of preference and this determines
their likeliness to ‘crossover’ with another individual to produce the
30
next generation of solutions. Individual genes within a solution
have an opportunity to ‘mutate’ to a random feasible solution and
thus ensure that the optimisation does not get stuck in local
20 minima. Some solutions are defined as ‘elite’ individuals, these
represent the top percentage of solutions that pass to the next
%

generation unchanged by crossover or mutation.

The process of producing new generations continues until a pre-
10
defined stopping criterion has been met. This could relate to a
maximum time, maximum number of generations, or related to the
change in optimal solution over time. The MATLAB optimisation
0 toolbox GA function was used in this paper, so it could be simply
coupled with the ANN which were also developed in MATLAB. The
exact parameters of the GA used in this paper is shown in Table 2.
Note that the maximum number of generations is set to the MAT-
-10
1 2 3 4 5 6 LAB default of 100 multiplied by the number of decision variables
which in this instance is relatively high. This allows the GA to exit in
Zone Number
most cases by reaching the function tolerance which ensures that
Fig. 3. Statistical measurements of ANN prediction for energy consumption. the GA has fully converged rather than forced to exit prematurely.

4.1. Objective function and fitness evaluation

2
CVRMSE
MBE The objective of this optimisation strategy is to minimise the
energy consumption whilst maintaining thermal comfort by
1.5 selecting the optimal temperature set point schedule, each hour, for
each zone. The set point is free to vary between 12  C and 24  C
during unoccupied periods and 20  Ce24  C during occupied times
1 as these were the temperature bounds requested by the occupants
to maintain thermal comfort. Whilst the setting of these bounds
%

forms a large part of ensuring thermal comfort is met a further

internal penalty function is included. If the indoor temperature
0.5 predicted by the ANN is below 20  C or greater than 24  C when the
zone is occupied, then the energy consumption during that time-
step is set at 100 kWh effectively excluding that solution from being
0 competitive in the fitness evaluation and hence discarded. This
penalty function is mainly necessary during the first occupied hour
of the day where it is conceivable that the zone set point temper-
ature would be above the lower bound of 20  C but the indoor
-0.5
1 2 3 4 5 6 temperature would remain lower than this during the first hour
while the zone warms up.
Zone Number
The fitness evaluation procedure developed in this paper is
Fig. 4. Statistical measurements of ANN prediction for indoor temperature. displayed in Fig. 5. The input variables are combined into one
matrix with the appropriate structure to be inputted into the ANN.
These include the outdoor temperature, solar irradiance, hour of
consumption is partly due to the nature of the data. Heating energy the day, occupancy, temperature set point and previous indoor
consumption is much more variable in nature with a few, large temperature. Once the inputs are collated, they are fed to the zone
peaks but an overall low mean. This is significantly harder to pre- ANN which predicts energy consumption and indoor temperature
dict than the gradual evolution of indoor temperature. for that timestep. Then follows the thermal comfort check to ensure

4. Optimisation strategy
Table 2
Genetic algorithm parameter settings.
As elaborated in the previous section, a GA is used to optimise
each zone's set point temperature for the next 24 h. This section GA Parameter Setting used
will provide finer detail of the optimisation process. A GA is a Number of Variables 24
population based, meta-heuristic, searching algorithm inspired by Population Size 200
the process of natural selection [35]. These have commonly been Creation Function Uniform
Selection Function Tournament
applied to building problems due to their ability to cope with non-
Crossover Function Scattered
linear characteristics often found within building control as well as Elite Count 5%
their tendency to converge to the global optimal rather than local Mutation Function Uniform
optimal solutions [36]. A GA contains a population of solutions, Mutation Rate 0.1
each individual solution contains ‘genes’, which in our case are 24 Max Number of Generations 2400
Function Tolerance 1  105
set point temperatures between 12  C and 24  C representing one
734 J. Reynolds et al. / Energy 151 (2018) 729e739

in reality, instead of using the EnergyPlus simulation model, you

would simply record the measured indoor temperature in each
zone before carrying out the optimisation. The procedure is dis-
played in the diagram shown in Fig. 6.

