SPECIAL SECTION ON CYBER-PHYSICAL SYSTEMS

Received February 19, 2019, accepted March 15, 2019, date of publication March 19, 2019, date of current version April 5, 2019.
Digital Object Identifier 10.1109/ACCESS.2019.2906311

A Probabilistic Algorithm for Predictive Control

With Full-Complexity Models in
Non-Residential Buildings
JUAN GÓMEZ-ROMERO 1 , CARLOS J. FERNÁNDEZ-BASSO1 , M. VICTORIA CAMBRONERO2 ,
MIGUEL MOLINA-SOLANA 3 , JESÚS R. CAMPAÑA1 , M. DOLORES RUIZ1 ,
AND MARIA J. MARTIN-BAUTISTA1
1 Department of Computer Science and Artificial Intelligence, Universidad de Granada, 18071 Granada, Spain
2 Acciona Ingeniería, 28108 Alcobendas, Spain
3 Data Science Institute, Imperial College London, London SW7 2AZ, U.K.
Corresponding author: Juan Gómez-Romero ([email protected])
This work was supported in part by the Universidad de Granada under Grant P9-2014-ING, in part by the Spanish Ministry of Science,
Innovation and Universities under Grant TIN2017-91223-EXP, in part by the Spanish Ministry of Economy and Competitiveness under
Grant TIN2015-64776-C3-1-R, and in part by the European Union (Energy IN TIME EeB.NMP.2013-4), under Grant 608981.

ABSTRACT Despite the increasing capabilities of information technologies for data acquisition and pro-
cessing, building energy management systems still require manual configuration and supervision to achieve
optimal performance. Model predictive control (MPC) aims to leverage equipment control–particularly
heating, ventilation, and air conditioning (HVAC)–by using a model of the building to capture its dynamic
characteristics and to predict its response to alternative control scenarios. Usually, MPC approaches are based
on simplified linear models, which support faster computation but also present some limitations regarding
interpretability, solution diversification, and longer-term optimization. In this paper, we propose a novel
MPC algorithm that uses a full-complexity grey-box simulation model to optimize HVAC operation in
non-residential buildings. Our system generates hundreds of candidate operation plans, typically for the next
day, and evaluates them in terms of consumption and comfort by means of a parallel simulator configured
according to the expected building conditions (weather and occupancy). The system has been implemented
and tested in an office building in Helsinki, both in a simulated environment and in the real building, yielding
energy savings around 35% during the intermediate winter season and 20% in the whole winter season with
respect to the current operation of the heating equipment.

INDEX TERMS Model predictive control, simulation, control, building energy management system.

I. INTRODUCTION Directive on the Energy Performance of Buildings addressing

Buildings account for more than one third of the worldwide the target of a 30% increase of energy efficiency by 2030 [5].
primary energy consumption [1] and they are an equally There are several complementary strategies to reduce
important source of CO2 emissions [2]. In western countries, energy consumption in existing buildings. Renovation works
non-residential buildings consume between 30-40% of the and retrofitting, making the most of affordable and clean
energy, mostly during the operational stage and by the HVAC sources, are essential, and to be effective, they must be
(heating, ventilation, and air conditioning) systems [3]. These accompanied by suitable operation protocols to optimize
figures are expected to increase in the future due to ineffi- energy management [6]. As a matter of fact, selecting daily
ciency of aging infrastructures, impact of climate change in optimal setpoints for the HVAC equipment is estimated to
weather, and economic growth in China and India [4]. At the lead to savings up to 35%, depending on the climate [7].
same time, technological advances offer great opportunities New approaches to building energy management
to achieve energy savings in new and old buildings. For the systems (BEMS) offer interactive and real-time building
latter, the European Union issued in 2016 an update of the monitoring and remote control, and provide support for
simulation and optimization [8]–[10]. Still, a great deal of
The associate editor coordinating the review of this manuscript and the decision-making is left to the operators, who must analyze
approving it for publication was Mengchu Zhou. available data, estimate energy demand, and propose control
2169-3536
2019 IEEE. Translations and content mining are permitted for academic research only.
38748 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 7, 2019
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

rules to be implemented in the BEMS. Common a priori technologies, which provide support for massive data man-
control strategies include optimized start/stop of equipment, agement for continuous model calibration, distributed execu-
chiller and boiler optimization, adaptive control, and optimal tion of simulation software, accurate prediction of building
energy sourcing [11]. conditions, and remote operation.
In the last decade, several proposals for automating the The core of the system is the intelligent operational plan
generation of operational plans based on Model Predictive generator (OPG) module, an MPC-like control scheduler
Control (MPC) have been presented [12], [13]. MPC uses supported by a cloud-based extension of the IESVE2 simula-
a simulation model of the building to capture its dynamic tion software. The OPG algorithm calculates an operational
characteristics and predict its response to alternative con- plan (OP) for a future period (typically the next day) after
trol scenarios. It pursues a (conflicting) dual target: reduc- simulating hundreds of candidate plans under the forecasted
ing energy consumption thanks to pre-emptive control and state of the building (i.e. considering weather and occu-
anticipation of the building state while keeping users’ com- pancy estimations) in order to minimize energy consumption
fort. By establishing a complete sequence of instructions for while guaranteeing occupants’ comfort. Eventually, the OP
the building equipment –i.e. the (daily) operational plan–, setpoints are automatically applied to the equipment with-
it overcomes the limitations of homeostatic controllers, which out direct involvement of the operator. To the best of our
cannot guarantee long-term optimal operation: the ahead time knowledge, this is the first proposal using an off-the-shell
and the timespan of the control instructions can expand to full-complexity model for predictive control.
several hours, leading to plans entailing more uncertainty – In this paper, we describe the OPG algorithm design,
because of the use of forecasted building conditions (e.g. implementation, and evaluation in the Sanomatalo commer-
weather, occupancy)– and more complexity –because of the cial building located in Helsinki (Finland). The control strate-
exponential increase of possible plans–, but also more effi- gies for this building focus on optimizing the air supply
cient –because of the exploitation of the inertial effects of temperature setpoint and the airflow volume setpoints. The
HVAC equipment. main contributions of this research work are the following:
MPC is formulated as a combinatorial optimization prob- • The OPG algorithm, based on probabilistic search,
lem, in which a search algorithm must find the best actua- directly provides operational plans for HVAC equipment
tion plan, in terms of thermal comfort and overall building including on/off and numerical setpoint values that are
consumption, in a solution space including all the possi- directly applied through the BEMS –no additional trans-
ble setpoint combinations for a given future period [14]. lation from demand estimations into actions is needed.
Nevertheless, most works tend to simplify the models (e.g. • We extend the control horizon compared to usual MPC
by reducing the model differential equations to linear com- approaches. The OPG considers setpoints up to a 1-day
binations) or to reduce the search space (e.g. by limit- period, which fits better to the usual building operation
ing the control to a small part of the building equipment, (e.g. the operator can validate control for the whole day)
and by incorporating manually-extracted expert knowledge). and offers more opportunities for longer-term energy
This results in short-scope, limited-extensibility and low- saving policies.
performance solutions involving a great deal of manual work • We use a full-complexity simulation model out of the
The departing hypothesis of our research work is that box, decoupled from the optimization algorithm and
we can exploit the increasing capabilities of massive and directly interpretable by experts and operators. The sim-
parallel data processing technologies to run a large amount ulation model self-recalibrates by using data directly
of simulations with full-complexity physical models and to measured from the building and runs on a cloud-based
assess multiple hypothetical control scenarios to obtain the distributed version of IESVE.
appropriate setpoints in terms of efficiency and comfort. • We carry out an evaluation of the system in the simula-
Availability of sensor data allows us to develop more accurate tion environment and in the real building; in the latter
models, since data can be used for calibration, calculation of case, over a longer period of time than related works
better predictions of relevant contextual factors (e.g. occu- (30 days), in line with the recommendations in [16].
pancy), and detection of control performance decline. At the Comparison with the base control, performed according to
same time, physical models are more interpretable and easier the International Performance Measurement and Verification
to extend; actually, we can use physical models and model Protocol (IPMVP) [17], yielded energy savings above 20%,
development tools out of the box, such as TRNSYS, Energy- with peaks above 40% at the end of the winter season.
Plus or IESVE [15]. The remainder of the paper is organized as follows. Next,
In the Energy IN TIME project,1 we developed an we describe several related works, most of them centered
advanced BEMS for optimized HVAC operation in non- in the use of simplified simulation models. In Section III,
residential buildings. This BEMS is powered by Big Data we describe the pilot building, the simulation model, and the
1 The Energy IN TIME project (Simulation-based control for energy effi- evaluation methodology. In Section IV, we detail the design
ciency building operation and maintenance) was funded by the European of the OPG algorithm and its features. Section V presents the
Commission within the 7th Framework Programme in 2013-2017. See [66]
for a brief description of the overall project results. 2 https://www.iesve.com/VE2018