5. Optimisation results

In this section the GA-ANN, zone level, heating set point

scheduler will be applied during a test week in February using
Fig. 5. Flowchart of the fitness evaluation procedure. actual, 2016, weather data from a nearby weather station in Cardiff
which was converted to an epw file for use in EnergyPlus. To pro-
vide a comparison, a baseline scenario has also been developed.
that during occupied hours the indoor temperature is predicted to This uses the current heating set point strategy of the building
be above 20  C and below 24  C. If this is not the case the energy which is 21  C during the occupied hours (08:00 to 19:00) and 12  C
consumption for that time step is changed to 100 kWh. Unless all during unoccupied hours in all 6 conditioned zones. First, the
24 h have been calculated, the process loops around to repeat the optimisation will be run as day ahead scheduling and then as MPC
calculation for the next timestep using the internal temperature with a 1-h timestep and 24-h control horizon. Note that in both
prediction from the previous hour as an input. Once all 24 h have cases the schedules resulting from the optimisation will be put back
been completed, the energy consumption is summed over the 24 h into EnergyPlus to validate the results. This allows fair comparison
and the resulting number is the solutions' fitness. A GA using the with the baseline scenario as both simulation models are identical
described procedure is completed for all 6 conditioned zones. This (including weather conditions) apart from the heating set point
procedure can be accomplished in parallel to reduce optimisation schedule of the zones. This also removes any influence ANN pre-
time as each zone optimisation is independent and not reliant on diction errors may have to allow true evaluation of the effect of the
inputs from other zones. optimised set point strategy.
Two optimisation scenarios were run, one where there is a
standard, flat pricing tariff and one using a time of use, TOU, tariff.
4.2. MPC adaptations In the first scenario, the optimisation will aim to minimise energy
consumption and hence cost will also be minimised. When the time
The optimisation procedure described in the previous sub- of use tariff is used the optimisation will aim to minimise electricity
section can be run once at midnight and produce a schedule for cost for heating which will not necessarily minimise energy con-
the following day provided it has 24-h weather and occupancy sumption. The optimisation requires minimal adjustment to ach-
predictions and the initial zone temperatures. However, in this ieve this objective change. Given that the energy consumption at
study, the effect of implementing this strategy as MPC will also be each hour is already calculated during the optimisation, this is
assessed. When implemented 24-h ahead without MPC, errors in simply multiplied by the price per kWh at that particular time of
temperature prediction at earlier timesteps can lead to compound day which is the same for every weekday. Whilst popular in con-
errors later in the day. Once set, the entire heating set point tinental Europe and parts of America, TOU tariffs have still yet to
schedule would be enacted regardless of any unforeseen changes in achieve significant penetration in the UK. However, this is pre-
circumstances. However, if implemented as MPC, the optimisation dicted to change after a government backed roll-out of smart me-
would be run every hour, still with a 24-h time horizon. This would ters and the first widely available TOU tariff is now available from
allow feedback from the building system of internal temperatures Green Energy UK, [38]. The energy prices from their TIDE tariff is
allowing the controller to react to any prediction errors or receive a that used in Section 5.2 and the price variation is shown in Fig. 7.
more up-to-date weather forecast. Running as MPC means the 24-h Energy is cheapest, £0.0499/kWh, from 23:00 to 06:00 and has
set point schedule is updated and changed every hour but only the peak prices of £0.2499/kWh between 16:00 and 19:00, all other
first hour of each optimisation is ever enacted. hours are an intermediate price of £0.1199/kWh.
As this is a simulation based case study, the ‘real’ building is
replicated by an EnergyPlus simulation model thus a method of
5.1. Standard energy tariff
automatically linking the EnergyPlus model and the MATLAB
optimisation procedure was required. The Building Controls Virtual
The optimisation strategy was run for each day from the 15th to
Test Bed, BCVTB [37], middleware software was used to achieve
the 19th of February and the subsequent schedules were compiled
this. The data interchange, facilitated by BCVTB, was set up so that
on the hour the indoor temperatures of each zone are recorded
from the simulation model. Using these initial values, the optimi-
sation procedure could run and generate a 24-h set point schedule
for each zone that was sent back to BCVTB to be implemented in the
EnergyPlus model. The simulation model would then continue for
the next hour with the first set point values. Once this hour was
complete the temperature was again recorded by BCVTB, passed to
MATLAB and the optimisation is run again with the updated, ‘real’,
temperatures from the building. Therefore, only the first hour of the
optimal set point schedule is ever implemented but the optimisa-
tion time horizon remains at 24-h to give it the foresight to plan
ahead and allow the possibility of pre-heating or turning off early.
Note that the optimisation procedure that takes place each hour is
identical to that described in Section 4.1, however it occurs every
hour rather than just once at the beginning of each day. If deployed Fig. 6. MPC procedure using BCVTB.
 J. Reynolds et al. / Energy 151 (2018) 729e739 735

30 lower than that of the baseline scenario. In summary, both opti-

misation modes have shown significant energy savings can be
made by allowing a smart scheduler to have the freedom to vary set
25 point temperatures between pre-defined bounds and by actively
Electricity Price GBP Pence / kWh

considering occupancy and external weather conditions. This en-

ergy saving comes with no impact on the thermal comfort for oc-
20
cupants as the temperatures remain above the 20  C lower bound.
However, it has not demonstrated the value of MPC over a simpler
day ahead scheduling approach.
15

5.2. Time of use tariff

10
Both the day ahead optimisation and the MPC optimisation were
run again to include the TIDE TOU tariff and altered to minimise
5 cost of heating. The same week was studied using the same
weather conditions. The same baseline scenario is used which does
not make any attempt to adjust to the new pricing regime. The
0 results of these optimisations are shown in Table 4. As is clear,
0 2 4 6 8 10 12 14 16 18 20 22 24
under the TOU tariff optimisation the energy savings are lower
Hour of the Day
compared to that shown in Section 5.1. This is because the opti-
Fig. 7. TIDE tariff electricity price on weekdays. misation objective function is now related to minimising the cost of
energy, not the energy consumption itself. In terms of cost, the
savings compared to the baseline are around 27% and once again
Table 3
there is very little difference (£0.32 over the week) between the
Optimisation results using a standard energy tariff.
MPC optimisation scheme and the day ahead control strategy. Fig. 9
Zone Energy consumption/kWh Savings vs baseline/ is a representative example of the approach the optimisation tried
%
to take during the TOU scenario. The figure shows that the opti-
Baseline Day ahead MPC Day ahead MPC misation attempts to pre-heat between 05:00 and 06:00 which is
Downstairs Office 30.82 27.05 27.17 12.22 11.84 the last time period where the electricity price is at its lowest.
Kitchen 16.16 1.84 1.79 88.62 88.90 Furthermore, there are smaller energy consumption spikes at 13:00
Reception 42.08 39.84 39.14 5.32 6.98 and 15:00 with the aim of reducing the energy consumption during
Meeting Room 50.92 7.66 7.39 84.96 85.48
PhD Office 121.92 101.31 102.08 16.91 16.27
the on peak price period of 16:00e19:00 which it successfully
Researcher Office 58.25 57.93 63.73 0.54 9.41 achieves when compared to the baseline strategy.
Whole Building 320.15 235.63 241.31 26.40 24.63
5.3. Discussion