VOLUME 7, 2019 38749

 J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

experimental setup and the results obtained in the simulation As an alternative to MILP and related techniques,
environment and in the real building compared to the baseline Katsigarakis et al. [46] created a surrogate building model
operation. In Section VI we discuss the contributions of our by applying Machine Learning techniques. This surrogate
proposal in terms of energy savings and comfort achievement, model is automatically learnt from pre-computed outcomes
as well as possible improvements to the system. Finally, of the real model by using a regression technique (e.g. support
we summarize the conclusions of the work and introduce vector machines), and optimization with it is significantly
prospective directions for future research. faster than in MILP. Unfortunately, it can be inaccurate or
unfeasible if the building state is difficult to model; i.e. when
II. RELATED WORK the control scope is too broad, there are too many outputs
MPC was introduced by Mahdavi in 2001 [18], and was to estimate, or the variables have complex interdependen-
initially used offline to derive an optimized control law from cies. Analogously, Casals et al. used Bayesian networks to
sensor measurements and simulations, and to validate pre- simplify the simulation model of a subway station, obtaining
defined control strategies [19], [20]. Associated small-scale good prediction accuracy [47]. Their system does not provide
experiments, most of them carried out in the simulation envi- long-term operation plans –and consequently, it does not opti-
ronment, showed that the application of MPC can effectively mize HVAC operation–, yet it achieves considerable energy
accomplish a reduction in energy consumption [21]. Further savings in ventilation and lighting systems –thanks to the use
studies characterized and performed a preliminary evaluation of sophisticated Computer Vision techniques for real-time
of HVAC-related energy management actions that can be occupancy estimation. Manjarres et al. trained a predictive
exploited in MPC [22]: outside air economizer cycle, pro- black-box model using Random Forests that reproduces the
grammed start and stop lead time, load reset, and occupied daily behavior of the building and replaces the physical
time adaptive control strategy. Additionally, other authors model of the building; however, the control strategies are
emphasized the need for considering subjective comfort mea- limited to switching on and off the HVAC systems [48].
sures beyond indoor temperatures and humidity thresholds, Kontes et al. created a surrogate model with support vector
such as predicted mean vote (PMV) [23]. machines (SVM) to optimize radiator operation with similar
In contrast, current MPC-powered BEMS are not limited promising results [49].
to only apply a plan elicited from expert knowledge and A subsequent problem of MPC is the accuracy of the
confirmed suitable after simulation. They can dynamically simulation model, particularly if a simplified version is
generate control instructions by searching an operational plan required [50]–[52], or if there is uncertainty in the expected
that, according to the simulation model, satisfies the expected building conditions; e.g., weather forecast and occupancy
energy demand while minimizes consumption. Nevertheless, estimations [53]–[55]. In this regard, Kwak et al. proposed
the calculation of the fitness of a plan by simulation is com- exploiting parallel co-simulation, which is the execution
putationally expensive [24]. of several simulation models under different conditions to
Bianchini et al. [25] addressed this issue by replacing the minimize the errors due to uncertainty in input data and
full model of the building by a simplified linear model. The unexpected occupancy variations. The authors implemented
linear model is afterwards solved by using different heuris- a general-purpose enthalpy controller that generated control
tics that reduce the search to a computable mixed integer signals starting 15 and 30 minutes later [56], and a daily
linear programming (MILP) problem. Although this solution controller [57]. For the combination of the simulation mod-
considerably reduces the capability of the algorithm to find els –in EnergyPlus and MATLAB–, they used the Building
unknown solutions, it proved to yield good results in a sim- Controls Virtual Test Bed (BCVTB) suite. The system was
ulation environment when tested for a delimited section of tested during one day in severe weather conditions in a real
the building. Different proposals using linear and non-linear building, showing energy savings around 2% in the best case.
programming, having different degree of complexity, appli-
cation scope, evaluation comprehensiveness and achieved III. MATERIALS AND METHODS
energy savings, can be found in the literature, in particular A. SANOMATALO BUILDING AND PILOT AREA
for non-residential buildings [26]–[34]. Sanomatalo3 (Sanoma house, ‘house of the press’) is a
Similarly, MPC solutions have been successfully applied to multi-purpose building situated in Helsinki and inaugurated
optimize the use of different energy sources in buildings with in 1999. It was designed by Jan Söderlund and Antti-Matti
mixed supply systems [35]–[37] and to achieve distributed Siikala, featuring a double glass façade with a steel frame
control [38], [39] –enabling extensions to minimize com- structure to reduce the need for heating. In its 9 floors and
munication between network components [40]. To increase 8227,56 m2 , it houses the offices of the Sanoma media group
the capabilities for solution diversification, other search tech- and offers 2 floors of covered public space. The building is
niques have been applied to optimization in MPC, such as managed by Caverion,4 a Finnish construction and mainte-
genetic algorithms [41]–[43] and particle swarm optimiza- nance company.
tion [44]. To address the stabilization of the control process,
nonlinear MPC solutions with varying horizon have been 3 https://sanoma.fi/en/sanoma-house/
proposed [45]. 4 https://www.caverion.com/

38750 VOLUME 7, 2019

J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

available at the beginning of the project in 2013, but they were

acquired in 2015-2017.
For demo purposes, we identified a pilot area of
2,748.60 m2 encompassing floors 6th to 8th, which
include small-size offices, meeting rooms, and open poly-
valent spaces. The use of the pilot area is the expected
one for an office building, with flexible working hours
between 6am–18pm and an overall floor space factor of
26.2 m2 /person. Total electricity consumption in the pilot area
in 2017 from January to April was about 60 MWh, while
district heating consumption was about 35 MWh in the same
area and period. These floors are served by a single not-shared
air handling unit (AHU), which is configured by means of a
temperature setpoint. This piece of equipment was the main
parameter of the energy optimization strategies (see section
III.C). In addition, we adjusted the air volume setpoint of
three variable air volume (VAV) units serving 8th floor.

B. SIMULATION MODEL AND CALIBRATION

The accuracy of the simulation model is a crucial aspect
of MPC approaches to avoid the generation of control
instructions under wrong assumptions [58]–[60]. To this aim,
control-oriented models must effectively catch all the inter-
actions between HVAC equipment (radiators, heat pumps,
etc.) [61]. This is however a difficult and costly process [62].
Grey-box models have showed good performance and
cost-benefit ratio [63], [64], even with relatively simple for-
FIGURE 1. Sanomatalo building: (a) general view; (b) detail of the façade mulations and few input variables [65]. This kind of models
(source: FUNIBER for the Energy IN TIME project). rely on the existing corpus of expert knowledge to model ther-
mal behaviour by using differential equations encoding the
physical principles of mass, energy and momentum transfer;
The building is connected to the district heating network and they apply statistical models to tune model outputs based
and rooms are heated by waterborne radiators and fan coil on historical and live data.
units. There are four heat exchangers, one of them dedicated A canonical grey-box model –namely, the operational
to the AHU heating network (power = 550 kW). All areas model– was created at system design time with the IESVE
in the building have mechanical ventilation, which adjusts software by IES energy experts with the support of Caverion’s
airflow based on room temperature and CO2 concentration. building operators. IESVE comprises a series of individual
The BEMS is provided by Schneider Electric and allows con- components including climate, geometric modelling, solar
trolling ventilation, heating, and cooling sub-systems from a shading, energy and carbon, lighting, airflow, thermal mass,
centralized console. It enables about 2.000 inspection points, value/cost and egress modules that are linked by a single
as well as an OPC (OLE for Process Control) module that Integrated Data Model (IDM) through a Common User Inter-
allows remote setpoint writing. face (CUI). By combining these modules, we can model and
The main challenge in Sanomatalo is minimizing energy simulate all aspects of a building’s construction, location,
consumption (and costs) while guaranteeing comfort (indoor geometry, climate, usage, sub-systems and thermal perfor-
temperature and CO2 concentration) during the heating sea- mance.
son –usually between September and May, being the period The simulation model developed for Sanomatalo included:
from January to March the coldest one. Indoor temperatures (a) the passive components of the building (façade, claddings,
can be retrieved in real-time through the BEMS, whereas CO2 solar irradiation, etc.), created with the ModelIT and the Sun-
sensors cannot be remotely accessed –data must be down- Cast modules; (b) the active components (anything producing
loaded offline. Heating consumption is monitored every hour or consuming electricity, especially in relation to the HVAC
by a separated sub-system. District heating prices are fixed system), created with the ApacheHVAC, MacroFlo and Vista
for each season, amounting to approximately 50e/MWh modules; (c) the expected building conditions (predicted
in the harsh winter period (Jan-Feb), and 45e/MWh in occupancy and weather forecast). Simulation was performed
the remainder of the winter period (Mar-May, Nov-Dec). by the ApacheSim module, which dynamically simulates the
Electricity price is about 77 and 79e/MWh, respectively. interaction between all of the active and passive elements over
No detailed historical records of sensor measurements were a selected period of time, taking into account the external