The results shown in Section 5.1 and 5.2 clearly indicate that
into a week-long schedule (note the optimisation does not run on
implementing a smarter, more context aware building controller
weekends). The energy consumption of each zone under both
can lead to improvements over traditional static control. Optimis-
optimisation strategies is shown in Table 3. There is a very minor
ing at a zone level rather than setting a building level strategy can
difference between the day ahead optimisation and the MPC. In
lead to significant energy and cost savings. The ANN surrogate
fact, the day ahead optimisation slightly outperforms the MPC with
models developed in this paper have been proven to be accurate
the main difference coming in the researcher office. Both optimi-
enough to replicate the simulation model in this case study. How-
sations show the potential for around 25% energy savings over the
ever, future work will aim to implement this control strategy on a
course of this test week. The main source of the energy savings
real case study building in the future to validate this conclusion.
come from the kitchen and the meeting room which are sporadi-
The optimisation strategy has proven to be flexible to a changing
cally occupied but are currently heated all day reflected in the
energy environment. It was simply adapted to take into account a
baseline scenario. The zones that achieve the lowest energy sav-
TOU tariff. Further adjustments could simply be made to factor in
ings, the reception and the researcher office, are directly adjacent to
local renewable resources or demand response events as part of a
the meeting room. Therefore, the lower energy savings in these
district heating network potentially benefitting the energy provider
zones are not necessarily a failure of the optimisation but due to the
as well as the consumer.
lack of heat gain from the meeting room which is now only heated
Throughout both tariff scenarios, the results show negligible
for a fraction of the time. Furthermore, zones such as the down-
difference between the day ahead optimisation and the MPC opti-
stairs office achieve energy savings over the baseline strategy
misation. This contradicts results published in many other state of
whilst still having the same 08:00e19:00 occupied hours. To un-
the art building control papers. However, this may be due to the
derstand these savings, Fig. 8 shows the set point schedule, indoor
lack of uncertainty in the testing scenarios presented in this paper.
temperature and energy consumption for the downstairs office on
Both occupancy and weather conditions are assumed known in
the 15th of February for both the baseline and optimal scenario. As
advance and these forecasts are assumed 100% accurate which
this figure shows, the optimal strategy chooses to more gradually
would not be true in practice. Therefore, future work will introduce
heat the room with some heating between 07:00 and 08:00, it then
forecasting uncertainty and assess the impact on the two optimi-
targets a lower temperature just above the 20  C bound during the
sation scenarios. The hypothesis being that the MPC optimisation
morning, both of which result in a much lower morning peak.
will adjust to these uncertainties better than the day ahead
Between 12:00 and 15:00, when the solar gains are higher, the
prediction.
optimal strategy chooses to heat the building to a higher temper-
An additional point of future work will aim to create a mecha-
ature to make the late afternoon energy peak, at around 19:00,
nism by which each zones' optimisation can influence adjacent
736 J. Reynolds et al. / Energy 151 (2018) 729e739

2.5
Optimised

Energy Consumption / kWh

Baseline
2

1.5

0.5

0
0 2 4 6 8 10 12 14 16 18 20 22 24
Hour of the Day
24
Ti Optimised
Ti Baseline
22
Tsp Optimised
Temperature / Deg C

Lower Bound
20

18

16

14

12
0 2 4 6 8 10 12 14 16 18 20 22 24
Hour of the Day
Fig. 8. Comparison between MPC optimisation and the baseline; for energy consumption (above) and indoor temperature (below) on the standard energy tariff, in the downstairs
office (Feb 15th).

Table 4
Optimisation results using a TOU energy tariff.

Zone Baseline scenario Energy savings/% Cost savings/%

Energy/kWh Cost/£ Day ahead MPC Day ahead MPC

Downstairs Office 30.82 4.335 1.55 5.13 9.87 6.84

Kitchen 16.16 2.443 90.01 88.18 92.32 91.11
Reception 42.08 6.001 5.44 6.13 8.20 10.18
Meeting Room 50.92 7.409 83.46 79.50 87.62 86.60
PhD Office 121.92 18.296 15.49 14.73 19.16 17.66
Researcher Office 58.25 8.386 6.96 9.53 0.88 0.04

Whole Building 320.15 46.869 23.31 21.28 27.94 27.26

zones. In this study, each zone is optimised separately. This was a sensors and direct control of heating units. Currently, there is a
conscious decision to allow each zone optimisation to run in par- significant surge in interest and availability of smart home devices
allel, hence reducing the total optimisation time to the order of controlled by a central AI coordinator using the paradigm of IoT. It is
10 min. Despite the lack of interaction between the zone optimi- therefore feasible and indeed probable that most future (and some
sations the proposed procedure was able to achieve significant current) buildings, both commercial and residential, will have the
energy savings with no loss to thermal comfort. This was likely due capability to control individual room set points and devices
to the set point schedules not deviating significantly day-to-day, through an integrated system. The proposed optimisation proced-
the optimisation altered set points only somewhat. Therefore, the ure would sit above these physical systems requesting and sending
heat transfer from zone to zone did not vary enough to have a relevant information (set points and indoor temperatures) taking
significant impact and prevent the optimisation working. Future advantage of existing physical and network infrastructure. It is
work will aim to pre-screen case study buildings in order to assess envisaged that this control scheme would be more applicable to
closely coupled zones and develop a method by which decisions commercial buildings initially. This is due to occupancy patterns
made in one zone are transmitted to the second. being more clearly defined and predictable within office buildings
To be able to practically deploy this solution to a real building and the fact that occupants do not necessarily expect to have direct
would require a reasonably small amount of additional hardware. control over the heating systems.
The optimisation procedure would require zone level temperature The most significant challenge in the application of this control
 J. Reynolds et al. / Energy 151 (2018) 729e739 737

2.5
Optimised
Baseline

Energy Consumption / kWh

2

1.5

0.5

0
0 2 4 6 8 10 12 14 16 18 20 22 24
Hour of the Day
24
Ti Optimised
Ti Baseline
22 Tsp Optimised
Lower Bound
Temperature / Deg C

20

18

16

14

12
0 2 4 6 8 10 12 14 16 18 20 22 24
Hour of the Day

Fig. 9. Comparison between MPC optimisation and the baseline; for energy consumption (above) and indoor temperature (below) on the TOU energy tariff, in the downstairs office
(Feb 16th).

strategy is the development of the surrogate models for the pre- for improved knowledge mining and feature extraction. For
diction of energy consumption and indoor temperature. The example, the system could assess wider environmental variables to
approach used in this study was to train an ANN based on large investigate correlation between these and energy saving actions. It
amounts of simulated data. However, accurate simulation models could also lead to more advanced reasoning to allow better pre-
are not widely available for most buildings. It is theorised that diction or measurement of building occupancy.
building simulation models are likely to become more available in
the future, driven by government legislation aiming at reducing
energy consumption from buildings and improving retrofitting 6. Conclusion
procedures. This is leading to increased prevalence of Building In-
formation Modelling, BIM, which are increasingly including energy This paper has shown the development of a GA-ANN, zone level,
analysis modules. Researchers are working on methods to capture heating set point scheduler. A simulation model produced a bank of
existing building information, convert to a digital representation, training data from which zone level ANN could be trained. These
from which generate a building energy simulation model and took weather, occupancy, set point schedule, and previous indoor
calibrate the model based on existing historical data [39]. Alter- temperature as inputs to predict the energy consumption and in-
natively, if the case study in question has developed a significant door temperature at the next timestep. A GA then used the ANN as
log of historical energy consumption and temperature data, ma- an evaluation engine to calculate the energy consumption over 24 h
chine learning models could be directly generated from this. To as the fitness function. GA optimisation of each zone took place in
model at an hourly or sub hourly temporal scale the authors' parallel to create bespoke set point schedules for each zone, each
believe that specific ANN would be required for each building as day.
generic ANN based on broader building categories would not be The optimisation was run in two modes, day ahead optimisation
able to capture the intricacies if an individual building. and MPC. In day ahead mode, the optimisation was carried out once
The authors also argue that any such building energy optimi- at the beginning of the day whereas the MPC strategy re-optimised
sation strategies should be performed within a semantically every hour with updated information. Furthermore, two scenarios
enriched environment. A semantic model should encompass as- were considered, one using a standard flat pricing electricity tariff
pects relating to energy consumption and management both and the other using a TOU tariff. Using the standard tariff, the
within the building and beyond to wider local energy networks. optimisation reduced energy consumption by around 25% in both
This method of linking data combined with the increased sensing modes. With the TOU tariff the objective was altered to minimise
capability that could be provided by the IoT could provide the basis cost and the optimisation achieved a cost reduction of around 27%
for both modes successfully shifting load to cheaper pricing
738 J. Reynolds et al. / Energy 151 (2018) 729e739

periods. Environ 2010;45:739e46. https://doi.org/10.1016/j.buildenv.2009.08.016.