VOLUME 7, 2019 38751

 J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

influences (i.e. weather and occupancy) and the internal ther- (2) The airflow of 3 VAV devices (VAVairflowi ) in floor
mal behavior. The results of the simulation were viewed in 8th , in the range [50, 200] l/s. The choice of selecting these
the VistaPro module for analysis of heating and cooling loads, 3 VAVs was the limited availability of CO2 sensors at the
energy consumption, internal temperatures, thermal comfort, beginning of the project: only the area affected by these
etc. 3 VAVs was monitored.
The details of the Sanomatalo model are not public and In pre-OPG operation, Tsupply values were manually set
fall out of the scope of this paper. Nevertheless, this should by the operators and VAVairflow values were automatically
not be seen as a limitation of our proposal. On the contrary, set by using presence sensors.
our approach is agnostic to the underlying simulation model, The comfort requirements for the new system in the heating
as far as it allows setting operational profiles as input. period were the following:
The parameters of the operational model were con- • Indoor air temperature (IAT) must be in the range [20.5,
tinuously adjusted to fit live data measurements with 21.5] ◦ C during office hours 6:00–18:00. A flexible
the simulation output. Calibration was implemented as a margin in [20, 22] ◦ C is considered acceptable. This
semi-automatic procedure encompassing two iterative steps: temperature was represented by 25 output simulation
(1) measuring the model accuracy by comparing simulation variables, corresponding to 25 sensors spread across the
outputs with measured building data; (2) modifying model 3 floors directly accessible through the BEMS.
parameters to reduce model errors. In addition, IES carried • CO2 concentration (Con) upper limit is 850 ppm dur-
out an entropy analysis to detect which parameters have ing office hours. This concentration was represented by
the greatest influence in the model output, and therefore 4 output simulation variables, corresponding to 4 sensors
should be firstly modified. Overall, the calibration procedure for which there were no live measurements through the
resulted in a simulation model yielding errors below 5% [66]. BEMS.
The target variables to optimize were the heat and the fan
C. ENERGY OPTIMIZATION STRATEGIES power consumption meters of the pilot area –one of each for
the whole pilot area–, which we will call Heat and Fan. They
Following the Energy IN TIME terminology, control strate-
were represented by two output variables in the simulation
gies specify the setpoint values allowed for each piece of
model. There were no corresponding physical sensors for
actionable equipment. Strategies can denote single setpoint
these variables, but their values can be directly derived from
restrictions (e.g. setpoint variable range, frequency of change)
the BEMS temperature and air flow measurements.
or cross-parameter restrictions (e.g. two setpoints cannot have
specific values at the same time). Besides, strategies can
vary depending on the season. Energy optimization strategies D. EVALUATION METHODOLOGY
are strategies enriched with heuristic information aimed at Following the International Performance Measurement and
improving the energy efficiency and maintaining comfort. Verification Protocol (IPMVP), our evaluation methodology
That is, energy optimization strategies define additional set- compared energy savings achieved by the OPG with respect
point constraints that can help to reduce energy consumption to a base case in which it is not used. This process was
(e.g. reasonable length of the pre-heating period). Energy carried out both in the simulation environment and in the real
optimization strategies can be seen as the instantiation of the building:
Energy Management Control functions proposed in [22] for • Evaluation in the simulation environment: We selected
a particular building. 3 days in the 2016-2017 period, respectively corre-
During the plan generation process, the operational model sponding to a prototypical average (12-Jan-2016), cold
is cloned and reconfigured according to the forecasted occu- (21-Jan-2016), and warm day (30-Jan-2017) of the win-
pancy and weather conditions –namely, the independent pro- ter season. The baseline was the real operation of the
file variables. As introduced in Section IV.A, the occupancy building for the same days. These data were collected
was measured as the room occupancy % from the building at the beginning of the project. More details of this
agenda, and the weather was a set of variables including procedure are described in Section V.A.
outdoor air temperature (OAT), solar irradiance, etc. • Evaluation in the real building: The OPG was activated
Therefore, to run a simulation, we specify the operational in the building during a 30-day period in the late winter
input profiles –i.e. the equipment setpoint sequences to be season, from April 19th to May 19th 2017. The reason of
tested in the simulation– and the independent profiles –i.e. this choice is that we identified in the simulation envi-
the occupancy and the weather time series–, in order to get ronment that the OPG can achieve better results in the
the predicted profiles –i.e. the value sequences for indoor transitions between seasons –usually, the heating season
temperatures, CO2 concentration, and energy consumption. in Sanomatalo ends in the second week of May. For the
Energy optimization strategies for the Sanomatalo experi- baseline, we built a regression model from historical data
ments with the OPG solution encompassed: which estimates the energy consumption of the HVAC
(1) The supply temperature of the AHU in the pilot area system without the OPG from the weather and the occu-
(Tsupply), in the range [17, 23] ◦ C; pancy values, following the recommendations in [67].

38752 VOLUME 7, 2019

J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

With this model, we obtained a reliable approximation the setpoint means reducing the IAT and the energy consump-
of the consumption that would have been measured if the tion. We will also center the explanation in type A situa-
system without the OPG had been used during the real tions (see below). Nevertheless, the same principle applies
test period. More details of this procedure are described to problems involving multiple variables and situations B
in Section V.B. and C. The explanation can be easily extended to more than
We also studied comfort in terms of the indoor tempera- one (independent) variable.
tures (IAT) and CO2 concentration (Con) mentioned above, The overall functioning of the algorithm is depicted
checking that the simulated and measured values were within in Fig. 2, and its details are covered in the following
the acceptable ranges. subsections.

IV. PREDICTIVE CONTROL ALGORITHM 1) IDENTIFICATION OF SITUATIONS OF INTEREST

A. CONTROL SYSTEM ARCHITECTURE In Fig. 3, we show an optimization scenario in which the IAT
The Operational Plan Generation module is the core of the is controlled by a single Tsupply setpoint, as in our building.
control system in Energy IN TIME. It encompasses three Given an initial plan for Tsupply values, we can simulate it
main stages: and identify opportunities for energy optimization:
1) Collection of forecasted building data: The OPG 1) 8:00 – 12:00: Heating control results in an IAT above
retrieves the weather forecast and the occupancy pre- the upper comfort threshold. The previous Tsupply set-
diction for the operation period, usually the next day. points must be reduced to guarantee comfort.
Within our project, the weather forecast was obtained 2) 7:00 – 13:00: Heating control results in an IAT within
from the Weather Analytics API,5 and occupancy pre- the comfort range, but it may be possible to reduce the
dictions were obtained by using an agenda, which iden- previous Tsupply setpoints while keeping the tempera-
tified working days and average room occupancy per ture above the lower bound of the comfort threshold.
hour. Note that in both situations setpoint decrement may not be
2) Generation, simulation and evaluation of candidate possible if it is already at the minimum value allowed by the
plans: The OPG runs several simulations to reproduce equipment.
the expected building behavior, in terms of energy con- Analogously, we can identify one situation in which more
sumption and comfort, under different operation plans energy is required, since the comfort requirements are not
and according to the forecasted conditions retrieved in satisfied:
the previous stage. The best plan in terms of energy effi- 1) 15:00 – 18:00: Heating control results in an IAT below
ciency and comfort satisfaction is selected. We explain the upper comfort threshold. The Tsupply setpoint must
in Section IV.B how these candidates to best plan are be increased
generated and assessed. Note that in this case setpoint increment may not be pos-
3) Storage and execution of best plan: The best plan is sible if it is already at the maximum value allowed by the
stored in a database and made available to the setpoint equipment.
writing component, which eventually sends the OP
control instructions to the BEMS. This database also
2) GENERATION OF CANDIDATE PLANS
stores the context associated to each selected OP, i.e.
Let us consider a time instant t, and the corresponding set-
the forecasted building data used by the OPG algorithm
point value at this time st . For example, in Fig. 3, let us
and the simulation results. This information is useful to
suppose t = 9:00; hence, st = 23 for Tsupply setpoint. We
explain the rationale of an OP to the building managers,
notate setpoint values at time t − 1t as st−1t ; e.g. if 1t = 4,
who can revise and modify the setpoint values in real
then st−4 = 20, considering 1-hour intervals for simplicity’s
time as well.
sake.
Let us notate the modification of a setpoint value s as
B. OPERATIONAL PLAN GENERATION ŝ = s ± 1s; e.g. decrementing st in 1s = 0.5 give us
Creation of alternative plans is performed by an iterative ŝt = 23 − 0.5 = 22.5. The sets {1t} and {1s} are discrete
algorithm based on a greedy heuristic and extended to balance and ordered. We define a time horizon 1t max = max {1t} to
diversity and local optimization. In this section, we explain limit the temporal window of the modifications, as well as a
the main steps of this stage: (1) identification of situations of maximum setpoint change value 1smax = max {1s} .
interest for energy savings or comfort improvement; (2) gen- The candidate plan generation process starts from the
eration of candidate plans to address these situations; (3) can- current best plan, which at the beginning can be prede-
didate plans simulation and selection of the best one. fined or roughly estimated from outdoors temperatures. Next,
To illustrate the processing in the OPG, we will assume it detects a situation of interest by analyzing the simulation
a case with only one Tsupplysetpoint, in which decrementing of the current best plan; e.g. in our example, situation A
at t = 9:00. For this t, the algorithm will propose a few
5 http://dev.weatheranalytics.com
candidate plans by decrementing previous setpoint values.

VOLUME 7, 2019 38753

 J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

FIGURE 2. Overall functioning of the OPG algorithm, including main stages: (1) identification of situations of
interest (Section IV.B.1); (2) generation of candidate plans (Section IV.B.2); (3) plan simulation (IAT outside the
comfort range during office hours is marked with ) and assessment (Section IV.B.3). The second candidate
plan is selected, because it has the best comfort ranking.

To model all possible combinations of setpoint modifica- 4) The transition probability at each step from ŝt 0 to ŝt 0 −1
tions in situation A, we define a lattice graph like the one is given by the following function (Eq. 1):
in Fig. 4. Each vertex of this graph represents a setpoint (
δ if 1st 0 = 1st 0 −1
modification at a given previous instant: ŝt−1t = st−1t − p(ŝt 0 → ŝt 0 −1 ) = (1)
1−δ
1st−1t . Each directed edge connects a setpoint change with |{1s}|−1 otherwise
the following setpoint change in reverse time order. with the diversification parameter δ ∈ [0, 1]. This function
From this graph, the setpoint modifications that form a balances two choices: maintaining the same previous setpoint
candidate plan are modeled as the result of a random walk6 change (δ) and selecting any setpoint change (1−δ). If δ = 0,
through this graph w = hŝt , ŝt−1 , ŝt−2 , . . . , st−1t max i, with the transition probabilities at each step are the same for each
the following properties: allowed direction. If δ  0, the setpoints will tend to decrease
1) The walk starts at (0, 0) node, representing the current in the same amount.
setpoint at starting time t (i.e. the current setpoint is not An identical graph is built in situation B. An analogous
modified) graph and a corresponding probability function are defined
0
2) Each step goes from t 0 to t −1 for any t 0 in the sequence in situation C to represent setpoint increments.
(i.e. always moving from right to left in the graph) In Fig. 5, we depict two examples of random walks and the
3) The length of each path is |{1t}| (i.e. each path is a resulting setpoint modification sequences w1 , w2 .
sequence of setpoint changes from t to t − 1t max ) To reduce the number of possible alternatives, we can
introduce an additional restriction to the walks:
1) Only moves to closest nodes in horizontal, vertical
6 A random walk is a path consisting of a sequence of random steps on
and diagonal directions are allowed (i.e. differences
a mathematical space. Formally, it can be defined as a sum of a sequence
of independent, identically distributed random variables representing move between time instants of changes of 1s, if any, are
directions, or as a Markov chain over the subjacent state space [68]. small)

38754 VOLUME 7, 2019

J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

FIGURE 3. Identification of savings opportunities and discomfort in a simulated plan: indoor temperature vs
Tsupply setpoint values . The comfort range in [20, 22] ◦ C is also shown (dashed line).