[18] Asadi E, Da Silva MG, Antunes CH, Dias L, Glicksman L. Multi-objective opti-
Future work will introduce weather and occupancy forecasting
mization for building retrofit: a model using genetic algorithm and artificial
uncertainties and assess how the two optimisation modes deal neural network and an application. Energy Build 2014;81:444e56. https://
with this. Future work will also aim to integrate an optimisation doi.org/10.1016/j.enbuild.2014.06.009.
strategy like this as part of a wider district or microgrid setting. [19] Ascione F, Bianco N, De Stasio C, Mauro GM, Vanoli GP. CASA, cost-optimal
analysis by multi-objective optimisation and artificial neural networks: a
Finally, once robust enough we aim to implement this control new framework for the robust assessment of cost-optimal energy retrofit,
strategy on a real case study rather than a simulated building to feasible for any building. Energy Build 2017;146:200e19. https://doi.org/
validate the results in a real-world trial. 10.1016/j.enbuild.2017.04.069.
[20] Magalha ~es SMC, Leal VMS, Horta IM. Modelling the relationship between
heating energy use and indoor temperatures in residential buildings through
Acknowledgements Artificial Neural Networks considering occupant behavior. Energy Build
2017;151:332e43. https://doi.org/10.1016/j.enbuild.2017.06.076.
[21] Benedetti M, Cesarotti V, Introna V, Serranti J. Energy consumption control
The authors would like to acknowledge the financial support of automation using Artificial Neural Networks and adaptive algorithms: pro-
EPSRC (Engineering and Physical Sciences Research Council) and posal of a new methodology and case study. Appl Energy 2016;165:60e71.
https://doi.org/10.1016/j.apenergy.2015.12.066.
BRE (Building Research Establishment) as well as the European
[22] Chae YT, Horesh R, Hwang Y, Lee YM. Artificial neural network model for
Commission as part of the Horizon2020 THERMOSS (Project Id: forecasting sub-hourly electricity usage in commercial buildings. Energy Build
723562) and PENTAGON (Project Id: 731125) projects. Information 2016;111:184e94. https://doi.org/10.1016/j.enbuild.2015.11.045.
on the data underpinning the results presented here, including how [23] Afram A, Janabi-Sharifi F, Fung AS, Raahemifar K. Artificial neural network
(ANN) based model predictive control (MPC) and optimization of HVAC sys-
to access them, can be found in the Cardiff University data cata- tems: a state of the art review and case study of a residential HVAC system.
logue at http://doi.org/10.17035/d.2018.0047045867. Energy Build 2017;141:96e113. https://doi.org/10.1016/j.enbuild.2017.02.
012.
[24] Papantoniou S, Kolokotsa D, Kalaitzakis K. Building optimization and control
References algorithms implemented in existing BEMS using a web based energy man-
agement and control system. Energy Build 2015;98:45e55. https://doi.org/
[1] European Parliament. Directive 2010/31/EU of the European Parliament and of 10.1016/j.enbuild.2014.10.083.
the Council on the energy performance of buildings. Off J Eur Union 2010;53: [25] Lee YM, Horesh R, Liberti L. Optimal HVAC control as demand response with
40. on-site energy storage and generation system. Energy Procedia 2015;78:
[2] Guan X, Xu Z, Jia Q. Energy-efficient buildings facilitated by microgrid. IEEE 2106e11. https://doi.org/10.1016/j.egypro.2015.11.253.
Trans Smart Grid 2010;1:243e52. https://doi.org/10.1109/TSG.2010.2083705. [26] Moon JW, Kim J-J. ANN-based thermal control models for residential build-
[3] Haniff MF, Selamat H, Yusof R, Buyamin S, Sham Ismail F. Review of HVAC ings. Build Environ 2010;45:1612e25. https://doi.org/10.1016/j.buildenv.
scheduling techniques for buildings towards energy-efficient and cost- 2010.01.009.
effective operations. Renew Sustain Energy Rev 2013;27:94e103. https:// [27] Prívara S, Jan S, Ferkl L, Cigler J. Model predictive control of a building heating
doi.org/10.1016/j.rser.2013.06.041. system: the first experience. Energy Build 2011;43:564e72. https://doi.org/
[4] Kubli M, Ulli-Beer S. Decentralisation dynamics in energy systems: a generic 10.1016/j.enbuild.2010.10.022.
simulation of network effects. Energy Res Soc Sci 2016;13:71e83. https:// 
[28] Siroký J, Oldewurtel F, Cigler J, Prívara S. Experimental analysis of model
doi.org/10.1016/j.erss.2015.12.015. predictive control for an energy efficient building heating system. Appl Energy
[5] Reynolds J, Rezgui Y, Hippolyte J-L. Upscaling energy control from building to 2011;88:3079e87. https://doi.org/10.1016/j.apenergy.2011.03.009.
districts: current limitations and future perspectives. Sustain Cities Soc 2017: [29] Oldewurtel F, Parisio A, Jones CN, Gyalistras D, Gwerder M, Stauch V, et al. Use
1e15. https://doi.org/10.1016/j.scs.2017.05.012. of model predictive control and weather forecasts for energy efficient building
[6] Shaikh PH, Nor NBM, Nallagownden P, Elamvazuthi I, Ibrahim T. A review on climate control. Energy Build 2012;45:15e27. https://doi.org/10.1016/
optimized control systems for building energy and comfort management of j.enbuild.2011.09.022.
smart sustainable buildings. Renew. Sustain Energy Rev 2014;34:409e29. [30] Molina D, Lu C, Sherman V, Harley R. Model predictive and genetic algorithm
https://doi.org/10.1016/j.rser.2014.03.027. based optimization of residential temperature control in the presence of time-
[7] Afram A, Janabi-Sharifi F. Theory and applications of HVAC control systems - a varying electricity prices. IEEE Trans Ind Appl 2013;49:7e13.
review of model predictive control (MPC). Build Environ 2014;72:343e55. [31] Moroşan P-D, Bourdais R, Dumur D, Buisson J. Building temperature regula-
https://doi.org/10.1016/j.buildenv.2013.11.016. tion using a distributed model predictive control. Energy Build 2010;42:
[8] Ahmad MW, Mourshed M, Yuce B, Rezgui Y. Computational intelligence 1445e52. https://doi.org/10.1016/j.enbuild.2010.03.014.
techniques for HVAC systems: a review. Build Simul 2016;9:359e98. https:// [32] Pisello A, Bobker M, Cotana F. A building energy efficiency optimization
doi.org/10.1007/s12273-016-0285-4. method by evaluating the effective thermal zones occupancy. Energies
[9] Mirakhorli A, Dong B. Occupancy behavior based model predictive control for 2012;5:5257e78. https://doi.org/10.3390/en5125257.
building indoor climatedA critical review. Energy Build 2016;129:499e513. [33] Erickson VL, Cerpa AE. Occupancy based demand response HVAC control
https://doi.org/10.1016/j.enbuild.2016.07.036. strategy. In: Proc 2nd ACM Work Embed Sens Syst Energy-Efficiency Build -
[10] Li X, Wen J. Review of building energy modeling for control and operation. BuildSys10; 2010. p. 7. https://doi.org/10.1145/1878431.1878434.
Renew Sustain Energy Rev 2014;37:517e37. https://doi.org/10.1016/ [34] Yuce B, Li H, Rezgui Y, Petri I, Jayan B, Yang C. Utilizing artificial neural