3) CANDIDATE PLANS SIMULATION AND SELECTION

Each w ∈ W produces a candidate OP, which is built by
replacing the appropriate setpoints of the current best plan by
ŝt−1 , ŝt−2 , . . . , ŝt−1t max . The remaining setpoints outside


the [t −1t max , t −1] interval are not modified. The algorithm
only selects a small random subset of candidate OPs C ⊆
W to be simulated. In the best case, all the OPs in C will
be simulated in parallel; therefore, the selection of C may
depend on the simulator capabilities (see the experimental
setup in Section V).
Finally, the algorithm picks the most efficient OP satis-
fying comfort requirements. Efficiency is calculated as the
total energy consumption of the plan, while comfort can
be measured in different ways; for example, by using the
root-mean-square deviation (RMSD) or the % of time with
FIGURE 4. Lattice graph representing possible setpoint modifications comfort-related values (e.g. IAT, Con) inside the comfort
(situation A, Tsupply). Time intervals are set to 1 hour, setpoint
modifications are multiples of 0.5 ◦ C.
range, maybe limited to a period of interest (e.g. office hours).
If there is no such plan, the OPG selects the closest one to
meet the requirements. To do so, OPs are firstly sorted by
This assumption considerably reduces the number of pos- comfort satisfaction, and secondly by energy consumption.
sible walks, as shown in Fig. 6. Note that this restriction The procedure is restarted to identify the next interesting
may prevent the algorithm to explore the complete range of situation (Section IV.B.1), now using the simulation of the
allowed {1s} modifications. Moreover, other heuristics could new plan as a reference. The algorithm iterates while there are
be incorporated to the process by means of additional walk remaining situations to process or when a maximum number
restrictions encoded in the transition probability function; e.g. of situations have been processed.
to limit how many different 1s can be used in the same walk.
Our implementation of the OPG considers two particu- 4) TRIGGERING THE OPG ALGORITHM
lar situations: pre-conditioning and post-conditioning. Pre- The OPG algorithm is usually launched before midnight to
conditioning is performed to achieve comfort at the beginning calculate the setpoints for the next day, allocating enough
of the working day, while post-conditioning is performed time to let the process finish before setpoints are due –
to save energy by relaxing the comfort requirements at the a few hours in most cases. The algorithm can run again
end of the working day and later. We apply predefined several times during the day, in order to create a new plan
setpoint change strategies for each variable during these for the remainder of the day using updated weather and
intervals, which allow us to reduce the number of required occupancy predictions and to recover from control deviations
simulations. and failures.

VOLUME 7, 2019 38755

 J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

note that such a receding horizon could be implemented

just by generating control instructions for a whole period
(e.g. 24 hours) each time the OPG algorithm is triggered,
instead of generating control instructions until the end of the
current day.
At the moment, the implemented recalculation process
only generates a baseline plan using predefined operation
curves when an updated weather forecast significantly differs
from the initially used one. Enabling a faster and maybe
simplified version of the OPG for quick recovery during the
day, triggered by different events –e.g. comfort degradation
is detected with live BEMS data–, remains as future work.

C. COMPUTATIONAL PROPERTIES
The OPG algorithm cannot guarantee a global optimum in
terms of energy consumption for two reasons: (a) only a
limited number of setpoint modifications are explored; (b) the
global plan is built from locally pseudo-optimal choices
focused on situations of interest A, B, C. Conversely, it yields
good solutions in a reasonable time, and allows easy incorpo-
ration of heuristics in the setpoint modification process.
Regarding (a), in the general formulation (Fig. 4 and 5),
the number of possible setpoint change sequences |W | for
each situation and independent variable is bounded by |W | ≤
|{1s} + 1||{1t}| . In the restricted formulation (Fig. 6), the
number of possible random walks is bounded by |W | <
3|{1t}| . The |W | for multiple-dimension random walks grows
exponentially [68]. In any case, only |C|  |W | candidate
OPs will be simulated at each iteration. Therefore, the overall
efficiency of the algorithm is bounded by the number of sit-
uations of interest processed multiplied by the time required
to run each batch of simulations of size |C|. Note that the
FIGURE 5. Samples of setpoint modification sequences obtained by using execution time of the OP generation process is insignificant
random walks, restrictions (A)–(D) apply.
compared to the simulation time.
The parallel cloud version of IESVE allows running a fixed
number p of parallel simulations without performance degra-
dation. To increase solution diversity, we can set |C| < p,
and our implementation will fill the remaining simulation
slots with other plans, namely: (a) random variations of the
current best OP at any time before t; (b) combinations of
previously discarded good OPs; (c) baseline OPs –e.g. for
Sanomatalo, OPs based on outdoors temperature. These plans
are compared against the plans obtained with the random
walks, and can be selected as best current plan for the next
iteration in the same conditions.
Regarding (b), under some realistic assumptions, the OPG
algorithm finds a good approximation to the optimal solu-
tion. Specifically, for Tsupply control in Sanomatalo we can
assume that external temperatures and internal occupancy
FIGURE 6. Simplified setpoint modification graph and random walk
values follow a bell-shaped curve. To guarantee comfort in
sample, restrictions (A)-(E) apply. the winter season, the optimal OP would entail increasing the
temperature setpoints in the early morning, then decreasing
This formulation slightly diverges from the receding them around noon, and maybe incrementing them again in
horizon control typically implemented in canonical MPC, the afternoon. Moreover, the low external temperatures favor
because the prediction horizon is not shifted. However, heat losses, which would in turn require supplying hot air

38756 VOLUME 7, 2019

J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

frequently. Therefore, it is safe to limit the search to local

setpoint increments and decrements.

V. EXPERIMENTS AND RESULTS

The OPG algorithm has been implemented in the Python
and R programming languages. For the experiments in this
section, it ran on a Supermicro SuperServer 6027R-TRF,
configured with 2 processors Intel Xeon E-2600 2.4GHz
(2 × 8 cores), 128 GB RAM, 2 × 600 GB magnetic storage.
The details of the cloud-based version of the IESVE simulator
are not disclosed by IES by confidentiality reasons.
Experimentation based only in the simulation environment
was performed in advance to test and tune the deployment
of the OPG used in the real building. After some prelimi-
nary tests and following the building requirements, the OPG
parameters were set to the following values:
• Simulation batch size: p = 50, resulting in simulation
times below 20 minutes
• Ahead period of the OPG: 1 day, no receding horizon
• OPG starting time: > 2 hours before the first setpoint is
due
• Maximum setpoint change frequency: 15 minutes
• Simulation output resolution: 15 minutes
• Only minor setpoint changes are allowed in random
walks
• {1t}Tsupply, VAVairflow = {30, 60, 90, 120} minutes
• {1s}Tsupply = {0, 0.5, 1.0, 1.5, 2.0, 2.5, 3.0} ◦ C
• {1s}VAVairflow = {0, 25, 50} l/s
• δ = 0.3 (random setpoint modifications are preferred)
• Tsupply and VAV setpoints are optimized independently
(first Tsupply and then VAVairflow)
• Comfort satisfaction is measured by using the RSMD
from the comfort interval
Additionally, the maximum number of simulation batches
was restricted in order to establish an upper limit to the execu-
tion time. Since an average simulation batch took 20 minutes
(with p = 50), we set the maximum number of batches
per plan to 6 in order to keep the execution time under
2 hours. This means that 6 A-B-C situations (Figure 3) can
be analysed in each run of the OPG. Excluding pre- and post-
conditioning, 4 out of 6 were reserved for Tsupply changes,
and 2 for VAVairflow changes. Situations are sorted by rele-
vance at each iteration of the OPG algorithm; e.g. for Tsupply,
situations A and C are more important than B.

A. SIMULATION ENVIRONMENT
As described in Section II.D, we selected three prototypical
days of the winter season: average (Standard), cold (Harsh)
and warm (Intermediate). Then, we simulated the behaviour
of the pilot area of the building according to the setpoints orig-
inally applied (i.e. the base plan) and the setpoints calculated FIGURE 7. Comparison of baseline and OPG plans in terms of setpoints
and comfort values (with comfort thresholds) for a Standard winter day,
by our algorithm (i.e. the OPG plan), in order to check how simulation environment: (a) Tsupply operation; (b) VAVairflowi operation.
they compare in terms of comfort and consumption.
Fig. 7 depicts the simulation results for a Standard day, of the base and the OPG plans for the Tsupply and the
corresponding to the most common conditions during the VAVairflowi operation –the main working hours are delimited.
winter season. In the top of the figure, we show the setpoints In both cases, the largest operation differences correspond to

VOLUME 7, 2019 38757

 J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

Table 1 includes the detailed numbers for the three refer-

ence days. To obtain the overall energy consumption, we have
approximated the integral of the power functions with the
area under the curve (AUC). AUC has been computed by: (1)
interpolation of data points with a spline; and (2) calculation
of the adaptive quadrature of the interpolated function [69].
It can be seen that the OPG reduces the power consumption
of the base operation of both the heating and the ventila-
tion subsystems. As expected, in the experiments the highest
heating savings are achieved in the warmer intermediate day,
when there is still room for adjustments. Broadly speaking,
the OPG dynamically adapts the operation to the particular
conditions of a specific day without requiring the operator
attention, which is convenient in less cold days in the winter
season or before transitioning to the spring season. Con-
versely, the HVAC system is already operating at (almost) full
power during the harsh days to achieve comfort, and therefore
there is little room for improvement during working hours.
A more detailed discussion on these features is included
in Section VI. On the other hand, fan power savings have
similar values in different working days. The bad results in the
intermediate day were the consequence of the misestimation
of the occupancy used by the algorithm.

B. ON-SITE TEST AND EVALUATION

The evaluation in the pilot area of the real building was
performed from April 19th to May 19th 2017. These days
mostly fit into the Intermediate category studied in the previ-
FIGURE 8. Comparison of power (kW) of baseline and OPG plans for a
STANDARD winter day, simulated environment: (a) heat meter; (b) VAV fan ous section, the one which yielded the highest energy savings.
power meter. The baseline for daily energy consumption was calcu-
lated by a generalized linear regression model (glmnet) [70],
a method based on lasso analysis (least absolute shrinkage
and selection operator). Other prediction techniques, such
the less crowded periods. Note that the Tsupply setpoints of
as linear regression or autoregression, could have also been
the base plan before 5:00 and after 21:00 are registered but not
explored. Source data for the model was obtained from build-
applied; control is managed by a human-operated switch. In
ing sensors (energy, OAT and occupancy) logged in the period
the bottom of the figure, we show the mean comfort values
February-May 2016.
obtained in simulation in terms of IAT and Con; these values
More specifically, we developed two baseline models for
lay within the comfort intervals (also included in the figure).
prediction of daily consumption of heating equipment and
Fig. 8 shows the power consumption calculated by the sim-
VAV fans, based on the expected heating demand and occu-
ulation model; respectively, the Heat and Fan power meters
pancy. Expected daily energy demand (hdd, in heating degree
values. We can observe that most savings are achieved at the
days) was calculated by using integration with base temper-
borderline hours, that is, at the beginning and at the end of the
ature set to 18 ◦ C and the BEMS OAT [71]. Estimated daily
working day. This is consistent with the pre-conditioning and
occupancy (occ, in %) was the maximum occupancy value of
post-conditioning provisions made by the OPG algorithm.
the office agenda. To build the prediction models, we firstly
pre-processed the data, discarding outliers and measurement
TABLE 1. Energy consumption (kWh) for the experiments in the errors.
simulation environment. Fig. 9 compares the energy consumption in February-
May 2016 and the values calculated by the heating and the
fan consumption prediction models. The parameters of the
regression models are given in Eq. 2 and Eq. 3 respectively,
yielding correlation coefficient values of R2 = 0.632 and
R2 = 0.234. Note that: (1) the heating baseline model slightly
overestimates consumption from mid-April to June, which
means that energy savings calculated in the next section are
slightly overestimated as well; (2) the fan power model has

38758 VOLUME 7, 2019

J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

FIGURE 9. Comparison of daily energy consumption (kWh) estimated by

the baseline models vs historical data in Feb-May 2016: (a) heating;
(b) VAV fans.

a low R2 value, which means that energy savings calculated

with this model should be considered with caution.
Heat ∗ (hdd, occ) = −9.122×hdd +0.703×occ+20.71 (2)
Fan∗ (hdd, occ) = −0.038×hdd +0.012×occ+5.015 (3)
Fig. 10(a) and 10(b) show the comparison of the values of
daily energy consumption in the pilot area obtained from the
BEMS (+) with the values estimated by the prediction models
(o) for the test period in the real building. Fig. 10(c) shows the
energy savings achieved in % of the (estimated) consumption
before optimization.
FIGURE 10. Comparison of daily energy consumption (kWh) during the
test period vs estimated by the baseline models: (a) heating; (b) VAV fans;
TABLE 2. Energy savings (kWh) achieved in the on-site test with the OPG (c) savings. Days in red italic font are weekend or holiday days.
control vs estimated by the baseline models.

and around 20% for the electrical subsystem. Weekends and

holidays offer opportunities for higher energy savings, since
the OPG adjust the operation to the building occupancy better
than the manual operation.
Savings have been achieved without compromising users’
comfort. Figure 11 shows the IAT and CO2 concentration
values in the pilot area in the evaluation period. The IAT
values were calculated as follows: (1) sensor measurements,
obtained from the BEMS temperature sensors (25), were
resampled and interpolated to match the setpoint change
frequency parameter (15 minutes); (2) sensor temperatures
As summarized in Table 2, the average savings per day were averaged at each timestamp; (3) maximum and mini-
are, respectively, around 40% for the thermal subsystem mum values of timestamps within the working hours were

VOLUME 7, 2019 38759

 J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

reduces energy consumption; particularly heating consump-

tion. Automatic control allows for more effective plans since
it enables a finer-grained and more frequent scheduling of
setpoint changes without the supervision of the building oper-
ators. The OPG algorithm and software offer a flexible and
configurable framework to generate more efficient operation
plans, predicting the building state and adapting energy usage
to more realistic demand estimations without compromising
users’ comfort. As expected, it has proved to be particularly
successful in optimizing temperature setpoints, in which a
longer control horizon, accounting for the inertia of the equip-
ment, is crucial. The building operators were satisfied by the
use of the system during the test period. One highlighted
system’s feature was the capabilities to validate the plans
in advance (and even to modify them) and to provide jus-
tifications of the algorithm decisions –by means of graphi-
cal depictions of the simulation results, in a similar way to
Fig. 11.
As already anticipated by the experiments in the simulation
environment (Section V.A), the highest energy savings can
be obtained for heating in the mid-season, when it is not
necessary to use the heating equipment at full, and particu-
larly, in the warmer days (Table 1, Intermediate). At the same
time, the system can react to isolated cold days. The on-site
FIGURE 11. Daily comfort values achieved in the on-site test with the evaluation in the Sanomatalo building, which was carried
OPG control, maximum and minimum sensor average values: (a) indoor out at the end of the 2017 heating season, confirmed these
air temperature (IAT, ◦ C), with comfort interval; (b) CO2 concentration
(Con, ppm), with comfort threshold. CO2 measurements were only
assumptions. Energy usage in colder days could be even more
available from April 19th to 14th May. Days in italic red font are weekend optimized by relaxing the comfort temperature restrictions to
or holiday days. permit OPs with some minor discomfort for a short period of
time. The advantage of our system is that it allows operators
obtained. Con values were retrieved from 4 offline sensors; to characterize and quantify this discomfort in advance, thus
the remainder of the procedure is the same as for IAT. supporting them to make more informed decisions. (Note that
IAT min values lie within the comfort range during the test this feature was not exploited in the experiments.)
period. Actually, it would have been possible to configure the Our system reduced the temperature setpoints given by
OPG to reduce Tsupply even more. However, as explained the normal operation of the building between 0.5 and 2 ◦ C.
at the beginning of this section, we prioritized optimizing During the on-site test, this meant savings in heating above
discomfort situations. IAT max values are over the comfort 40% (Table 2) while keeping comfort (see Fig. 11(a)). The
upper threshold by 1 ◦ C. A more detailed analysis of these algorithm adapted well to workdays and weekends, show-
values identified that discomfort was not sustained and only ing slightly better results in the former ones (Fig. 10(c)).
happened for short time periods (less than 1 hour). A possible explanation for this is that operators have lower
Similarly, CO2 concentration values are mostly below the availability in weekends and holiday days, and therefore it is
comfort threshold, although with some exceptions. After not possible for them to create customized plans. The airflow
more detailed analysis, we identified that the highest values consumption was also reduced in a 20% (Table 2) without
were measured by a single sensor, which in some cases compromising the CO2 concentration comfort (Fig. 11(b)),
exceeded 950 ppm. However, the levels calculated by the despite the lack of a proper model calibration and the smaller
simulation environment were considerably smaller. This is a number of simulations involving VAVs. Nevertheless, due to
clear example of the importance of having all the sensor data the lower accuracy of the baseline model, these results are
available through the BEMS. With the CO2 sensors offline, less precise and should be further analyzed; e.g. by using
our system was not able to recalibrate the simulation model autoregression to build the baseline [72].
–which would have led to better plans–, nor to detect in real In summary, although the Sanomatalo building was already
time that some plans were not guaranteeing comfort –which efficiently operated, and considering the limitations of the
would have triggered a correction action. baseline estimations, the overall savings figures in the inter-
mediate winter are in line with the 30% target of EU energy
VI. DISCUSSION directive [5] and the 35% savings estimations provided in [7].
The results presented in Section V show that the use of The experiments also revealed more opportunities for sav-
MPC in the offices of the Sanomatalo building significantly ings in the future, e.g. by improving the simulation model

38760 VOLUME 7, 2019

J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

with online CO2 data (for calibration) and more detailed in our experiments were mostly static, while occupancy
occupancy predictions (actual agenda data were not very monitoring and reconfiguration have shown effective in the
fine-grained). past [77], [78]. In this regard, the capabilities and limitations
for OP recalculation during the day should also be further
TABLE 3. Energy savings (heating, MWh) in the pilot area projected for explored. Four, a more comprehensive study of energy sav-
the whole year. ings with additional baseline models should be carried out in
order to quantify more precisely the return of investing in our
solution [79], in particular if only ventilation is addressed.
Last but not least, the building setup was relatively simple,
with district heating and almost fixed energy costs. It would
be interesting to study the applicability and the scalability
of the OPG approach to smart grids, including more control
variables –some of them affecting the production side– and
energy storage equipment; see for example [80], [81].

VII. CONCLUSIONS AND FUTURE WORK

This paper has presented the design and the implementation
In Table 3, we show a rough projection of energy savings in of an MPC-based control system aimed at reducing energy
the pilot area for the whole year using: (1) savings calculated consumption in non-residential buildings while guaranteeing
in Section V.A (only heating); (2) historical monthly con- occupants’ comfort. The main difference of our proposal with
sumption values provided by the building operators; (3) esti- respect to other approaches is that we use a full-complexity
mated distribution of day types. We assume that the heating simulation model, which runs in parallel in the cloud. This
system is not used during the summer, and therefore it does allows using more accurate models and facilitates commu-
not make sense to quantify savings in this period. It can be nication between computer scientists, building operators and
seen that the overall energy consumption reduction during the simulation developers, exploiting synergies of their joint
whole winter season is around 20%, larger than the consump- work. Comprehensive quantitative and qualitative compari-
tion of a standard winter month. We can also estimate savings son with MPC approaches using reduced-complexity simu-
of CO2 emissions: assuming a carbon factor of 206 kgCO2 / lation models would be useful to support decision-making
MWh for district heating energy in Finland [73], the new between different alternative approaches.
system applied in the pilot area can save more than 1.60 Tons Experimentation in the Sanomatalo building, located in
of CO2 per year. These figures could be directly adapted to Helsinki, both in the simulation environment and in the real
other estimation of day types (e.g. including savings in the building, has shown that important energy savings –up to 40%
summer) and extended to other sections of the building with at the end of the winter season– can be achieved, particularly
similar configuration. by optimizing the control of the heating equipment. Note
The implementation of the system in other buildings that our approach can be adapted to other scenarios, and
entails: (a) developing a specific simulation model, if not specifically, to cooling equipment. In our experiments we
available; (b) parametrizing the OPG algorithm, including did not consider the energy costs of running our system,
the definition of energy optimization strategies; (c) find- which should be deducted from the HVAC savings [82].
ing appropriate sources for weather and occupancy fore- These promising figures can give rise to disruptive models
casts; (d) adapting the setpoint writing component, if for energy service provision, as we explore in [83].
fully-automatic control is enabled; (e) deploying the compu- The OPG algorithm opens several opportunities for further
tational infrastructure to run these components. As a matter research. The current design relies on a variant of heuristic
of fact, in the context of the Energy IN TIME project we search, which can be hard to scale up if several variables
applied modified versions of the OPG to other scenarios, are to be optimized at the same time. In this regard, other
such as an airport and a hotel – achieving similar results, search and optimization techniques could be applied. Specif-
as briefly described in [66]. Among the tasks required to ically, genetic algorithms allow balancing diversification and
extend the system to other buildings, developing and tuning intensification of solution search by adjusting their param-
the simulation model is the most time-consuming one. eters. Another possible extension of the OPG would be to
The collaboration with the Sanomatalo building operators incorporate means to define imprecise comfort ranges, thus
revealed some prospective improvements to the system. formalizing the notion of relaxed comfort into the proce-
First, it would be convenient to offer a better interface dure. It would also be interesting to study how to repre-
for configuration of the OPG and interaction with the sent energy optimization strategies in a machine-processable
generated plans, as in [74]. Second, users’ comfort should language, in such a way that the system could use them
be measured beyond thermal and CO2 concentration inter- for self-configuration. Moreover, self-configuration could be
vals, probably by using PMV, adaptive comfort models and supported by machine learning techniques able to identify
comfort standards [75], [76]. Third, occupancy estimations successful operation patterns from historical data, and to