j.rser.2014.05.056. network to predict energy consumption and thermal comfort level: an indoor
[11] Ahmad MW, Mourshed M, Rezgui Y. Trees vs Neurons: comparison between swimming pool case study. Energy Build 2014;80:45e56. https://doi.org/
Random Forest and ANN for high-resolution prediction of building energy 10.1016/j.enbuild.2014.04.052.
consumption. Energy Build 2017. https://doi.org/10.1016/ [35] Krarti M. An overview of artificial intelligence-based methods for building
j.enbuild.2017.04.038. energy systems. J Sol Energy Eng 2003;125:331. https://doi.org/10.1115/
[12] Ma J, Qin SJ, Li B, Salsbury T. Economic model predictive control for building 1.1592186.
energy systems. In: ISGT 2011. Anaheim, USA: IEEE; 2011. p. 1e6. https:// [36] Reynolds J, Hippolyte J, Rezgui Y. A smart heating set point scheduler using an
doi.org/10.1109/ISGT.2011.5759140. artificial neural network and genetic algorithm. In: Int. Conf. Eng. Technol.
[13] Ahmad MW, Hippolyte J, Reynolds J, Mourshed M, Rezgui Y. Optimal sched- Innov. Funchal, Portugal: IEEE; 2017. p. 726e32.
uling strategy for enhancing IAQ, thermal comfort and visual using a genetic [37] Wetter M, Haves P. A modular building controls virtual test bed for the
algorithm. In: IAQ 2016 Defin. Indoor air qual. Policy, Stand. Pract. Alexandria, integration of heterogeneous systems. In: SimBuild 2008, Berkeley, CA; 2008.
VA, USA: ASHRAE; 2016. p. 69e76.
[14] Papadopoulos S, Azar E. Optimizing HVAC operation in commercial buildings: [38] Green Energy UK. Tide tariff. 2017 (Accessed 01 August 2017), https://www.
a genetic algorithm multi-objective optimization framework. In: 2016 Winter greenenergyuk.com/Tide.
Simul. Conf. Washington DC, USA: IEEE; 2016. p. 1725e35. https://doi.org/ [39] Cho YK, Ham Y, Golpavar-Fard M. 3D as-is building energy modeling and
10.1109/WSC.2016.7822220. diagnostics: a review of the state-of-the-art. Adv Eng Inf 2015;29:184e95.
[15] Ascione F, Bianco N, De Stasio C, Mauro GM, Vanoli GP. A new comprehensive https://doi.org/10.1016/j.aei.2015.03.004.
approach for cost-optimal building design integrated with the multi-objective
model predictive control of HVAC systems. Sustain Cities Soc 2017;31: Nomenclature
136e50. https://doi.org/10.1016/j.scs.2017.02.010.
[16] Ascione F, Bianco N, De Stasio C, Mauro GM, Vanoli GP. Simulation-based
model predictive control by the multi-objective optimization of building en-
ergy performance and thermal comfort. Energy Build 2016;111:131e44. Abbreviation
https://doi.org/10.1016/j.enbuild.2015.11.033.
[17] Magnier L, Haghighat F. Multiobjective optimization of building design using
TRNSYS simulations, genetic algorithm, and artificial neural network. Build IoT: Internet of Things
 J. Reynolds et al. / Energy 151 (2018) 729e739 739

PV: Photovoltaics MBE: Mean Bias Error

DR: Demand Response BIM: Building Information Modelling
TOU: Time of Use
ANN: Artificial Neural Network Optimisation
MPC: Model Predictive Control
BMS: Building Management System
BCVTB: Building Controls Virtual Test Bed To: Outdoor Temperature
HVAC: Heating Ventilation and Air Conditioning Irr: Solar Irradiance
AHU: Air Handling Unit Occ: Occupancy
MILNP: Mixed Integer Non-Linear Programming Tsp: Temperature Set Point
PMV: Predicted Mean Vote Ti: Indoor Temperature
RC: Resistor Capacitance EC: `Energy Consumption
GA: Genetic Algorithm t: Current Timestep
CVRMSE: Coefficient of Variation of Root Mean Squared Error