VOLUME 7, 2019 38761

 J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

apply reinforcement learning to reward and reuse particularly t Time instant

efficient OPG plans. Tsupply AHU supply temperature setpoint value
Finally, we believe that combining interpretable white/ (◦ C)
grey-box models, like the one used in this work, and efficient VAV Variable air volume unit
black-box models, learnt from historical data, is one of the VAVairflowi Airflow setpoint value for VAV number
most prospective directions for future work. Faster simulation i (liters per second, l/s)
of such hybrid model would allow for the implementation w Ordered sequence of setpoint changes
of more sophisticated optimization and planning techniques. from t to t –1t max
Recent approaches to data-driven black-box models have W Set of all w
showed good accuracy, but only for short time periods [84].
Learning more general and precise models would require ACKNOWLEDGEMENTS
larger datasets, more computational power, and techniques The authors would like to thank Vicente Madero (Acciona
able to exploit them. Recent advances in the Deep Learning Construcción S.A.); Catherine Conaghan, Adalberto Guerra,
area suggest that this is a feasible goal. Stephen Earle (Integrated Environmental Solutions Ltd.);
Christian Beder (Cork Institute of Technology); and Jukka
Heino (Caverion Suomi Oy) for their support and assistance
NOMENCLATURE with this research work.
AHU Air Handling Unit
AUC Area under the curve REFERENCES
BCVTB Building Controls Virtual Test Bed [1] J. Laustsen, ‘‘Policy pathways: Energy performance certification of build-
BEMS Building Energy Management System ings,’’ International Energy Agency (IEA), Paris, France, Tech. Rep., 2010.
C Set of candidate plans considered in an [2] D. Urge-Vorsatz, K. Petrichenko, M. Staniec, and J. Eom, ‘‘Energy use in
buildings in a long-term perspective,’’ Current Opinion Environ. Sustain-
iteration of the OPG ability, vol. 5, no. 2, pp. 141–151, 2013.
Con CO2 concentration (parts per million, ppm) [3] L. Pérez-Lombard, J. Ortiz, and C. Pout, ‘‘A review on buildings
δ diversification parameter energy consumption information,’’ Energy Buildings, vol. 40, no. 3,
pp. 394–398, 2008.
1t Time increment / decrement [4] X. Cao, X. Dai, and J. Liu, ‘‘Building energy-consumption status
{1t} Set of time increment / decrement values worldwide and the state-of-the-art technologies for zero-energy build-
1t max Maximum time increment / decrement ings during the past decade,’’ Energy Buildings, vol. 128, pp. 198–213,
Sep. 2016.
1s Setpoint increment / decrement [5] Proposal for a Directive of the European Parliament and of the Council
{1s} Set of setpoint increment / decrement values Amending Directive 2010/31/EU on the Energy Performance of Buildings,
1smax Maximum setpoint value increment Eur. Commission, Brussels, Belgium, 2016.
[6] A Clean Planet for all—A European Long-Term Strategic Vision for a
/ decrement Prosperous, Modern, Competitive and Climate Neutral Economy, Eur.
Fan Energy consumption due to electrical Commission, Brussels, Belgium, 2018.
[7] A. Ghahramani, K. Zhang, K. Dutta, Z. Yang, and B. Becerik-Gerber,
subsystem (kWh) ‘‘Energy savings from temperature setpoints and deadband: Quantifying
Fan∗ Estimated daily fan energy consumption with the influence of building and system properties on savings,’’ Appl. Energy,
the baseline model (kWh) vol. 165, pp. 930–942, Mar. 2016.
[8] V. Marinakis, H. Doukas, C. Karakosta, and J. Psarras, ‘‘An integrated sys-
HDD Heating Degree Days tem for buildings’ energy-efficient automation: Application in the tertiary
hdd Estimated daily demand measured in HDD sector,’’ Appl. Energy, vol. 101, pp. 6–14, Jan. 2013.
(integrated) [9] A. Costa, M. M. Keane, J. I. Torrens, and E. Corry, ‘‘Building operation
and energy performance: Monitoring, analysis and optimisation toolkit,’’
Heat Energy consumption due to thermal subsystem Appl. Energy, vol. 101, pp. 310–316, Jan. 2013.
(kWh) [10] B. Swords, E. Coyle, and B. Norton, ‘‘An enterprise energy-information
system,’’ Appl. Energy, vol. 85, no. 1, pp. 61–69, 2008.
Heat∗ Estimated daily heating energy consumption [11] Y. Lu, S. Wang, and K. Shan, ‘‘Design optimization and optimal control
of grid-connected and standalone nearly/net zero energy buildings,’’ Appl.
with the baseline model (kWh) Energy, vol. 155, pp. 463–477, Oct. 2015.
HVAC Heating, Ventilation, and Air Conditioning [12] A. Afram and F. Janabi-Sharifi, ‘‘Theory and applications of HVAC control
IAT Indoor Air Temperature (◦ C) systems—A review of model predictive control (MPC),’’ Building Envi-
ron., vol. 72, pp. 343–355, Feb. 2014.
IPMVP International Performance Measurement and [13] A. Mirakhorli and B. Dong, ‘‘Occupancy behavior based model predictive
Verification Protocol control for building indoor climate—A critical review,’’ Energy Buildings,
MILP Mixed Integer Linear Programming vol. 129, pp. 499–513, Oct. 2016.
[14] G. Serale, M. Fiorentini, A. Capozzoli, D. Bernardini, and A. Bemporad,
MPC Model Predictive Control ‘‘Model Predictive Control (MPC) for enhancing building and HVAC
OAT Outdoor Air Temperature (◦ C) system energy efficiency: Problem formulation, applications and oppor-
occ Estimated daily occupancy (maximum) (%) tunities,’’ Energies, vol. 11, no. 3, p. 631, 2018.
[15] D. B. Crawley, J. W. Hand, M. Kummer, and B. T. Griffith, ‘‘Contrasting
PMV Predicted mean value the capabilities of building energy performance simulation programs,’’
RSMD Root mean square deviation Building Environ., vol. 43, no. 4, pp. 661–673, Apr. 2008.
[16] P. Rockett and E. A. Hathway, ‘‘Model-predictive control for non-domestic
st Setpoint value at time t buildings: A critical review and prospects,’’ Build. Res. Inf., vol. 45, no. 5,
ŝt Setpoint value at time t modified pp. 556–571, 2017.

38762 VOLUME 7, 2019

J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

[17] S. Meyers and S. Kromer, ‘‘Measurement and verification strategies for [41] H. Pombeiro, M. J. Machado, and C. Silva, ‘‘Dynamic programming
energy savings certificates: Meeting the challenges of an uncertain world,’’ and genetic algorithms to control an HVAC system: Maximizing thermal
Energy Efficiency, vol. 1, no. 4, pp. 313–321, 2008. comfort and minimizing cost with PV production and storage,’’ Sustain.
[18] A. Mahdavi, ‘‘Simulation-based control of building systems operation,’’ Cities Soc., vol. 34, pp. 228–238, Oct. 2017.
Building Environ., vol. 36, no. 6, pp. 789–796, 2001. [42] F. Ascione, N. Bianco, C. De Stasio, G. M. Mauro, and G. P. Vanoli,
[19] J. A. Clarke et al., ‘‘Simulation-assisted control in building energy man- ‘‘Simulation-based model predictive control by the multi-objective opti-
agement systems,’’ Energy Buildings, vol. 34, no. 9, pp. 933–940, 2002. mization of building energy performance and thermal comfort,’’ Energy
[20] S. Petersen and S. Svendsen, ‘‘Method for simulating predictive control of Buildings, vol. 111, pp. 131–144, Jan. 2016.
building systems operation in the early stages of building design,’’ Appl. [43] S. Wang and X. Jin, ‘‘Model-based optimal control of VAV air-conditioning
Energy, vol. 88, no. 12, pp. 4597–4606, 2011. system using genetic algorithm,’’ Building Environ., vol. 35, no. 6,
[21] M. Killian and M. Kozek, ‘‘Ten questions concerning model predic- pp. 471–487, 2000.
tive control for energy efficient buildings,’’ Building Environ., vol. 105, [44] C. D. Corbin, G. P. Henze, and P. May-Ostendorp, ‘‘A model predictive
pp. 403–412, Aug. 2016. control optimization environment for real-time commercial building appli-
[22] W. Z. Huang, M. Zaheeruddin, and S. H. Cho, ‘‘Dynamic simulation of cation,’’ J. Build. Perform. Simul., vol. 6, no. 3, pp. 159–174, 2013.
energy management control functions for HVAC systems in buildings,’’ [45] D. He, ‘‘Dual-mode nonlinear MPC via terminal control laws with free-
Energy Convers. Manag., vol. 47, nos. 7–8, pp. 926–943, 2006. parameters,’’ IEEE/CAA J. Autom. Sinica, vol. 4, no. 3, pp. 526–533,
[23] R. Z. Freire, G. H. C. Oliveira, and N. Mendes, ‘‘Predictive controllers Jul. 2017.
for thermal comfort optimization and energy savings,’’ Energy Buildings, [46] K. I. Katsigarakis, G. D. Kontes, G. I. Giannakis, and D. V. Rovas,
vol. 40, no. 7, pp. 1353–1365, 2008. ‘‘Sense-think-act framework for intelligent building energy management,’’
[24] D. Agdas and R. S. Srinivasan, ‘‘Building energy simulation and parallel Comput.-Aided Civil Infrastruct. Eng., vol. 31, no. 1, pp. 50–64, 2016.
computing: Opportunities and challenges,’’ in Proc. Winter Simulation [47] M. Casals, M. Gangolells, N. Forcada, M. Macarulla, A. Giretti, and
Conf., 2014, pp. 3167–3175. M. Vaccarini, ‘‘SEAM4US: An intelligent energy management sys-
[25] G. Bianchini, M. Casini, A. Vicino, and D. Zarrilli, ‘‘Demand-response tem for underground stations,’’ Appl. Energy, vol. 166, pp. 150–164,
in building heating systems: A model predictive control approach,’’ Appl. Mar. 2016.
Energy, vol. 168, pp. 159–170, Apr. 2016. [48] D. Manjarres, A. Mera, E. Perea, A. Lejarazu, and S. Gil-Lopez,
[26] J. Figueiredo and J. S. da Costa, ‘‘A SCADA system for energy man- ‘‘An energy-efficient predictive control for HVAC systems applied to
agement in intelligent buildings,’’ Energy Buildings, vol. 49, pp. 85–98, tertiary buildings based on regression techniques,’’ Energy Buildings,
Jun. 2012. vol. 152, pp. 409–417, Oct. 2017.
[27] J. Široký, F. Oldewurtel, J. Cigler, and S. Prívara, ‘‘Experimental analysis of [49] G. D. Kontes, C. Valmaseda, G. I. Giannakis, K. I. Katsigarakis, and
model predictive control for an energy efficient building heating system,’’ D. V. Rovas, ‘‘Intelligent BEMS design using detailed thermal simulation
Appl. Energy, vol. 88, no. 9, pp. 3079–3087, 2011. models and surrogate-based stochastic optimization,’’ J. Process Control,
[28] H. Huang, L. Chen, and E. Hu, ‘‘A new model predictive control scheme vol. 24, no. 6, pp. 846–855, 2014.
for energy and cost savings in commercial buildings: An airport ter- [50] S. Prívara, J. Cigler, Z. Vána, F. Oldewurtel, C. Sagerschnig, and
minal building case study,’’ Building Environ., vol. 89, pp. 203–216, E. Žáceková, ‘‘Building modeling as a crucial part for building predictive
Jul. 2015. control,’’ Energy Buildings, vol. 56, pp. 8–22, Jan. 2013.
[29] J. Ma, J. Qin, T. Salsbury, and P. Xu, ‘‘Demand reduction in building energy [51] D. Picard, J. Drgoňa, M. Kvasnica, and L. Helsen, ‘‘Impact of the controller
systems based on economic model predictive control,’’ Chem. Eng. Sci., model complexity on model predictive control performance for buildings,’’
vol. 67, no. 1, pp. 92–100, 2012. Energy Buildings, vol. 152, pp. 739–751, Oct. 2017.
[30] S. J. Kang, J. Park, K.-Y. Oh, J. G. Noh, and H. Park, ‘‘Scheduling-based [52] M. Gruber, A. Trüschel, and J. O. Dalenbäck, ‘‘Model-based controllers for
real time energy flow control strategy for building energy management indoor climate control in office buildings—Complexity and performance
system,’’ Energy Buildings, vol. 75, pp. 239–248, Jun. 2014. evaluation,’’ Energy Buildings, vol. 68, pp. 213–222, Jan. 2014.
[31] I. Hazyuk, C. Ghiaus, and D. Penhouet, ‘‘Optimal temperature con- [53] S. Petersen and K. W. Bundgaard, ‘‘The effect of weather forecast uncer-
trol of intermittently heated buildings using model predictive control: tainty on a predictive control concept for building systems operation,’’
Part II—Control algorithm,’’ Buildings Environ., vol. 51, pp. 388–394, Appl. Energy, vol. 116, pp. 311–321, Mar. 2014.
May 2012. [54] F. Oldewurtel, D. Sturzenegger, and M. Morari, ‘‘Importance of occu-
[32] R. De Coninck and L. Helsen, ‘‘Practical implementation and evaluation pancy information for building climate control,’’ Appl. Energy, vol. 101,
of model predictive control for an office building in Brussels,’’ Energy pp. 521–532, Jan. 2013.
Buildings, vol. 111, pp. 290–298, Jan. 2016. [55] B. Dong and K. P. Lam, ‘‘A real-time model predictive control for building
[33] S. C. Bengea, A. D. Kelman, F. Borrelli, R. Taylor, and S. Narayanan, heating and cooling systems based on the occupancy behavior pattern
‘‘Implementation of model predictive control for an HVAC system detection and local weather forecasting,’’ Building Simul., vol. 7, no. 1,
in a mid-size commercial building,’’ HVAC&R Res., vol. 20, no. 1, pp. 89–106, 2014.
pp. 121–135, 2014. [56] Y. Kwak, J.-H. Huh, and C. Jang, ‘‘Development of a model predictive
[34] K. Deng et al., ‘‘Model predictive control of central chiller plant with control framework through real-time building energy management system
thermal energy storage via dynamic programming and mixed-integer linear data,’’ Appl. Energy, vol. 155, pp. 1–13, Oct. 2015.
programming,’’ IEEE Trans. Autom. Sci. Eng., vol. 12, no. 2, pp. 565–579, [57] Y. Kwak and J.-H. Huh, ‘‘Development of a method of real-time build-
Apr. 2015. ing energy simulation for efficient predictive control,’’ Energy Convers.
[35] I. Sharma et al., ‘‘A modeling framework for optimal energy manage- Manag., vol. 113, pp. 220–229, Apr. 2016.
ment of a residential building,’’ Energy Buildings, vol. 130, pp. 55–63, [58] P. de Wilde, ‘‘The gap between predicted and measured energy perfor-
Oct. 2016. mance of buildings: A framework for investigation,’’ Autom. Construct.,
[36] B. Mayer, M. Killian, and M. Kozek, ‘‘A branch and bound approach for vol. 41, pp. 40–49, May 2014.
building cooling supply control with hybrid model predictive control,’’ [59] A. C. Menezes, A. Cripps, D. Bouchlaghem, and R. Buswell, ‘‘Predicted
Energy Buildings, vol. 128, pp. 553–566, Sep. 2016. vs. actual energy performance of non-domestic buildings: Using post-
[37] S. Salakij, N. Yu, S. Paolucci, and P. Antsaklis, ‘‘Model-based predictive occupancy evaluation data to reduce the performance gap,’’ Appl. Energy,
control for building energy management. I: Energy modeling and optimal vol. 97, pp. 355–364, Sep. 2012.
control,’’ Energy Buildings, vol. 133, pp. 345–358, Dec. 2016. [60] N. Li, Z. Yang, B. Becerik-Gerber, C. Tang, and N. Chen, ‘‘Why is
[38] R. Zafar, A. Mahmood, S. Razzaq, W. Ali, U. Naeem, and K. Shehzad, the reliability of building simulation limited as a tool for evaluating
‘‘Prosumer based energy management and sharing in smart grid,’’ Renew. energy conservation measures?’’ Appl. Energy, vol. 159, pp. 196–205,
Sustain. Energy Rev., vol. 82, pp. 1675–1684, Feb. 2018. Dec. 2015.
[39] T. Bai, S. Li, and Y. Zheng, ‘‘Distributed model predictive control for net- [61] H. Satyavada and S. Baldi, ‘‘An integrated control-oriented modelling for
worked plant-wide systems with neighborhood cooperation,’’ IEEE/CAA HVAC performance benchmarking,’’ J. Build. Eng., vol. 6, pp. 262–273,
J. Autom. Sinica, vol. 6, no. 1, pp. 108–117, Jan. 2019. Jun. 2016.
[40] X. Mi and S. Li, ‘‘Event-triggered MPC design for distributed systems [62] E. Źáčeková, Z. Váña, and J. Cigler, ‘‘Towards the real-life implementa-
with network communications,’’ IEEE/CAA J. Autom. Sinica, vol. 5, no. 1, tion of MPC for an office building: Identification issues,’’ Appl. Energy,
pp. 240–250, Jan. 2018. vol. 135, pp. 53–62, Dec. 2014.

VOLUME 7, 2019 38763

 J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

[63] Z. Afroz, G. M. Shafiullah, T. Urmee, and G. Higgins, ‘‘Modeling tech- JUAN GÓMEZ-ROMERO received the degree in
niques used in building HVAC control systems: A review,’’ Renew. Sustain. computer science and the Ph.D. degree in intel-
Energy Rev., vol. 83, pp. 64–84, Mar. 2018. ligent systems from the Universidad de Granada,
[64] A. Foucquier, S. Robert, F. Suard, L. Stéphan, and A. Jay, ‘‘State in 2004 and 2008, respectively.
of the art in building modelling and energy performances predic- He was a Lecturer with the Applied Artifi-
tion: A review,’’ Renew. Sustain. Energy Rev., vol. 23, pp. 272–288, cial Intelligence Group, Universidad Carlos III de
Jul. 2013. Madrid, from 2008 to 2013, and a Research Asso-
[65] B. Gunay, W. Shen, and G. Newsham, ‘‘Inverse blackbox modeling of the ciate in the EU FP7 Project Energy IN TIME with
heating and cooling load in office buildings,’’ Energy Buildings, vol. 142, the Universidad de Granada, from 2013 to 2017.
pp. 200–210, May 2017.
He was also a Visiting Researcher with the Data
[66] A. Conserva et al., ‘‘Energy in time project: Summary of final results,’’
Science Institute, Imperial College London, from 2016 to 2017. He has
in Proc. 12th Conf. Sustain. Develop. Energy, Water Environ. Syst., 2017,
been a Senior Research Fellow with the Computer Science and Artificial
Paper SDEWES2017-0867.
Intelligence Department, Universidad de Granada, since 2016. He has partic-
[67] Z. Li, Y. Han, and P. Xu, ‘‘Methods for benchmarking building energy
consumption against its past or intended performance: An overview,’’ Appl. ipated in more than 20 projects in security, ambient intelligence, and energy
Energy, vol. 124, pp. 325–334, Jul. 2014. efficiency. His research interests include the use of semantic representation
[68] G. F. Lawler and V. Limic, Random Walk: A Modern Introduction. models and machine learning techniques to perform automatic reasoning
Cambridge, U.K.: Cambridge Univ. Press, 2010. towards higher-level information fusion.
[69] R Core Team. (2018). R: A Language and Environment for Statistical Dr. Gómez-Romero is the Principal Investigator of the projects BIGFUSE:
Computing. [Online]. Available: http://www.r-project.org/ Semantics for Big Data Fusion and Analysis: Improving Energy Efficiency
[70] J. Friedman, T. Hastie, and R. Tibshirani, ‘‘Regularization paths for gen- in Smart Grids and PROFICIENT: Deep Learning for Energy-Efficient
eralized linear models via coordinate descent,’’ J. Statist. Softw., vol. 33, Building Control.
no. 1, pp. 1–22, 2010.
[71] M. S. Al-Homoud, ‘‘Computer-aided building energy analysis tech-
niques,’’ Building Environ., vol. 36, no. 4, pp. 421–433, 2001.
[72] W. Liang, R. Quinte, X. Jia, and J.-Q. Sun, ‘‘MPC control for improving
energy efficiency of a building air handler for multi-zone VAVs,’’ Building CARLOS J. FERNÁNDEZ-BASSO received the
Environ., vol. 92, pp. 256–268, Oct. 2015. degree in computer science and the M.Sc. degree
[73] B. Koffi, A. Cerutti, M. Duerr, A. Iancu, A. Kona, and in data science from the Universidad de Granada,
G. Janssens-Maenhout, ‘‘CoM default emission factors for the member in 2014 and 2015, respectively, where he is cur-
states of the European union—Version 2017,’’ European Commission,
rently pursuing the Ph.D. degree in computer sci-
Joint Research Center (JRC), Brussels, Belgium, Tech. Rep., 2017.
ence and energy efficiency.
[74] J. Cigler, P. Tomáško, and J. Široký, ‘‘BuildingLAB: A tool to analyze per-
He was a Lead Developer in the EU FP7 Project
formance of model predictive controllers for buildings,’’ Energy Buildings,
vol. 57, pp. 34–41, Feb. 2013. Energy IN TIME in the topics of building simula-
[75] J. H. Lim, J. T. Kim, S. H. Cho, and G. Y. Yun, ‘‘Development of the tion and control, data analytics, and machine learn-
adaptive PMV model for improving prediction performances,’’ Energy ing. He also collaborates with the Data Science
Buildings, vol. 98, pp. 100–105, Jul. 2015. Institute, Imperial College London, where he has carried out research stays,
[76] B. W. Olesen, ‘‘Indoor environmental input parameters for the design and from 2016 to 2018. He is currently a Research Assistant with the Computer
assessment of energy performance of buildings,’’ REHVA J., pp. 17–23, Science and Artificial Intelligence Department, Universidad de Granada.
Jan. 2015.
[77] T. Ekwevugbe, N. Brown, V. Pakka, and D. Fan, ‘‘Improved occupancy
monitoring in non-domestic buildings,’’ Sustain. Cities Soc., vol. 30,
pp. 97–107, Apr. 2017.
[78] A. Capozzoli, M. S. Piscitelli, A. Gorrino, I. Ballarini, and V. Corrado,
‘‘Data analytics for occupancy pattern learning to reduce the energy con-
sumption of HVAC systems in office buildings,’’ Sustain. Cities Soc.,
vol. 35, pp. 191–208, Nov. 2017. M. VICTORIA CAMBRONERO received the
[79] Q. Meng, M. Mourshed, and S. Wei, ‘‘Going beyond the mean: Distribu- degree in industrial engineering with a specializa-
tional degree-day base temperatures for building energy analytics using tion in construction and industrial facilities and the
change point quantile regression,’’ IEEE Access, vol. 6, pp. 39532–39540, M.Sc. degree in energy technology for sustainable
2018. development from the Polytechnic University of
[80] C. Wang, B. Jiao, L. Guo, Z. Tian, J. Niu, and S. Li, ‘‘Robust scheduling Valencia, in 2008 and 2009, respectively.
of building energy system under uncertainty,’’ Appl. Energy, vol. 167, She was a Project Developer and a Researcher
pp. 366–376, Apr. 2016. on building management and energy efficiency
[81] H. Thieblemont, F. Haghighat, R. Ooka, and A. Moreau, ‘‘Predictive con- with the Institute of Energy Engineering, Valen-
trol strategies based on weather forecast in buildings with energy storage cia, before joining Acciona Infrastructures R&D
system: A review of the state-of-the art,’’ Energy Buildings, vol. 153, Department, in 2010. Since 2018, she has been a Project Manager with
pp. 485–500, Oct. 2017. Acciona Ingeniería. She has a wide experience in research and innovation
[82] E. Feller, L. Ramakrishnan, and C. Morin, ‘‘Performance and energy projects and has collaborated and managed several FP7 and H2020 projects
efficiency of big data applications in cloud environments: A Hadoop
in renewable energies, HVAC systems and storage integration, and control
case study,’’ J. Parallel Distrib. Comput., vols. 79–80, pp. 80–89,
strategies for energy efficiency improvement in buildings and districts; e.g.,
May 2015.
2DISTRICT, EnergyINTIME, LowUP, Flexynets, FC-DISTRICT, MESSIB,
[83] J. Gómez-Romero, M. Molina-Solana, M. Ros, M. D. Ruiz, and
M. J. Martin-Bautista, ‘‘Comfort as a Service: A new paradigm for residen- EINSTEIN, CommONEnergy, and COST-EFFECTIVE. She is currently an
tial environmental quality control,’’ Sustainability, vol. 10, no. 9, p. 3053, Industrial Engineer specialized in energy efficiency and building simulation.
Aug. 2018. Ms. Cambronero holds a Project Management Professional certification
[84] F. Ferracuti et al., ‘‘Data-driven models for short-term thermal behaviour granted by the Project Management Institute and a Certified in Measurement
prediction in real buildings,’’ Appl. Energy, vol. 204, pp. 1375–1387, and Verification Protocol certification granted by the Association of Energy
Oct. 2017. Engineers and Efficiency Valuation Organization.

38764 VOLUME 7, 2019

J. Gómez-Romero et al.: Probabilistic Algorithm for Predictive Control With Full-Complexity Models

MIGUEL MOLINA-SOLANA received the degree M. DOLORES RUIZ received the degree in math-
in computer science and the Ph.D. degree from ematics and the European Ph.D. degree in com-
the Universidad de Granada, in 2007 and 2012, puter science from the Universidad de Granada,
respectively. in 2005 and 2010, respectively.
From 2012 to 2015, he was a Research Asso- She held a non-permanent teaching positions
ciate with the Universidad de Granada in the with the Universities of Jaén, Granada, and Cádiz.
FP7 Project Energy IN TIME. He was a Research She has participated in more than ten projects,
Associate on visualization with the Data Science including the EU FP7 Projects ePOOLICE and
Institute, Imperial College. He is currently a Marie Energy IN TIME. She is currently a Research
Curie Research Fellow with Imperial College Lon- Associate with the Computer Science and Artifi-
don, U.K. His research interests include applied work in machine learning cial Intelligence Department, Universidad de Granada. Her research interests
and knowledge representation in diverse domains such as music, energy include data mining, information retrieval, energy efficiency, big data, cor-
management, and business. relation statistical measures, sentence quantification, and fuzzy sets theory.
Dr. Molina-Solana is the Principal Investigator of the H2020 Project She has organized several special sessions about Data Mining in inter-
DATASOUND–Understanding Data With Sound. national conferences and was part of the organization committee of the
FQAS’2013 and SUM’2017 conferences.
Dr. Ruiz belongs to the Approximate Reasoning and Artificial Intelligence
Research Group and the Cybersecurity Lab, Universidad de Granada. She
has been the Principal Investigator of the project Exception and anomaly
detection by means of fuzzy rules using the RL-theory. Application to fraud
detection.

MARIA J. MARTIN-BAUTISTA received the

degree in computer science and the Ph.D. degree
from the Universidad de Granada, in 1996 and
2000, respectively.
She has been a Professor with the Computer Sci-
JESÚS R. CAMPAÑA received the M.Sc. and ence and Artificial Intelligent Department, Uni-
Ph.D. degrees in computer science from the Uni- versidad de Granada, since 2018. She was the
versidad de Granada, where he has been a member Principal Investigator in the FP7 European Project
of the Intelligent Databases and Information Sys- Energy IN TIME with the Universidad de Granada,
tems Research Group, Department of Computer from 2013 to 2017. She has also been a Principal
Science and Artificial Intelligence, since 2005. Investigator of several international and national projects and knowledge
From 2013 to 2018, he was a Lecturer with the transfer contracts with private companies. She has published more than
Universidad de Granada. He has been a member 100 papers in international journals and conferences. Her current research
of several research projects related to fuzzy data interests include intelligent systems, big data, and knowledge representation
representation in databases, data mining, text min- with applications to energy, security, and health.
ing, and energy efficiency, including the FP7 Project Energy IN TIME. His Prof. Martin-Bautista is a member of the IEEE Society and the EUSFLAT
research interests include fuzzy databases, XML, knowledge representation, Society.
semantic web, data mining, and text mining.

VOLUME 7, 2019 38765